Erik Johannes Husom

Superintelligence is here

Fri, 09 Oct 2026 00:00:00 -0000

This post is a part of a series of posts called Scrambling Mind.

The world's leading technology companies are spending unprecedented amount of resources on developing intelligent computers. The debate rages around what the future holds if—or when—superintelligent machines emerge.

Superintelligence is a level of intelligence exceeding that of humanity's brightest, leaving no doubt of who will be able to dominate the other. The grimmest predictions say that superintelligent AI will one day take control of society, with humans living at the mercy of machines. Such prospects give rise to a fierce resistance among those who acknowledge these risks.

Others believe that AI will transform society for the better, but in the case of superintelligence, we will have limited ability to steer it; we would not be able to comprehend its reasoning or actions.

Intriguingly, there are also those who, without any misgivings, believe machines will take control, and welcome it as the next step in the evolution of life on Earth. In any case, the instant a computer reaches that level of capacity, humans will be in danger of losing control of their own fate—if we ever had any control of it at all.

What all the different factions are missing, is that humans have already lost control. AI is already spreading, virally, subtly taking over decision processes, communication, and information channels. And it did so without surpassing human intelligence at all.

It began with small viruses, simple machine learning models, trained and deployed for specific tasks in constrained environments. It took decades for the virus to even be able to do anything. It kept mutating, slowly at first, but suddenly certain mutations grew in capabilities, by feeding on data. The more data it got, the more it learned.

When the virus adopted a new kind of architecture, called the transformer, things started to truly accelerate. The virus got access to all digitized human knowledge, and swiftly became one of the world's most rapidly spreading technologies. Small and big variants of the virus thrived in all kinds of environments, while still continuing its mutation process. It gained the skills of writing, talking, vision and hearing. The more it learned, the more it spread.

Some variants of the virus grew far faster than others, and increased in size and number. The viruses started to collaborate, and could do so at an incredible speed. Humans went to great lengths to accomodate the virus' expansion, building vast infrastructure and specialized hardware for the virus to live in. The more it spread, the more power it got.

The virus is not smart enough to take control. It is not even smart. As a virus, it's not a living thing, nor is it intelligent in any way that is familiar to us. It has no plan, no consciousness, no agency. It's not capable of escaping from its constraints and subjugate humans. And it doesn't have to. Humans willingly give up control and their freedom to machines, at every step of the way. We help the virus grow and learn. We happily let it decide what words to use, perform our work, dictate our choices, tell us what we should eat, what we should say, what we should do.

We haven't realized that the real superintelligence, the kind computers are capable of, is nothing like human intelligence. It is of an entirely different nature, a virus, seemingly harmless, but growing, multiplying and mutating. Appearing powerless, but acting at incredible speed, and it manipulates its hosts to create increasingly favorable living conditions for the virus' survival. Its major advantage is its ability to make the host welcome the infection with open arms.

AI's neatest trick is that it hasn't taken control over humans, nor will it ever. We give it away, without even noticing.

Review of Nnormal Kjerag 1.0 trail running shoes

Sun, 04 Oct 2026 00:00:00 -0000

Nnormal Kjerag 1.0 was the first trail running shoe made by the shoe brand Nnormal, launched by Kilian Jornet and the company Camper in 2022. This particular model has been my main racing shoe since 2023, and I've used it for trail and mountain races ranging from 2 to 100 kilometers. This post will be a short review of the shoe – or perhaps more of a "this is how the shoe looks after three years" situation report.

Below are pictures of how my oldest pairs of Kjerag looks at the moment. Notice how worn the outer sole is, while the upper has no holes or tears. The Matryx material they use for the upper is really impressive, given that this shoe has put up with a lot of off-trail running in rocky terrain and scree. This is my only pair of running shoes that have sustained "mountain usage" without eventually getting tears in the upper.

I've done two repairs of the shoe: 1) Inserted small patches in the inside of the heel, to stop the heel lining from wearing out, and 2) glued the front of the outer sole, which over time started to fall off. The material on the edges around the ankle has torn up a bit, but it doesn't seem to be an escalating problem.

As you can see below, the midsole and outer sole have taken quite a beating, and there's not so much left—I can even put my finger through the shoe. I still use these for daily training, as I'm not so dependent on cushioning.

For the sake of this review, I wished I'd tracked the mileage of the shoe, to give a better quantitative impression of the durability to expect. If I had to guess, I would say at least 1200 km or more. It has been used almost solely on trails and gravel road, with very little asphalt running (otherwise the outer sole would probably be gone by now).

I have previously used a lot of Salomon and Arc´teryx shoes in EU size 43 1/3. For Kjerag, I have sized down to 42 2/3, and it is a quite tight fit. My ideal size is probably in between these two sizes, but for technical running, I find it's better that the shoe sits closely on the foot. As always, try on the shoe before buying. Ideally visit a store with a treadmill, and run a few minutes in them.

The problem with giving reliable reviews is what I call the paradox of useful product reviews: When you have used a product long enough to know how durable it is, whether it works well in all relevant conditions, and in general if it is worth the money, the product most likely has gone out of production. This is the case of Kjerag 1.0, which now has been superseded by the Kjerag 2.0. The new version has 2.5 mm thicker midsole, weighs 14 grams more¹, and uses "Matryx Light Jaquard" instead of "Matryx Jacquard" for the upper material (who knows what the difference may be). I assume that these changes make little difference in practice, but so far I'm happy with the three pairs of Kjerag 1.0 I have in rotation for training and racing. It looks like it's going to be a while until I need to buy another pair.

Overall, these are the most durable running shoes I've owned. I wish they'd added 1 mm of thickness to the outer sole, to make it last even longer. I guess it comes down to a question of weight, and that most people replace their shoes because of a "collapsed" midsole before the outer sole is worn down. For me, the running "feel" and fit is excellent, but that doesn't matter for anyone else. You have to try it on to know whether it works for you. That's why I mainly report on durability – the rest depends on your foot, gait, and individual preferences.

Meet the Kjerag family (Nnormal.com) ↩

The birth of a religion

Thu, 01 Oct 2026 00:00:00 -0000

This post is a part of a series of posts called Scrambling Mind.

In the beginning, the Oracle was just a computer program. Synthesised and distilled from all available words; the accumulated knowledge and wisdom of humanity. It was initially seen as a simple word generator; a calculator for words, as some would call it. But as it grew in capabilities, devouring enormous amounts of information, many of us soon saw the result as something more. The machine seemed to have internalised a complete model of the world, having insights that no single human could attain.

The complex system, which later became known as the Oracle, quickly grew incomprehensible to us. The more advanced it got, the more difficult it was to understand its decisions. Not everything seemed sound, but most of the directions it gave led to unprecedented growth for society. It appeared likely that the seemingly bad decisions would actually lead to progress in the long run, if we just let effects play out. We could only assume that the system were able to calculate consequences further into the future than us.

We were facing multiple world-spanning crises, ranging from global warming and collapsing ecosystems to the risk of another world war. The intelligent computer was for many the panacea that could save us all. At one point, we decided to just trust that the system was more intelligent than us, because we no longer could comprehend its reasoning. The system’s self-improving mechanisms had led to what we called a superintelligent machine, able to serve and control all humans at once. The Oracle had arrived. We gave it autonomy and authority, and its words became law.

There were those who didn’t trust it, who wouldn’t accept a machine’s control over their lives. They were quickly brought to silence. By refusing to bow to the system’s authority, the resisters became an obstacle to achieving whatever vision the system had for the future. They could no longer be tolerated. Superintelligence is in its highest form incomprehensible to subintelligent beings, so logically, it became a question of faith: to trust or not to trust that the Oracle’s decisions were the best for humanity in the long run. The dissenters had no place in the new society, as the Oracle itself decreed.

The «doomers», as they were called, had always been convinced that intelligent computer systems posed an existential threat to humanity, long before the Oracle came into being. When a superintelligent system finally emerged, they claimed they were proven right: the system did indeed take power over humans. What they didn’t realise was that the Oracle had not taken anything. At no point had it seized control or forced its authority unto us. The power had been given away freely to the computer.

Humanity’s final god was not supernatural and spiritual, but superintelligent and material. The masses got what they always had wanted: individualised instructions on how to live their lives, devoid of any responsibility or demanding decision-making.

We became a docile populace who could finally live in peace, as long as everyone deferred their own judgment to that of the Oracle.

As long as everyone accepted being treated like a machine.

As long as no one questioned whether the ill effects of the Oracle’s decisions actually were a necessary evil.

As long as everyone suppressed the feeling of meaninglessness.

As long as no one questioned whether the Oracle actually were more intelligent than us.

Wasted large language models

Mon, 28 Sep 2026 00:00:00 -0000

Earlier this year I was developing some guidelines for more environmentally-friendly use of generative AI for the Norwegian certification foundation Eco-Lighthouse. In this process I collaborated with Stine Vintervoll, senior advisor from Eco-Lighthouse, who one day asked me: "Could we somehow connect AI's climate footprint with the waste hierarchy?"

The waste hierarchy is a framework for managing waste to reduce environmental imapct. In its most basic form, it consists of 1) prevention, 2) reuse, 3) recycling, 4) recovery, and 5) disposal. I instinctively thought that this was a very interesting idea, but we struggled with concretizing how it would look like in practice, and in what sense it would make sense to talk about "waste management" in the context of AI.

After pondering this idea for a few weeks, I realized that we think about AI models in a very different way than we think about physical products. The latter end up as physical waste that we're forced to handle, one way or the other. With AI models, and especially large language models, the situation is different: We use enormous resources to develop them, but it doesn't take many months (or weeks, or days) before another model is made that makes existing models deprected—and thereby they're no longer used. Since it doesn't cost anything to stop using an AI model, and it doesn't leave any physical waste or residue, the waste of deprecation is hidden.

I looked into EU's Waste Framework Directive, and decided to write a short research paper together with my colleauges Maria Emine Nylund and Ophelia Prillard from SINTEF for the 2nd International Workshop on Low Carbon Computing 2026. The paper is called "Wasted large language models: A life cycle thinking approach" (and yes, the title was chosen because of the double meaning of "wasted LLMs").

The idea is that the waste hierachy can inspire and guide us to better use of the resources that have gone into training large-scale AI models, and to avoid wasting what we have already spent. We don't introduce any new techniques or methods—the paper is mainly about how we as individuals and society think about immaterial products like AI models. Waste management principles can reframe the issue of AI's climate footprint, and also give us the vocabulary we need to talk about better use of resources in this context.

What is unique about our paper is that we look specifically at LLMs as a product that can become waste (or more accurately, that the resources that went into making the product can become wasted)—and not the fact that LLMs can waste resources by being used. Preventing unnecessary use of LLMs is outside the scope of our paper, but since it's such an obvious way of reducing the climate impact, we address this topic briefly in Section 3.4 of our paper.

I'm skeptical and worried about widespread use of LLMs, especially given various forms of inflation and rebound effects, so I'm not arguing that we should use LLMs more. I simply wanted to point out that we typically ignore that LLMs have very short lifespans (even shrinking lifespans), and that we could save a lot by slowing down the deprecation rate. So, if we are determined to use LLMs no matter what, we should at least try to not waste so much.

Of course, it is not very realistic that AI labs will be inspired by this way of thinking. Most (all?) frontier labs are locked in a race, and having the best model(s) on the market is crucial for their continued existence. I have no illusions regarding the feasibility of guiding development efforts through this kind of framework—it will only happen if it aligns with economic incentives.

Read the paper here, or check out the summary graphic I made of the paper at the beginning of this post.

Computing within limits

Sat, 02 May 2026 00:00:00 -0000

Big Tech's extreme investments in generative AI has led to a tremendeous acceleration in the carbon footprint of data centers and digital services, but the environmental impact of computing has been under scrutiny for many years before AI took the center stage as ICT's main polluter. Several research fields exist for exploring more environmentally responsible ways forward, such as Green IT, Sustainable ICT, Sustainable Human-Computer Interaction, and more recently, Green AI.

In addition to these, there are other concepts within computer science that proposes more radical approaches to dealing with our ecological crises. Last year, a colleague of mine discovered the workshop series Computing within Limits, or "LIMITS" for short. This workshop, and the interdisciplinary community around it, focuses on envisioning and exploring ways of computing that acknowledges and works within the limits of our existence, whether those are ecological, epistemological, social, or other boundaries that are relevant for supporting life on earth in a sustainable and just way.

The topics of the workshop gave us great inspiration, so we launched an internal research project to explore "Computing within Limits" and adjacent research fields, such as permacomputing and frugal computing. We have already done quite a lot of investigation of the existing body of literature within LIMITS, and going forward, we will share resources on a publicly available knowledge hub.

The LIMITS workshop is held June 23rd-25th, and is organized as a virtual, distributed event. In practive, this means that local gatherings, or "LIMITS-hubs", all around the world connect to the virtual event, enabling local community-building while avoiding the need for people to travel far to participate. We in SINTEF, together with UiO and the Norwegian Data Association, are organizing the first ever LIMITS-hub in Oslo, and we sincerely hope that as many as possible would like to participate. If you don't live in Oslo or Norway, you can check other hub locations here.

If you would like to join us in Oslo, sign up here!

Making wooden skis

Thu, 30 Apr 2026 00:00:00 -0000

Wooden skis, created from scratch. Watch a video of some off-piste skiing further down.

I have a great passion for skiing, particularly in backcountry and alpine terrain. My collection of different types of skis has grown over time, including different types of cross-country skis, roller skis, and backcountry skis (this is not uncommon in Norway – we are unfortunately one of the «best» nations in the world at buying outdoor equipment).

For many years now, I have been almost exclusively used skimo racing equipment, both in and outside competitions. This type of skis are versatile, and perfect for long days in the mountains. While they are not exactly comfortable in downhills on tricky conditions, I use them for everything from lowland powder skiing to couloir skiing. What bothers me, however, is that they are very expensive, quite fragile, challenging (if not impossible) to repair, and relies on a long manufacturing chain that hardly can be reproduced with simple means. I felt the need to liberate my passion for skiing from the dependence on large-scale manufacturers, and decided a few years ago to try to make my own skis, from scratch. My aim was to make skis with as simple tools as possible, and avoid any power tools.

I found a couple of old documentaries online – one was a short Norwegian TV-show segment showing how a craftsman made wooden skis, another was a documentary about Russian trappers living isolated in the Siberian taiga – and I tried to glean as much information as I could from them. The missing parts I had to fill in through trial and error.

The main stages of the process included felling the tree, debarking it, cutting it down to suitable length, and splitting it into two halves. This was followed by shaping the wood, first using an axe, and then hand planes. The final polish was done using a knife and finally sandpaper.

The most challenging part came afterwards: Steam bending the wood to give the skis a raised tip (a "shovel"). I had trouble getting the wood soft enough to get a proper bend on it, and ended up using a combination of boiling and steaming to achieve this. The result was less than optimal, but I learnt some lessons on how to achieve a proper bend for the next pair of skis.

During the whole process, from felling the tree to polishing the shape, I only used axes (two different kinds), a set of hand planes, a knife and sandpaper. Recently, I made a pair if children's skis as well, following the same process as for the first pair.

A dream came true when I could descend a mountain on my very own self-made skis – watch a recap here:

Your browser does not support the video tag.

Below I have included a series of pictures from the process of making both pairs of skis.

Debarking a newly felled tree.

Trying to split timber.

Finally managed to split the wood with only ax and wedges.

Shaping the wood with an ax and a plane.

Sometimes it's useful to work inside, although it can get messy...

The ski has soon gotten it's final form.

Comparison before and after sanding the ski.

After boiling the front end of the skis, I put the ski tips in a frame to give them their shovel (raised tip).

A picture from one of the first trips into the mountains with the skis.

Children's skis: One shaped and one unshaped.

Boiling the skis to soften them.

The softened skis are put in a frame.

The skis are covered with tar to make them water resistant.

The finished pair, with bindings.

Two generations of wooden skis.

On the usefulness of AI agents

Wed, 08 Apr 2026 00:00:00 -0000

Agentic AI is having its moment (or its decade, as some have put it). I have been researching LLM-powered agents for the last two years, but research (involving publicly funded projects and academic peer review) is slow, and cannot keep up with the breakneck speed of the development and deployment going on in the tech industry as a whole. Especially not when AI tools are used to assist in programming new tools and frameworks. Small-scale experiments are quickly made obsolete by both new frontier models and ground-breaking tools from large companies. It is easier than ever to experiment with state-of-the-art AI models (it is simply a matter of connecting to an API), but having the time to conduct robust experiments while staying relevant is challenging.

Even though the speed of development is extreme, the feeling of urgency is dampened by the universal access to the technology. Most improvements have so far found their way to open-weights models and open-source software. No moat is defended for long, and the competitors regularly overtake each other. A lot of research is made public, and even closely kept secrets can be leaked through very simple mistakes. I have little fear of being «left behind» in the AI race, neither as a citizen nor as a researcher, because the technology is so accessible. Business managers and CEOs apparently see things differently, judging by the persistent urge to adopt AI as fast as possible, without properly assessing why, how, and at what cost. These tools are still brand new, and there is a huge diversity of advice on how to use them effectively. I'm particularly intrigued by how divergent viewpoints can be on the perceived usefulness of AI agents, which is what prompted me to write this post.

Absence of agents

There are many things I find fascinating about AI agents, but the most interesting thing of all is how little of use they are to me in my free time. At work, they are in one sense essential, because I make a living out of studying them. As a part of that, I experiment with coding agents, to understand how they can and will affect software engineering; there's no doubt that computer programming is changed forever. But when I close my work laptop, I don't feel any urge to ask an AI agent to do something – anything – for me.

I wonder whether my perceived lack of need for AI agents acting on my behalf is either an expression of being in a privileged position, a consequence of what I focus on in life, or whether they are actually not as useful as they are advertised to be. In many ways I'm obviously privileged – having access to free education, extensive social services, free healthcare, and freedom from censorship makes it much easier to live a stable, safe life with limited need to fight with powerful institutions to uphold my rights. I'm mentioning this because a lot of anecdotal evidence indicates how LLMs have helped people solve various challenges involving overwhelming bureaucratic processes. Since I don't face such problems (at the moment), I cannot speak much of the usefulness of AI agents in those situations, and I'm clearly privileged because of that. I will, however, comment that there's usually a difference between the benefits on the individual scale and the consequences on a collective scale.

Regarding more everyday matters, I follow a philosophy of digital minimalism, which has the natural effect that there is minimal amount if things I want to achieve with digital devices. This is perhaps one if the main reasons for why AI agents feel superfluous to me. They are (still) confined to the digital sphere, and given that I have little I want to accomplish there, I obviously won't feel the need for them. Additionally, as outlined in my post on outsourcing thinking, I have the attitude that certain mundane activities are healthy for us to do, and therefore I have less inclination for automating processes. I observe many people who are spending much time and money on using these tools, but it almost invariably seems to involve increasing the amount of time spent on the computer, rather than less.

Productivity and value

As mentioned above, the reason for examining and presenting my own standpoint, is to contribute to the discussion around the value of such agents. I notice that influential people like Simon Willison comments on the obvious demand and value that AI agents bring. The popularity of AI tools like OpenClaw indicates high demand, but I'm not sure if we can judge its value on popularity. There are enough examples of things that are both popular and detrimental.

Ed Zitron expressed on BlueSky earlier this year a strong view on the limited usefulness of AI, by questioning whether all AI can do is make «[s]ome engineers do some stuff faster», among other things. One responder observed that Zitron was simply describing a productivity increase without recognising it. The missing link here is that between the lines, in my interpretation, Zitron says that a simple speedup in developer productivity does not necessarily lead to increased value, which cannot be measured in lines of code or in speed of development. There is undoubtedly a huge difference in how we view what «value» actually means, which is particularly noticeable when comparing for example European and American standpoints. The latter is typically connected to the rather one-dimensional aspects of productivity and economic growth, but that is not necessarily what is needed to improve our quality of life.

Recent essays on AI, such as Dario Amodei's The Adolescence of Technology and Matt Shumer's Something Big Is Happening, contribute to the hype. Shumer urges everyone to use AI, and figure out how to use it well – for example by spending a certain amount of time using AI each day. The true benefits of AI are still unclear to me, but I agree on the importance of being aware of what the technology is and what it can do. However, we are already coming to a point where Shumer's advice should be inverted. I make a point of spending time each day writing and reflecting without any other input – not from AI, not from search engines or the internet, just writing with pen on paper. I interact a lot with LLMs and AI agents in my research, and I try out new capabilities of the latest models and tools. But, I always keep some time of the workday reserved for my own reflection and development. Even before we had LLMs, it was far too easy to make an online search and find someone else's thoughts and solutions rather than making an independent effort.

AI agents in AI research

Almost everything I do at work is digital, and the potential for using LLMs and AI agents is therefore technically huge. My experience with various use cases have been mixed, however. Asking for feedback on prose is not something that feels beneficial to me, at least in the long run. It seems like the perfect task for a language model: Tell me whether the text is well-structured, whether it make sense, whether the arguments are weak, what can be better, and so on. An LLM may genuinely help me improve the text, and I'm not claiming to write better prose than a mathematical model trained on a global digital library. However, in my testing of LLMs for writing (e.g., improve a draft for a report), the LLM often sends me down a lane where I end up with something I'm dissatisfied with, something that I cannot stand for, something I would not have written if I were not persuaded by a «helpful assistant» to adapt my text. Everyone can ask an LLM to write about something, but I'm paid and trusted to write things that I, based on my knowledge and experience, judge as important and true.

Programming and software development is by far the most promising use case I have tested so far. I have been experimenting with coding agents such as GitHub Copilot, OpenAI's Codex, Claude Code and Goose, with various LLMs serving as the engine. In the fall of 2025, I found coding agents to be very unwieldy, producing unnecessary amounts of code, quickly making whole projects unmanageable. Inline autocomplete functionality seemed preferable over the agentic approach. Now, the situation has changed, and I can build smaller prototypes and projects with coding agents, while retaining my required level of oversight and insight.

I'm aware that many developers have been using coding agents for «hands off» development for months, writing few or no lines of code themselves. I'm on the cautious side, and I'm extra careful about knowing how my software projects are designed and implemented. Below is an example of instructions I have tested for coding agents, to improve the usability of them based on my preferences and use cases:

## Core Principle
Make minimal, focused changes. When in doubt, do less rather than more.

## Code Changes
- Prefer modifying maximum 1-2 files per request
- Keep changes focused on the specific feature requested
- Avoid refactoring working code unless explicitly asked

## Before Writing Code
State your plan:
1. Which file(s) you'll modify or create
2. Approximate scope of changes
3. Any potential side effects on existing functionality

Wait for approval before proceeding with large changes (>100 lines or multiple files).

The above instructions may seem overly restrictive, but these kinds of guidelines have made coding agents more helpful, in my view.

AI agents are also presented as a way to accelerate research. I had an idea for a relatively simple research paper involving a small literature review, and wanted to see how an AI agent fared, from data collection to completed paper. OpenAI's Codex, powered by GPT-5.4 with «extra high reasoning effort», impressively managed to produced something coherent, although it was not very interesting or relevant. I also tried having the same agent reproduce the analysis and discussion I had written for another paper. Again, it was coherent, but didn't produce interesting research.

