Reflections on AI, May 2026

[Skip to bullet list of my view of the state of AI in May 2026]

Early takeoff

In 1956, when Dutch computer scientist Edsger W. Dijkstra invented his famous algorithm for finding the shortest path between cities on a map, computers (the machines) were widely regarded as useless. There were few everyday problems that computers could help with, to such a point that “programmer” was not considered a real job (while notably “Computer” was still a job title). Yet, to a small number of people it was apparent that this technology had massive value and would revolutionize the world – clearly finding the shortest path between any two points on a map is useful.

May 2026, as I begin writing this text, I see certain parallels with AI. To me and my colleagues, the utility is extremely noticeable: all the things I did when I started my job one year ago are today done by agents, and humans merely ask the questions. In many ways, the job I started does not exist anymore. I used to be an experimental physicist; it is not normal to go from a physicist with a bit of coding experience to one of the most productive software engineers at a software company in just a year. I am at least twice as productive as I was a year ago, and I have agents that work for me day and night doing as much work as I was manually doing before. What I am experiencing is continued acceleration in that direction. Some call this early takeoff: the start of an inevitable societal transformation.

Yet, when I talk to people outside of tech, their belief in AI capabilities is… limited. Many still question the usability of the technology and some even speak of an impending bubble bursting. This discrepancy in perception is stunning! To be clear: the utility achievable today is not limited to writing code. I already do my taxes, business admin and bookkeeping using Claude. It just works. In fact, it does a better job than the consultants I used to hire for thousands of dollars per year.

And somehow it is true that my friends speak of three-year timelines until AI can even start doing any kind of bookkeeping reliably. Why? Diffusion is much slower than I had expected. That is, the time it takes from when something becomes possible to when it becomes widely adopted. And in a sense this should come as no surprise: if you do not use frontier AI at least every few weeks, you will be far behind in your mental image of what is possible. Humans are bad at internalizing exponential growth (as highlighted by e.g. the Covid-19 pandemic). The fact is still that most people have never paid for AI, have never used frontier models, and have not spent 20 hours trying to solve their problems with AI. And why would you, if they seem mostly useless? It is hard to even come up with what tasks to give it if you have no intuition for AI capabilities. These things take time to learn.

Sometimes tech people like to say things like “It will just take 5 minutes to do with AI”. This is a bit like saying “Solving this complex integral will take just 5 minutes”, neglecting to mention the need for hours of studying calculus, including all the little tricks to reach the answer faster. Learning to use capable tools takes time because we need to build intuition for the capabilities, and all the small things a seasoned user takes for granted.

I notice the gap in perceived AI capabilities even internally at my company (which by all standards is at the very forefront of AI adoption). Even just a slight difference in capability perception leads to a massive difference in possible output when it comes to certain tasks. For example, we recently beta-tested an “agentic coding interview” where we wanted to allow all types of AI tools. I was given 45 minutes. I just copy-pasted the interview task into my local AI setup and went to grab some snacks. After 17 minutes it had completed the task flawlessly and even built out an extensive demo and an explainer of exactly what it had done. When others tried it, it took them anywhere from 14 to 65 minutes to solve the task – while having access to the exact same tools!

This is just to illustrate the variance of output even within a very AI-forward organization. We had to rethink the interview.

Being “AGI-pilled”

There is this fascinating concept of being AGI-pilled. It basically describes someone who believes that Artificial Intelligence is going to be human-level at most things, and that it will rapidly reshape society, the economy and the way we work and live.

It is a bit presumptuous to implicitly claim that you are one of few who see the world as it truly is. To top it off, the term “AGI” is extremely vague. Yet, it is kind of useful as a shorthand.

Am I AGI-pilled? I’m starting to believe that it is probably wishful denial to believe that humans will be better at making any decisions about code at all in the very near future. The question is rather how far beyond code that will be true, and how soon. I believe this because I have seen it happen bit-by-bit and with a clear trajectory in my own work. For example, I recently built a cost-monitoring agent for Lovable. For my first version, I built out a clever (and complex) anomaly-detection framework using things I’d learnt from signal processing. I then told an agent to just run that program every few hours and report anomalies. It was terrible. It triggered on random noise and daily fluctuations. It failed to notice slow but sustained changes, akin to someone making a small but bad change (the stuff I wanted to catch). What did I do instead? I just told the agent: “Fetch the recent data and look at it. If anything looks off, page me and investigate possible causes.” Yes, seriously – that worked flawlessly! By the time I get to my computer or phone, it has usually already found the root cause. My ego is still recovering (though it feasts on the knowledge that the agent is saving us millions of dollars, and it realizes this will continue to happen).

Where we used to have to prompt models hard to use best practices for coding (refactoring code, building small extensible parts, making sensible abstractions) we might soon find ourselves prompting the opposite: “Please make changes in chunks humans can understand”, “Don’t do cleanup for simple tasks”, “Don’t overthink future extensibility”. To use the manager analogy: I think we are moving from “Give clear instructions and inspire” to “Get out of the way and trust” – if our goal is maximum output. What this means for humans is profound and hard to predict. I strongly believe we are severely underestimating the impact this will have on work. We will need to spend much more time writing specifications and deciding what to do, instead of how to do it.

I have started giving some of my close family and friends Claude subscriptions to help bridge this gap. What I notice is that it is still hard to get going; the time investment needed to see value is still high. I have hundreds of colleagues who teach me all the latest workflows, showing me the things that work now that did not work 4 months ago. I constantly have to update my beliefs.

Closing the loop

I must briefly mention a recent article published by The Anthropic Institute (a part of Anthropic): “When AI builds itself”. It talks about recursive self-improvement: AI accelerating AI development. In the article they present evidence of this feedback loop. I recommend reading the full article. In my opinion it is well written and realistic.

One thing I want to highlight is their mention of Amdahl’s Law: the authors basically say that as one part of a system becomes very efficient, something else becomes the bottleneck. Writing code is now almost free, but reviewing it is still expensive. This does to some extent explain the slow diffusion we are observing: human adoption is bottlenecked by human timelines – how often we talk to each other, try new things, and update our beliefs. Something that struck me is that AI systems appear remarkably good at identifying bottlenecks and automatically addressing them. I do believe an argument like this is the strongest argument against AGI in the near future: that we may hit some bottlenecks that are simply slow to get past (Compute? Energy? Metal atoms?) However, I also think this risks being wishful thinking – at least I would not rely on it.

When I first learnt of the model that has come to be known as “Claude Mythos”, it became clear that this level of capability would become generally accessible within 6 months (that is, within at most a few months now). Because even if Anthropic does not release it, someone else will have caught up by then. If nothing dramatic changes, this will continue to happen: progressively more powerful models will keep being released.

The question most people seem to still be asking is “Can it ever?”. I instead urge everyone to internalize the current state of AI and try to get a sense of how fast things are still developing. Test things! Experiment! Talk about it! This could very well be 100 times more important than “business as usual”. If you think about these issues today, you will have an incredible opportunity to shape the future.

The question we collectively need to urgently answer is: What do we want the world to look like when AI development continues to accelerate and becomes widely adopted? Do we actually want this continued acceleration? What does a world with full AGI actually look like?

Benjamin Verbeek
Stockholm, Sweden
May 2026

Below is a concrete list of observations I have made in May 2026 about the state of AI. It may help give some more substance to why I observe continued acceleration.