What Three Months Look Like Now

Every three or four months Axelera, the non-profit association I founded fourteen years ago to spread the understanding of exponential technologies, holds a periodic update. Ours was on 23 September. I had asked four different models to prepare the slides, describing what they judged to be the most important developments of the previous three or four months in artificial intelligence, agents, energy and space. They produced beautiful presentations and I disliked every one of them. I had no appetite for correcting them, so I spoke without slides.

Jolting, Not Exponential

For years I have been publishing the paradigm of jolting technologies, where what grows is the rate of acceleration itself. We are used to Moore’s law, fifty years of doubling roughly every two years. In AI the intervals themselves are collapsing, from two years to one, from one year to six months, from six months to three. Follow the current rate of increase in the acceleration and you reach the projection of a significant frontier release every single day. I do not expect that projection to hold in practice. The measurements accumulating now already justify dropping the word exponential in favour of jolting, or super-exponential.

China Closes, and the Summit Will Not Slow It

The open-weight models arriving from China have narrowed the performance gap to roughly four and a half months, from around six earlier in the year, and parity around the end of the year is the reasonable expectation. The American and Chinese presidents met in Washington on 23 and 24 September to work out how to proceed in a competition both of them believe they are controlling. I do not expect the meeting to change the pace of development at all. When people say slow down, they are not describing a car approaching a bend and braking, nor a driver lifting a foot as the light turns amber while still pressing the accelerator. They are asking for the increase in the acceleration to be smaller. Whatever agreement is announced, I expect no practical effect.

The Model That Is Not a Language Model

A release from the days just before our meeting was not a large language model at all. Jev, from TypeSafe AI, is a classifier: fast, cheap, and structurally old, since these approaches have existed in machine learning for decades and are only now being fed modern training capacity. Pair one with a model from OpenAI or Anthropic and you get results at a trivial incremental cost, because the input tokens are priced very low and the classification decisions themselves are given away. We spent several years so fixated on LLMs that we forgot other architectures existed, with the partial exception of the diffusion models behind image generation, which many people assume work the same way. That period is ending, and a profusion of specialised models is arriving. Marco De Rossi, who founded the e-learning platform WeSchool and then spent time at MetaMask, is now building one of them at Levanto, called Sage.

Agents, and the Rooms They Work In

Agents began the year as a curiosity and have become the centre of the field. OpenClaw opened everyone’s eyes, and its lessons were absorbed everywhere. Claude now launches a dozen agents in parallel in the background for a single task and coordinates them without mentioning it. Grok Bot pushes in the other direction, towards specialisation, giving each agent long-term memory, an awareness of the kind of work it exists to do, and the ability to coordinate explicitly with the others. Meta’s Muse will probably reach a scale comparable to ChatGPT or beyond it, because it is decent enough and it has Meta’s distribution behind it. Google is inexplicably far behind, and Apple further still.

The step very few absorbed yet is the one after that. It is not a person with several agents. It is a space where people and agents work together. Slack arrived first, which was natural given that people already live inside it. The one that interests me more is Buzz, built by Block, open source, so you can run it on your own machines or in the cloud, and deeply programmable. I expected to need a bridge between Buzz and the agents I run elsewhere, and there was nothing to build: Buzz is a Nostr workspace, every participant holds their own key, and a Grok Bot already holds one. Add its key as a member and it reads and writes in the channel like anyone else. Claude and the rest joined the same way, alongside my team. A swarm of agents available not to one person but to a whole team is going to matter enormously at the level of the firm.

Recursive Self-Improvement, and Who Instructs Whom

Further out sits recursive self-improvement, which every frontier lab is now pursuing and every one of them hints it is close to. A platform designs and builds its own successor, knows which direction to go, can verify that the improvement is real, and instantiates it by replacing itself. If that is ever completed we reach something it will be natural to call a phase transition, because the systems become autonomous in their own trajectory. That is exactly why the unsolved problems of alignment and safety need to be solved on the way there rather than after arrival.