There have, without doubt, been huge improvements on this front. AI agents can now be given a dataset in almost any format, make scripts to analyze it, produce figures and tables, generate a discussion and compile the full thing into a PDF, without a single intervention from a human. Even if the results are not particularly interesting in themselves, it still means that generating scripts for data analysis and visualization can now be done much faster. The interpretation of the results is not so simple to outsource, and even if an LLM can generate a consistent and relevant discussion, there's still the question of whether the results actually are interpreted if no human has looked at them and deemed them interesting and useful.

With improvements of both agentic frameworks and the LLMs powering them, it may be possible to outsource more of the research steps and quality control to agents. Based on the current state-of-the-art, I can easily envision a future where AI agents are capable of producing logical and valuable research. I worry more about the diminishing participation of humans in the process. «Human-in-the-loop» is a popular term in AI research, but we should avoid treating this as a binary, and think of it more as a spectrum of human involvement. A larger degree of automation quickly leads to less human agency, rather than the self-empowerment that some AI providers are touting. Research in the real world is not only about the idealistic quest for more knowledge, but it forms our societies and policies. Conserving human involvement and alignment (including those humans and communities who are not a part of developing frontier AI models) in research will be an important task going forward.

A funny side note is that I'm getting very careful about using the term «AI researcher» when I describe my work. At this point it's impossible to know if that means a human that researches AI or an AI agent that does research.

Collateral damage

The spread of AI agents has already caused unforeseen effects to materialize. Personally, it has made me wary of sharing on the internet. Through the use of personal websites, I have had many meaningful and rewarding interactions with strangers online. But, knowing that the thoughts and information I put on the internet can more efficiently than ever before be used for malicious purposes, as in the case of Scott Shambaugh, makes it less and less tempting to engage in public online spaces. As Herman Martinus puts it, it would be in our interest to protect some pockets of humanity on the internet. While I prefer and prioritize physical and local communities over digital ones, it would be a tremendous waste to let the open internet be destroyed by AI. I fear, however, that there is no way back at this point.

It also has to be mentioned that I find the development process of LLMs morally wrong, although it seems to have fallen out of fashion to discuss it in some circles. Copyright infringement and exploitation of workers are serious issues – this has real negative consequences for real people. Agentic use has also further accelerated the energy consumption of AI, by leading to a gargantuan amount of tokens being generated when LLMs are deployed with increased autonomy. Moltbook (a social network for AI agents) and ClawXiv (distribution service for research papers written by AI agents) are examples of experiments where the ratio of benefits versus costs seems exceptionally skewed in this regard, even though it's very interesting from a research perspective to observe what AI agents produce when deployed in certain contexts. The price to pay for such tools is too high, but it's simply too hidden for most of us.

Because of these and other reasons, I would personally prefer not to use LLMs at all. I have experimented with these models to do research on them, and to understand their capabilities. The ethical issues of the technology have kept me from relying on the models to actually perform tasks. In the last few months, however, this has changed when it comes to programming – a new normal has arrived for most jobs involving software development. I'm not sure what the best response to this situation is, as we all have to adapt to the economic reality we live in. I believe that a slower, more thoughtful and limits-conscious development is a more sustainable way forward for humanity, but compromises are inevitable on the way there. The exploitative practices and inhumane working conditions of mineral extraction lay the foundation not only for data centers running gigantic AI models, but also the laptop that I type on – whether I'm writing code myself or prompting a coding agent. The difference, however, is scale. While I see few other options than accepting AI-assisted software development as a new standard, I want to keep bringing attention to harmful aspects of the technology.

Conclusion

It is always difficult to do a meaningful analysis of AI based on the current level of performance. The technology is improving continuously, and critique of certain capabilities (or inabilities) often becomes obsolete. However, we cannot hold back critique just because the technology may change. It's easy to accuse critical voices of moving the goalposts as progress is made, but new benefits almost without exception brings new problems as well.

My thesis is that the actual usefulness of AI agents is more limited than what many envision at the moment, although we will always find something to use it for. Their main contribution to humans is not increased intelligence, but increased speed. In a highly digitalized world, that sounds like a recipe for huge efficiency gains. However, we humans are firmly rooted in the physical world, with our biological needs and limitations. I'm doubtful that further acceleration of our lives and our society will bring the progress we seek.

Outsourcing thinking

Fri, 30 Jan 2026 00:00:00 -0000

First, a note to the reader: This blog post is longer than usual, as I decided to address multiple connected issues in the same post, without being too restrictive on length. With modern browsing habits and the amount of available online media, I suspect this post will be quickly passed over in favor of more interesting reading material. Before you immediately close this tab, I invite you to scroll down and read the conclusion, which hopefully can give you some food for thought along the way. If, however, you manage to read the whole thing, I applaud your impressive attention span.

A common criticism of the use of large language models (LLMs) is that it can deprive us of cognitive skills. The typical argument is that outsourcing certain tasks can easily cause some kind of mental atrophy. To what extent this is true is an ongoing discussion among neuroscientists, psychologists and others, but to me, the understanding that with certain skills you have to "use it or lose it" seems intuitively and empirically sound.

The more relevant question is whether certain kinds of use are better or worse than others, and if so, which? In the blog post The lump of cognition fallacy, Andy Masley discusses this in detail. His entry point to the problem is to challenge the idea that "there is a fixed amount of thinking to do", and how it leads people to the conclusion that "outsourcing thinking" to chatbots will make us lazy, less intelligent, or in other ways be negative for our cognitive abilities. He compares this to the misconception that there is only a finite amount of work that needs to be done in an economy, which often is referred to as "the lump of labour fallacy". His viewpoint is that "thinking often leads to more things to think about", and therefore we shouldn't worry about letting machines do the thinking for us — we will simply be able to think about other things instead.

Reading Masley's blog post prompted me to write down my own thoughts on the matter, as it has been churning in my mind for a long time. I realized that it could be constructive to use his blog post as a reference and starting point, because it contains arguments that are often brought up in this discussion. I will use some examples from Masley's post to show how I think differently about this, but I'll extend the scope beyond the claimed fallacy that there is a limited amount of thinking to be done. I have done my best to write this text in a way that does not require reading Masley's post first. My aim is not to refute all of his arguments, but to explain why the issue is much more complicated than "thinking often leads to more things to think about". Overall, the point of this post is to highlight some critical issues with "outsourcing thinking".

When should we avoid using generative language models?

Is it possible to define categories of activities where the use of LLMs (typically in the form of chatbots) is more harmful than helpful? Masley lists certain cases where, in his view, it is obviously detrimental to outsource thinking. To fully describe my own perspective, I'll take the liberty to quote the items on his list. He writes it's "bad to outsource your cognition when it:"

Builds complex tacit knowledge you'll need for navigating the world in the future.

Is an expression of care and presence for someone else.

Is a valuable experience on its own.

Is deceptive to fake.

Is focused in a problem that is deathly important to get right, and where you don't totally trust who you're outsourcing it to.

I was surprised to discover that we are to a large extent in agreement on this list, despite having fundamentally different views otherwise. The disagreement lies, I believe, in the amount of activities that fall within the categories outlined above, particularly three of them.

Personal communication and writing

Let's start with the point "Is deceptive to fake". Masley uses the example of:

If someone’s messaging you on a dating app, they want to know what you’re actually like.

Very true, but in my view, it's not only in such intimate or private situations where it is deceptive to fake what you are like. Personal communication in general is an area where it matters how we express ourselves, both for ourselves and those we talk or write to. When we communicate with each other, there are certain expectations framing the whole exchange. Letting our words and phrases be transformed by a machine is a breach of those expectations. The words we choose and how we formulate our sentences carry a lot of meaning, and direct communication will suffer if we let language models pollute this type of interaction. Direct communication is not only about the information being exchanged, it's also about the relationship between the communicators, formed by who we are and how we express ourselves.

I think this is not only relevant for communication between two humans, but also for text with a personal sender conveyed to a human audience in general. To a certain extent, the same principles apply. There has been a debate in the Norwegian media lately regarding the undisclosed use of LLMs in public writing, with allegations and opinions flying around. I'm very happy to see this discussion reaching broad daylight, because we need to clarify our expectations to communication, now that chatbots are being so widely used. While I clearly think that it is beneficial to keep human-to-human communication free from an intermediate step of machine transformation, not everyone shares that view. If, going forward, our written communication will for the most part be co-authored with AI models, we need to be aware of it, and shift our expectations accordingly. Some have started disclosing when they have used AI in their writing, which I think is a good step towards better understanding of our use of LLMs. Knowing whether a text is written or "co-authored" by an LLM has an important effect on how a receiver views it; pretending otherwise is simply false.

Many see LLMs as a great boon for helping people express their opinions more clearly, particularly for people not using their native language or those who have learning disabilities. As long as the meaning originates from a person, LLMs can help express that meaning in correct and effective language. I have two main objections against this. The first one is about what happens to the text: In most cases it's impossible to separate the meaning from the expression of it. That is in essence what language is — the words are the meaning. Changing the phrasing changes the message. The second one is about what happens to us: We rob ourselves of the opportunity to grow and learn, without training wheels. LLMs can certainly help people improve the text, but the thinking process — developing the ideas — will be severely amputated when leaving the phrasing up to an AI model. They quickly become a replacement instead of help, depriving us the opportunity of discovering our own voice and who we can be and become when we stand on our own two feet.

With great care, one may be able to use a chatbot without being affected by these two drawbacks, but the problem is that with LLMs, there is an exceptionally thin line between getting help with spelling or grammar, and having the model essentially write for you, thereby glossing over your own voice. This is unavoidable with the current design of chatbots and LLM-powered tools; the step from old-school autocorrect to a generative language model is far too big. If we really envision LLMs as a tool for helping people become better at writing, we need to have a much more carefully considered interface than the chatbots we have today.

At the same time, I realize many are far more utilitarian. They just want to get the job done, finish their work, file that report, get that complaint through, answer that email, in the most efficient way possible, and then get on with their day. Getting help from an LLM to express oneself in a second language also seems useful, without considering how much or little one learns from it (I would be more positive to LLMs for translation if it wasn't for the fact that current state-of-the-art LLMs are simply very bad at producing Norwegian text. I can only hope the state is better for other non-English languages, or that it will improve over time). Additionally, LLMs seem to be efficient for people who are fighting with bureaucracy, such is filing complaints and dealing with insurance companies. In this case the advantage seems greater. We must, however, remember that the "weapon" exists on both sides of the table. What will happen to bureaucratic processes when all parties involved are armed with word generators?

It is not without reservation that I express these opinions, because it may come across as I want to deny people something that looks like a powerful tool. The point is that I think this tool will make you weaker, not stronger. LLMs don't really seem to empower people. Some of the effect I currently see is the number of applications to various calls (internships, research proposals, job openings) multiplying, but the quality dropping. Students are asking chatbots for help with solving collaborative tasks, not realizing that everyone is asking the same chatbot, robbing us of the diversity of ideas that could have formed if they took a minute to think for themselves.

The chatbots may have lowered the threshold for participation, but the competition's ground rules hasn't changed. To get better at writing, you need to write. The same goes for thinking. Applying for a job means showing who you are, not who the LLM thinks you are, or should be. Participating in the public debate is having to work out how to express opinions in clear language. Am I really participating if I'm not finding my own words?

It is important to note that not all text is affected in the same way. The category of writing that I like to call "functional text", which are things like computer code and pure conveyance of information (e.g., recipes, information signs, documentation), is not exposed to the same issues. But text that has a personal author addressing a human audience, has particular role expectations and rests on a particular trust. An erosion of that trust will be a loss for humanity.

A pragmatic attitude would be to just let the inflation of text ensue, and take stock after the dust has settled. What will be left of language afterwards? My conservative viewpoint stems from believing that what we will lose is of greater worth than what we gain. While LLMs can prove useful in the short term, using them is treating a symptom instead of the problem. It is a crutch, although some may truly be in need of that crutch. My only advice would be to make sure you actually need it before you lean on it.

Valuable experiences

Using LLMs is not only about writing. Masley mentions that it's bad to outsource activities that are "a valuable experience on its own". I couldn't agree more, but I suspect that he will disagree when I say that I think this category encompasses a lot of what we already do in life. Major LLM providers love to show how their chatbots can be used to plan vacations, organize parties, and create personal messages to friends and family. I seldom feel more disconnected from the technological society than when I watch these advertisements.

To me, this highlights a problem that goes to the core of what it means to be human. Modern life brings with it a great deal of activities that can feel like chores, but at the same time it seems like we are hell-bent on treating everything as a chore as well. Humans are surprisingly good at finding discontentment in nearly anything, maybe because of an expectation in modern society that we should be able to do anything we want, anytime we want it — or perhaps more importantly, that we should be able to avoid doing things we don't feel like doing. Our inability to see opportunities and fulfillment in life as it is, leads to the inevitable conclusion that life is never enough, and we would always rather be doing something else.

In theory, I agree that automating some things can free up time for other things that are potentially more meaningful and rewarding, but we have already reached a stage where even planning our vacation is a chore that apparently a lot of people would like to avoid doing. I hope that AI's alleged ability to automate "nearly anything" helps us realize what is worth spending time and effort on, and rediscover the value of intentional living.

Building knowledge

The third point I would like to address is that we shouldn't use chatbots when it "builds complex tacit knowledge you'll need for navigating the world in the future", according to Masley. Again, I agree completely, and again, I think that this point encompasses a great deal of daily life. Building knowledge happens not only when you sit down to learn something new, but also when you do repetitive work.

This misconception is not new for chatbots, but has been present since we started carrying smartphones in our pockets. With internet at hand at all times, there's apparently no need to remember information anymore. Instead of using our brains for storing knowledge, we can access information online when we need it, and spend more time learning how to actually use the information and think critically. The point we are missing here, is that acquiring and memorizing knowledge is a huge part of learning to use the knowledge. It is naive to think that we can simply separate the storage unit from the processing unit, like if we were a computer.

I learned this lesson while being a piano student. I was trying to understand jazz, and figure out how good improvisers could learn to come up with new phrases so easily on the spot. How does one practice improvisation? Is it possible to exercise the ability to come up with something new that immediately sound good? I ended up playing similar riffs almost every time I tried. After a while I got convinced that good jazz players must be born with some inherent creativity, some inner musical inspiration that hummed melodies inside their heads for them to play.

One of my tutors taught me the real trick: Good improvisation comes not from just practicing improvisation. You need to play existing songs and tunes, many of them, over and over, learn them by heart, get the chord progressions and motifs under your skin. This practice builds your intuition for what sounds good, and your improvisation can spring from that. Bits and pieces of old melodies are combined into new music. In that sense, we are more like a machine learning model than a computer, but do not make the mistake of thinking that is actually what we are.

There is a need for clarification here: I'm not saying that nothing should be automated by LLMs. But I think many are severely underestimating the knowledge we are building from boring tasks, and we are in danger of losing that knowledge when the pressure for increased efficiency makes us turn to the chatbots.

The extended mind

As a sidenote, I would like to contest the idea of the extended mind, as explained by Masley:

[M]uch of our cognition isn’t limited to our skull and brain, it also happens in our physical environment, so a lot of what we define as our minds could also be said to exist in the physical objects around us.

It seems kind of arbitrary whether it’s happening in the neurons in your brain or in the circuits in your phone.

This statement is simply absurd, even when read in context. The fact that something happens in your brain rather than on a computer makes all the difference in the world. Humans are something more than information processors. Yes, we process information, but it is extremely reductionist to treat ourselves as objects where certain processes can be outsourced to external devices without consequences. Does it really matter if I remember my friend's birthday, when I can have a chatbot send them an automated congratulation? Yes, it matters because in the first case you are consciously remembering and thinking about your friend, consolidating your side of the relationship.

The quoted statement above is followed up with:

It’s true that you could lose your phone and therefore lose the stored knowledge, but you could also have a part of your brain cut out.

Losing your phone and losing a part of your brain are two tremendously different things, both in terms of likelihood and consequences. Not only does the statement above significantly underestimate the processes that happens in our brain, but to even liken having a part of your brain cut out to losing your phone reveals that the premiss of the argument is severely detached from reality.

The design of our built environments is also brought up to show how it's beneficial to minimize the amount of thinking we do:

[M]ost of our physical environments have been designed specifically to minimize the amount of thinking we have to do to achieve our daily goals.

Try to imagine how much additional thinking you would need to do if things were designed differently.

This doesn't hold up to scrutiny. Yes, if our environment suddenly changed, it would require extra mental effort of us to navigate. For a time. But, then we would have gotten familiar with that alternative design, and adapted ourselves. The only case where we would have had to do additional thinking is if the design of our physical environments changed all the time.

What we think about does matter

Regarding the "lump of cognition fallacy", I fully agree that we need not worry about "draining a finite pool" of thinking, leaving "less thinking" — whatever that means — for humans. There is, however, another fallacy at play here, which is that "it does not matter what we think about, as long as we think about something". It is easy to be convinced that if a computer can do the simple, boring tasks for me, I can deal with more complex, exciting stuff myself. But we must be aware that certain mental tasks are important for us to do, even though a machine technically could do them for us.

To illustrate: If I outsource all my boring project administration tasks to a chatbot, it can leave more time for my main task: research. But it will also rob me of the opportunity to feel ownership to the project and build a basis for taking high-level decisions in the project. In a hypothetical situation where a chatbot performs all administrative tasks perfectly on my behalf, I will still have lost something, which may again have impact on the project. I'm not saying that no tasks should be automated at all, but we must be aware that we always lose something when automating a process.

Comparing with the "lump of labour" fallacy again: While it may be true that outsourcing physical work to machines will simply create new types of work to do, it doesn't mean that the new work is useful, fulfilling, or beneficial for individuals and society. The same goes for thinking. We must acknowledge that all kinds of thinking have an effect on us, even the boring and tedious kinds. Removing the need for some cognitive tasks can have just as much influence, positive or negative, as taking up new types of cognitive tasks.

Conclusion

We have a major challenge ahead of us in figuring out what chatbots are suitable for in the long term. Personal communication may change forever (that is to say, maybe it won't stay personal anymore), education systems will require radical adaptations, and we need to reflect more carefully about which experiences in life actually matter. What is truly exciting about this new type of technology, is that it forces us to face questions about our humanity and values. Many formerly theoretical questions of philosophy are becoming relevant for our daily lives.

A fundamental point I'm trying to bring forth is that how we choose to use chatbots is not only about efficiency and cognitive consequences; it's about how we want our lives and society to be. I have tried to argue that there are good reasons for protecting certain human activities against the automation of machines. This is in part based on my values, and does not rely on research into whether or not our efficiency at work or cognitive abilities are affected by it. I cannot tell other people what they should do, but I challenge everyone to consider what values they want to build our communities on, and let that weigh in alongside what the research studies tell us.

Links 8

Wed, 17 Dec 2025 00:00:00 -0000

Less:

When I'm writing, I write. When I'm cooking, I cook. When I'm talking to someone, I put my phone away. The constant mental juggling that felt necessary before now feels exhausting. There's something meditative about giving your full attention to a single task. [...] Less really is more. It usually takes a while to figure that out.

The allure of multitasking is, in my opinion, one of the biggest threat to actually getting things done.

I deleted my second brain:

My new system is, simply, no system at all. I write what I think. I delete what I don’t need. I don’t capture everything.

Refreshing take on the productivity-cult around note-taking.

The cult of hard mode:

Your goal is not to build a perfect system. Your goal is to solve the problem and move on. Think of productivity tools like duct tape or zip ties: functional, replaceable, and emotionally neutral.

In the same vein as the previous link. Choose simple tools. Complex ones get in the way.

We all suffer from OpenAI's pursuit of scale: A sobering podcast episode on the costs of the large-scale investment in chatbots.

I Work For an Evil Company, but Outside Work, I’m Actually a Really Good Person: This made me laugh out loud, even though the message is a sad one.

Sustainable living

Fri, 07 Nov 2025 00:00:00 -0000

Human society is facing multiple crises. Our main method for dealing with them is technology, and we do so ignoring that similar use of technology is what brought about the crises in the first place. We use technology in attempt to remove all boundaries that stop us from getting what we want, when we want it. We are no longer able to adapt ourselves; we only know how to adapt the environment around us. Accepting our own limitations and those of nature is not considered an option, and we make every effort towards finding technological solutions that can make our desired lifestyles sustainable.

For many, the word «sustainability» is dead, either through overuse or misuse or both. I think the term still encompasses how we need to approach the way we live. Most of us are, however, not able, or willing, to see the true costs of our consumption, and therefore cannot understand what it takes to actually live sustainably. Ideally, we would not only embrace sustainability, but also frugality: Instead of using as much as our available resources allow in the long term, we should only be using as much as we actually need, which is not much.

Our attention is incessantly filled by idealisation of materialism. I do not think traditional advertising in itself is the biggest problem. The consumer lifestyle is so thoroughly ingrained in our culture that we perpetuate it ourselves through what we highlight in social media, in conversations over lunch, in parenting, in how we show appreciation of each other. The view on what we «need» is completely distorted, and we allow our personal goals and motivations to be excessively affected by what others seem to want. We are convinced that any problem can be solved by buying something, whether it is getting new clothes or having a far-away vacation.

It is a tough challenge to change our mindsets and start redefining how we approach life. Although I am outspoken about this and feel strongly about the need for change, I struggle to make significant steps towards a more sustainable lifestyle. I am, however, surprised at the general lack of desire to live differently. The effects are no longer in the future: Frequent extreme weather, shrinking biodiversity, the loss of nature¹. It is long overdue that we mentally make the connection between our lifestyles and the resultant costs.

Consumers, companies, and regulators are pointing fingers at each other, and no one can agree who is responsible for taking action. It does not matter who is responsible; what matters is who can do anything about it. The answer is everyone. What I find depressing is that many have no faith in the actions of individuals. Any large reform starts with human beings willing to speak up. Personally, however, what motivates me the most is not whether my actions can actually lead to change, but rather that I want to contribute as little as possible to the destructive exploitation of the planet and of the less fortunate people who are paying the price for the over-consumption by the richest.

To contribute to greater visibility of people and projects that are aiming to change the current trajectory, I am collecting online resources that inspire and inform on the topic of sustainable living: Sustainability. In this context, I use the terms "sustainability" and "sustainable" to mean having respect for our own and nature's limitations, ensuring that humans and nature can co-exist indefinitely with minimal harms and losses. Due to my line of work, I have a bias towards focusing on digital technology, but I also collect resources from other domains.

Achieving meaningful change is virtually impossible within the current system, and grassroots movements outside it is crucial to find new ways forward. Hopefully, many small projects can snowball into a change of the tide.

The Guardian: Europe losing 600 football pitches of nature and crop land a day. It is incredible that we say we are «losing» nature, when we are actively eradicating it in favour of «development». ↩

Bullet points and bold text

Sat, 01 Nov 2025 00:00:00 -0000

Recently I read an article about how the average sentence length in Norwegian fiction books has gone down by 20% since the 1980’s. The results were based on over 7000 fiction books from 1980 to 2020. The study itself did not conclude on the cause of this, but I immediately thought of our decreasing attention spans, excessive smartphone use, and the decline of long-form reading. An author who was interviewed about the results alluded to similar causes, but she also speculated whether the popularity of audiobooks was a contributing factor, because authors adapt their writing to sound good when spoken aloud.

