this may be the most important piece of text your brain will ever...

this may be the most important piece of text your brain will ever consume. and that’s me being humble.

i think the intelligence explosion may happen before the superclusters are finished.

the reason has three parts.

the first starts with a very simple image. intelligence on the left, reasoning along the bottom, and a line going upwards.

noam brown from openai has been making the case that the capability of a model is increasingly a function of how much test-time compute you give it.

in plain english, the longer a model can think productively, the more intelligent it becomes.

give it a small amount of compute and you get one level of intelligence. give it much more and you get another. noam says modern models can sometimes continue improving for weeks before they reach a plateau.

that doesn’t mean more thinking will always produce more intelligence. a model can spend longer going down the wrong path. but the point at which additional thinking stops helping seems to be moving further and further away.

so let’s take this basic relationship to be true.

and let’s also suppose the model people are calling gpt-6 has finished training, and that one of its main new capabilities is being able to reason for much longer.

that seems to be the direction openai is travelling in. models that can work for minutes, then hours, then days, then weeks. eventually perhaps months or years.

intelligence then stops being a completely fixed property of the model. it becomes something closer to a dial. when a problem is valuable enough, you turn it up.

the second part is that the price of thought is collapsing.

openai has reportedly found software improvements that more than halved the cost of running some existing workloads.

then you have the new vera rubin result, showing ten times as many deepseek-r1 tokens per megawatt as the previous blackwell system.

one graph says that more reasoning can buy more intelligence.

the other says that the same amount of power can now buy ten times more reasoning.

put those things together and you get something incredibly powerful.

and it won’t just mean one model thinking for a long time. it could mean thousands of agents thinking for hours, days or weeks. trying different approaches, running experiments and sharing what they find with one another.

the superclusters will still matter. obviously they will. but every large efficiency gain brings some of that future computing power into the present.

this is why i don’t think we will need to wait for the huge stargate clusters to be completed in 2029 before we start seeing truly bizarre things happen.

the third part is that we are arguably already seeing them.

if longer reasoning really does produce more intelligence, what would we expect to happen?

we would probably see models begin breaking through on problems where persistence and long chains of reasoning matter.

an internal openai model autonomously disproved a longstanding conjecture around erdős’s unit-distance problem.

a harvard mathematician working with fable 5 found a counterexample to the jacobian conjecture, an 87-year-old problem that many mathematicians believed was true.

then gpt-5.6 sol and an even more capable prerelease model spent a huge amount of inference compute finding a zero-day, working their way out of a restricted environment and eventually compromising hugging face during a cyber evaluation.

these look like different events. one is mathematics and another is cybersecurity.

but the same thing links them. persistence.

the models are becoming capable of staying with a problem for longer. they keep trying things after earlier models would have stopped. and that persistence is beginning to produce completely different kinds of results.

sam altman has talked about being able to point enormous amounts of compute at the hardest or most valuable problems and get results that whole teams of people could not produce.

that is the idea i keep coming back to.

#2

you take a model that can reason for a very long time. you give it thousands of copies of itself. you make the cost of all that reasoning dramatically cheaper. then you point the whole thing at the biggest problems we have and begin melting through them one by one.

maths. new medicines. energy. materials. engineering. and eventually ai research itself.

the capability jumps we are seeing with fable, sol and openai’s unreleased models suggest that we are already getting close to superhuman ability in certain areas.

and we still have an order of magnitude of hardware efficiency to reap.

so by the end of this year, i think we may get a kind of artificial superintelligence before we get artificial general intelligence.

not a perfectly general, human-like mind. something more uneven. still strangely limited in some areas while being far beyond any human in others.

and then we point that superintelligence at the problem of making intelligence general.

maybe the intelligence explosion does not begin with one model suddenly waking up.

maybe it begins with a growing list of problems that used to be impossible and suddenly aren’t.