Anthropic has published how it classifies the working relationship between its researchers and its agents, which is to say who instructs whom. As of August 2026 Claude leads 26% of the company’s AI research and development, against under 1% in February, meaning it can take most of a task end to end from a high-level prompt while a human supervises. More than 90% of the work sits at collaboration or above. None of it is fully autonomous, yet. The share that the agent leads is rising.

Eighty-Eight Hours on a Century-Old Problem

OpenAI announced a solution to the Navier-Stokes problem in which, in finite time and with finite energy input, velocities inside a vortex go to infinity. Alex Wissner-Gross described it as stirring sugar into your coffee and creating a black hole, which is a fun way of saying the model does not correspond to physical reality. As mathematics it stands, checked by the Lean theorem prover, and it was produced by about ten thousand concurrent agents burning about 130 billion output tokens, arriving at the resolution some 88 hours after the first of them was launched. Two years ago the models were doing secondary school problems. A year ago they were doing university problems. Anyone extrapolating exponentially expected doctoral problems next, not the resolution of something no one on Earth had managed for a century. Mathematicians have reacted with something close to despair at seeing their territory taken, and the question of what a pure mathematician should now do with a career is a reasonable one. Whatever is written in protest will have no effect on the availability of these tools or their application.

There are credible rumours that both Anthropic and OpenAI hold results just as startling and have chosen not to release them, because there is no advantage in doing so once you have demonstrated you can clear the field. They are building verticals instead, where value can be captured: multi-billion-dollar agreements with consultancies announced in May, biotechnology at Anthropic, a model for law firms at OpenAI. The original OpenAI investment documents promised a maximum return of one hundred times capital, on the reasoning that without such a cap the company risked capturing the entire wealth available in the future light cone of the universe. That is their own wording, and they are proceeding with few visible reservations.

The Energy Story Is Missing Its Denominator

The noise about data centres and their rising share of energy consumption is an incomplete narrative of what is happening. Nobody setting out those consumption figures sets them against the performance that is improving. They report the power drawn and not what the power produces. When the training cost of a model is presented as alarming, nobody asks how much energy it takes to raise a child to twenty, or thirty if you are unlucky, before they become autonomous. We are proud of our brains, which we describe as the most complex object in the known universe, and we like to point out that twenty watts does everything we do. Do the arithmetic and AI has already beaten us. Boris Power of OpenAI put it at five IQ points per watt for humans against seven to forty for AI, while noting that joules per equally good completed task is the better measure. A data centre absorbing a great deal of energy is producing a density of intelligence out of all proportion to the input, because the models and their efficiency keep improving.

A Trajectory to Alpha Centauri

My favourite example from the last few months is that we are about to launch the first interstellar mission to Alpha Centauri. Voyager is leaving the solar system, but it was never aimed, and where it goes next was no part of its design. This probe, the Fermi Explorer, is meant to pass close to the target star. The trajectory is counter-intuitive, because it begins by flying towards the Sun rather than away from it, circling closer for twelve years before the final pass flings the probe outward. It would have required years of human effort to find, with millions of dollars of investment, and nobody tried. An AI physics system ran for about three days and found it, as the first announced result of a company called Physical Superintelligence, whose purpose is to push the frontiers of physics and materials science. The launch will cost no more than fifteen million dollars. The probe will take eighty thousand years to arrive, and the people involved would be surprised if it were not overtaken several times on the way, with a welcoming ceremony held by descendants who have already colonised the planets around Alpha Centauri.

Ten Thousand Agents

I wrote recently asking what we would do with ten thousand AI agents if we had them. OpenAI used that many on Navier-Stokes. My answer was that we can do whatever we want, and all we have to do is want it. If ten thousand agents for a given number of hours costs too much today, that cost will fall by a factor of millions, and what cost a million dollars will cost one. So the urgent question is what to do with our attention and our time, because the tools to do anything at all will be there.

The questions that followed took the evening somewhere the update had not gone, and I have written them up separately. They arrive tomorrow.