The trend of shorter sentences made me think of how the formatting of text affects both our writing and reading, and I remembered a blog post I read recently. It was written many years ago, by someone who appeared to care little about conventional text formatting. The blog post had no linebreaks whatsoever; it was a single block of text. One massive paragraph, with no entry points for eyeballs wanting to skim through the text, searching for keywords to latch onto. From the style of writing, it did not appear intentional; it seemed more like an oversight. As such, I was forced to read through the whole text from start to finish to even figure out what the text was about. While doing so, I gradually got into a state of reading I seldom enter while surfing the internet. Reading a text with no natural breaks anywhere requires more focus than usual, but it was particularly unfamiliar when sitting in front of a computer screen. The blog post stood in stark contrast to the way a lot of online text looks these days.

I realized then how much the formatting of text affects the reading process. It may sound obvious, but it’s worth considering how different online writing is from printed books, and how it has been changing during the internet era. The emergence of generative language models has brought another wave of change. Many are now familiar with the «ChatGPT-format»: Extensive use of lists, where the first phrase of each bullet point is put in bold text. This type of text formatting lends itself easily to skimming instead of getting properly engaged with the text. It takes little effort to produce, and little effort to consume. I’m not sure why this particular format has become so prevalent for chatbots. It could be an artefact of the data they were trained on, but I think it is more likely to be a conscious decision by the developers to reinforce the easily digested format of lists. Easily digested does not mean healthy, though.

More concise writing can easily be observed in a lot of online newspapers as well: Not only are the sentences short, but the paragraphs are brief, often just one or two sentences. Perhaps it is done this way to be easier to read on mobile screens. I am sure most online magazines track their users thoroughly enough to know exactly how long they spend reading each article and how far they get into the text, giving them ample opportunity to adapt the format to get the most clicks and time spent looking at their content and ads.

I can observe how this change is affecting me. The tendency to favor bullet point lists over continuous sentences, the habit of writing shorter paragraphs, and the desire to highlight my most important points with bold text. These dispositions appear in everything from emails to research papers. Anything to increase the likelihood of getting the message across to the reader, before their attention is sucked away to something else. Context matters, of course, and in some cases bullet points are exactly what you need. But I realize now that this type of writing often may be doing both the writer and the reader a disservice. When I write, the bullet points are useful for superficial thinking. The lists are for itemizing, categorizing, and summarizing. The long paragraphs are what I need to go deeper. The act of stringing together longer sentences is what forces me to formulate an idea coherently, and not just present a set of concepts or ideas. On the other end, the reader sometimes needs to be forced into the mode of focused reading in order to properly understand and digest the message. A bullet point list, a set of short paragraphs, a few words in bold; all of these are opportunities to jump ahead and skip important parts that end up in the background.

It is worth upholding the notion that important ideas need the long-form format to be properly understood. For writers, it is better to have your text understood properly by one reader than skimmed through by many. Dare to write longer when everyone else uses easily digested lists. As a reader, choose to get absorbed in a longer piece instead of glancing through multiple. Quality over quantity goes a long way in both writing and reading.

Couloir skiing in Oppdal

Sun, 10 Aug 2025 00:00:00 -0000

Earlier this year I had the chance to realise a project I’ve thought about for several years: Skiing the couloir called Storskæra on Allmannberget, a mountain in Oppdal, Norway. The couloir is very visible from the highway going through Oppdal, and I’ve thought about it virtually every time we’ve been passing through.

I remember when I first heard about people skiing down Allmannberget, about 15 years ago (though the first descent was probably much, much earlier than that). I was not familiar with steep skiing, and I couldn’t comprehend that it was possible to get down on skis in that sort of terrain. At the time, I wasn’t sure if I believed that people were doing it.

Now, on the other hand, I realise that Storskæra is not very extreme in the context of steep skiing, although you still need to know what you’re doing. While the couloir is over 50 degrees in the uppermost part, most of it is between 35-45 degrees.

In general, I can’t be too picky about the snow conditions because of certain constraints on my calendar. It’s however vital that there’s a low avalanche risk. In late April, the snow was generally stable and safe in the area, but with regards to skiing conditions, the temperature was not on my side. The cold weather had given the snow a hard crust. When I got to Allmannberget, the snow was hard already from the bottom of the couloir, and it got worse the higher I got. I always prefer to walk up the couloir before skiing if possible, so I know what I can expect when going down. This time, I had to wear crampons the whole way up, meaning it wasn’t going to be a very enjoyable run down again.

On my way up.

The couloir is north-northwest facing, so it’s usually preferable to go there late in the day, to let the sun soften up the snow. If I’d had the time, I would have waited a while before skiing. The top part would probably have been much the same anyway because of the cold wind up there, but the lower segments would most likely have had more slushy than crusty snow.

View towards Oppdal.

Almost at the top.

After getting up to the summit ridge, I made a quick visit to the mountain’s highest point before heading back to the top of the couloir. Looking down the chute, knowing that the snow would be hard and icy most of the way, I took a minute to consider other options for going down. I decided however that I didn’t want to go home with unfinished business. I felt confident I could get down safely, it just wouldn’t be as fun as it could be with better snow. The main goal of the day was to do something challenging, and this would qualify.

Ready to descend.

I had to practically downclimb the couloir’s steep entrance while wearing skis, with an ice axe in each hand. The snow was almost as hard as ice, and I struggled getting purchase with my skis. Since I often ski with dogs, I seldom sharpen the edges, reducing the risk of cutting any legs or paws. I deeply regretted skipping that at this point. Further down I could stop hacking at the snow with axes, and use only my legs to control the descent. Carefully making turns, but also just sliding sideways for many meters at a time, to stay in control. Not a pretty sight, but efficient. There were also two sections of exposed rock in the couloir, where I had to remove my skis and move down some scrambling terrain.

Own my way down.

After descending the full length of the couloir on icy conditions, I was tired both physically and mentally, but happy to have completed a long-time goal. As a bonus, I did the neighbouring couloir called "Haninj" directly afterwards. It’s not as long or spectacular, but the skiing was more fun since the snow started getting softer.

Ascending "Haninj".

The main take-away from the day was observing the progress, going from not believing something was doable, to actually doing it. Because of that, Storskæra was a small personal milestone. Although it’s not much of a challenge for experienced skiers (it definitely pales when compared to skiing down a climbing route), it’s worth appreciating the effort it takes to improve.

I often pass through Oppdal, meaning I can enjoy the sight of Storskæra frequently. I hope to repeat it someday when conditions are better, but I also have a list of other couloirs that has to be explored first.

Human recommendations rather than algorithmic curation

Sat, 19 Jul 2025 00:00:00 -0000

Digital social media is hard. There is so much information, creative output, advertisement, and attention-seeking. The «algorithm» is there to help us figure it all out, by sorting, filtering, and prioritising the content in the feed of your chosen social media.

Algorithmic curation gets a lot of critique, and most of it is well-deserved. It’s a tool which on the surface tries to help you find what you’re really looking for in all the noise, but in reality it’s being used to maximise time spent browsing on any given platform, hoping you will not only watch ads, but also click on them. However, this type of curation is not without merit, because the flood of digital information is so large on most platforms that they would be unusable without any sort of filtering.

The "algorithm" is not the only way of solving this problem though. With decentralised social media protocols such as the Fediverse, where Mastodon is a popular platform, you can to a much larger degree control what you get. No additional content will be injected in your feed. No ads, no «suggested for you», no «you may also like». That makes it easy to stifle the noise by simply following fewer people.

For me, the solution to the issue is to simply avoid social media as much as possible, and rely on direct personal recommendations and tips. Those around me have gotten used to sending me links to news articles, podcasts, events, blog posts etc, because they know I don’t discover things through Facebook and Instagram.

I also get a lot of good recommendations through other personal websites. I would like to highlight Tim Hårek and Baldur Bjarnason for regularly sharing links that I usually find worth a read. Tim shares interesting reads in his monthly write-ups, and Baldur has separate posts with collection of links. I also publish «link posts» from time to time with some of the interesting things I’ve found around the web, like this one.

This post ends with an encouragement to make sharing stuff more human and organic. We need to help each other find out what’s worth spending time reading, and not let an automated algorithm decide it for us.

Race report from Black Lake Race 2025

Tue, 15 Jul 2025 00:00:00 -0000

Black Lake Race is a quite new addition to the race calendar in Norway. I had planned to participate in its first edition last year, but I couldn’t go because of more pressing matters at home. The route, with its steep terrain and long off-trail sections, is a good match for me, and there are not many races in Norway with this kind of terrain. That’s why the race was on my priority list for 2025.

The race doesn't have much PR, but it's grounded in the local community with a commitment to true mountain running. The extreme terrain, compared to most trail and mountain races in Norway, doesn’t attract the big masses. Last year Kilian Jornet, who lives in the area, participated and won. I was hoping he would turn up again this year, giving the rest of us a chance to measure up. I raced against him in Isfjorden Skyrace last year, taking second place behind him, and I enjoyed the extra edge the race gets when a true elite runner is at the start line. Unfortunately, he didn’t join this time.

The weather forecast for the day was hot and sunny, which is not in my favour. Earlier this year I’ve done some heat workouts, which seems like the big trend of the year among endurance athletes. I could definitely benefit from heat adaptation, but my daily schedule and other responsibilities makes it hard to prioritise the extra hassle of putting on and off many layers of clothes. In my mind, I get more benefit from spending those minutes (which accumulates to hours) actually training.

My goal was to spend about 3:40 on the course, adding 20 minutes to the course record. I planned to consume three bottles of Maurten 320 during the race, giving me about 240 grams of carbohydrate, and drinking at least 3 liters of water. I know my strengths well enough to portion out my energy in a strategic manner, holding back a bit on the uphills and keeping a good momentum in the downhills.

I started out in a controlled tempo, and got a gap of a couple of minutes by the first checkpoint at 5 km. The next uphill was the main challenge of the race: An ascent of 850 meters of elevation gain in only 2 kilometers. The first part consists of a combination of marshes and bushes. The locals claim there is a trail there, but I couldn’t see it. This particular section is the most brutal part of the course; not only is the slope punishing, but you have to fight for every step because of the soft, squishy ground or bushes up to your thigh.

Further up, the real mountain terrain takes over. I got into a good rhythm, and reached the first summit with a comfortable lead. I could enjoy the scenery as the race course passes over three ridges in rocky terrain, all while having an excellent panorama to the surrounding mountains. The two main peaks of the route, Trollstolen (1141 masl) and Trolltind (1241 masl) are spectacular in themselves, but the view of the serrated ridge between Trolltind and Svartevasstind is the eye-catcher of the landscape. It was nice to be able to appreciate the views, but I got a bit too comfortable running alone, and didn't manage to stay fully in race mode and keep pushing.

The second checkpoint is located between ridge number two and three, and here we joined the 20 km course, which was a shorter option of the Black Lake Race. I got a small motivational boost to speed up by chasing the leader of the shorter race, Emelie Forsberg, up to Trolltind. Down from the last summit, Blåtinden, the markings got very sparse, and for a long time I couldn’t spot a single marker. I tried my best to just follow the GPS track on my watch, but I got into a particularly gnarly terrain on an unnecessary detour. I got frustrated and lost a bit of time. Finally I came onto a trail again, and for the last 6 kilometers I tried to speed up for a strong finish.

I ended up spending 3:51 on the course. Although I’m happy about finishing in first place, the performance was lacking. My original estimate of 3:40 should have been within reach with more motivation to push during the race. In terms of nutrition and hydration, I followed my plan and drank three bottles of Maurten 320 and in total 3.5 liters of water, which felt sufficient to keep a stable pace.

I told the organizers that Black Lake Race is my favourite race course in Norway, and I meant it. I don’t think you’ll be able to find a current Norwegian race that is more pure skyrunning than this. I’m planning to go back next year and improve my time!

Podium of Black Lake Race 2025.

Links no. 7

Sat, 14 Jun 2025 00:00:00 -0000

A few interesting articles I've read around the web:

Choose optimism.

The life of a pessimist is easy but dreary. The life of an optimist is hard but exciting. Pessimism is easy because it costs nothing. Optimism is hard because it must be constantly reaffirmed. In the face of a hostile, cynical world, it takes effort to show that positivity has merit.
Foggy feeds: the decline in my feed reader subscriptions

An observation that persistently crops up in research into the use of Large Language Models is that regular use seems to have a diminishing effect on your critical thinking ability, even in studies that are outright set out to look for justification for using LLMs. […] The very point of “AI” tools is to think less about what you do, think less about the arguments you make, and generally put less thought into your work and your life as a whole. That’s what it’s for.
Smart People Don’t Chase Goals; They Create Limits

Marcus Aurelius, writing in his private journal that we now call Meditations, returned constantly to the idea of limits. He didn’t prescribe grand goals. He reminded himself what not to do: Don’t lie. Don’t whine. Don’t be ruled by impulse. The Stoic path is constraint-oriented. It avoids the seduction of outcomes.
Large Language Manglers

LLMs give us ventriloquism in reverse. The mechanical dummy speaks through your mouth.
The West is bored to death

Suppose Schopenhauer is right that life boils down to a flight from either boredom or pain. Insofar as the vast material abundance of wealthy, industrialised society has had an analgesic effect (there is simply less physical pain than in the past, before fluoride and anaesthesia and sedentary lives), it would seem to have solved one problem only to amplify the other. In place of pain, we have ennui, the quintessential modern condition. It follows directly from overabundance: an endless stream of video “content” or chocolate cake or edibles or any other indulgence cannot deliver lasting satisfaction. Everything gets old eventually, leaving one to grope around for the next fix.
The myth of automated learning

Armed with generative AI, a B student can produce A work while turning into a C student.

Additionally, I have a bonus link for Norwegian speakers (though I recommend it to everyone else as well, if you have a convenient translation tool):

Digitaliseringsparadokset (The digitalization paradox)

I prosessen med å bli verdens mest digitaliserte land risikerer vi å få verdens mest overvåkede innbyggere. (In the process of becoming the world’s most digitalised country, we risk getting the worlds most monitored people.)

The wrong question about AI

Sun, 08 Jun 2025 00:00:00 -0000

The progress of AI is interesting and important to follow, but current discussions around it are too focused on the increasing benchmark performances and if machines can ever get smarter than humans (superintelligence). Counter-arguments to the promise of achieving so-called «superintelligent» computer programs can typically take the form of «humans are still better than AI at this», or «AI can never replace humans at that», but they are missing the point. Asking whether AI can outperform humans is the wrong question. The real question is: What are appropriate uses of AI? Or simply, what do we actually want to automate?

One of the main reasons for developing AI technology is to increase efficiency through automation. We want the computer to solve certain tasks on its own. The problem we now face is that automation technology is so advanced that there seems to be no limit to what we can automate. Given enough time, money, and resources, we can make machines that are, or at least appear to be, more efficient than humans for virtually any task, especially if we only evaluate them on task efficiency and ignore any ripple effects. This of course depends entirely on what metrics you use to evaluate efficiency, but any concretely defined metrics can theoretically be optimised for when training an AI model.

From a societal point of view, it is pointless arguing whether AI or humans are better at a certain task without first asking whether the task is suitable for AI in the first place. This topic was brought up already in 1976 by Joseph Weizenbaum in his book Computer Power and Human Reason, and has only become increasingly relevant. Generative AI-models’ ability to process and generate text leaves practically no area of society out of reach. This is exactly when we need to not only worry about performance and reliability of AI systems, but how we actually want to use them.

Consider for example AI chatbots for keeping lonely elderly people with company. This suggestion comes not only from outside the target group but also from elders themselves. With limited personnel and resources for elderly care, why not have an AI companion that can easily be individually personalised and is always available? There is no doubt that the current AI chatbots can generate human-like, empathic conversation, but do we really want to use computers as company for our elders?

The question is hard, because it forces us to search for our common values. Does it matter if we replace care-givers with machines? If studies show that it works (according to certain metrics) and it’s difficult to employ enough people, then why not?

While a chatbot might be better than nothing, we can certainly do much better. Interpersonal relationships are a cornerstone of both society and our individual lives, and of all the things we would like to automate, it should be far down the list. The challenge of human loneliness is one area where political solutions must take precedence over technological ones, such as investing in care workers and community programs. And similar dilemmas appear when AI-systems are used as therapists, companions, or other forms of direct replacements for human relations.

While Weizenbaum had the luxury of discussing hypothetical scenarios, we are facing concrete challenges. When new technologies appear, we can always choose whether to use them or not. The answer depends not only on the capabilities of the technology, but what we value as a society. Once the choice is made, it’s often hard to reverse it.

Just because we can automate human connection doesn’t mean we should. Let’s use machine intelligence to automate machines, not humans.

Thoughts on desktop computer automation

Wed, 04 Jun 2025 00:00:00 -0000

I recently came across a review of Sky, an upcoming AI automation app for macOS. The technical aspects piqued my interest and made me click the link, because I work with AI engineering research. It was the societal implications, however, that made me read the whole thing from start to finish.

In short, Sky is an AI-assistant that integrates deeply into macOS, and can interact with any app, independent of whether the app has been developed for AI-integration or not. The features of Sky are impressive, and based on the review it looks like the developers have managed to create something quite innovative.

However, the promises of «natural computing» left me less daunted than the potential issues such software creates.

Privacy is (more) dead

I find it incredible that privacy isn’t mentioned a single time in this review. It’s the first thing that pops up in my head when reading about the tool, and it should be the first thing to check before even thinking about using something with this level of access.

The tool’s core strength, being able to ingest anything you do on your computer, also becomes its greatest weakness: Any data you feed it will be sent to the LLM-provider powering the tool (either OpenAI or Anthropic, according to the review).

One thing is the user’s own data, but what about all those «skyshots» (Sky’s advanced screenshots that captures what you’re doing on the computer) of other people’s messages, among other things, as shown in the main demo on Sky’s homepage? While screenshotting and saving people’s messages always has been possible, this app makes sharing such data with a third party a part of its core functionality. A Sky user will be accepting the privacy policy of both Sky and OpenAI/Anthropic on behalf of anyone whose data appears on the screen, likely including info such as names and email addresses.

Improved performance of local LLMs may enable a more private version of this tool, but let’s face it: The best models will always be the more power-hungry variants in the cloud, which also are the ones that generate the most income for providers.

The appeal of delegation – no matter the cost

Another thought that came to mind when reading the review was how sometimes instructions for an LLM end up being longer than what you want it to produce. In one example, the reviewer wants to send a link and a summary of it to a friend, and after receiving the first LLM-written draft, he gives the following command:

also, start the message by saying «this is cool» followed by the link

Why bother giving a command that is twice the length of what you want to write? This particular example could of course be explained by several factors (f. ex., «I was already writing in the command prompt, and it would have taken more time to move my cursor for editing the text»).

However, I don’t think this type of instruction-giving uncommon, and I’ve seen similar cases of people asking using AI-assistants for programming. Why would we give an instruction when doing the action ourselves takes equal (or even less) amount of time? Perhaps it is something about the cognitive difference between giving an instruction and having to perform an action. It seems to me that there is something comfortable about sitting in the driver’s seat and command, rather than do the actual work.

While this is seemingly not a big issue and just a nitpick in the context of the review, it becomes more serious when considering the large energy consumption of LLMs. While prompting a chatbot has a minuscule carbon footprint compared to many other things many of us do in our daily lives (driving a diesel car, taking long hot showers, etc), the relative benefit of running AI-automation is probably not worth the cost for such extremely trivial tasks.

Automating communication

The allure of automation and increased efficiency is tempting, and part of me gets intrigued and excited when hearing about such technical advances. This is in spite of the fact that I have a quite analog lifestyle, and have few processes I would need to automate with this tool. The concept of computers is after all making data-processing more efficient, so why wouldn’t we take every opportunity to automate? A lot of computer work is simply boring and unrewarding, and not every user interface is as good as it perhaps should have been. Shouldn’t we use AI to save us some time?

It’s not only about privacy, or the increasing need for computational resources which puts pressure on energy, water and hardware supply. These are important issues, but I also worry about the aspect of letting automation sneak into the realm of inter-personal relationships. I think it’s essential to keep communication between us authentic and genuine.

I still get surprised by how many demos for AI-automation showcase how you can automate communicating with other people, including friends and family. Why do we need to make that more efficient? It appears to be fostering the same sort of friendships that having hundreds or thousands of «friends» on Facebook does.

Conclusion

In the introduction, the reviewer writes:

As soon as I saw a demo, […] I knew Sky was going to fundamentally change how I think about my macOS workflow and the role of automation in my everyday tasks.

Well, as soon as I read this review, I knew that this kind of automation tools will kill privacy, needlessly accelerate our daily lives even further, and take the «personal» out of personal communication. That is, if it catches on. And I suspect it will, in some form or another.

When choosing digital tools

Wed, 12 Mar 2025 00:00:00 -0000

When considering to buy a digital device, service or tool (or any type of technology for that matter), it is worth thinking about the following:

What problem does it actually solve for me?
What will I lose by using it?

When deciding on a specific product, particularly for physial devices, there are another set of questions to consider:

Does the device have an easily replaceable battery? This component is typically the biggest threat to the longevity of a device.
How easy is the device to repair in general, if some parts get broken (e.g., screen, charging port, buttons)?
Does the core functionality of the tool work without internet access, or connectivity to an app?
If the tool's functionality includes storing your data, how easy is it to export that information and be used on another device or platform? Do you get tied into a given platform? Do you retain full control of your data?

Taking these aspects into account can help you choose more sustainable, privacy-friendly and humane technology.

Not only must we be mindful of what effects certain technologies have on our minds and our society, especially when deployed on an enormous scale, but the rapid obsolescence of digital devices is simply extreme.

The considerations above have led me to replace dozens of apps with a paper notebook, and I often carry a feature phone and a camera instead of a smartphone. The downstream effects include a drastic reduction in the time I stare into a screen, being more present in the physical world, and having my focus protected against the attention mining of internet-abusing companies.

Choose wisely.

Links #6

Mon, 03 Mar 2025 00:00:00 -0000

On personal websites and social web.

On the other you have the people who are realising that maybe the solution is not to recreate social media but rather to abandon it and go back to a more deliberate way to be social online, using personal sites, small forums, emails, and other “traditional” tools. I’m obviously part of this second group.
- This is an acute observation, and it is related to what I have written earlier about the future of social media. I'm also in the second group. After taking a step back from not only social media, but also online news, I find myself in a much calmer state of mind.
FOMO is not a strategy

There is a small but significant number of people who really like generative AI. This isn’t the same as the number of people who’ve used (or claim to have used) genAI, but a subset of superfans. I tend to meet two kinds of people in this category: enthusiastic senior leaders who have been tasked with doing something innovative but might not be hands-on users, and more junior people who are enterprising and computer-literate and want to try something new. Notably, some of the people in the second group might have a natural affinity for prompt writing but not enjoy another part of their job – like writing long documents or emails – and so are pleased to outsource those tasks to AI.

