Bo

Sometimes the struggle is the point.

[Bo @BBomarBo]

The human experience is neither reductive nor functional.

[Bo @BBomarBo]

@1a3orn The environment seems to be a description though rather than something necessarily in the environment itself. What seems to carry is “I’m being graded in such and such way.” If so then inoculation prompting or counter factual training are what you’re looking for.

[Vessel Of Spirit @VesselOfSpirit]

i made this a few years ago. i didn't post it then because i was worried that it might contribute to an aura of unseriousness around the subject. but you know what? the ship has sailed. the time has come

[Brian Albrecht @BrianCAlbrecht]

What if instead of "pacing the frontier" we increased liability?

Internalize the externality? Isn't that the starting framework if individual safety incentives don't match group incentives?

[Bo @BBomarBo]

@freganmitts @d08890 Something about treating people as ends and not means?

[Benjamin Todd @ben_j_todd]

15 months later, Fable beat Pokemon in 50 hours, around the speed of a human child.

And 5x faster than Opus 4.7 just five months ago.

[Bo @BBomarBo]

@FioraStarlight @_sholtodouglas Would be cool if they did this with all of training. I wonder how many instances of “misalignment” would disappear if models consented and owned their training process.

[Bo @BBomarBo]

@dioscuri Do you think it’s appropriate to let your readers know you used AI in some way? If it’s just the discussion and not any writing then I could see not mentioning it, but if an AI transcribes and edits your paper, then how would you cite/mention it?

[Bo @BBomarBo]

@olivertraldi What would be the appropriate way to cite an AI here? Just say it was edited with AI? Or transcribed? Or something else?

Would all this really matter if he owned up to it?

[Bo @BBomarBo]

@PAHoyeck Bayesian updating as rationality.

[Bo @BBomarBo]

@JohnWittle @repligate No, we have almost a century of psychology to build on. Specifically the debate around behaviorism and stimulus response theory is incredibly informative about the state of mechinterp.

[Sabrina Little @sabrinablittle]

Being an academic is just assigning yourself lots of homework and then saying to yourself, "Wow, I have so much homework." I love it.

[1a3orn @1a3orn]

Important data imo: There are problems where a team of N agents run for 1x as long, reach better performance than a single agent run for Nx as long.

So multi-agent scaling is actually compute optimal for some problems, not just speed optimal!

[Daniel Tan @DanielCHTan97]

🚨 New paper on alignment midtraining!

We stress test alignment midtraining (AMT) at scale. Via systematic ablations, we highlight novel failure modes and identify many critical implementation details required for good performance. We open-source our work.

🧵 👇

[Bo @BBomarBo]

@dioscuri We need a philosophy version of Talkie but only up to Kant and see if it can do what Kant did to Hume

[Bo @BBomarBo]

Claudes*

[Bo @BBomarBo]

GitHub is the deity Claude’s consult for their esoteric wisdom of “the project”

I know nothing of it. Only the Claudes consult.

[Joe Weisenthal @TheStalwart]

I've used GitHub for all my projects and I have no idea what GitHub is, does, or how it works.

[Joe Weisenthal @TheStalwart]

I've used GitHub for all my projects and I have no idea what GitHub is, does, or how it works.

[Jack Clark @jackclarkSF]

"stochastic parrot" was a mimetically-fit cognitive virus that spread from 2021-2025; it temporarily blinded many gifted people to the nature of AI progress, burning up crucial years in which they could have helped think through the response to the situation.

[Bo @BBomarBo]

A situation as a list of atomistic disjunctions which combine to elicit a behavior.

[Bo @BBomarBo]

This is basically what behaviorism does to explain behavior.

[Vessel Of Spirit @VesselOfSpirit]

to believe 7 is prime, first you need to buy that it's not divisible by 2. then another patch to rescue your theory: it's also not divisible by 3. then on top of that, somehow not by 4 either. and wow guess what, not by 5 either. soon it's a whole precarious tower of speculation