General
When Acceleration Accelerates
The Presentation
On November 18 I presented to Axelera an update on exponential technologies, covering the three months from mid-August to mid-November 2026. The presentation was prepared by Astra 6.2, which OpenAI had released the day before, from my input. The result is respectable, except that it is too accommodating: it equivocates and tries to keep a foot in every shoe it can find, and I don’t know how many feet it has.
Astra carefully separates what the sources report, what is interpretation and what is subjective scenario. I, on the other hand, wanted to say some things provocatively, almost conspiratorially, because I believe the moment has come when it is better to err on the side of too much paranoia than of too much common sense. That is why the text is in two parts. This one tells what has changed, keeping the facts as the sources report them distinct from my reading. The second gathers the more unsettling readings, in order of increasing alarm.
Space and Energy
Space exploration is increasingly intertwined with the development of AI. Starship Flight 15 launched on November 13, and on November 14 SpaceX completed the catch of the upper stage after an orbital flight, something nobody in the world had ever done. The booster, this time, aborted its own catch. The next test is rapid and cheap reuse: how much work the refurbishment requires, how much time passes before the next flight, how much it costs.
My reading: with this, SpaceX is well on its way to demonstrating rapid and full reuse of the Starship system. At that point the cost of a launch becomes the cost of the propellant, today about one million dollars, and SpaceX expects to bring it down to half a million. The investments in research and infrastructure remain, of course, but in bare terms putting 200 tonnes into orbit will cost between one million and half a million dollars. Two hundred thousand kilos for a million dollars makes 5 dollars per kilo. It is an epochal change, and I believe that November 14, 2026 will be remembered as the founding date of our spacefaring civilizations.
Meanwhile the merger negotiations between SpaceX and Tesla are fairly advanced, and could reach a vote of the two companies’ shareholders as early as this year. The merger benefits both. I am surprised that they have apparently overcome the cross-conflict between SpaceX’s military contracts and Tesla’s Chinese factory. I don’t know how they managed it or what it will mean, but for now it does not seem to stand in the way of the merger. The synergies will be enormous: the production of the Optimus humanoid robots and many other initiatives can have a catalytic effect.
In the United States, stationary energy storage has exceeded 200 GWh of cumulative capacity, according to SEIA. In the first nine months of 2026, 48.3 GWh were installed, 20% more than in the same period of 2025. In the third quarter installations were 14.3 GWh, in line with the previous quarter.
Biology and Robotic Labs
AI is entering biology directly, and it is doing so in a way that guarantees the scarcity of GPUs and memory for us poor mortals. Both OpenAI and Anthropic have formed alliances or launched their own initiatives to apply AI to biology. Anthropic in particular has stated that it will tackle rare diseases, the ones pharmaceutical companies ignore because under traditional models they do not generate sufficient revenue.
The most important step, in my view, is the transfer of AI’s discoveries into the physical world, because that is where the timelines of experimentation lie. Dark labs, fully robotized, are a reality today. Robots do the work that used to require armies of PhD students and postdocs: building them is expensive, running them costs a fraction, and the results arrive much faster. The big AI labs are moving in this direction, and the companies founded to bring these technologies into health are raising capital in the billions of dollars. Through experimentation in robotic labs, AI reaches our health.
On November 10 Beam reported positive interim results from the BEACON study, which uses base editing, a technique that modifies a single base of DNA, against sickle cell disease. With data as of October 20, 49 patients have been treated, and 31 of them have completed the follow-up required for the primary endpoint: at least twelve consecutive months without severe vaso-occlusive crises, in the interim analysis specified by the protocol. The therapy is still experimental and awaits approval.
World Models and Autonomous Driving
World models, the models that generate interactive simulated worlds, are bringing AI into physical environments. Genie 4 is in preview for a restricted group of testers. On October 21 Nvidia added new tools for physical AI to Cosmos. Waymo expanded its service to new cities and, since October 21, to Las Vegas airport for customers who are already registered.