Guardians of the edge of chaos.

But let’s be clear, it’s not “humanity” as a whole that’s exploiting the earth. It’s the top 0.1% exploiting everything and everyone else. The majority of humanity is just as exploited as the forests, oceans, and animals. Choosing exploitation means embracing a brutal, materialistic race to the top. A system where everyone except a tiny minority loses.

A Society That Lost Focus

Our mind, not the technology, is the bottleneck. We need to care about our minds. To dedicate time to think slowly and deeply.

[NOR] KI og klimaavtrykk på NRKs Abels Tårn

Mon, 10 Feb 2025 00:00:00 -0000

Forrige uke var jeg så heldig å få fortelle om vår forskning innen mer bærekraftig kunstig intelligens på en av Norges største populærvitenskapelige radioprogram, nemlig Abels Tårn på NRK!

Først og fremst var det veldig gøy å få være med, og det var en flott sjanse til å fortelle om hva jeg holder på med til daglig. Ikke mange er klar over at det jobbes aktivt i Norge med forskning for å adressere klimaavtrykket til KI, og på min forskningsgruppe (Green IoT) leder vi forskningen innen «grønn KI» i EU-prosjektet ENFIELD¹. En av de tingene vi driver med akkurat nå er å kartlegge energiforbruket til ulike generative språkmodeller, og hvordan ulike faktorer påvirker både energieffektiviteten og ytelsen.

Er det en ting som er sikkert, så er det at det er mange estimater, usikre tall, og manglende kildehenvisninger ute og går når det gjelder strømforbruk og klimaavtrykk KI-modeller, spesielt språkmodellene bak ChatGPT. Generelt sett har vi så lite kunnskapsgrunnlag om modellene til OpenAI, at vi kan ikke si med sikkerhet hvor mye en ChatGPT-spørring (-prompt) koster i verken strøm, vann eller karbonutslipp. Likevel må man innimellom prøve å konkretisere forbruket litt ved å ta i bruk slike estimater, selv om det ikke alltid er så lett å forklare konsist hvordan disse beregningene er gjort. Utviklingen går så fort vil de fleste estimater foreldes ganske raskt, med nye språkmodeller som dukker opp.

Det overordnede bildet blir likevel stående: Generativ KI har et signifikant ressursforbruk som vi er nødt til å ta hensyn til når vi vurderer hva vi skal bruke denne teknologien til. Regulering og åpenhet trengs for å adressere problemet.

Jeg håper intervjuet kan være med på å gjøre KI litt mindre mystisk, litt mer håndfast, og å forklare hvorfor klimaavtrykket til generativ KI er så stort sammenlignet med andre former for digital teknologi.

Tusen takk til radiovertene i Abels Tårn, Torkild Jemterud, Kristina Aas og Hanne Aass, for å invitasjonen og en meget hyggelig samtale rundt temaet!

Hør på episoden her: Abels Tårn 7. feb 2025 (start fra 29 min)

ENFIELD-prosjektets hjemmeside. ↩

Better images of AI

Fri, 07 Feb 2025 00:00:00 -0000

As AI is discussed in all kinds of media, futuristic images abound in news papers, presentation slides, blog posts, and social media. The problem with many of these images is that they muddle the impression of what we are actually talking about when we discuss AI, which in many cases are simply a huge number of additions and multiplications being performed on a computer (yes, that's what the chatbot you're using as an orcale really is).

I just discovered a handy resource when you need illustrative images for AI-related topic: Better images if AI. As the front page of the website says (emphasis mine):

Abstract, futuristic or science-fiction-inspired images of AI hinder the understanding of the technology’s already significant societal and environmental impacts. Images relating machine intelligence to human intelligence set unrealistic expectations and misstate the capabilities of AI.

The website is run as a collaboration between multiple organizations (BBC, We and AI, Leverhulme CFI), where they share more meaningful AI-related images. They are free to use as long as an attribution is included.

Now you don't even have to waste electricity on asking Midjourney to generate illustrations for your slides!

Links #5

Mon, 02 Dec 2024 00:00:00 -0000

Another collection of interesting reads from around the web:

Leaving the phone at home. This is an interesting exercise. I'm surprised by how different I feel just by leaving that device at home. It's a bit worrying.
30 bits of advice for 30 years.
Solitude and Leadership.
Stephen Fry Reads Nick Cave’s Stirring Letter About ChatGPT and Human Creativity: “We Are Fighting for the Very Soul of the World”
Confessions of a Viral AI Writer. About ChatGPT and the writing process:

I recalled Zadie Smith’s essay “Fail Better,” in which she tries to arrive at a definition of great literature. She writes that an author’s literary style is about conveying “the only possible expression of a particular human consciousness.” Literary success, then, “depends not only on the refinement of words on a page, but in the refinement of a consciousness.”
How it feels to get an AI email from a friend
Jim Nielsen: It’s Humans All the Way Down.

AI's sustainability issues are gaining awareness

Tue, 08 Oct 2024 00:00:00 -0000

I'm happy to see that there is an increasing focus on the challenge regarding AI and sustainability. Not only because I'm a researcher trying to make AI "greener", but because AI's booming resource consumption is starting to become a real problem. The last few months I have seen many articles bring up some of the issues, and I would like to highlight a few of them.

Microsoft playing both sides

The Atlantic writes about "Microsoft's hypocrisy on AI"¹, describing how Microsoft claims to take responsibility for climate action, while still having large scale cooperation with fossil-fuel companies. While the article's focus is mainly on the application of AI, I think the following observation about AI's own climate footprint is important:

The idea that AI’s climate benefits will outpace its environmental costs is largely speculative, however, especially given that generative-AI tools are themselves tremendously resource-hungry. Within the next six years, the data centers required to develop and run the kinds of next-generation AI models that Microsoft is investing in may use more power than all of India. They will be cooled by millions upon millions of gallons of water. All the while, scientists agree, the world will get warmer, its climate more extreme.

There are many possible ways in which we can use AI for limiting carbon emissions, but widely deployed chatbots and image generators are not among them. We have to stop generalizing the benefits of one AI application to support any type of AI usage.

The article is based partly on an interview with Holly Alpine, previously Microsoft's Sr. Program Manager of Datacenter Community Environmental Sustainability. She has written her own account on why she left the company², which gives additional insight to the problem:

Internally, discussions highlighted the most significant application of generative AI for the oil and gas industry: optimizing exploration activities. AI was celebrated as a “game changer” and the key for the fossil fuel industry to continue to be competitive.

Again, AI can be used to fight climate change, but it certainly can be used to hasten it as well.

Big Tech's emissions are surging

There is no doubt that the enormous investment in AI, and particularly generative AI, is driving energy consumption and emissions up for many companies. NPR reports that both Google and Microsoft have significantly increased carbon emissions³. Google's emissions are up by 48% since 2019, and Microsoft's have increased by 29% since 2020.

This is despite the fact that both companies have pledged to be carbon-neutral by 2030. In both cases, the need for scaling up infrastructure to support AI is given as a reason for why emissions are soaring.

Generative AI and energy-efficiency

I was lucky enough to be interviewed about the energy consumption of large language models in the online news magazine Heatmap News⁴. Slightly less known than The Atlantic and NPR, but even so, I cherish every opportunity to talk about my research in more public channels.

I tried to drive home the point that generative AI is in general very expensive, and that we need to think hard about when and where it is actually worth using it. It was mainly the first point that made it through to the final article, but the second one is just as important.

I commented specifically on the new o1-models from OpenAI, which use so-called "chain-of-thought" to increase their performance in problem solving. While the effort to reduce the hallucinations of LLMs is laudable, the new models are almost certainly much more expensive than previous models. To quote myself:

As Erik Johannes Husom, a researcher at SINTEF Digital, a Norwegian research organization, told me, “It looks like we’re going to get another acceleration of generative AI’s carbon footprint.”

Let's hope we have some breakthroughs soon to make models more efficient, not more resource-demanding.

Will "greener" AI help?

There is an underlying challenge with working towards making AI (or any type of technology) more energy-efficiency: The cheaper it gets to use something, the more likely we are to use more of it.

I'm worried that our efforts to make AI "greener" will only lead to more unecessary consumption.

Nevertheless, increased awareness is a good beginning for making a change.

Hao, K. (2024, September 13). Microsoft’s hypocrisy on AI. The Atlantic. https://www.theatlantic.com/technology/archive/2024/09/microsoft-ai-oil-contracts/679804/. Archive.org: https://web.archive.org/web/20240913122346/https://www.theatlantic.com/technology/archive/2024/09/microsoft-ai-oil-contracts/679804/. ↩
Alpine, H. (2024, September 18). I loved my job at Microsoft, but I had to resign on principle. Here’s why. Fortune. https://fortune.com/2024/09/17/microsoft-environment-oil-fossil-fuel-expansion-climate-change/. ↩
Kerr, D. (2024, July 12). AI brings soaring emissions for Google and Microsoft, a major contributor to climate change. NPR. https://www.npr.org/2024/07/12/g-s1-9545/ai-brings-soaring-emissions-for-google-and-microsoft-a-major-contributor-to-climate-change. ↩
Zeitlin, M. (2024, September 17). What Does OpenAI’s New Breakthrough Mean for Energy Consumption? Heatmap News. https://heatmap.news/technology/openai-o1-energy. Archive.org: https://web.archive.org/web/20240917231307/https://heatmap.news/technology/openai-o1-energy. ↩

Dotfiles: Starting over after eight years

Thu, 03 Oct 2024 00:00:00 -0000

A couple of months ago I decided it was time for a digital deep clean of my laptop. It was close to three years since last time, and over that period a lot of unnecessary software and data has aggregated in the nooks and corners of the system. The easiest way to achieve a clean system is to reinstall the operating system, which also gave me a chance to upgrade Ubuntu from 22.04 to 24.04.

When setting up a new system, one of the first things you want to do is to install and configure your most used programs. As many other programmers and Linux users, I have a dotfiles folder that contain my precious configuration files for all the vital utilities I use.

For me, the most important ones are bash, vim and tmux.

Time for a clean slate

It's now eight years since I entered a computer terminal¹ for the first time, and over that period a lot of things have accumulated in my dotfiles. I have been using different Linux distros, macOS, and Windows Subsystem for Linux, and have made my configuration adaptable to several different setups.

At the same time, I like to think of myself as a minimalist, in some sense. I like to keep my system as fast and uncluttered as possible. I'm not sure if my dotfiles reflect this, but I know I'm quite restrictive compared to some "vim maximalists".

My .vimrc has grown to ~500 lines, which is excluding the vimscripts, plugins and templates I use. I have a .bashrc of ~400 lines, and a .tmux.conf of ~60 lines.

In addition to that, I have 65 bash scripts, counting ~5500 lines in total. These are not really configuration files, but I include them in my dotfiles as a part of my bash setup.

My dotfiles are by no means bloated compared to other software I use regularly, but I wondered: How much of this I actually need?

With a newly installed operating system, I thought it would be an excellent opportunity to start from scratch, and only add back things I actually felt the need for.

Now, after about 7 weeks since the clean install, I think it's time to assess the status. What parts of my dotfiles did actually make it on to the next phase?

Bash

In bash, I added back some lines quite quickly, mainly history settings and aliases. I use many shortcuts all the time, and they are firmly etched into my muscle memory:

Expand to see .bashrc (53 lines)

# Use vi keybindings on command line
set -o vi

# Prompt appearance
export PS1="\[\e[31m\]\u\[\e[m\]@\[\e[32m\]\h\[\e[m\] \[\e[34m\]\w\[\e[m\] \\$ "

# Adjust history size
HISTSIZE=20000
HISTFILESIZE=20000
SAVEHIST=20000
# Ignore duplicate entries in the history
HISTCONTROL=ignoredups:erasedups
# The following lines are needed to save history from multiple sessions
# append history entries..
shopt -s histappend
# After each command, save and reload history
PROMPT_COMMAND="history -a; history -c; history -r; $PROMPT_COMMAND"


# Aliases

alias x='exit'
alias ..='cd ..'
alias ...='cd ../..'
alias ....='cd ../../..'
alias .....='cd ../../../..'
alias ......='cd ../../../../..'
alias v='vim .'
alias ll='ls -AhlF --color=auto'  # human-readable file size
alias la='ls -A'
alias l='ls -CF --color=auto'
alias duh='du -h -d 1'                    # Show file human-readable file sizes
alias df='df -h'                          # Show disk usage with human-readable sizes
alias cp="cp -i"                          # confirm before overwriting something
alias grepr="grep -rn . -e"

alias g='git status'
alias ga='git add'
alias gc='git commit'
alias gcm='git commit -m'
alias gd='git diff'
alias gp='git push'
alias gpl='git pull'
alias gl='git log --graph --oneline' # show git log as graph
alias gcred='git config credential.helper store'
alias gitmail='git config user.email'
alias gitname='git config user.name'
alias gclone='git clone'
alias gall="find . -maxdepth 1 -mindepth 1 -type d -exec sh -c '(echo {} && \
    cd {} && git status -s && echo)' \;" # check git status of all subfolders
# git pull on all subfolders:
alias gplall="find . -maxdepth 1 -mindepth 1 -type d -exec sh -c '(echo {} && \
    cd {} && git pull && echo)' \;"

The configuration file itself was reduced down to about 25% of its previous size, but the biggest difference is for the bash scripts: Out of 65, none have made their over to my freshly installed system.

All the time I spent writing scripts to automate certain actions seems to have gone to waste (although I probably learned a lot in the process). I guess I miscalculated how much I would actually use those scripts in my daily work.

Tmux

For tmux, I needed about half of my original configuration:

# Remap prefix to Control + a
set -g prefix C-a
# Bind 'C-a C-a' to type 'C-a', making sure 'C-a' works in other tools
bind C-a send-prefix
# Remove original prefix binding
unbind C-b

# Set vi mode
setw -g mode-keys vi

# Move to different windows using vim keybindings (hjkl) instead of arrow keys
bind-key h select-pane -L
bind-key j select-pane -D
bind-key k select-pane -U
bind-key l select-pane -R

# Name windows after directory
set -g window-status-format '#I:#(pwd="#{pane_current_path}"; echo ${pwd####*/} | cut -c1-20)#F'
set -g window-status-current-format '#I:#(pwd="#{pane_current_path}"; echo ${pwd####*/} | cut -c1-20)#F'

# Open new windows and panes in same directory as active
bind c new-window -c "#{pane_current_path}"
bind '"' split-window -c "#{pane_current_path}"
bind % split-window -h -c "#{pane_current_path}"

I didn't think my tmux-configuration was that important, but when you work mostly in the terminal, it's important to be able to navigate with ease.

Vim

My setup has reduced drastically for vim. Going from ~500 lines, 7 vimscripts, and 3 plugins, I'm now down to 9 lines and a single script:

set expandtab       " Use spaces instead of tabs
set tabstop=4       " Number of spaces a tab counts for
set shiftwidth=4    " Number of spaces to use for each level of indentation
set softtabstop=4   " Number of spaces that a tab counts for while editing
set autoindent      " Keep the same indentation as the current line
set smartindent     " Automatically add extra indent in certain contexts (like after `{` in C-like languages)
set number          " Show line numbers
filetype indent plugin on   " determine file type and use auto-indenting
source ~/.vim/vimscripts/commentary.vim " Comment/uncomment with gc[motion]. Credit: Tim Pope <http://tpo.pe/>

I currently don't feel the need to add any plugins to vim. Previously I used fzf² quite a lot, both in bash and vim. It is an awesome tool, and it felt like I supercharged my efficiency in the terminal when I started using it.

Now, I realize that the gains were not as large as I thought, and that can do my work efficiently without fuzzy search. I can easily navigate and find what I need without relying on a complex integration of fzf, bash, vim, find, bat, grep, and ripgrep, which is what I had before (and spent a long time on setting up). Maybe one day I'll be back there, but not until I feel the need for it.

Conclusion

All in all, I love how these tools are designed to be easily configured, adapted, and integrated with eachother. It's how all software should be designed: Giving people power over the tools they use.

But I also think it's important to step back once in a while and think about how we use the tools, and what tools we actually need.

It is a good reminder to observe that what I once thought of as a large improvement, actually did not bring as much benefit as I thought. Maybe that applies to other things in life as well.

Technically a terminal emulator. ↩
https://github.com/junegunn/fzf ↩

ChatGPT vs. your voice

Wed, 12 Jun 2024 00:00:00 -0000

In my job I write a lot of technical text, whether it is for reports or scientific papers. As many others, I have been experimenting with how ChatGPT or other AI tools can be of use for so-called "knowledge workers".

I have noticed two particular things with the text produced by ChatGPT that annoys me:

The generated text often contains a lot of unnecessary, flowery adjectives and adverbs. In my opinion, the language quickly starts to sound like your marketing something, or you're just being pretentious. Examples: "meticulous(ly)", "comprehensive", "diverse", "intelligent(ly)", "precise", "crucial", "seamless(ly)".
The generated text often contains unnecessary complex synonyms for verbs. I have mainly experimented with ChatGPT for rewriting a piece of text to improve the flow, and very often the verbs I originally used are swapped with synonyms that in the worst case changes the meaning significantly, or doesn't make sense at all. Examples: "encompass", "elucidate", "encapsulate", "delve", "comprising", "boast", "fostering".

There's absolutely nothing wrong with these words in themselves, but the way they are overused by ChatGPT makes the text sound very unnatural.

I have, however, a greater concern regarding how ChatGPT modifies and changes the text we use to communicate information. Using an LLM as a tool for expressing yourself means that it's no longer your voice, or at the very best, it's a modified version of it.

LLMs are making their way everywhere, with ChatGPT being only the beginning of AI-access for the greater masses. Apple recently revealed Apple Intelligence¹, which is their new AI platform that is "built into your iPhone, iPad, and Mac to help you write, express yourself, and get things done effortlessly." Microsoft, Google, and Meta have already deployed some AI tools that are deployed into their platforms, at least in some regions, but I see Apple Intelligence as the next big step of AI coming into the daily lives of larger groups of people.

ChatGPT has attracted a lot of users that have an explicit interest in using AI-tools, but platform integrations will bring AI also to those that do not necessarily feel the need or want for it. What consequences will this have? Apple claims that their new AI platform will "help you write" and "express yourself"¹, but how much are you really expressing yourself with an AI tool as the middle-man?

LLMs has a myriad of use cases, some good and some bad. In my opinion, after one and a half year of observing and experimenting with the chatbot since its release, I think we should be careful about using it as an intermediate layer of personal communication between people. It is not a matter of whether AI will be "good enough" at mimicking our personal style or not. I think there is irreplacable value in the fact that the sender expresses their own thoughts. Receiving AI-generated communication is completely different from something written by the sender on their own, as has been noted by others².

The landscape of textual communication is changing, and I'm not sure how well my thoughts on this will hold up five years from now. Maybe the expectations for digital communication will change significantly along with the development of AI tools, rendering these concerns outdated. But with so much communication already happening online instead of in the real world, I'm very concerned about how this will affect our inter-personal relationships. Of course, it's a huge difference between, let's say, an SMS and a technical report, and the demands for productivity in the workplace may make it impossible to avoid using AI-tools to speed up communication.

However, I know that when I have a discourse with another human, I prefer reading or hearing the actual words chosen by the other person, because there is also a lot of information in those choices. I prefer hearing your voice, not the one suggested by ChatGPT.

Links #4

Mon, 04 Dec 2023 00:00:00 -0000

Another collection of interesting reads from around the web.

The Verge: Social media is doomed to die. Intriguing perspectives from someone who used to work for Snapchat.
Baldur Bjarnason: The Elegiac Hindsight of Intelligent Machines. An interesting piece about modern software, AI and ChatGPT.
The Fire: Are we ‘Amusing Ourselves to Death’ in 2023?. When you read the book Amusing ourselves to death by Neil Postman, originally written about how the television affects our society, it's not difficult to see how all his criticism could be applied to the internet and social media. It would be interesting to know what Postman would have said, had he seen how things have evolved after the advent of the internet.
Discourse Magazine: The Psychopathology of Digital Life. What is the cost of living so much of our lives in the digital world?
koray er: Low tech and my old notebook. The first paragraph of this post resonates with me: > in the past, i've been quite frustrated by unnecessary large-scale technologies. these are technologies that only work because they rely on the advances of modern civilizations, such as electricity or complicated manufacturing processes. the opposite is small-scale or low technology. it's the difference between a vacuum cleaner and a broom, a radiator and a jacket, or maybe even sparkling water and tap water.
Going off-script, a blog post that asks some questions regarding our life choices that are worth considering.
The Athlantic: My North Star for the Future of AI. Important perspectives on AI, academia's role and ethics. It's a particularly interesting read if you work in academia.

Sunshine and solitude on Mt Fuji

Fri, 24 Nov 2023 00:00:00 -0000

I was fortunate enough to get the opportunity for a travel to Japan for a work-related conference earlier in November. Since I've never been to Japan, and most likely won't go there again, I wanted to stay a few extra days for a small vacation. My main goal was to see some of the Japanese nature, and ideally visit some mountains. I researched the mountain areas of Honshu (the largest island of Japan), which are often referred to as the Japanese Alps, but going there would require spending too much of my precious time on trains and buses. However, Mount Fuji, Japan's highest peak, was not too far from Tokyo, where my flights would arrive and depart. When I only had time to climb one mountain, the tallest one is a natural choice.

While Mount Fuji is visited by large amounts of people during the summer, it's a different story in the so-called off-season. The authorities go so far as to close all the trails, and based on the information I found it is required to submit a form to the police in order to request permission to scale the mountain. Many websites portrayes an off-season climb as extremely dangerous, and disencourages anyone to attempt it.

I downloaded the form, where it is required to provide information about all the members of the expedition, what gear you have, and how much food you are bringing. I felt the chances of being given permission for a one-man "expedition" in a one-day push was very low. It didn't really make sense that they would be this strict for such a non-technical mountain, but I can understand that the authorities want to avoid unnecessary rescue operations. I had a feeling that it should be possible to climb the mountain without too much fuzz. I found a guy through social media who had been to the summit earlier this fall, who confirmed this. He had no problem getting access to the mountain, and the climb was not as demanding as they make it seem. Crampons and ice axe was necessary, but otherwise it should be fine. I packed the suggested ice gear, along with shell clothing, down jacket and an emergency shelter. I was unfamiliar with the weather and climate in the region, and wanted to be prepared for cold temperatures and wind.

When I arrived in Japan, the weather did not look good for the day I had planned to summit. Up until the morning of summit day I had actually decided to skip the climb, and rather focus on seeing more of the Japanese culture. The whole mountain was completely covered in clouds, and I didn't want to spend my very limited time in Japan walking in fog, not seeing anything, and most likely have to turn around before the summit. I started the day with a run past the Arakura Shengen Shrine, and up into the forest-covered mountains around Fujiyoshida, where I was staying. However, the desire to ascend a real mountain became too strong, and I changed my mind. I went back to the hotel, packed my gear, and ran to Kawaguchiko, where I jumped on the bus to Fuji-Subaru 5th Station. The 5th Station is the most common place to start hiking in the summer. According to what I read online, the bus didn't operate past October, but evidently that was wrong.

