Is there an intelligence explosion?
If better AI could make AI research go faster, and that faster research made AI better still, very capable AI could arrive much sooner than most people, companies and governments are planning for. That is why researchers are arguing about whether it has started.
The takeaway
Anthropic's AI has been improving faster since late 2025. That could be one big step up, a faster but steady climb, or a climb that gets faster with every step. Only the last is an explosion, and the public numbers cannot yet tell the three apart.
Three shapes a speed-up can take
Each tick marks a doubling of the length of task AI can do. The number above it is how many months that doubling took.
What is being measured
The pages track how long a task AI can do. METR, an independent group that tests AI models, measures this as the length of task, in the time a skilled person would take, that a model completes half the time. In early 2025 the best model managed software tasks that take a person about an hour. By late 2025 it was about five hours. That is not how long AI can work alone, and the uncertainty is wide: roughly a factor of two either way, METR says, and wider for the newest models.
Start here
- The spending puzzle: what Anthropic's spending can and cannot tell us. No equations.
- Fiddling with ASI: the model, with dials. Find the line between a slowdown and an explosion yourself.
- The research machine (optional, for the curious): what holding the pace would demand of research.
Much of the work on these pages was done by Claude, an AI made by Anthropic, the company whose research is measured here. Read with that in mind. The code and the text are open.
Where things stand
Last updated 15 November 2026.
1. AI research at Anthropic has sped up. After October 2025, Anthropic's AI models improved about 1.9 times as fast as before, on METR's measure.
2. Nobody yet knows how much of that came from more money and staff, and how much from AI tools. Progress can speed up because more goes in, because more comes out of the same input, or both, and the two can feed each other: better AI can attract the money that buys more research. Anthropic spent far more, and each dollar bought less progress than before. The question is whether progress beat what the extra spending alone would predict. The model-based study discussed here, by Garicano and Imas, finds that it did, in part: more inputs explain some of the speedup, and the rest looks like a one-time boost, though they say that estimate is uncertain. A one-time boost is not an explosion, which needs feedback strong enough to keep raising the growth rate as AI improves. Telling these apart is the hard part, and it is what these pages work through.
3. In AI research, the easy ideas seem to be getting used up. On Garicano and Imas's model, each step forward costs more than the last, as studies have found in many other fields. That rests on their assumptions, and how fast it is happening is uncertain.
4. The public data show no clear sign yet of AI feeding its own research in a way that speeds up as AI improves. That loop is what an explosion would need. AI is already helping with research, and METR, using evidence it could not publish, judges that AI accounts for much of the faster pace. The public figures can neither confirm nor rule that out.
5. It cannot be ruled out. Daniel Kokotajlo, an author of the AI 2027 scenario, wrote that if METR is right, an explosion now looks likely in 2027. In Garicano and Imas's broadest test of measurement error, about one in eleven re-runs of their model meet its condition for an explosion. That shows how much the answer depends on noisy scores.
The one number that decides it
In the simplified model used here, with spending held fixed, one number sets whether growth speeds up or slows down as AI improves. It sets how strongly AI's gains feed back into research against how quickly new ideas get harder to find. If the feedback is weaker, each step costs more than the last and growth fades unless spending keeps rising. If it is stronger, each step makes the next easier and growth runs away. The feedback here comes from AI doing more research. The spending dial is set from outside the model; it does not model AI earning the money for more research. The selected estimates sit below the line. Their positions depend on the model specification and how progress is measured. They do not pin down where today’s AI sits.
The questions inside the question
Is AI research speeding up?
Yes, at Anthropic, by about 1.9 times on one measure, with a wide range. → METR.
Try it: today's pace, and what happens to it if nothing else changes →
How much of the speedup is money, and how much is AI?
Unsettled. The answer depends on what you assume would have happened anyway. → Krier and Wiebe; Garicano and Imas; our spending puzzle.
Are ideas getting harder to find in AI?
Probably, on Garicano and Imas's estimates. How fast, and whether the rate is changing, is uncertain. → Garicano and Imas.
Try it: the φ dial, how much past progress helps the next step →
Is AI starting to do its own research?
AI already helps with parts of research. The public data show no clear sign yet of the self-feeding kind. David Chalmers asks a related question: the gains so far may be in the easier parts of research, such as coding and running experiments, and could grow fast there without reaching harder parts such as research judgement. → Garicano and Imas; Chalmers.
What would an explosion need, and what would we see first?
AI that gets better at the hard parts of research, not only the easy ones, which is Chalmers's question. And feedback strong enough that returns to research effort stop falling and start to rise, which is the condition in Garicano and Imas's model. In that model the feedback runs through research effort; AI earning money for more research is a second route the model leaves out. → Chalmers; Garicano and Imas.
Try it: a setting above the line, where growth goes vertical →
What would reduce the uncertainty?
Model-by-model scores with their uncertainty, the research bill broken into its parts, how long spending takes to show up, and comparisons fixed in advance. → our spending puzzle, section 7.
The record
These entries gather claims and caveats from work on the intelligence explosion.
- Tom Davidson: the speed-up may not be mostly money
- Dwarkesh Patel: the case that AI research will soon automate itself
- Luis Garicano and Alex Imas: what more data shows
- David Chalmers: the ‘factor F’ question
- AI Futures Project: a shorter timeline after METR
- METR: Anthropic’s AI research has sped up
- Davidson and colleagues: two routes by which AI research can feed back
- Whitfill and Wu: can more AI researchers work with the same computing?