Skip to content

Progress

We've seen some exciting developments in the past few weeks:

But technological progress is not the same as actual progress.

It's a big deal

Demis Hassabis, co-founder and CEO of Google DeepMind, says:

"AI is going to be 10 times bigger than the Industrial Revolution, and maybe 10 times faster."

— Demis Hassabis in "A Framework for Frontier AI and the Dawning of a New Age"

The reasoning here is that the impact of AI is going to really scale once it's coupled with robotics. And only the US and China have the industrial capacity to develop robotics (including massive quantities of humanoids) at scale. This is not hypothetical: we're already seeing very rapid progress in this area. Just watch the video of Gemini Robotics 2 demonstrating whole body robot intelligence. Then have a look at the sheer scale of the new Terafab chip factory announced by SpaceX and Tesla just days ago: a massive $16.8 billion plant that will be the biggest building in the world by a large margin. At the same time, Terence Tao, one of the world's leading mathematicians, is calling for a rethink of what it means to be a mathematician.

And then there is the particularly gloomy Europe 2031 scenario. In it, a group of AI researchers explain that Europe has not yet absorbed the pace and magnitude of the change that is coming. I agree with that analysis; of the three recent developments I described above, mainstream Dutch media only ran an article announcing that free ChatGPT users won't hit usage limits any more (source).

The AI researchers from Europe2031 go on to detail a plausible scenario where the EU fails to properly invest in training its own frontier-class AI models, or in data centers with significant compute. As a consequence, the EU falls so far behind that it ends up being sold off in pieces to the US and China.

Noema recently published an article titled Why Is Everyone In Tech So Sad? It's about knowledge workers starting to realize their work is pointless and wanting to pick up things like goat farming.

Scenarios

So all these developments can lead to a weird mix of gloom and excitement. When things change, there is anxiety. Luckily no one can predict the future, and the world is complex, with many developments that could change the course of history.

It could be that we realize open-source and open-weights models hit a sweet spot on the price/intelligence frontier, good enough to power the vast majority of LLM workloads. But you still need to power those models. The major AI labs are very aware that the model is not the moat; the real advantage America has is in the chips and the data centers used to run the models. Dario Amodei, Anthropic CEO, is very clear on that in his post Our position on open-weights models.

The mass layoffs aren't happening just yet. It could very well be that our jobs are safe. The Jevons paradox states "when technology makes the use of a resource more efficient, total consumption of that resource usually increases rather than goes down". Build more highways, and traffic goes up. With better AI the cost of software goes down, enabling entire new areas of custom software and automations that were previously not cost-effective. The net result is higher demand for software and for people who can build it. And AI adoption within corporates is still very low (see AI Adoption is a Myth).

If we get general intelligence too cheap to meter, this entirely rewrites the rules of the game. That would be an existential crisis. Like discovering aliens are real. Time Magazine did publish an article a couple of days ago stating America Is Finally Taking Extraterrestrials Seriously. Heck, maybe 2027 will be the year we see both superior artificial and extraterrestrial intelligence.

AI progress could stall. On the road to artificial general intelligence, some researchers believe we are stuck in "LLM Valley": we're gradient descending our way into what will prove to be a local optimum. Many believe the real breakthroughs are most likely to happen when neural networks are combined with reinforcement learning, like the emergent intelligence we saw with DeepMind's AlphaGo Zero. We're already starting to see that happen, with the Prime Intellect harness that beat ARC-AGI 3 and of course the Gemini robotics models from DeepMind. Meanwhile, investment is ramping up in companies researching the promising next step: World Models. And of course, LLM progress is not stalling; it's still advancing rapidly. So I believe it's unlikely AI progress will stall anytime soon.

All the while, technological progress in other areas could also affect AI development. A few examples around power generation: new technology for capturing energy from sea waves, progress in nuclear fusion and advances in nuclear power like the gravity nuclear reactor. And AI is accelerating science, at least the kind that requires extensive search. The paper LLMs can't jump (Zahavy, 2026) argues that AI cannot generate the novel explanatory hypotheses required for discovery, and is capable only of induction and deduction. Even if true, given the excitement around the recent advancements in math, perhaps induction and deduction will take us a long way.

What's next?

Nobody knows for sure what's next. But things are changing rapidly. And with change comes uncertainty. I think the best thing to do is to embrace innovation, to continue experimenting and building. To think big.

David Heinemeier Hansson puts it well: the age of agents has given us endless execution, and it's nirvana for people with endless ideas. "AGI is a nefarious concept to pin down, but I'm not sure how different whatever definition we eventually settle on will look from what I'm already experiencing on the daily". It's crazy what we can make computers do now.

I'm hopeful and excited: we're witnessing rapid technological improvements that have serious potential to become actual progress for humanity, in our generation.