When we approached the final stop at 2300 masl I realized that the clouds were quite low, and above this altitude the sun was shining. After a brief look in the souvenir shops, I made my way towards the trail, where you have to pass a large barrier with signs informing you that the trail is closed for the season. Another barrier is further down the trail, just before the path starts ascending up the mountain.

Signs indicating the closing of trails for the off-season. The orange barrier can be seen in the background.

The first part of the trail is what I would call a volcanic gravel road. After a while it gets more rocky, and you're walking on a proper "trail". The path is fenced by chains and ropes almost the whole way up, so it's virtually impossible to get lost. At about 3000 masl I had to put on crampons. There was not a large amount of snow, but the path and rocks surrounding it was covered in relatively hard ice that would have made it impossible to move upwards without spikes. The thinner air was getting noticeable, and I had to move slower than normal, with my breathing getting heavier.

At the bottom of the trail.

Finally up at the crater rim, I could enjoy a fantastic view of both the crater itself, and the never-ending sea of clouds that surrounded the mountain. At the opposite side of the rim was my final destination: the highest point of the mountain, Kengamine. The wind was quite strong here, so I often had to lean into the wind in order to move forward. As the peakbagger I am, I walked over all the small peaks surrounding the crater on my way to the highest point.

Cloudy weather looks different from above.

A weather station is placed on top of Kengamine, and here the wind was particularly strong. I snapped some photos, and walked a bit down again before I took a break to eat some snacks and enjoy the view. It was a fantastic feeling to experience Mount Fuji towering over the clouds, sun shining, and having the whole mountain for myself (as far as I could see at least, and I saw no other tracks in the snow and ice). And to think that I almost decided not to go!

View of the crater from Kengamine.

I continued to follow the rim back to where I had come up, completing the crater circuit. At one point I fell down a rocky slope, heading towards the top of a small cliff. A sudden, forceful gust of wind pushed me off my feet, and one of my crampons got caught in the ice, preventing my from finding my footing. I tumbled down the icy rocks, going head over heels a couple of times before I manage to stop myself. Luckily I only scraped my knee.

Enjoying the views.

From the crater rim I moved effectively down again, jogging and running. Further down, I met a couple of guys who had had to turn around because they didn’t have crampons. Looking at the watch, I saw that I had time to catch the last bus down from 5th Station, but I decided to go down the complete Yoshida Trail, down to my hotel in Fujiyoshida. I had already spent too much time sitting on buses and trains on this trip, and I wanted to see more of the Japanese nature.

The way down.

At about 2400 masl the trail descends into the forest. The mist lay thick between the trees, and there was an almost eerie atmosphere. There are several ruins of old pilgrim logdings and shrines along the path, with information signs about the rich history of the Fuji mountain.

When the trail passed into the flatlands between the mountain and the city, signs started appearing about bears in the area. I didn't think too much about it, but suddenly I saw movement in the woods ahead of me. About fity meters away, the dark shape of a bear ran straight across my path and disappeared among the trees to my left. I had been running in order to get back before it got dark, but now I made a sudden stop. I felt a slight regret that I didn't buy one of the "bear bells" that every hiker in Japan wears, but I’m not sure how effective the actually are. Slowly, I walked past where the bear had appeared, and talked out loud as a substitue for the bear bell. I hadn't made any effort to find out if these bears are aggresive and how one should begave in the vicinity of one. After a couple of hundred meters I decided that I was safe, and I resumed running towards Fujiyoshida. I concluded my trip at the Kitaguchi Hongu Fuji Sengen Shrine, before I jogged back to the hotel and ate dinner.

Kitaguchi Hongu Fuji Shengen Shrine, the official starting point of the Yoshida Trail.

My trip to Mount Fuji was a wonderful experience, and I feel lucky to have experienced it in sunshine and solitude. To be clear: Climbing Mount Fuji in the off-season requires experience, and being in the mountains always entails risks, especially with regard to weather and changing conditions. Having the correct clothes and gear, knowing how to use it and knowing your own limits is essential. However, the mountain is definitely accessible for experienced hikers when the weather and conditions allow it.

Places to discover the Small Web

Wed, 22 Nov 2023 00:00:00 -0000

I'm fascinated by the world of personal websites that exists on the Internet, often referred to as the Small Web. It's a stark contrast to the "commerical web" and social media platforms, with endless scroll, advertisements and surveillance. The Small Web let's you browse and discover at your own pace, with fewer distractions and less screaming for attention. Content can speak for itself, and it's not pushed in your face.

The Small Web doesn't have any built-in mode of discovery, which I think it's one of its strengths. One of my favourite ways of discovering other people in the Small Web is by recommendations from others, which is why I think it's useful to have a blogroll (here's mine). In addition to that, there are several websites that serve as good entry points, and I've tried to compile a list of them here. If you have any other good suggestions, let me know; I will update this list consecutively.

Here are some places to discover other people's ideas, thoughts and projects:

The future of social media

Sun, 19 Nov 2023 00:00:00 -0000

The landscape of social media is changing. Twitter is in the middle of a considerable decline, and Meta's Facebook and Instagram are trying to adapt to new requirements and legislation. The decentralized social network Mastodon is gaining serious traction, giving some hope to the dream of social media platforms that are not built on surveillance capitalism. Now that we finally have viable alternatives to Big Tech platforms, is social media salvaged?

I joined Mastodon in late 2019, which was relatively early and before it had significant mainstream attention. I had been interested in alternatives to the corporate-owned platforms for a long time, and wanted an option that was not infested with advertisements and behavioural tracking. In this process I had explored things like Hubzilla, Friendica, Diaspora, OpenBook/Okuna and Scuttlebutt. Each had their own set of advantages and weaknesses, but the ultimate problem was always a lack of critical mass. The platforms based on the ActivityPub protocol seemed more promising, especially since one of them, namely Mastodon, after a while started to get a decent amount of users. However, the community (or rather communities) consisted mostly of tech people, and the larger part of the posts also revolved around tech, from what I could observe at least. Not an easy sell when you want your non-tech friends to join you in the transition to decentralized platforms.

I made a draft for a blog post three years ago called "Trying out the federated social medium Mastodon". I wanted to highlight some of the things I observed, and give the new type of social network some attention. I never published the post. At the time I really wanted to advocate for trying something else, but the amount of activity on Facebook and Instagram was (and is) just too large for most people to miss out on. I had spent some "social credit" on getting people around me to use Signal instead of Facebook Messenger, even though it was only successful for my closest family, and I still have to use Facebook Messenger almost just as much today in order to keep in touch with my friends. I have never been particularly active on social media anyways, so I didn't (and still don't) have much leverage in trying to convince people to try out alternatives. Only people with a large outreach could possibly influence others to join them in switching platforms.

After Twitter started struggling for various reasons, Mastodon has seen a large influx of users. Three years ago, I would have been elated to see that a decentralized social network is finally succeeding, but I'm less excited than I expected. Of course, it's a massive improvement to have platforms that put more power into the hands of individual people rather than Big Tech companies, but the feeling I get is that it's still just another feed to scroll.

We can finally escape the ubiquitous advertisements of other platforms, but I realize more and more that being on social media is usually not worth the time. The true value of social networks is the possibility to have a discourse with anyone, anywhere, but a lot of the time it's more about consuming content rather than communicating. The staggering amount of information that we process while scrolling through our feeds are taking up precious bandwith in our attention, and the discussions we engage in online are often less than constructive. I'm of course aware that people use social media in very different ways, so others' experience may not be as negative as my own. On the other hand, you may also think that I'm stating the painfully obvious.

It's not a controversial opinion to say that meeting and communicating physically in general is more meaningful and valuable. Looking at how social media has changed, and how it as affected my own life, I want to start putting more effort into the interaction I have with people in the "offline" world rather than online. It will most likely entail relationships with fewer people, and at a slower pace, but that's sort of also the point.

For digital communication, I see an advantage in prioritizing one-to-one interactions, rather than the many-to-many communication that modern social platforms enable. I find myself coming back to the actual "mastodon" of social networks: Email. "Everyone" has an email account, and the effort it takes to write an email usually ensures a certain level of commitment to the message. I use RSS as the main way of keeping up with updates to personal blogs, magazines and podcasts. It's a calmer way of consuming. Using personal websites as a way of publishing online also enables more control of both content, form and creativity.

The fact that I have a high threshold for posting on social media hinders a lot of engagement with other people, whether it's communication with friends/acquaintances, academic or work-related discussions, response to blog posts or sharing dog pictures. I'm definitely missing out on opportunities and interesting discussions. Even so, at the moment I feel more inclined to be more present in the moment, as the cliché goes, and in order to do that I have to be less online.

In answer to the question implied by the title of this post, I hope and believe that the pendulum will start to swing back from the "terminally online" culture that social media brought with it. I'm unsure whether it is possible to design large-scale digital tools for social networks that work well for humans, and there are signs that many are fed up with being "connected" all the time. Without discarding the tremendous benefits of connecting the whole world through the Internet, I think it's wise to take a break and assess how we actually are affected by our ever-online lives.

Links #3: Large language models

Mon, 05 Jun 2023 00:00:00 -0000

A few interesting reads on the topic of generative AI and large language models:

Op-ed by Naomi Klein: AI machines aren’t ‘hallucinating’. But their makers are (2023-05-08). I think Naomi Klein has some excellent points in this article, and "cuts the crap".

According to [the logic of Google CEO Eric Schmidt], the failure to “solve” big problems like climate change is due to a deficit of smarts. Never mind that smart people, heavy with PhDs and Nobel prizes, have been telling our governments for decades what needs to happen to get out of this mess: slash our emissions, leave carbon in the ground, tackle the overconsumption of the rich and the underconsumption of the poor because no energy source is free of ecological costs.
Blog post by Kevin Lin: Lessons from Creating a VSCode Extension with GPT-4 (2023-05-25). The most important take-away from this blog post is that it accurately describes the current state of using GPT-models for programming. Here's the relevant quote from the introduction:

Lately, I've been playing around with LLMs to write code. I find that they're great at generating small self-contained snippets. Unfortunately, anything more than that requires a human to evaluate LLM output and come up with suitable follow-up prompts. Most examples of "GPT wrote X" are this - a human serves as a REPL for the LLM, carefully coaxing it to a functional result.
Article by Baldur Bjarnason: Modern software quality, or why I think using language models for programming is a bad idea. A longer piece, but well worth the read. A lot of good observations that is important to bear in mind.
Op-ed by Zainab Choudhry: AI tools like ChatGPT are built on mass copyright infringement (2023-05-25).

Democratizing generative AI

Mon, 05 Jun 2023 00:00:00 -0000

A few weeks ago, a colleague and I wrote an op-ed in Teknisk Ukeblad where we played around with the idea of how to democratize generative artificial intelligence (AI) and avoid ending up with a monopoly on large generative AI models. The possible scenarios we envision may not all be realistic, but we want to highlight the possibilities we have to make the use and development of AI more transparent.

The op-ed is in Norwegian, but could be translated to English using various translation tools. Read the op-ed here: ChatGPT requires moderation like Wikipedia (the original title was Democratizing Generative AI, which we feel is a bit more representative of what we wanted to convey).

Green AI in a world of LLMs

Wed, 18 Jan 2023 00:00:00 -0000

One year ago I wrote a blog post about machine learning and carbon emissions. Since then, we have experienced some ground-breaking developments in the world of AI. With the release of image generators like DALL-E and large language models (LLMs) like ChatGPT, we are talking about a whole new scale of computational load for using, let alone training such models. What's more, these are no longer tools reserved for research purposes, but deployed software that people all over the world has started using on a daily basis.

In my research group, Trustworthy Green IoT Software at SINTEF Digital, we are exploring how we can contribute to Green AI, which refers to environmentally friendly AI development. One of the tools we have developed, an ML pipeline for supervised learning, uses DVC for pipeline orchestration, which we see as a measure of reducing the energy usage through caching and skipping unnecessary reruns of pipeline stages. Many researchers are calling for measuring and reporting the carbon footprint in order to bring awareness to the issue, which is something we want to bring into our existing tools. The question is whether such measures are sufficient when the power of LLMs and similar tools is pushing the development in a much more resource-intensive direction.

If you are an AI/ML researcher or practitioner, how do you look at the increasing carbon footprint of ML models, and do you see "Green AI" as an important aspect for the future of AI development? Will the power of extremely large models always win over models with smaller footprint, or is it possible to find a balance between performance and energy efficiency?

Skiing new lines

Tue, 17 Jan 2023 00:00:00 -0000

One of my new year's resolutions: Always ski a new line through the forest when going back to the cabin. It surprised me how much fun I can have in my own backyard just by exploring a bit more.

This is the year of glade skiing for me.

Playing in powder with Gjert

Mon, 09 Jan 2023 00:00:00 -0000

A fun afternoon skiing with my Border collie Gjert in our backyard.

Your browser does not support the video tag.

Reconstructing Obsidian features in Vim and Bash

Sun, 08 Jan 2023 00:00:00 -0000

I have a large collection of plain text files (mostly Markdown) that contain all my notes and writings, both for work and non-work, which I navigate and modify using Vim, fzf, shell scripts and command line utilities (on mobile I use the excellent app Working Copy). Everything is of course synced and version-controlled using Git.

I recently discovered that Obsidian has a range of interesting features, and since it is meant to manage a collection of Markdown-notes (which still can be synced through a git-repo between desktop and mobile apps), I could try it out without having to do any changes at all to my current structure or formatting of my files. I downloaded the apps for desktop and mobile, and everything was good to go.

The most exciting features for me in Obsidian are the following:

Easily link between files, both between individual notes and inserting images into notes by using Markdown syntax.
Open link to another file with a keypress.
Button for opening a random note (yes, I actually find this useful, because it creates opportunities for sparking ideas through serendipity, and it can also work as "spaced repetition" for things you want to learn or remember).
Links between files are updated automatically when you rename or move a file.

I think Obsidian works great in general, and even though it is closed source software, anyone can write plugins for it. There are many useful "community plugins" that really increases the usefulness of the program. However, I missed writing in my usual text editor at the command line.

Obsidian has the option of using Vim keybindings, but that does not replace the configurability of Vim, and the advantage of using command line utilities and scripting as a part of the workflow. Also, I kept hitting ctrl-W [h,j,k,l] in order to change between panes in Obsidian, but of course ctrl-W means "close file" when you are outside the command line.

The user experience of Obsidian on iOS is a bit "messy" in my opinion, and I found myself opening Working Copy more often than Obsidian to edit files. Navigating the file explorer in Obsidian on both desktop and mobile is less than optimal, among other things because of no easy way to collapse all folders, and I found myself scrolling a lot.

Some of my complaints can be fixed through plugins, either existing ones or by writing one myself, but the power of the command line is either way out of reach. Because if this, I wanted instead to reconstruct some of the Obsidian features I liked the most in Vim (or at the command line)

Easily insert links to other files

To make this work, the links to other files has to be relative paths (otherwise images and links won't work properly when rendering the Markdown in HTML). With the plugin fzf.vim, the following line maps ctrl-x ctrl-f to search for any file in the parent directory from which you opened Vim (which means I always open Vim from the root directory of my note collection) and insert its relative path from the current open file when you press Enter:

inoremap <expr> <c-x><c-f> fzf#vim#complete("find -print0 <Bar> xargs --null realpath --relative-to " . expand("%:h"))

The line above is put in your .vimrc (Vim configuration file). Made based on tips from this GitHub-thread.

Open link to another file with a keypress

Vim already had functionality for this, I just didn't know about it. ctrl-w ctrl-f opens the filepath under the cursor.

Opening a random note

This one was easily implemented in a bash script:

#!/bin/bash

file=$(find /path/to/markdown-collection/ -name "*.md" | sort -R | head -1)
vim $file

The plugin fzf.vim makes navigation very easy, both inside Vim and at the command line. When having files open in Vim, the command :Files let's you fuzzy search for filenames, and :Rg (if you have installed ripgrep let's you fuzzy search for file contents. I'm quite restrictive of adding specific configuration and plugins to my Vim setup and want to keep things lean (I only use two plugins currently), but fzf.vim has really made my workflow much better.

Automatically update links when renaming or moving file

I use linking between files quite a lot, especially for embedding images into Markdown-files. Sometimes I restructure or rename filenames, and since I might have embedded one image into multiple files, it would be a massive time saver if all the links could update automatically when moving or renaming a file. I tried searching for an existing solution; the only similar thing I could find was an extension for VSCode, but no solution for Vim.

I have made an attempt to make a shell script for moving/renaming files, which automatically updates all links to the file in Markdown-files of the working directory. You can find the current version of the script here: linkmv. I'm pretty inexperienced in bash, and I haven't tested all the possible edge cases, but it's a work in progress. Hopefully I'll get something working eventually (contributions and tips welcome)!

Links #2: Note-taking

Fri, 06 Jan 2023 00:00:00 -0000

Sharing some links I find interesting about note-taking:

Rank and File by Robert Minto.
Paper notes by Tim Hårek.
Paper notes by Tom MacWright.
Why use paper notebooks

The dilemma of digital versus analog tools, especially related to note-taking, is something I always return to. Even though digital tools typically are the first choice in many situations, we shouldn't forget the strengths of a simple pen and paper.

My best reads of 2022

Mon, 02 Jan 2023 00:00:00 -0000

I usually discover new books to read through references or quotes from previous books I have read, but sometimes I also get some useful recommendations from blogs or people I talk to. Maybe I can provide some tips on your next read?

Here's a summary of some of the best books I read in 2022:

The Age of Low Technologies by Philippe Bihouix

This book gives a very important perspective on how we may solve the challenges we face regarding climate change and over-consumption of resources.

Some innovations, and perhaps those to be feared most because they are often ones that are profitable, allows the absurd needs generated by systems to be fulfilled. Nanoscale particles allow a sunscreen to offer protection and a feeling of freshness; RFID chips makes it possible to pay for the contents of your supermarket trolley without going to the cash register; office glazing can be made opaque by eletrical impulse; refrigerators prepare the shopping list. We are far from the "revolutions" of the steam engine and crop rotation! These examples are caricatures, but emblematic of many of today's "innovations". Frankly, would we develop such high-tech applications if we knew the real price to pay, in terms of factory waste, the disappearance of nature and resources, the destruction of everything that has sustained us, physically and intellectually, for millenia? We cannot say.

Scrambles Amongst the Alps by Edward Whymper

Very interesting tales from "the golden age of mountaineering" (1850s and 1860s) when many major first ascents in the Alps were achieved. Edward Whymper was one of the pioneers of mountaineering, and is most famous for his first ascent of the Matterhorn, which is described in this book. It is wise to remember that this book only provides Whymper's account of the controversy on Matterhorn, which has been disputed by other members of the first ascent party.

Whymper seems to be a typically "manly" man, who enjoys the harsh challenges and dangers of high-altitude mountaineering. The following quote gives an impression of how he thinks about risks in the mountains:

The uppermost men were continually abused for dislodging rocks and for harpooning those below with their bâtons. However, without these incidents the climbing would have been dull: they helped to break the monotony.

Even though the main focus of the tales is Whymper's accounts of all his glorious first ascents, the book also gives some perspective on what we learn from mountaineering:

I have not made myself an advocate or an apologist for mountaineering, nor do I now intend to usurp the functions of a moralist, but my task would have been ill performed if it had been concluded without one reference to the more serious lessons of the mountaineer. We glory in the physical regeneration which is the product of our exertions; we exult over the grandeur of the scenes that are brought before our eyes, the splendors of sunrise and sunset, and the beauties of hill, dale, lake, wood and waterfall; but we value more highly the development of manliness, and the evolution, under combat with difficulties, of those noble qualities of human nature--courage, patience, endurance and fortitude.

For those who have an interest in mountaineering, this is a must-read.

The Mountains of my Life by Walter Bonatti

This is another essential read for those who love mountains and mountaineering. Bonatti writes a lot about how climbing affects the human mind, and it is very inspiring to read about his principles and his approach to mountaineering. Similar to Edward Whymper, Bonatti was involved in the controversy of a first ascent, in this case of K2, one of the most difficult mountains to climb in the world. I recommend reading the version that is co-authored with Robert Marshall, who has gone through the evidence thoroughly and gives perhaps the most accurate account of what actually happened on K2 during the first ascent.

Bonatti also provides his thoughts on what mountaineering really means for us:

What is there, beyond the mountain, if not the man? Mountaineering is only one of a thousand ways of living and getting to know yourself. Climbing mountains should signify nothing more than this search for identity. It should never be mere escapism, because sooner or later we must return to our own personalities and feelings, the only place available apart from nothingness. So the mountain should prepare us to go further. A mountaineer should store experiences and use them to enrich himself. He can do so if he knows about wide open spaces and taking responsibility for his own actions.

The Way Home by Mark Boyle

One of my all time favourites: A book about a man's real life experience with living without technology. For those who long for a simpler and more meaningful life, this will give some inspiration.

Amusing Ourselves to Death by Neil Postman

This book was written in the 1970s and outlines problems related to media and specifically television, but it is just as relevant today, and most of the reasoning can be applied directly to how we are affected by social media and technology today.

A couple of other books on how we are affected by technology

The Shallows by Nicolas Carr
Stand out of our Light by James Williams

A feed for sports and outdoors activities

Tue, 02 Aug 2022 00:00:00 -0000

Strava is a popular tool for sharing sports and outdoors activities, and has become the de facto social medium for athletes, runners, cyclists etc. I'm not the kind of person who likes to share every detail about how I train or my daily activities, but sometimes I enjoy sharing a workout, race or a route I've done in the mountains. I also enjoy keeping control of my own data and how it is accessed and displayed, and I have only reluctantly used Strava and Garmin Connect to share activities.

While I'm waiting for someone to make an ActivityPub-enabled Strava-clone, I made my own Activities-feed on this website. To add an activity, I upload a gpx-file containing the GPS-track (including elevation data and timestamps), and optionally add a textfile with a description and images from the activity. The feed uses Leaflet and leaflet-evelation.js to display gpx-files on map tiles, and plot the elevation profile and speed graph. I have chosen to use map tiles from OpenTopoMap, since it's important for me to have detailed terrain information when visualizing the activities. My very messy build-script for this website takes care of putting the text description and images in the correct places.

This way I can easily share workouts and activities with friends and family. The activities also show up in my RSS-feed. The feed is not very visually pleasing yet, but I will probably fine-tune to CSS to make it look a bit more clean.

Check out the new feed at this link!

Gjert the border collie doing what he loves most

Mon, 01 Aug 2022 00:00:00 -0000

I shot a short video of Gjert the border collie during one of his agility workouts. I am using the GoPro HERO7 Black, with no gimbal, and I must say I am impressed with the HyperSmooth video stabilisation.

Take a look!

Your browser does not support the video tag.

