1. I think “this is a political move to drive up hype” is definitely a factor. The fact that they’re concretely anchoring to 2026-2027 does downweigh this explanation, however: that’s not a very good political move, they should be keeping it more vague.[1] So...
2. … I think they do, themselves, mostly believe it. Which is to say, they’re buying their own hype and propaganda. That is a standard dynamic, both regarding hype (if you’re working on X, surrounded by people working on it and optimistic about it, of course the optimisms end up reinforcing each other) and propaganda (your people aren’t immune to the propaganda you’re emitting, and indeed, believing your own propaganda makes it much more authentic and convincing).
3. I’m very much not on the “they have secret internal techniques light-years ahead of the public SotA and too dangerous for the public eye”/”what did Ilya see?!” train. I think what they have are promising research directions, hopeful initial results, the vision to see that research through, and the talent they believe to be sufficient for it. This is what fuels their optimism/self-hype. Which is fine, I’m hyped for my own research too. But, of note:
Anthropic’s reasoning models were hyped up as scary, but what we got is a decidingly mediocre (as far as reasoning goes) Sonnet 3.7. SotA-good at programming? Yes. Scary? No. Well, perhaps they have even scarier models that they’re still not releasing? I mean, sure, maybe. But that’s a fairly extraordinary claim, and all we have for it is vague hype and scary rumors.
Satya Nadella was an insider, and he recently bailed on OpenAI and implied he’s skeptical of near-term economically transformative LLM effects. Sure, maybe he specifically is a pathological AGI disbeliever[2]. But it does put a sharp limit on how convincing their internal evidence can be.
I don’t buy it in general. AGI labs are competing for funding and attention, it’s a rat race, I don’t think they have the slack to sandbag, nor the ability to competently coordinate on sandbagging. Especially with defectors like DeepSeek breathing down their neck.
Though I do note that in the actual text submitted to the USG, they say “could emerge as soon as late 2026 or 2027”, not “will emerge in late 2026 or early 2027″, as they say in the blog post.
And in some other statements, Dario states “possibly by 2026 or 2027 (and almost certainly no later than 2030)”; which is to say, P(AGI by 2027) > 0, P(AGI by 2030) = ~0.99. Much weaker, and also, I note that 2030 is past the current US administration’s expiration date.
Anthropic’s … mediocre Sonnet 3.7. Well, perhaps they have even scarier models that they’re still not releasing? I mean, sure, maybe. But that’s a fairly extraordinary claim
Base model for Sonnet-3.7 was pretrained in very early 2024, and there was a recent announcement that a bigger model is coming soon, which is, obviously. So the best reasoning model they have internally is better than Sonnet 3.7, even though we don’t know if it’s significantly better. They might’ve had it since late 2024 even, but without Blackwell they can’t deploy, and also they are Anthropic, so plausibly capable of not deploying out of an abundance of caution.
The rumors about quality of Anthropic’s reasoning models didn’t specify which model they are talking about. So observation of Sonnet 3.7′s reasoning is not counter-evidence to the claim that verifiable task RL results scale well with pretraining, and only slight evidence that it doesn’t scale well with pure RL given an unchanged base model.
1. I think “this is a political move to drive up hype” is definitely a factor. The fact that they’re concretely anchoring to 2026-2027 does downweigh this explanation, however: that’s not a very good political move, they should be keeping it more vague.[1] So...
2. … I think they do, themselves, mostly believe it. Which is to say, they’re buying their own hype and propaganda. That is a standard dynamic, both regarding hype (if you’re working on X, surrounded by people working on it and optimistic about it, of course the optimisms end up reinforcing each other) and propaganda (your people aren’t immune to the propaganda you’re emitting, and indeed, believing your own propaganda makes it much more authentic and convincing).
3. I’m very much not on the “they have secret internal techniques light-years ahead of the public SotA and too dangerous for the public eye”/”what did Ilya see?!” train. I think what they have are promising research directions, hopeful initial results, the vision to see that research through, and the talent they believe to be sufficient for it. This is what fuels their optimism/self-hype. Which is fine, I’m hyped for my own research too. But, of note:
Anthropic’s reasoning models were hyped up as scary, but what we got is a decidingly mediocre (as far as reasoning goes) Sonnet 3.7. SotA-good at programming? Yes. Scary? No. Well, perhaps they have even scarier models that they’re still not releasing? I mean, sure, maybe. But that’s a fairly extraordinary claim, and all we have for it is vague hype and scary rumors.
Satya Nadella was an insider, and he recently bailed on OpenAI and implied he’s skeptical of near-term economically transformative LLM effects. Sure, maybe he specifically is a pathological AGI disbeliever[2]. But it does put a sharp limit on how convincing their internal evidence can be.
I don’t buy it in general. AGI labs are competing for funding and attention, it’s a rat race, I don’t think they have the slack to sandbag, nor the ability to competently coordinate on sandbagging. Especially with defectors like DeepSeek breathing down their neck.
Though I do note that in the actual text submitted to the USG, they say “could emerge as soon as late 2026 or 2027”, not “will emerge in late 2026 or early 2027″, as they say in the blog post.
And in some other statements, Dario states “possibly by 2026 or 2027 (and almost certainly no later than 2030)”; which is to say, P(AGI by 2027) > 0, P(AGI by 2030) = ~0.99. Much weaker, and also, I note that 2030 is past the current US administration’s expiration date.
Which seems likely, I (using DR/Grok 3) haven’t been able to find any evidence he believes in the Singularity at all.
Base model for Sonnet-3.7 was pretrained in very early 2024, and there was a recent announcement that a bigger model is coming soon, which is, obviously. So the best reasoning model they have internally is better than Sonnet 3.7, even though we don’t know if it’s significantly better. They might’ve had it since late 2024 even, but without Blackwell they can’t deploy, and also they are Anthropic, so plausibly capable of not deploying out of an abundance of caution.
The rumors about quality of Anthropic’s reasoning models didn’t specify which model they are talking about. So observation of Sonnet 3.7′s reasoning is not counter-evidence to the claim that verifiable task RL results scale well with pretraining, and only slight evidence that it doesn’t scale well with pure RL given an unchanged base model.
Fair.