In the days around September 28th 2026 I had my AGI moment. I do not mean that someone announced artificial general intelligence. I mean the point where an AI did the kind of work I do better than I would have done it myself. This post is about how I got there, what happened in those days, and what I think it means for me from now on.

Years of sparks

This moment did not come out of nowhere. I have been exploring AI for years and was interested in the theory underlying learning and intelligence for a long time. Therefore, I was naturally excited about the capabilities of large language models (LLMs), AI systems trained on huge amounts of text, like ChatGPT, which I have used since 2023, at work and privately. I followed every big step: the first reasoning models, o1 in 2024 and then o3, and in November 2025 the introduction of agents. An agent does more than answer in a chat window. It works through a task on its own, reads and writes files, runs programs and checks its results. So by now I have more than three years of experience with these systems, the good and the bad.

My milestones, 2023 to 2026
  • 2023, ChatGPT. OpenAI launched ChatGPT in November 2022 on GPT-3.5, and I started using it in 2023. Impressive text, but wrong too often to rely on.
  • March 2023, GPT-4. Clearly better. Researchers at Microsoft called their paper on it “Sparks of Artificial General Intelligence”, and that was my feeling too: sparks, not yet fire.
  • Maths. I tested it on maths problems. Sometimes it found a nice argument, often it wrote a confident proof that was wrong.
  • Work that did not work. Whenever I gave it a bigger task in my job, I spent more time correcting the result than doing it myself would have taken. Lots of frustration.
  • But always the sparks. Again and again there were moments where it solved something I did not expect it to solve, and that kept me coming back.
  • September 2024, o1. The first model that “thinks” before it answers. Maths got noticeably better.
  • December 2024 to April 2025, o3. Announced in December, released in April. Again a clear step up in reasoning.
  • Until spring 2025, copy and paste. I copied code and errors back and forth between my scripts and ChatGPT.
  • Spring 2025, Claude Code. I tested Claude Code, an agent that works directly in your files and terminal, with my own money.
  • Summer 2025, at work. From then on I could use Claude Code in my job.
  • November 2025, agents. With the models released at the end of November, coding agents such as Claude Code and Codex could work through a task on their own for a long time. I switched from writing code myself to directing agents.
  • January 2026, Karpathy. Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, wrote on X that in November 2025 he still wrote about 80 % of his code himself, with language models as not much more than autocomplete. A few weeks later about 80 % was written by agents. I experienced exactly the same. I had tried it often before, it failed and I got frustrated, but November 2025 was also the point where it started working for me. Reading his post I had to smile. One of the best in the field had the same experience as me, which felt like a validation, and I did not feel left behind.

All along I had the feeling that almost nobody outside the field sees the change that is coming for the whole society and economy. I heard very intelligent people argue against the capabilities of those AI systems. Also some people simply used them less at work than I did and said the models had not really improved over the last year, but I think they never really tried them on long tasks, which is exactly where the progress really is visible.

It is true that a year ago the agents were not that good. But even then they could get a lot of things done, because not every task in the economy needs much intelligence or knowledge. Much of it simply needs to get done, and we can now automate far more of it than we could ten years ago. For me this has been clear for years, and still the big wave is not seen. Even if the models don’t improve at all from now on, incorporating them in workflows of the economy would be very disruptive. And they did improve. What can be done now is really insane, and I cannot even imagine all of it.

September 28th 2026

I have tried to be objective about AI. I do not like agents just because they are AI; I just want to get things done. That is also why they became useful for me so early: I gave it simple tasks and kept the complex ones for myself.

In the days around September 28th this changed. It started with the release of Claude Opus 5.5. I worked with Claude Code as usual, but I never had to correct it, and instead of one occasional spark it was sparks all the time. The share of its results I could accept without changes rose dramatically. It did things better than I had expected, and better than I would have done them myself, even with much more time. And all that at a low cost. Simply insane.

Naturally, this moment is subjective, in two ways. First, the moment an AI surpasses your own ability comes earlier for people who know less about their field and later for people who know a lot about it. Second, my definition of AGI is subjective: for me it is a system that does the broad range of my work tasks better than I do, and is general across them. So this is my AGI moment, not a statement about the world.

What changes now

From now on it seems that the only thing that matters is to have agents running for you all day long, and every second without an agent running is a wasted one. That is stressful, and it stresses me out as well. However, I will try to embrace it. I am excited about the upsides of AI and want to see what the future holds.

Two analogies help me think about it. The first is the calculator. If you have to multiply two big numbers and a calculator lies right next to you, it would be insane to do it by hand. For me, not using an agent feels exactly like that. And just as the calculator only costs the electricity it uses, an agent only costs the tokens it uses (the pieces of text a model reads and writes, which is what you pay for).

The second is the car. With a car you can travel much further than on foot. If you have a car and want to explore, you would not just leave it standing and walk. I see AI the same way: it lets me explore ideas much further than I could without it, and I think not using it makes just as little sense.

Maybe this even reaches the big open questions, e.g. the question about P vs. NP. It asks whether every problem whose solution can be checked quickly can also be solved quickly. A good and trustworthy friend of mine, who is a mathematician, once told me that he always thought he would never see a resolution of this question in his lifetime, no way. Now his opinion changed. In his view, the probability of seeing a resolution thanks to AI systems rose a lot.

An open future

The future is very open, and two things about it stay on my mind.

The first comes from the book What Is Intelligence? by Blaise Agüera y Arcas. He describes intelligence as something that helps replicators replicate, and I really like this way of thought. A replicator is anything that makes copies of itself, from a gene to a living cell. Today’s AI agents are not replicators. But I think they already expand our intelligence, the intelligence that helps us humans and our genes to replicate, and maybe one day they become replicators themselves. I find this a very interesting topic, which is worth exploring much more.

The second point I really want to stress. I think the AI revolution is a revolution of its own, like the Industrial Revolution or the computer and internet revolution. But there is one big difference: it happens within five to ten years. For the first time in history, a revolution is much shorter than a human working life. People at school or university today really do not know whether what they learn will become obsolete. This is really different from earlier revolutions, and I think we are facing much more uncertainty now.

To deal with uncertainty I like the books of Nassim Taleb. He says we should be antifragile: not just robust, so that shocks do not break us, but positioned so that we gain from randomness and disorder. I think we can position ourselves in our private lives to be antifragile to the randomness and uncertainty of the future.

Everybody becomes a manager

If everybody has AI agents, I think everybody becomes a kind of boss of a big company, a manager. So management skills become much more relevant, because we can distribute our workload much more than before.

This also changes an old heuristic. Sam Altman wrote that focus is a force multiplier on work, and it was always wise to focus on one thing to get it done. I think this heuristic changes now. You have to parallelize and work on many things, because you can get all of them done.

Sources