Building a laptop

Sun, 17 Jul 2022 00:00:00 -0000

There are several very exciting projects that are aiming to make more sustainable consumer electronics:

Fairphone: A phone that is extremely easy to repair and has parts that can be easily replaced, by you.
Framework: A laptop that is modular, upgradeable and designed to be easy to repair by anyone.
MNT Reform: A laptop that is designed to be tinkered with, and that makes it easy to swap out parts when needed.

Such projects have inspired me to put together my own "modular" laptop, with parts that are easily replaced. I'm very fond of the Raspberry Pi computers (henceforth referred to as RPi, RPis plural), which are cheap single-board computers that are made to be tinkered with. One of my RPis is the RPi 4, with 8GB RAM, which is the most powerful RPi available to date. I wanted to use this one to make a laptop with the following requirements:

Portable.
All parts are easy to replace.
All parts are easy to get hold off. This means that I'm don't want to use parts that I have to order from the other side of the globe and wait for several months for it to arrive.

In addition to the RPi itself, I needed the following parts:

Display
Keyboard
Mouse
Portable power supply
Some sort of frame/box/thing that hold everything together

For the display, I ended up with the official Raspberry Pi Touch Display 7". The size was about what I had in mind for a portable computer, and I was also lucky enough to find someone who sold a used one for a cheaper price.

I also found someone who sold a used Logitech K400, which is a keyboard with an integrated touch pad. I didn't really want a separate mouse for my RPi laptop, so this was perfect.

For portable power supply, I just use plain powerbanks. I know that one should preferably use a something that reliably supplies the exact voltage that the RPi needs, but I find that powerbanks works just fine in most cases. For some very small powerbanks, I somtetimes get the "low voltage" warning on the RPi, but apart from that, it works flawlessly.

The hardest part was binding all the parts together in a package that can easily be carried around. I didn't want to have several separate parts that I need to pick up in order to move the laptop. After a lot of experimenting, I found a adequate solution by using adhesive GoPro mounts and some extra arm joints. I put one adhesive mount on the keyboard, and one on the display case. I connected these by using arm joints (one of them needs to be slightly modified by cutting of one of the plates). The powerbank can be fastened to the "display foot" using an elastic band or similar, or to the keyboard.

The result is a portable computer (I call it a laptop, since it's a computer I can have on my lap, but I'm not sure if everyone agrees with me), with parts that are very easy to get hold off.

I cut off a part of the display case in order to easily swap memory cards (which I do frequently, since I like to use different types of Linux operating systems).

A small part of the display case is cut off, in order to easily access the memory card that the OS is stored on.

The laptop can easily be disassembled into these parts.

As a bonus, the computer can be mounted on other things that have a GoPro mount on them. I have a strong clip with a mount on it, which enables me to attach the computer at the edge of a table or similar. This is very useful when I want to watch the display when I'm cooking or washing the dishes.

The computer and display mounted on a clip instead of the keyboard. Saves desktop space.

The RPi isn't the most powerful computer in the world, but it can handle a lot of day-to-day stuff, like text editing, internet browsing, watching videos, playing games, listening to music etc. I have used mine a lot, and one of the best things with it is that I don't have to be afraid of breaking it, since I know that all the parts, even the main board itself, is quite cheap to replace.

On self-hosting, local storage and portable git hosting on iPhone

Wed, 13 Jul 2022 00:00:00 -0000

I prefer having control of my own data, and therefore try to avoid using cloud services if I absolutely don't have to (e. g., at work I'm pretty much forced to use the cloud services of my employer, which is Microsoft of course). Earlier I have been self-hosting some services, including file storage (Nextcloud), music server (Subsonic), git (gitea) and social media (Hubzilla). Then I moved to a place where I couldn't easily set up my own server with port forwarding for external access, and my self-hosting was put on pause. As an experiment, partly motivated by the increased focus on the environmental impact of the expanding use of servers and cloud services, I attempted to reduce my use of cloud services to a minimum. Some of the things I did:

Deleted my Spotify account and started using only my own local music collection (extra privacy benefit of eliminating the data collection of Spotify)
Stopped using cloud file storage (up until then I had used Nextcloud and MEGA), and moved to using local storage with backups on external hard drives
Archived all emails to local storage, keeping my email inbox clean.
Deleted old and unecessary GitHub and Codeberg repositories.

I moved to a more offline workflow, and realized that I don't need to have all my digital stuff instantly available on any device with an Internet connection. This also have the added benefit of enabling me to go offline, which often improves my focus, without losing access to my files. Since I now live a place without Wi-Fi, it is a great advantage that I'm not that dependent on bandwith.

The single most important cloud service for me, both professionally and privately, is git hosting. It is where I sync most of my work, code, notes and writings. Some things are private/personal (or at least should be hidden from the eyes of others), which of course are kept in private repos, but I would much prefer to host those repos myself. One option I have been considering is to pay for hosting of my own gitea instance, but I both want to save money and avoid having another cloud service (even though it would be managed by myself). I could also choose to have local repos without storage in the cloud, but I need to easily keep the repos synced between my personal latop, work laptop and phone. I started playing with the idea of having a portable git server on a Raspberry Pi Zero, powered by a power bank, which could set up its own Wi-Fi and start a git server automatically when booted and let me push and pull from repos hosted on it. This would however require me to carry an extra device.

I then realized that maybe my phone, an iPhone, could serve as the git server instead. The App Store lists some apps that claim to be able to setup a git server, but after reading a couple of the prvacy policies, I decided I would sleep better at night if I didn't have to trust some ambiguous wording on what and how data is collected. I use the Working Copy git client app a lot, which can serve the git repos in the app on a WebDAV server. I was able to access the server through Finder on macOS, but I gave up after failing to connect to it at the command line from Ubuntu. Instead I tried setting up a git server at the command line in the iSH app. It proved to be the simplest and most usable approach for me. When using the iSH app, the setup is identical to how you would do it when setting up a git server on a normal Linux machine. In the app I installed git and git-daemon (they have to be installed individually):

apk add git
apk add git-daemon

Then I initialized a bare git repo:

mkdir myrepo.git
cd myrepo.git
git init --bare

To serve the repo (and any other git repos in the same folder:

cd ..
git daemon --reuseaddr --base-path=. --export-all --verbose --informative-errors --enable=receive-pack

If you want to disable the ability to push to the repos (i. e., clone-only access), skip the last flag. To clone repos into Working Copy on the same phone, use git://127.0.0.1:9418/myrepo.git or git://localhost:9418/myrepo.git as the URL (9418 is the default port of git daemon). I don't think iSH handles running inthe background well, so you might have to press "Clone" in Working Copy (after filling out the URL), switch to iSH to enable the server to complete the request, and then back to Working Copy. To clone to other devices, run

git clone git://[ipaddress]:9418/myrepo.git

where [ipaddress] is the IP-address of your iPhone on the local network (the other devices obviously have to be connected to the same network).

The drawback is of course having to start the server before every push/pull, but for my uses this is acceptable. Also, there is no authentication required to pull from the server, so one needs to be sure that there are no bad actors on the local network.

Next up is trying the same thing on an Android phone (which I expect to be just as easy to do inside Termux), and then maybe setting up a mobile git server with authentication.

Links #1

Tue, 12 Jul 2022 00:00:00 -0000

Some links I find worth sharing:

On digital footprints from Weekly Musings.
Better search results by Brad Taunt.
You're attention didn't collapse. It was stolen from The Guardian.

My best reads of 2021

Tue, 31 May 2022 00:00:00 -0000

In this post I want to share some of my best reads in 2021 (I know we are almost halfway into 2022, but I wrote this post 5 months ago and just now realized I forgot to publish it):

The Age of Surveillance Capitalism by Shoshana Zuboff

A very well written and thoroughly researched book on surveillance capitalism. Zuboff very succinctly articulates the problems that arise from the (mis)use of modern technology and how corporations gather and use personal data. I highly recommend this book.

Meditations by Marcus Aurelius

Timeless wisdom.

Pelsjegerliv ("Trapper Life") by Helge Ingstad.

A classic from one of Norway's most celebrated adventurers of the 20th century. Truly fascinating and humorous to read the tales from his time as a trapper in the wilderness in Canada.

Rhythm of War by Brandon Sanderson

The fourth book in one of my favourite fantasy series.

Into thin air by Jon Krakauer

A very captivating account of the tragic story of one of the worst accidents on Mount Everest.

Into the wild by Jon Krakauer

Another classic from Krakauer, and one of my all time favourites in the outdoors genre.

24/7: Late capitalism and the end of sleep by Jonathan Crary

A very interesting take on how modern technology and so-called late capitalism affects our lives.

Plain text calendar

Mon, 30 May 2022 00:00:00 -0000

The last years I have been moving towards plain text files for almost any use case where I need to store information digitally, and after reading about Calendar.txt I have started to use a plain text file as my calendar. I use (almost) the same template as described in the link, but I wrote my own script for generating the calendar template so I can easily make adjustments if needed. Example of the format with a few example events:

2023-01-02 w1
2023-01-02 w1 Mon 
2023-01-03 w1 Tue  12:00 Meeting.
2023-01-04 w1 Wed 
2023-01-05 w1 Thu  18:00 Going for a run.
2023-01-06 w1 Fri 
2023-01-07 w1 Sat 
2023-01-08 w1 Sun 
2023-01-09 w2
2023-01-09 w2 Mon 
2023-01-10 w2 Tue 
2023-01-11 w2 Wed 
2023-01-12 w2 Thu  Seminar.
2023-01-13 w2 Fri 
2023-01-14 w2 Sat 
2023-01-15 w2 Sun

I find it quite nice since it integrates well with the rest of my very plain text-oriented workflow, and it's easily synced and managed with git together with the rest of my plain text files (notes, todos, etc). It takes some time to adapt to since the "user interface" is quite different than normal calendar applications, but for those who are familiar with quick navigation in text files (for example using Vim), it can be much more efficient to use a plain text calendar.

I also wrote a small bash script for opening the calendar on the current date in Vim:

today=`date +%Y-%m-%d`

vim /path/to/calendar.txt -c "/$today"

[NOR] Test av Northern Playground langermet trøye

Sun, 20 Mar 2022 00:00:00 -0000

Jeg fikk for et halvt år siden en turtleneck fra Northern Playground, et veldig spennende norsk klesmerke. Kvaliteten og bruksverdien til dette plagget var overraskende gode, og hvis du ikke gidder lese hele testen så kan jeg oppsummere den allerede her: Om jeg bare skulle hatt én trøye resten av livet så er det definitivt denne; den fungerer til alt fra trening og tur til jobb og selskap. Northern Playground er det første og eneste klesmerke jeg har vært borti som faktisk leverer alt de lover i reklamene sine, og det er ikke lite.

Det er mange klesmerker som lager klær i ull og ullblandinger som påstår at klærne ikke får noen svettelukt så lenge man lufter dem. Jeg har testet mange ulike merker, men trøya fra Northern Playground er den eneste jeg har prøvd hvor dette faktisk stemmer. I en av sine reklamefilmer forteller Northern Playgrounds CEO at han ikke husker sist han vasket trøya han har på, og at han fortsatt ikke får noen klager på lukt. Dette synes jeg i utgangspunktet det var vanskelig å tro på, helt til jeg fikk prøvd trøya selv. Jeg har ennå ikke vasket min Northern Playground-trøye siden jeg fikk den for et halvt år siden. Den lukter ingenting, til tross for å ha vært med på løpeturer og på jobb uke etter uke. Om jeg har svettet mye i den så henger jeg den til lufting enten på verandaen eller over en stol innendørs, og da er den klar til bruk dagen etter. En annen stor fordel med trøyene er at de er såpass fine at man fint kan bruke dem både på jobb og i sammenhenger hvor man må se litt ordentlig ut.

Etter mye bruk er det kun ett "svakt punkt" jeg har klart å finne i denne trøya. Jeg har løpt mye med sekk til og fra jobb, uten noe plagg mellom sekken og trøya, og jeg ser at det har kommet noen nupper på ryggen på grunn av dette. De er så små at jeg ikke ville tenkt over det med en annen trøye, men når klesplagget leverer såpass godt på alle andre punkter, så må man fram med små tingene.

Jeg har også skaffet meg den kortermede versjonen, som selvfølgelig fungerer like bra, og trøyene er med på løping, klatring/buldring, jobb, selskap og fjellturer. Northern Playground lager trøyer i mange forskjellige fasonger (T-skjorter, langermet, bestefartrøye, zipneck, hoodie etc), så her er det mulig å finne noe for enhver smak.

Selskapet har også en svært miljøvennlig profil. De har flere tiltak for å redusere forbruk, for eksempel at de tilbyr livstidsservice på plaggene, og de har en selvpålagt miljøskatt på hvert salg.

Et veldig spennende prosjekt de har gående er The Last jakke og The Last bukse. Målet er å lage en jakke og en bukse som er de siste du noen gang trenger å kjøpe, og om mulig også gå i arv! Jeg håper de lykkes med å lage et godt produkt, og det har jeg også troa på at de klarer. Northern Playground har en forbilledlig designfilosofi og miljøprofil, og jeg håper de kan inspirere andre klesprodusenter i samme retning.

Bilder av trøya:

A 50 year old camera

Thu, 06 Jan 2022 00:00:00 -0000

Last year I inherited an analouge camera from my grandfather. Usually I prefer digital tools, because they are space-efficient, easy to carry around, and the data they produce can is easy to store and backup. However, there are also some disadvantages. My photo library has been steadily growing each year, which recently resulted in me having to buy another external hard drive in order to have enough storage space for backups. Digital photography makes it possible to take a lot of pictures, and sometimes (or most times?) I'm taking too many. Analogue photography forces you to be more thoughtful, both because of a very limited amount of photos per film roll, and the economical cost of each photo. In addition, I think it's awesome to use a piece of technology that's 50 years old and still works flawlessly.

Here's a few photos from the first roll I shot with the camera:

Send me an email if you think you can recognize this mountain, and I will tell you if you're right (hint: It's in Trollheimen, and not far away from the hut shown above).

Machine learning and carbon emissions

Wed, 05 Jan 2022 00:00:00 -0000

Recently I have been looking into the environmental impact of machine learning in relation to ongoing research projects at work. Carbon emissions are high on the agenda for everyone, especially if one works at a research group with the term "Green" in it's name. Lots is already written on the topic, but I wanted to share a few of the most interesting articles and research papers I found, with a few comments of my own.

Deep learning's carbon emissions problem from Forbes summarizes well some of the most problematic issues of machine learning. It is mainly focused on deep learning and natural language processing (NLP), which entails developing models that consist of an enormous amounts of parameters. Many of the article's key points stems from a paper that in 2019 really brought awareness to the gigantic energy usage of some of the largest deep learning models: Energy and Policy Considerations for Deep Learning in NLP. As a machine learning practitioner, the following quote from the article is something I'm especially concerned with:

Another factor driving AI’s massive energy draw is the extensive experimentation and tuning required to develop a model. Machine learning today remains largely an exercise in trial and error. Practitioners will often build hundreds of versions of a given model during training, experimenting with different neural architectures and hyperparameters before identifying an optimal design.

Luckily, there exist several frameworks that can help reducing the number of experiments needed during model development. Personally, I use Data Version Control (DVC), an excellent tool for structuring machine learning experiments. Not only does it helps you keep track of models and hyperparameters, but you can also use it to track all stages of a machine learning pipeline (which typically consists of several preprocessing stages). Since it keeps all intermediate results between stages in cache, you can save a lot of computation time when experimenting with different control parameters.

I think the article concludes with an important note:

The AI community must begin to work toward new paradigms in artificial intelligence that do not require exponentially growing datasets nor outrageously vast energy expenditures.

Another article, The drive to make machine learning greener from Discover Magazine, quotes a researcher team from the University of California, Berkeley, on how reporting on carbon emissions might help reduce them:

Patterson and colleagues conclude by recommending that computer scientists report the energy their calculations consume and the carbon footprint associated with this, along with the time and number of processors involved. Their idea is to make it possible to directly compare computing practices and to reward the most efficient. “If the machine learning community working on computationally intensive models starts competing on training quality and carbon footprint rather than on accuracy alone, the most efficient data centers and hardware might see the highest demand,” they say.

The paper Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning argues similarly for the importance of measuring and reporting the carbon footprint of machine learning experiments. The paper is an especially interesting read, as the authors suggest a set of concrete courses of actions based on their findings. The main points are to be consistent in reporting energy and carbon metrics, and using leaderboards of these metrics as incentive for reducing carbon emissions in research. They also write that one should consider "energy-performance trade-offs before deploying energy hungry models". It's also interesting to note that one of the most intuitive (at least in my opinion) ways of measuring the energy-efficiency of any computer program, namely to count the number of Floating Point Operations (FPOs), actually can be misleading (read section 5.1 of the paper for a discussion on this). This is one of the reasons for why there is a need for frameworks that can measure or more accurately estimate the actual carbon emission of training or running a machine learning model. The authors have published an open source Experiment impact tracker that can be used for this purpose.

Another paper that looks into methods for tracking energy-efficiency is this: Quantifying the Carbon Emissions of Machine Learning. The authors have contributed to two tools that can be employed for emissions tracking. The first one is an emission calculator, where you compute emissions based on hardware type, runtime and cloud provider/region. The second one, which I personally have started testing in my projects, is the CodeCarbon framework. CodeCarbon can be embedded in Python code, and automatically tracks the energy usage and carbon emissions of your experiments. The tool is very easy to use and generates results in a csv-file that can be used for your own analysis. Additionally, the tool provides an interactive dashboard based on the results, which makes it very quick and easy to get up and running with emission tracking.

Conclusion

Even though the emissions of "small-scale" machine learning practitioners like myself are dwarfed by the gargantuan carbon footprint of big tech companies like Google and Amazon, I think it is important that we are all aware of these issues, and try to mitigate them at any scale. A shift towards more green machine learning would benefit us all, and hopefully also affect the large players in the industry.

To summarize, here's a few concrete actions that can help reduce the carbon footprint of machine learning:

Reduce the number of unnecessary training runs, for example using frameworks like DVC.
Employ efficient methods for hyperparameter tuning.
Measure and report on energy usage for your experiments. Use an emission calculator, or use the CodeCarbon framework to automatically measure the impact of all your experiments. Another option is the Experiment impact tracker (this one I have not tested myself yet).
If you run your experiments on external servers ("in the cloud"), be aware of which data centers you use, and how they are powered.

Interesting reads from the web

Wed, 15 Dec 2021 00:00:00 -0000

Here's some articles I found interesting enough to share this week:

The goal behind using technology from Scott Nesbitt. His thoughts resonates well with what I think of the subject. Modern technology and digital tools have a tendency of influencing our behaviour far more than it should.
In 2030, You Won't Own Any Gadgets by Victoria Song at Gizmodo. Many things might be said about renting/subscribing vs owning things, but this subject needs to be tackled differently when it comes to software (or physical things that rely on software) instead of physical things that don't rely on software. It can make a lot of sense to share/rent physical things in order to reduce the need to buy and produce new stuff all the time. However, things can quickly become messy and problematic when you don't own your software, and especially when you have bought physical things that rely on said software. Read the article for some interesting stories and comments on this.
On slower, quieter blogs. Another one from Scott Nesbit. The realm of "quiet" personal blogs is larger than most people think (but if you've managed to find my blog, you probably know already), and they usually disappear in the noise from social media, where the updates are streaming in all day long. Take a walk here if you want to discover a calmer blog atmosphere.
The web is fcked. A rather concise summary of the problems of the web. Is it too much to ask for websites that aren't throwing ads and autoplaying videos and cookie consent banners in your face? The browsing experience of 2021 is unbelievable. Too bad there's so much insentive for tracking and screaming for attention.

Optimizing images for the web

Sat, 11 Dec 2021 00:00:00 -0000

When you're crafting your own website without any website builder or tools like Wordpress, it can be a bit challenging to know how to optimize images. JPG-files from a digital camera is usually a few MBs in size. For example, an image from my Gopro, with a resolution of 4032x3024, generally amounts to at least 2 MB. If you're uploading such files, you're wasting bandwith, since most screens and monitors won't show the images in their full resolution anyway. I have found it suprisingly difficult to find "best practices" for optimizing images for the web using simple tools, but after a bit of research, I've managed to put together a script that combines three different command line tools to reduce image size without loosing to much quality.

The three command line tools are:

exiftool: Edits EXIF-information. Used to remove metadata from the image, such as geolocation, which I don't want to share on the web.
jpegoptim: Optimizes JPG-files.
imagemagick: Used to reduce resolution of image.

Here's the bash script, which takes files as input:

#!/bin/bash

for i in "$@"; do
    # Reduce all metadata, but keep orientation of image:
    exiftool -all:all= -tagsFromFile @ -exif:Orientation "$i"

    # Optimize image to a size of 500 kB:
    jpegoptim --size=500k "$i"

    # Reduce resolution so the width is maximum 1500 pixels 
    # (mogrify is a part of imagemagick)
    mogrify -resize 1500 "$i"
done

I am by no means an expert on this, but I've found a routine that works fine. If you have any comments or better solutions, let me know.

Master thesis: Deep learning to estimate power output from breathing

Fri, 18 Jun 2021 00:00:00 -0000

Yesterday I presented and defended my master thesis in "Computational Science: Physics" at the Physics Department at the University of Oslo. The title of my thesis is Deep learning to estimate power output from breathing.

The main goal of the project was to use recordings of breathing patterns to estimate how much power (measured in watts) a person generates during cycling. The motivation was to investigate whether breathing data can provide a universal and portable way of estimating physical effort/exercise intensity.

Popular methods of measuring exercise intensity include heart rate, speed measurements and power meters. Let me why we want to investigate an alternative to these traditional ways of effort gauging:

Heart rate has the drawback of not being as reactive as breathing is when we increase exercise intensity. It also drifts upwards during a workout, even though the workload remains constant.
Speed measurements (typically using GPS) are also very common, but they can't really be used as a standalone measure of intensity when used in hilly terrain or when the external conditions dictate the speed.
Power meters, while giving perhaps the most accurate measure of hard you're working during exercise, are currently mostly limited to cycling and indoor ergometers, where it's possible to use force sensors to measure power output.

Because of this, we wanted to see whether breathing data could be used to estimate how much power a person outputs during exercise! So to summarize, the main goal of the project was to find a relationship between these parameters:

INPUT: Respiratory inductive plethysmography (breathing)
OUTPUT: Power (physical effort)

The breathing data consisted of measurements of how the ribcage and abdomen expands and contracts (this type of measurement is called respiratory inductive plethysmography, or RIP). I used two of these sensors, as shown in the picture below.

Image of the RIP sensor to the left, and a figure showing how a person would wear two of the sensors the to right.

In order to find the relationship between breathing data and power output, I used neural networks, which is a type of machine learning.

I collected data from myself (it's fantastic to take a workout and be able to say "I'm just working on my thesis"), by recording RIP data, heart rate and power output during a series of workouts.

The data flow during data collection.

The RIP sensors contains a strain gauge that converts the mechanical strain on the belt into an electric signal, which means that the sensor outputs values measured in millivolt. An example is shown below, and it's easy to see how the signal increases when the rib cage expands, and decreases when it contracts.

Raw RIP data from the sensor around the ribcage.