I have started using world models and I am having a lot of fun. A few months ago I published a video about my agents in which I used the metaphor of a city to represent their relationships and the rivers of data flowing from one to another. Now I have used the World Labs Atlas world model to give each agent a room. One of the rooms is the study where R. David Orban, my digital twin, sits.
I have always argued that representing the physical world digitally is beautiful and useful, but that representing something alternative is much more interesting. The example I often give is learning to juggle. If you can set lunar gravity, the balls fall slowly: without understanding much, you move your hand, catch and throw. Then you raise the gravity little by little, and when you reach Earth’s gravity you take off the 3D headset and you know how to juggle, having learned more easily and more quickly. We will see whether the applications of world models will serve mainly the physical world, or whether those for alternative worlds will prevail.
In Texas a service of fully autonomous robotaxis, without steering wheel, mirrors or pedals, has started: Tesla’s Cybercab. It is a new model of transport, and it will change cities as the spread of the automobile changed them a century ago. In Europe, for now, its autonomous driving is authorized only in supervised form, with no robotaxis on the roads. Yet Tesla feels encouraged enough to announce an investment of 3 billion euros in its German factory, where it plans to produce up to one million Cybercabs a year starting in 2029. A robotaxi is used much more than a private car, so a million robotaxis equal at least five million traditional cars.
Compute and Quantum
Compute for AI is spreading beyond a single supplier. On September 30 Nvidia’s Vera Rubin systems arrived at CoreWeave, on October 12 Microsoft put the Maia 200 into production for Copilot, on November 9 Google presented new TPU configurations and on November 12 AMD announced developments of the MI400. The limits on useful compute are now energy, cooling, memory and interconnects.
Among the processors, I was struck by Google’s new TPUs, now in their eighth generation, which support so-called confidential computing. You can run an encrypted AI model, in which the model weights and the input and output data are accessible only to you, as the holder of the private key that manages the process. The mathematical basis is complicated, and I expected it would take longer to get there. The TPUs of this generation have just been announced and, as far as I know, are not yet in production. For many applications, for regulatory reasons or because strong confidentiality is required, until now the only alternative was on-premise AI, running on one’s own servers. This also opens the way to AI in the cloud, with the necessary confidentiality guarantees.
On October 15 Google reported, with Willow, the cultivation of magic states with a logical fidelity of 99.99% for the T state and an error about 40 times lower. Magic states are a resource required for fault-tolerant quantum computation. It is a concrete step toward the fault-tolerant quantum computer.
Verifiable Mathematics and Physics
In the same three months AI produced new knowledge that can be verified. On October 6 OpenAI reported results of mathematical research, many of which have a formalization in Lean, a language in which a program checks every step of a proof. On October 23 and November 6 PSI extended the method to quantum theory, with a proof of the Aaronson-Ambainis conjecture checked in Lean. The conjecture concerns the speedup a quantum computer can achieve on certain queries.
The advantage of mathematics and theoretical physics is that the results can be verified objectively: Lean checks every step, and the judgment that follows is not in doubt.
The reactions of some mathematicians, indignant that AI dares to enter their territory, struck me as ridiculous. I have seen this scene before. In 2007 I was present at the launch of D-Wave, the first commercial quantum computer. The company was accused of fraud, because it was stepping on the toes of an academic establishment that wanted progress under its own control.
For quantum computing the application is immediate. These results improve algorithms and help us understand in which directions quantum computers will be useful and superior, and therefore to abandon the areas where the presumed quantum superiority could be refuted by a classical algorithm.
Agents and Work: What the Data Measure
METR, one of the most closely followed research institutes on these topics, measures the length of the tasks that AI agents complete successfully at least half the time. This length has gone from a few seconds to a few minutes, then to a few hours. The domains evaluated include software, machine learning and cybersecurity.
