I operate by Crocker’s rules.
I try to not make people regret telling me things. So in particular:
- I expect to be safe to ask if your post would give AI labs dangerous ideas.
- If you worry I’ll produce such posts, I’ll try to keep your worry from making them more likely even if I disagree. Not thinking there will be easier if you don’t spell it out in the initial contact.
See Wizard Power.
How’s this for a crux: How much of the variance in whether person A disrupts person B is explained by each variable?
If one employee gets zero/one work done without/with AI then the geometric mean is infinity.
You could keep any export that nobody downloaded in the airgapped archives, against some future day when you find a better point on the tradeoff curve.
Do you mention Claude Opus 4.6′s 80.8% because your analysis has one free parameter and you set it to fit that 80.8%? How well does your analysis translate other model’s percentage scores to their time horizons?
Could Dario sell an option to turn 2 OpenAI shares and $1000 into 1 Anthropic share?
(Then if OpenAI wants to slow down to satisfy Anthropic’s condition, they can buy that option as insurance and deterrent against Anthropic ignoring their commitment and pulling ahead enough to make the option worth exercising.)
How might they benefit from faking this?
The Department of
War’s records show that it designated Anthropic as a supply chain risk because of its “hostile
manner through the press.”
(2) is problematic on model welfare grounds, can we find another benchmark please?
it looks like the Department of War is trying to punish Anthropic for trying to bring public scrutiny to a contract dispute.
huh, that’s the first I hear of it looking like that, has anyone argued this before?
Then I am very confused: How do you get from “text extrapolated from base models depends on the sampler” to “base model circuits are less composeable”?
I’m surprised you diagnose this an optimizer issue—we’re talking about SGD/Adam/..., right? Sure, I would believe that what text a base model produces is heavily influenced by training details, since a base model is trained on internet text, while RL trains on what text the model produces. But whether the one circuit produces data that the other circuit can read sounds like a question of whether a model is good at the job it was trained for. So, in-distribution, I still expect the circuits that pre-training produced to be composeable.
the circuits are selected by reinforcement learning, especially RLVR, to be composeable
I predict that you would also find such selection in the base models.
Why only whole-body?
I guess during signup you could require authors to say what existing research bureaucracy they are working for, and only if they click the I am an independent researcher. link are they introduced to staking.
What happens if AI labs offer to support you, expecting that you make it more prestigiuous to publish capability evals?
In order to not drown in slop submissions you could require each author to stake as much money as the reviewers would be happy to be paid to reject their submission as slop.
Why did this get downvoted??
I think using AI as an arbiter is fine if both parties agree. If one party doesn’t agree and therefore looks unreasonable, that’s a feature. Check whether this document is the office-political immune response to that prospect. Recall prediction markets:
If you are resigned to or intend that immune response, say something, I might have an idea that makes everyone better off.