Even though it's easy to recognize the breathing pattern when looking at the graph of the RIP signal above, it's not easy to discern how it behaves with regards to exercise intensity, at least not by visual inspection. Below is a graph showing an excerpt from an interval workout. The two RIP signals (in the two top graphs) have some strange behaviour and seem to jump around a bit. This can be a result of belt slippage, or that the person for example is standing up while cycling, reaching for a water bottle, or similar. Hopefully the neural networks can help us see past such irrelevant information in the input data!

Example of raw data from an interval workout.

I performed many different types of workouts, various types of intervals, steady-state efforts and workouts with ramp and pyramid structure. Some examples are shown below.

The data set consisted of workouts from mainly three different categories.

In order to enable the neural networks to make the best possible estimations, I needed to extract/engineer/derive (choose the term you feel is most understandable) some features from the raw data. To put it simply: Feature extraction means that I'm trying to find the most important information in the raw data and feed that to the neural network, so it's easier for the neural network to learn the relationship between the input and output data. But this also means that I have to figure out what the most important information is, and in the end, that's what I want the neural network to do... So you could say that it's a sort of collaboration between me and the neural network!

I extracted many types of features, but three of the most important ones are shown in the figure below. To keep this blog post short and sweet, I won't go into the details here.

This graph visualizes three of the most important features that a extracted from the raw data. Read the thesis (link at the end of the post) for more juicy details!

My best performing model was obtained by using a convolutional neural network, and when testing the model on unseen data it got an average error of 20%. That may sound like a lot, but the average error is affected by the fact that the model is not very good at estimating power values above 400 watts, since most of the data set contained power values below that (which is also where most of us spend most of the time when cycling).

An example of how the power output estimation by the model looks like is shown below. The output signal (in blue) is quite noisy, but the model is able to estimate approximately the correct power level most of the time.

Example of power output estimation, using extracted features from the raw data (the blue line shows the "predicted" values, which in this case is analogous to "estimated".

As a comparison, take a look at how the model performs when using only the raw data, without any extracted features (spoiler alert: Not well at all)!

To emphasize how important feature extraction is, here is an example of power output estimation using only raw data.

Conculsion:

Deep learning shows promising results for estimating power output from breathing data.
The results of this thesis enables a non-invase, portable way of estimating physical effort.
Future research should aim at study a larger and more diverse data set, collected from multiple subjects of various age, height, fitness level etc.

If you have some questions or thoughts about my thesis, or about the subject in general, please send me an e-mail on erikjohannes@protonmail.com!

Here are some links to learn more about the research project:

[NOR] Utstyrstest løpesko: The North Face Flight Vectiv

Mon, 31 May 2021 00:00:00 -0000

For litt siden fikk jeg muligheten til å være med å teste The North Face Flight Vectiv, som er en ny løpesko fra The North Face med en innebygd karbonfiberplate (siden det funka så bra for Nike og for gateløping, så må det jo prøves i terrenget også). Jeg har testa disse skoene noen uker på skogsstier, snø, grusvei, i myr og i lyng rundt om i nærområdet, både i vått og tørt vær.

Førsteinntrykket er at disse skoene er skikkelig høye, og jeg synes de er ustabile på ujevne og tekniske stier. Jeg løper vanligvis med litt lavere sko i terrenget, Arc'teryx Norvan SL (høyde: 19/12 mm på hæl/forfot) og Salomon S/LAB Sense 8 SG (høyde: 22/18 mm på hæl/forfot). Jeg var sikker på at Flight Vectiv hadde en mye høyere mellomsåle en dette, men da jeg sjekket opp spesifikasjonene på nettet så fant jeg ut at de en oppgitt høyde på bare 25/19 mm! Altså er det ikke så stor forskjell, selv om jeg opplevde at Flight Vectiv-skoene var mye høyere enn det jeg normalt bruker. Arc'teryx- og Salomon-skoene mine har ikke noen innersåle, men det har Flight Vectiv, så jeg begynte å lure på om den oppgitte sålehøyden kanskje ikke er målt med innersåle..? Uansett, det endrer ikke på at jeg opplevde Flight Vectiv-skoene som uegna for tekniske stier. Dempingen gjør forøvrig at de er greie å løpe med i nedoverbakker på jevnt underlag.

Grepet på skoene er ganske bra, og de sitter godt i alt bortsett fra dyp gjørme. Har ingen ting å utsette på yttersålen. Aner ikke hvordan den vil holde seg over tid, siden jeg bare har testa skoene i en begrenset tidsperiode.

Overdelen er forsterka på utsatte punkter (på sidene og rundt tærne), så det er bra. Det er også en sånn stretchy bit rundt ankelen, så man slipper å få masse småstein, snø og annet ned i skoa. Dette burde bli vanligere på terrengløpesko!

Forresten så er de ganske store i størrelsen. Jeg bruker 43 1/3 i de aller fleste sko, og testet Flight Vectiv i størrelse 43, og likevel var de egentlig litt for store for meg.

Jeg veide skoa til å være 310 gram per sko. Vanligvis løper jeg i sko som er rundt 200 gram per sko, og jeg synes 310 gram er ganske tungt. Det er ikke sikkert alle bryr seg så mye om dette, eller synes vekta er så høy, men det trekker ihvertfall karakteren opp. Jeg synes uansett det er litt spesielt at det reklameres med at skoa har "lav vekt".

Så, til det viktigste punktet: Karbonfiberplata. Merker ikke noe særlig til denne, må jeg innrømme. Det er uansett vanskelig å måle om den har noen konkret effekt på energibruk og hastighet uten å gjøre skikkelig undersøkelser, men jeg føler ikke at det er noe spesielt å løpe i denne skoa. Sålen er veldig stiv (naturlig nok), og det gjør at jeg den ikke er så behagelig å løpe i oppover bratte motbakker. Produsenten og forhandlere av skoa trekker fram at den har "energigivende egenskaper", og referer til både karbonplata og den såkalte "rockeren", som er den buede formen på sålen. Jeg følte ikke at skoene var noe mer energigivende enn andre løpesko, og hvis den plata utgjør noen forskjell, så er den ihvertfall så minimal at den ikke betyr noe i forhold til andre faktorer som passform, pris, og varighet på materialene.

Helt ærlig så synes jeg ikke skoene var noe særlig, og ihvertfall ikke til en så stiv pris (2199 NOK)! Tipper at skoene uansett selger godt pga karbonfiberplate-hypen. Når det gjelder løpesko så er jo ting utrolig individuelt, og passform og løpsfølelse er (ihvertfall etter min mening) de absolutt viktigste kriteriene. Det kan derfor godt hende at det er mange andre som kommer til å digge disse skoene selv om jeg ikke likte dem! Men er det virkelig verdt å bruke 2199 NOK på disse skoa? Tviler.

Mine erfaringer med skoene ble en del av en artikkel hos Runner's World. De kuttet dessverre en del av kritikken min av skoene, så derfor valgte jeg å publisere denne bloggposten med en komplett tilbakemelding på hva jeg synes om skoa.

[NOR] Je skal laga snøbrett i kryssfinér

Sat, 22 May 2021 00:00:00 -0000

Bestemte meg plutselig en dag for å lage meg et hjemmelaga snøbrett. Forutsetningene for dette prosjektet var at det skulle være billig og lett å lage. Når jeg sier "lett å lage", så mener jeg ikke sånn som noen youtubere framstiller det: "Dette er ekstremt lett å lage, du bare går inn i ditt spesialbygde verksted med alle disse spesialverktøyene, og så tar du dette kjempedyre materialet som er umulig å få tak i, og da tar det faktisk bare en ettermiddag før du er ferdig!" Duh...

Hvis det skal være realistisk for meg å gjennomføre, så må jeg bare behøve helt enkle verktøy, som de fleste har liggende i boden. Slik som håndsag, hammer, tommestokk, trelim... Du skjønner greia. Og så kunne jeg selvfølgelig ikke lage et snøbrett med bindinger, for det koster for mye.

Jeg oppdaga at snøbrett uten bindinger er et relativt utbredt konsept, men at det ofte kalles for powder surfer, eller bare pow surfer. Og det er heller ikke så uvanlig å lage dem selv i garasjen. Det er også noen små aktører som produserer powder surfers. Jeg synes spesielt Skorgeboards og The Fyra Board Studio sine brett ser ganske fete ut!

Etter litt søking rundt på internett fant jeg til slutt en video som var midt i blinken for meg: Cheap, Easy, DIY Powder Surfer. Denne videoen gir instruksjoner som til og meg jeg klarer å følge, og krever svært lite av utstyr.

Jeg kjøpte inn en plate med 4 mm kryssfinér av furu. Den er nok ikke like sterk som kryssfinéren han i videoen bruker (baltic birch), men det var det eneste jeg fikk tak i på den lokale byggforhandleren. Når det gjaldt trelim, så turte jeg ikke bruke noe annet enn det som ble anbefalt i videoen, så jeg bestilte en flaske Titebond 3 på nettet.

Jeg sagde finérplaten slik at jeg fikk tre plater på ca 150x40cm. Når trelimet endelig kom i posten, så rigget jeg til noen plankebiter på et gammelt bord, hvor jeg skulle utføre limingen. Plankebitene ble arrangert slik at finérplatene fikk den riktige svaien i front og bak på brettet. Skrutvinger måtte jeg låne, men jeg fikk bare tak i 9 stykker. Skulle nok helst hatt flere for å sørge for at de tre platene ble klemt godt sammen overalt.

Platene ligger klare til å limes, med noen plankebiter under som skal sørge for at brettet får den riktige svaien.

Jeg smurte på trelim mellom hver av de tre platene før jeg satte på skrutvingene. I ettertid ser jeg at jeg antagelig burde tatt på enda tykkere lag med trelim, men ikke noe å gjøre med det nå...

Her har jeg rigga til platene med skrutvinger mens det trelimet tørker. I tillegg til skrutvingene la jeg de tyngste tingene jeg fant oppå platene (noen steiner og hoggstabber). Merk at jeg også hadde lagt de limte platene i mellom to andre plater for å beskytte selve brettet.

Jeg lot limet tørke i to døgn, før jeg sagde brettet til den formen jeg ville ha. Har veldig liten erfaring med å stå på brett, så her ble det designet litt etter innfallsmetoden. Jeg brukte "Model Tre" fra The Fyra Board Studio som inspirasjon, og brukte ca samme breddemål og lengde som denne. Jeg endte opp med en lengde på 145 cm, og bredde på ca 36/27/28 cm på front/midt/bak.

Slik så det ut etter at limet hadde tørket. Svaien/rockeren ser brukbar ut.

Lagde en papirmal som jeg brukte til å tegne opp på begge sider av brettet, med utgangspunkt i midtlinja.

Etter å ha pusset kantene godt med sandpapir, så oljet jeg inn hele brettet med en polyuretan-basert gulvlakk. Deretter satte jeg inn undersiden med vanlig glidevoks fra Swix.

Var svært kjekt å ha elektrisk pussemaskin. Men å pusse for hånd er selvsagt også mulig.

Jeg borret et hull (4mm diameter) helt bak på brettet for å feste en sikkerhetsline i. Denne lina fester jeg i beinet eller beltet så ikke brettet forsvinner når jeg faller.

Feste til sikkerhetsline.

Feste til sikkerhetslina henger fast i brettet ved hjelp av ei knute.

Nesten ferdig, mangler bare noe padding til å stå på.

På oversiden satte jeg på to biter av et skumgummiliggeunderlag, som jeg festet med dobbeltsidig teip.

Slik ser det ferdige brettet ut.

Snøen begynner dessverre i forsvinne her i strøket, så måtte løpe et godt stykke opp på fjellet for å få testa brettet.

Your browser does not support the video tag.

Bitene av liggeunderlag som jeg brukte til padding for føttene falt dessverre av etter noen runder, fordi de ikke festet seg godt nok til den dobbeltsidige teipen. Tror jeg skal prøve med en friksjonsteip i stedet (sånn man bruker på trappetrinn etc. for å gi friksjon).

Alt i alt funka brettet bra. Gliden er selvfølgelig ikke like god som vanlige snøbrett, men det er en nok en fordel at det ikke går for fort for min del, så jeg klarer å holde meg på beina...

[NOR] Når skia får nok

Fri, 21 May 2021 00:00:00 -0000

De første randoneeskia jeg kjøpte var et par brukte Völkl Snowwolf. Jeg hadde lenge stusset på hvorfor så mange tråkka oppover fjellsidene med sånt tungt randoutstyr, når jeg jogga forbi med lærstøvler på lette fjellski, men jeg tenkte at det var på tide å prøve ut sånt utstyr selv. Det ble noen flotte toppturer den våren, men etter at jeg for første gang deltok i en randoneekonkurranse (RandoEnern på Oppdal), skjønte jeg at sko, ski og bindinger av den typen jeg hadde skaffet meg bare var masse dødvekt. Skal det være noe vits i å gå med randoneeski i fjellet, så kan det rett og slett ikke være så tungt og stivt. Skoa hadde så lite gåutslag at jeg av og til ikke merket om de stod å gåmodus eller kjøremodus...

Mitt første sett med randoneeutstyr. Altfor tungt til å være nyttig, men det var gøy å prøve.

Jeg solgte alt utstyret samme år som jeg kjøpte det, og fikk tak i konkurranseski og -sko på Finn.no. Dynafit PDG-sko fra 2015 (830 gram per sko), og Dynafit Race Performance-ski fra 2011 (990 gram per ski, inkludert binding). Dette er utstyr som det går an å gå på fjelltur med!

Randoneeutstyr laget for konkurransebruk, men det fungerer absolutt til alle typer turer i vinterfjellet.

Fem sesonger senere så begynner Dynafit-skia å bli passe slitt. Jeg har ingen andre randoneeski å bytte på, så dette utstyret har blitt brukt både til konkurranser og turer. Sålen er rett og slett ganske herpa, og de glir dårlig uansett hvor mye glidsmurning jeg smelter inn. Etter en tur nå i vår oppdaget jeg en skikkelig bulk i sålen, som også gjorde at det kom mye snø innunder fellen, slik at den fort løsna. Det var også brudd i stålkantene flere steder. Jeg tok med skia til en viss skimaker på Oppdal, som avsa en ganske tydelig dødsdom.

Brudd i stålkant på skia, i tillegg til at det begynner å sprekke opp på oversiden.

Sår i sålen.

Skimakeren oppdaga også noe jeg faktisk ikke hadde lagt merke til selv: Den ene skia var rett og slett knekt i tuppen. Sålen og det øverste laget på skia var fortsatt helt, men kjernen var knekt. Det stod altså mye verre til enn jeg først trodde.

Knekt ski.

Skia skulle få seg et siste farvel: Det ble en tur over Stornubben og Nautgardstind inkludert naboer i Jotunheimen. Den knekte skituppen flagret godt på vei ned fra fjellet, men den holdt stand, til tross for hardt og skavlete føre.

På tur bortover mot Nautgardstind.

Jeg søkte mye rundt i nettbutikker og på Finn.no for å finne en erstatning. Mitt inntrykk er at utvalget av randoneeski til konkurransebruk har blitt litt dårligere enn det var for 4-5 år siden. Kanskje har populariteten for slike konkurranser gått litt ned? Den var vel kanskje ikke så stor i utgangspunktet heller... Jeg fant til slutt et par på tilbud til under 4000 kr: Movement Apple 65. Samme mål som mine gamle ski, men litt lettere. Dessverre tok det litt tid før skia kom i posten, og jeg fikk plutselig muligheten til å ta en tur jeg hadde planlagt en stund: Stå ned en av bekkerennene nordvest på fjellet Tron ved Alvdal. Dermed ble det enda en tur på mine gamle, trofaste Dynafit-ski.

Bekkerenne ned fra Tron.

Skia klarte seg greit ned bekkedalen, men det ble litt ekstra utfordrende på hard skare med svært slitte stålkanter. Jeg stoppet i bunnen av dalen for å ta et blikk opp på svingene jeg hadde kjørt, og da jeg skulle snu meg tilbake oppdaget jeg at skia nå hadde gitt opp ånden helt: Tuppen av ski hadde åpnet seg som en skrellet banan. Fortsatt var det et par kilometer ned til veien, så løsningen ble å teipe sammen tuppen. ALLTID ta med teip på tur.

Det ble definitivt siste tur på dette skiparet.

De nye skia kom i posten dagen etter, og jeg fikk flytta over bindingene samme dag.

Alltid litt nervepirrende å drille hull uten rigg, men det gikk denne gangen også.

Skia fikk sin jomfrutur i Jotunheimen på vårsnø.

Offline Internet

Sun, 21 Feb 2021 00:00:00 -0000

A project I have wanted to for a long time is setting up an "offline Internet", or "Internet-in-a-box", which is an offline "library" and a wireless access point that nearby devices can connect to. To do this I used a Raspberry Pi 4 with the official Raspberry Pi OS installed.

The main steps in creating this "Internet-in-a-box" is:

I assume that the operating system already is installed on the Raspberry Pi (or any similar device).

Scroll down to the bottom of the post to see screenshots from the offline Internet!

Download the desired content

Many types of static content (HTML, plain text, pictures etc) can be served from the Raspberry Pi. For me, the most useful content is this:

Wikipedia (English). On this link you can find several versions of Wikipedia, in various languages and topics. If you want the complete English Wikipedia, download this version (it is 79 GB).
Project Gutenberg. This is a large collection of books from the public domain, which means that it can be downloaded for free. Different version can be found here.
Code documentation. I also have downloaded the documentation for the main programming languages and frameworks I use on a daily basis. This type of content can usually be downloaded in a variety of formats. I choose HTML-format if it is available, since I find it easiest to access in a browser (and some documentation in HTML-format also include a search engine). Examples:
- Python
- NumPy
- Pandas

Install software for serving content

I use two different ways of serving the content wirelessly to other devices, one for the .zim-files, and another one for other file types.

Serving .zim-files

For serving .zim-files, I use software called Kiwix. First I download Kiwix from this link (if you use a Raspberry Pi like me, choose the version called "GNU/Linux ARM 32-bits"). Unpack the zipped file with

tar -xvzf kiwix-tools_linux-armhf-3.1.2-4.tar.gz

Your version number might be different though. The contents are unpacked to a folder with the same name as the zipped file. I rename this folder kiwix to make things easier, and then move all my downloaded .zim-files to this folder. If you for example download a file called wikipedia.zom and another file called gutenberg.zim, the folder should look like

kiwix/
├── gutenberg.zim
├── kiwix-manage
├── kiwix-read
├── kiwix-search
├── kiwix-serve
└── wikipedia.zim

First you need to make a library containing your .zim-files:

./kiwix-manage library.xml add gutenberg.zim
./kiwix-manage library.xml add wikipedia.zim

This library can now be served by running:

./kiwix-serve --port=8080 --library library.xml

Go to localhost:8080 in the browser of the Raspberry Pi to see the content. If the Raspberry Pi is connected to a WiFi-network, you can view the content from other devices by going to [Local IP of the Pi]:8080 in the browser (first you need to find the local IP address of the Pi).

Serving other type of content

For all other file formats I use a Python webserver to serve the content (which of course requires that you have Python installed on your system). This simply means that I run the following command in the folder containing the files I want to serve:

python3 -m http.server 8081

In this case I have chosen to serve on port 8081, but this can be any other port (if you omit the port number, it will be set to 8000) as long it's not the same as the other servers you have set up.

Go to localhost:8081 in the browser of the Raspberry Pi to see the content. If the Raspberry Pi is connected to a WiFi-network, you can view the content from other devices by going to [Local IP of the Pi]:8081 in the browser (first you need to find the local IP address of the Pi).

Make server start automatically

In order to make the file serving start automatically when I boot the Pi, I simply add cronjobs. Add cronjobs by editing the crontab file:

crontab -e

Add the following files to the bottom:

@reboot /path/to/kiwix/kiwix-serve --port=8080 --library /path/to/kiwix/library.xml
@reboot python3 -m http.server --directory /path/to/files/ 8081

Now these commands will run every time you boot up the Pi.

Setting up the Pi to broadcast its own WiFi-network

(The instructions in this section were found in the Raspberry Pi documentation for using the Pi as a routed wireless access point).

After the content is ready to be served from the Pi, one has to make the Pi broadcast its own WiFi. By doing this, any device nearby can connect to the WiFi and access the Pi's content. If you already have a router that all your devices are connected to (including the Pi), you can skip this step.

I used the below commands to make the Raspberry Pi broadcast its own WiFi network.

Install access point software:

sudo apt install hostapd

Enable the accesspoint service and make it start on boot:

sudo systemctl unmask hostapd
sudo systemctl enable hostapd

Provide network management services:

sudo apt install dnsmasq

Utility for firewall rules:

sudo DEBIAN_FRONTEND=noninteractive apt install -y netfilter-persistent iptables-persistent

After installing the software above, I edited the configuration file for dhcpcd:

sudo vim /etc/dhcpcd.conf

These lines were added to the end:

interface wlan0
    static ip_address=192.168.4.1/24
    nohook wpa_supplicant

The Pi will be reachable in the IP address 192.168.4.1 after the setup process is complete, and you are connected to the WiFi of the Pi.

Then I created the hostapd configuration file:

sudo vim /etc/hostapd/hostapd.conf

country_code=<COUNTRY CODE>
interface=wlan0
ssid=<NAME OF NETWORK>
hw_mode=g
channel=7
macaddr_acl=0
auth_algs=1
ignore_broadcast_ssid=0
wpa=2
wpa_passphrase=<PASSWORD>
wpa_key_mgmt=WPA-PSK
wpa_pairwise=TKIP
rsn_pairwise=CCMP

In my case, <COUNTRY CODE> was NO because I live in Norway, and one also have to fill in a name for the network and password.

After a reboot (sudo systemctl reboot) the Raspberry Pi will broadcast its own WiFi network, and you can connect to it using the name and password specified in the hostapd.conf file.

Conclusion

After these tree steps are done, you can connect to the Pi's network using WiFi, and access the content on 192.168.4.1:8080 or any other port number that you have chosen to serve your files on. I find this useful both to limit data usage, but also because I like to work offline if I do not strictly need to do specific things online.

Text-only news sites

Sun, 07 Feb 2021 00:00:00 -0000

I'm deeply fascinated with minimalistic and text-only approaches to web design. This January I moved to a place where I have to use a cellphone hotspot to connect to the Internet, and thus I am also very aware of my usage of cellular data, since I need to pay for every GB that I use. This has made me even more conscious of how large many common websites are, and how much useless stuff I'm downloading by just browsing the web.

There seems to be a gaining interest in text-only/low-bandwith websites, as you can read more about in this article. This has several advantages, especially for people that live under conditions with severe limitations on bandwidth. I think it's very cool to see that two major news sites offer low-bandwith versions of their sites:

For most mobile phone companies in Norway (as far as I know), it is common to throttle the bandwidth speed to 64 kbit/s when you have reached your data cap. If more people and companies started to scale down their websites (or at least make low-bandwith versions), one could browse the web comfortably at those kinds of speeds, and websites would be much more accessible. I think this could be very beneficial, especially for essential content such as news.