On November 5 METR published an assessment based on an internal Anthropic capability benchmark. The recent pace is 1.9 times the pace before October 2025, and according to METR assistance from AI contributes to the change.
The Remote Labor Index measures something else: the share of selected remote work projects that agents automate entirely. As of November 17 the leaderboard shows 41.25%, with an estimated doubling every 3.3 months over the period considered.
On November 17, over dinner, a friend told me he had cloned one of the most complex and valuable proprietary software products in the world, with a swarm of agents that worked for eleven weeks. He then verified that the result is functionally equivalent to the product its vendor sells for millions of dollars. I will not say which software it is, but the story gives an idea of what a horizon of weeks means.
In my view, whether we measure task length or doubling time, we are measuring the same thing, and we have left the era of constant acceleration: acceleration itself is increasing. If you work at a computer, soon the things you do will be done by AI agents in a way equivalent to how you do them, and I hope you do not become replaceable. There will be barriers, regulations and choices of preference, and we will see how they are handled, but every choice of this kind will be paid for by giving up the economic superiority of the AI agent. There will be nations, companies and sectors that choose to be less efficient. And there will be new-generation competitors, in every activity with a strong digital component, who will answer: we do not want to be less efficient. They will have a knife between their teeth and will use everything AI can offer to the full.
The Jolting Technologies Hypothesis
For me these data are the most recent demonstration of what I call the Jolting Technologies Hypothesis. In physics the jolt is the third derivative of position, that is, the change in acceleration. The hypothesis asks whether capability doubling times are getting shorter, because ever shorter intervals would indicate an acceleration that goes beyond a stable exponential.
An ordinary exponential already has a positive third derivative, so the sign to look for is a sustained shortening of doubling intervals, on a measure of capability that stays consistent over time.
Astra 6.2 and AI Self-Improvement
The theme that connects these data is RSI, recursive self-improvement: the improvement that artificial intelligence systems bring about in themselves. The AI 2027 scenario placed it around 2028. Whether it will happen, and whether we will want to postpone or slow it, are the open questions. Meanwhile many results point toward ever more advanced RSI, and the labs are reaching ever higher alert levels.
OpenAI rated Astra 6.2 at the High level of its own scale for AI self-improvement capabilities. It is the first externally released model to reach it. OpenAI describes the impact with a comparison, in its assessment: compared with a 2024 baseline, it is as if each of its researchers were paired with an additional, highly capable research engineer.
Alignment and Takeoff Scenarios
On alignment, research is trying new paths. The November 2 study on weight-space programming directly modifies the weights of small networks such as Qwen3.5-0.8B, and a November 16 reply already disputes its mechanism. We are far from a method that guarantees the alignment of a frontier model, and even further from one that guarantees it across successive generations of models.
Then there is the question everyone asks: when does it take off? By takeoff I mean the transition of AI-driven research and development from a frontier human level to a decidedly superhuman capability, within weeks or months. I discussed it with Dot, and these are its subjective estimates of the probability that takeoff is already under way: 3% by the end of the first quarter of 2027, 6% by the end of the second, 10% by the end of the third and 15% by the end of the year.
The mechanism is simple to describe: each generation of models accelerates the next, and humans limit themselves to approving the compute. According to Dot, the probability that development at the two leading labs is already AI-assisted exceeds 90%, and the probability that one of the two has already entered sustained and largely autonomous self-improvement lies between 5% and 15%. The signals to watch are new model generations evaluated externally, complete development cycles that get shorter, and reliable orchestration of agents.
And if it happens, how do governments react? If by mid-2027 credible evidence emerges of accelerating recursive self-improvement, my analysis puts the probability that at least one major power imposes binding limits by the end of the year at over 60%, and the probability of a verified international agreement covering most frontier training at 25%. Imposing restrictions is easier than coordinating, but without coordination the slowdown does not come.
How politics, the military and the agents themselves will react, and what makes sense to do in the meantime, is the subject of the second part.