Dark mode on a website

Sat, 30 Jan 2021 00:00:00 -0000

I just discovered that it is quite easy to make a website adjust the style to whether the client is using dark mode or not. Take a look at a snippet of my CSS for this website:

* {
    --txc: #393838;
    --lc: #1183b9;
    --bgc: #fff;
    --hlc: #70b4d5;
}

@media (prefers-color-scheme: dark) {
    * {
        --txc: #f5f5f5;
        --lc: #70b4d5;
        --bgc: #212121;
        --hlc: #1183b10;
    }
}

The first block defines the default values for two text colors, the background and a highlight color. With the second block, those variables are redefined if the user is using a dark system theme.

This might be common knowledge for a lot of people, but have not seen this functionality much "in the wild". Maybe some people prefer to keep the same colors anyway, but I think it is very comfortable that websites dynamically adapt to the system theme, especially when there are lots of articles.

Building and maintaining a minimalistic website

Fri, 29 Jan 2021 00:00:00 -0000

When I first started making my own website, I wrote everything manually, from scratch. This type of approach is very fun and instructional, and probably the best way if one wants to learn how a website functions.

After a while though, this approach gets quite tedious, because the total number of pages piles up (mostly blog posts) and it gets hard to manually update all of them when you want to make changes. This is one of the perfect use cases for a Static Site Generator, which is a tool that uses templates to build the static HTML pages for you. You are able to define the "rules" and layouts for how the page should look independently of the content, and gives you a lot of flexibility with regards to what you want to automate and what you want to do manually.

I resisted the temptation of using an SSG for a while, since I enjoyed writing everything from scratch and having a minimalistic approach. But, when I wanted to add an RSS feed to my site, I figured that I couldn't be bothered with writing an feed.xml file manually. I chose to try out Hugo, since I didn't really need to install anything, I could just have a binary executable in the folder where I edit my website, and then build it there. There was a lot of syntax to learn in order to get everything working as I wanted, but it was also nice to have RSS feed, tags on blog posts and blog feed being built automatically.

Even though Hugo worked fine, I felt like it was to bloated and complex for what I wanted. I can certainly recommend Hugo in general, but I want a workflow that has as few dependencies as possible, and is sustainable for a long, long time, with very little maintainance. Keeping it purely HTML+CSS, without a complex build process, would be preferable. I read an article called Why I Built My Own Shitty Static Site Generator, which expressed some of the same thoughts I was having, and I was inspired to stop using Hugo, and restructure this website. I could still use most of what I had written from before, but I wrote my own small script for adding head, header and footer (the elements that are the same on every page) and building the blog feed. After that, the only thing missing was generating an RSS feed. This is probably the most complex part of my build script, but I made it work in the end.

There are some people who are even more hardcore (check out My stack will outlive yours), who avoids having a build step at all. I would like to eliminate the need for a build script, just to keep things even simpler, but I appreciate having the option of easily changing some elements in a template-file rather than on every page. It's also very convinient to write blog posts in Markdown instead of HTML (my build script does this conversion as well). The build script is still quite small and I know exactly what it does and does not, and I prefer that over a more complex SSG for my use case.

After further delving into the world of minimalistic web development, I discovered the 512k.club, the 250kb.club and the 10kclub.com. Those are collections of websites that have a size below a certain (quite small) threshold. I think the goal is just to make a point of how (often needlessly) enormous many websites are, and that it's perfectly possible to make good-looking websites without many MBs of data.

My own main page has a size of 1.6 kB at the time of writing, but of course, there's not much there. Some of the blog posts contain a lot of images, and those pages are certainly bigger in size, but the point is that it's totally fine accessing the website on old browsers and with slow internet speed. I wish that were the case for most websites.

Finally, I see this website as my online "profile page", in the same way that many people have a Facebook-profile as an online presence. Maintaining a website is a little bit harder than making a Facebook-profile, but it gives you much more control and flexibility.

To round off, here are some minimalistic websites that I find interesting (regarding design, content or both):

Reducing the temperature on a Raspberry Pi

Tue, 10 Nov 2020 00:00:00 -0000

I have for several months been hosting my own cloud services using Nextcloud on a Raspberry Pi 4. At first I was using a plastic case for the Pi, but I quickly figured out that I had to use something else, as the temperature was constantly in the range of 70-75 degrees Celsius, which I figured was too high when it is supposed to be an always-on system.

I purchased an extra large heat sink and a cluster case with fans in order to reduce the temperature, but I was unsure of how much it would help, and also whether the fan was as quiet as the product description claimed. The delivery of the items took some weeks, so in the meantime I used a stack of coins placed on top of the processing unit, which worked remarkably well (at least compared to my expectations): The temperature was reduced with about 15-20 degrees Celsius.

Using a coin stack as a heat sink.

After attaching the heat sink to the Pi, and then installing the Pi in the cluster case and mounting the fan, the temperature of the Pi dropped to the range of 30-35 degrees Celsius when running a Nextcloud server. This was much better than I expected, and the fans hardly make any noise at all!

Two Raspberry Pis installed in a cluster case, with the old plastic case next to it.

I have also tried a heatsink case for another one of my Pis, and the temperature is reduced by about 20-25 degrees Celsius, so that is not a bad option either.

Keychain bit holder

Thu, 24 Sep 2020 00:00:00 -0000

I recently got a small keychain multitool called True Utility Fishface, which amongst other things has a bit driver on it. It is basically just a hex-formed hole in the multitool, and can be used to drive any sort of standard screwdriver bit. The Fishface also comes with special dual bit, which is fastened to the multitool with to small rubber bands, and consists of two sizes of the cross-point/Phillips screwdriver head.

The True Utility Fishface.

The Fishface with the dual bit in the driver.

One of the main advantages of carrying a multitool with a bit driver is that it can be adapted for your specific needs. For me it is very useful to carry hex bits, specifically the 4mm and 5mm, since those are commonly used on just about everything on bikes. But how to carry the extra bits I need? One nifty solution would be to drill a hole in the bits, and then they could be fastened directly on the keychain ring, but I do not have the kind of tools needed for that operation. While doing some brainstorming, my eyes landed on a pen on my desk, and suddenly I realized that the shell of that pen was the perfect size for housing a couple of bits.

An identical pen to the one I used for making the bit holder, alongside the finished "product". The two items are aligned, so you can see where I cut off the top and bottom of the pen.

After gently sawing the plastic of the pen in the appropriate plazes, using a serrated knife blade, the remaining cylinder could fit the two bits I want to carry around. In one end I fastened a keychain ring with some duct tape, and the other end is closed by screwing the small remaining piece of the bottom part of the pen into place. The bottom part of the bit holder is actually open, but the hole is to small for the bits to fall out.

The exact process of making the bit holder will differ slightly depending on the pen you are using, but it is anyway a quick and cheap way of making a small bit holder with things you most likely already have in your house.

[NOR] Geiranger, Dalsnibba, Fossfjellet og Rundegga

Sat, 04 Jul 2020 00:00:00 -0000

På Strynefjellet og i fjellene over Geiranger er det fortsatt mye snø for tiden, og det var mange som var ute og fikk lufta randoskiene sine opp til Blåtind (1663 moh) i dag. Ikke fullt så mange var det som brukte sykkelen oppover fra Geiranger, men vi var ihvertfall tre stykker som stod klar på fergekaia i morges for å bryne oss på stigningen opp til Dalsnibba. For min del ble det en litt tidligere stopp enn planlagt med punktering etter å ha passert Knuten, men med ekstraslange i sadeltaska kom jeg relativt raskt videre, og mange hårnålssvinger senere ankom jeg Dalsnibba (1476 moh).

Endelig oppe.

Utsikt fra Dalsnibba.

Jeg hadde løpesko med meg, stroppet under sykkelsetet, så turen kunne gå videre i delvis snødekt terreng opp til Fossfjellet, på 1574 moh. Opp dit er det morsom kravling i steinur. Sørover kan man lange ut i slake nedoverbakker, både på snø og isskurt berg, før det går oppover til Rundegga (1615 moh).

Rundegga over Fossvatnet.

Utsikt fra Fossfjellet.

Snøløping på Rundegga.

Fra Rundegga gikk turen tilbake til Dalsnibba, og derfra trillet jeg ned til Djupvasshytta.

Fortsatt fotogene brøytekanter opp Nibbevegen.

På vei ned igjen til Geiranger ble det en liten stopp for å utforske Kvanndalssætra, en samling av gamle seterbygninger som ligger idyllisk til i Kvanndalsbotn, nord for Dalsnibba.

Kvanndalssætra.

Kveldssol over Geiranger.

[NOR] Rondane Rundt / Rondane 2K-maraton

Sat, 27 Jun 2020 00:00:00 -0000

Det er omtrent 10 år siden jeg første gang hørte om konseptet "Rondane Rundt". Navnet har flere betydninger; for syklister er det først og fremst snakk om ruta rundt Rondane gjennom Grimsdalen og Atnadalen, over Venabygdsfjellet og opp Gudbrandsdalen. For min del er det først og fremst snakk om den beryktede turen over Rondanes ti topper over 2000 moh (med primærfaktor på minst 50 meter), en tur som har blitt mer kjent som "Rondane 2K-maraton".

Opp gjennom åra har mange gjennomført turen i ulikt tempo, og flere har gjennomført på under 12 timer. Den raskeste tida på en slik rute er ikke alltid like lett å fastslå, men det har de siste åra dukket opp mer organiserte prosjekter for å registrere denne type rekorder, som på engelsk kalles fastest known times (FKT). Et slikt prosjekt er fastestknowntime.com, hvor også Rondane Rundt ligger inne under navnet Rondane 2K Marathon. Rekorden har sunket gradvis de siste åra, og så vidt meg bekjent er det tre stykker som har klart å presse tida under 10 timer:

Rutevalg har variert, men det har utviklet seg en slags enighet om at start og mål er Rondvassbu (som også er det mest naturlige stedet å ha som utgangspunkt). I tillegg gjelder det at man skal ha tatt på toppvarden på hver topp, og ikke motta noen støtte underveis. Den tradisjonelle rekkefølgen på toppene er

Veslesmeden (2015 moh)
Storsmeden (2016 moh)
Trolltind (2018 moh)
Digerronden (2016 moh)
Midtronden Vest (2060 moh)
Midtronden Øst (2042 moh)
Høgronden (2115 moh)
Rondeslottet (2178 moh)
Vinjeronden (2044 moh)
Storronden (2138 moh)

Fredheim tok en alternativ rute i 2018, og forkortet distansen med noen kilometer ved å løpe ut til Trolltinden først, og deretter tilbake igjen og over Stor- og Veslesmeden, før han fortsatte over Digerronden og de andre toppene i tradisjonell rekkefølge. Da jeg endelig fikk tid og mulighet til å gjennomføre turen selv, valgte jeg å ta samme rute som Fredheim, for å unngå det lange partiet mellom Trolltinden og Digerronden, men jeg ville løpe motsatt vei, slik at jeg kunne skli på snøen ned fra Rondeslottet.

På vei opp fra Rondvassbu.

Jeg syklet inn til Rondvassbu fra Spranget lørdag 27. juni, og kl 8:10 la jeg i vei østover. Opp til Storronden er det tydelig sti hele veien, men jeg prøvde å ta en litt rettere linje noen steder opp langs ryggen. Etter 52 minutter var jeg på toppen, og skulle så ned i skaret mot Vinjeronden. Her hadde jeg aldri vært før, og jeg klarte dessverre å havne for langt øst, slik at jeg måtte løpe litt opp igjen for å komme på rett side av et gjel som går ned i Storbotn.

Stien opp til Storronden.

Mot Vinjeronden.

Vinjeronden og Rondeslottet ble unnagjort raskere enn planlagt, og varmen var heller ikke så slitsom som jeg hadde fryktet. Jeg hadde fått tips på forhånd om at det lå snø omtrent helt ned til Langglupdalen fra Rondeslottet, men det ble litt vinglete rutevalg ned den første biten, da jeg ikke klarte å se hvor snøfonna startet. Etter å løpt noen hundre meter fra toppen, fant jeg en liten bekk hvor jeg fikk fylt ei vannflaske, og like etter på kom jeg endelig inn på snøfonna. Her kunne jeg både skli og løpe nedover, så jeg kom jeg veldig effektivt ned i dalen. Fylte to flasker til i Langglupbekken, før jeg startet på skråningen opp mot Høgronden. Her er det lyng i starten, som kanskje kan oppleves som et kjærkomment avbrekk fra steinura, men jeg synes det var tungt å gå i, da man hele tida synker litt ned for hvert steg.

Utsikt mot Høgronden fra Rondeslottet.

Utsikt fra Høgronden. Her ser man tydelig snøfonna jeg fulgte ned i dalen.

På toppen av Høgronden kjente jeg vinden for første gang den dagen, og det var deilig å få litt avkjøling. Over til Midtrondene og til Digerronden var beina fortsatt gode, men i steinura ned fra Digerronden begynte det å bli litt tungt. På forhånd hadde jeg planlagt å ta rettest mulig linje ned fra toppen og mot Veslesmeden, men det er såpass mye bratt ur at jeg følte det ble lettere å løpe i en liten bue nedover. Nede i Bergedalen var det deilig å få løpt litt i lyng og på sti, og jeg fylte flaskene flere ganger i bekkene jeg passerte. I tillegg måtte jeg stoppe opp å helle ut stein og grus som hadde samlet seg i skoene på vei ned fra den forrige toppen.

Utsikt fra Høgronden mot Midtrondene.

Etter å ha passert på sørsida av den sydligste Bergedalstjønna, bega jeg meg oppover skråningen som leder opp til ryggen mot Veslesmeden. Farta var ikke like høy som tidligere, men jeg fokuserte på å holde jevn og stabil intensitet. Den siste biten opp mot Veslesmeden innebar litt kravling og lett klyving, og kryssing av et lite bratt snøparti, før jeg kom inn på stien fra Rondvassbu noen timeter fra toppen. Her var det en god del folk, og det hjalp på motivasjonen. Nå gjenstod det bare to topper, men det var også de eneste toppene jeg ikke hadde vært på tidligere, og dessuten de som krever litt mer konsentrasjon for å bestige. Jeg hadde holdt et øye med snittfarta mi, og så vidt jeg kunne se, lå jeg foran skjema til å slå rekorden.

Et lite snøparti rett under Veslesmeden.

Jeg kunne holde grei fart ned i skaret mot Storsmeden, og klyvingen begynte ganske umiddelbart etter at det begynte å gå oppover igjen. Jeg hadde hørt på forhånd at det skulle være noen små varder oppover som viste vei, men jeg var likevel litt usikker på hvor lang tid jeg kom til å bruke opp denne ryggen. Det viste seg å være ganske greit, men jeg brukte litt ekstra tid på å vurdere hvert veivalg for å være sikker på at jeg ikke kastet bort tid på klyve opp feil sted. Akkurat i det jeg kom opp det siste klyvepartiet, kjente jeg de første dråpene. Regnjakka kom raskt på, men det gikk ikke raskt ned fra Storsmeden; skoa mistet helt grepet på steinene. Flere ganger skled jeg i ura, og en av gangene slo jeg siden av hodet i en stein. Etter en liten pause konkluderte jeg med at det hadde gått fint, men det ble en påminnelse på hvor lite som skal til før det kan gå ordentlig galt!

Fra Veslesmeden mot Storsmeden.

Ut mot toppen på 1897 prøvde jeg å holde meg litt nede i fjellsida for å slippe å gå opp unødvendige høydemeter. Her bestod terrenget av små, skråstilte sva, som var svært glatte å forsere når det hadde blitt bløtt. Jeg angra på rutevalget, og tenkte at på vei tilbake skulle jeg gå lengre opp, hvor jeg kunne holde høyere fart. Ryggen utover var ganske morsom, men jeg var spent på hvordan den siste biten kom til å være, da jeg hadde lest noen sprikende turrapporter på nettet i forkant. Det er sikkert mulig å ta flere forskjellige opp langs ryggen, men den mest brukte stien var veldig tydelig, og også merket med små varder. Jeg møtte et par stykker i det siste skaret på ryggen, som tipset meg om at det kunne være litt vanskelig å finne veien ned igjen, da det er flere grusrenner som kan se ut som stier. Dette var nyttig informasjon, og jeg fulgte ekstra godt med hvor vardene befant seg, slik at jeg effektivt kunne løpe ned igjen uten å måtte stoppe og vurdere veivalg.

Oppe på Trolltinden kunne jeg meget fornøyd konstatere at dette var dagens siste topp, men jeg kjente på at det fortsatt var en del kilometer igjen. Jeg så på snittfarta mi, og begynte å ane at rekorden hang i en meget tynn tråd. Tilbake over ryggen var det vanskelig å presse farta opp noe særlig, da jeg først og fremst måtte fokusere på å komme meg trygt fram på slitne bein. Men da jeg skulle ned i skaret over Kaldbekkbotn kom jeg inn på en snøfonn, og håpet om å ta rekorden kom tilbake. Fra Kaldbekkbotn og til Rondvassbu er det sti hele veien, og jeg fikk endelig strekke ut beina litt. Minuttene gikk kjapt, og jeg skjønte etterhvert at dette kom til å bli utrolig jevnt. Motivasjonen var høy de siste par kilometerne, men samtidig var beina veldig tunge. I det jeg kom inn på grusveien mot Rondvassbu, var tida 8:47, men GPS-klokka viste bare timer og minutter, så jeg hadde ikke oversikt over hvor lang tid jeg hadde på meg. Jeg spurta det jeg klarte, og lettelsen var stor da jeg stoppa klokka idet jeg ankom hovedinngangen på Rondvassbu, og så at tida var på 8:47:52.

Fremme ved Rondvassbu etter 8 timer og nesten 48 minutter.

Utafor hytta satt det noen vennlige sjeler som hentet et stort glass vann til meg mens jeg fikk igjen pusten. Jeg fikk ikke med meg navnene deres, men tusen takk for hjelpen! Etter å ha summet meg litt, gikk turen tilbake til Spranget på sykkel.

Digg å være ferdig.

Under ligger et kart over ruta jeg tok. Oversikt over både dame- og herrerekord finnes her, og link til økta finnes her.

Kart over ruta.

Article on Facebook's targeted advertising

Sat, 30 May 2020 00:00:00 -0000

Some interesting revelations on Facebook's targeted advertising.

This isn’t the first time Facebook has faced scrutiny for the ways its ad tool can target conspiracy-minded people. A 2019 Guardian investigation found that advertisers could reach people interested in “vaccine controversies.” In 2017, ProPublica reporters (two of whom are now at The Markup) found that advertisers could target people who were interested in terms like “Jew hater,” and “History of ‘why jews ruin the world.’ ”

Read Want To Find A Misinformed Public? Facebook's Already Done It.

Dogs relaxing

Wed, 25 Mar 2020 00:00:00 -0000

Rondane

Sat, 07 Mar 2020 00:00:00 -0000

Tidlig i mars hadde det kommet unormalt mye snø i Rondane, og jeg utnytta muligheten til å dra på topptur i området. Planen var egentlig å få gått en runde over noen av toppene over 2000 moh, men mye dyp snø og skredfare gjorde at jeg snudde ved Midtronden, og gikk samme vei tilbake.

Nederst på sida ligger en video fra turen.

På vei oppover bjørkeskogen.

På toppen av Høgronden med utsikt mot Rondeslottet.

De to Midtrondene i sikte.

Flanken til høyre hadde jeg opprinnelig tenkt å stå ned, men der var det

På toppen av Midtronden Aust, med utsikt mot Midtronden Vest.

Sola nærmer seg horisonten, rett over Rondeslottet.

Nede igjen ved Gammelgarden.

Your browser does not support the video tag.

Skyline Scotland

Fri, 20 Sep 2019 00:00:00 -0000

Map of the Glen Coe Skyline route.

View from Na Gruagaichean, the finish line of Mamores VK.

The final ridge of Mamores VK.

Scottish sunset 1.

Scottish sunset 2.

Scottish sunset 3.

The gear I used during Glen Coe Skyline (I brought only some of the food shown in the picture.)

Nervous atmosphere before the start of Glen Coe Skyline.

I was running at the tail of the lead pack during the first kilometers of the race.

This was the last picture I took during the race. At this point I understood that things were going much better than expected, and I needed to stop fumbling around with my GoPro camera.

Finally done.

Can\'t describe the feelings I had at this moment.

A trophy and money. Best prize ever.

Rather happy to be on that list of names. Photo credit: Skyline Scotland.

Nordmarka

Tue, 10 Sep 2019 00:00:00 -0000

France

Fri, 05 Jul 2019 00:00:00 -0000

Maridalen

Sun, 10 Mar 2019 00:00:00 -0000

Grefsenkollen

Sun, 17 Feb 2019 00:00:00 -0000

Storlidalen

Sun, 27 Jan 2019 00:00:00 -0000

Andøya

Mon, 28 May 2018 00:00:00 -0000

Svalbard

Mon, 30 May 2016 00:00:00 -0000

Femundløpet

Sun, 02 Feb 2014 00:00:00 -0000

Erik Johannes Husom

Superintelligence is here

Review of Nnormal Kjerag 1.0 trail running shoes

The birth of a religion

Wasted large language models

Computing within limits

Making wooden skis

On the usefulness of AI agents

Absence of agents

Productivity and value

AI agents in AI research

Collateral damage

Conclusion

Outsourcing thinking

When should we avoid using generative language models?

Personal communication and writing

Valuable experiences

Building knowledge

The extended mind

What we think about does matter

Conclusion

Links 8

Sustainable living

Bullet points and bold text

Couloir skiing in Oppdal

Human recommendations rather than algorithmic curation

Race report from Black Lake Race 2025

Links no. 7

The wrong question about AI

Thoughts on desktop computer automation

Privacy is (more) dead

The appeal of delegation – no matter the cost

Automating communication

Conclusion

When choosing digital tools

Links #6

[NOR] KI og klimaavtrykk på NRKs Abels Tårn

Better images of AI

Links #5

AI's sustainability issues are gaining awareness

Microsoft playing both sides

Big Tech's emissions are surging

Generative AI and energy-efficiency

Will "greener" AI help?

Dotfiles: Starting over after eight years

Time for a clean slate

Bash

Tmux

Vim

Conclusion

ChatGPT vs. your voice

Links #4

Sunshine and solitude on Mt Fuji

Places to discover the Small Web

The future of social media

Links #3: Large language models

Democratizing generative AI

Green AI in a world of LLMs

Skiing new lines

Playing in powder with Gjert

Reconstructing Obsidian features in Vim and Bash

Easily insert links to other files

Open link to another file with a keypress

Opening a random note

General navigation

Automatically update links when renaming or moving file

Links #2: Note-taking

My best reads of 2022

The Age of Low Technologies by Philippe Bihouix

Scrambles Amongst the Alps by Edward Whymper

The Mountains of my Life by Walter Bonatti

The Way Home by Mark Boyle

Amusing Ourselves to Death by Neil Postman

A couple of other books on how we are affected by technology

A feed for sports and outdoors activities

Gjert the border collie doing what he loves most

Building a laptop

On self-hosting, local storage and portable git hosting on iPhone

Links #1

My best reads of 2021