Henry Bigelow

Computational Biologist, Machine Learning Engineer and Researcher @AsteraInstitute. Formerly at Amgen and Broad Institute

[Henry Bigelow @hrbigelow]

@suchenzang This is interesting! I read it more not as AI welfare but human welfare. If bots allow verbal abuse, this will breed mental illness in the human user eventually, even if bots are never mean nor conscious.

[Henry Bigelow @hrbigelow]

@suchenzang @Noahpinion A (Caucasian) high school friend learned Mandarin well enough that he got offered a job cold-calling a company in China. They thought he was a local over the phone. He's lived there ever since.

[Henry Bigelow @hrbigelow]

@ChieroJoseph @nayaraelora good point 🧐

[Henry Bigelow @hrbigelow]

@fchollet pytorch is "batteries included", so the comparison has to be with jax + stateful layer. jax+equinox is my favorite. and it does seem like the better architectural decision to have the compiler act on pure functions and deal with statefulness outside of that.

[Henry Bigelow @hrbigelow]

@charles_irl 🤣

[Henry Bigelow @hrbigelow]

@fleetwood___ Very cool plots. This made me realize more that ML models have no idea whether they are trained or have just been born from random init.

[Henry Bigelow @hrbigelow]

I read it more that Dwarkesh intentionally teed up a widespread misconception for Sutton to attack. It's this lingering idea that is dead wrong and prevents us from grokking what RL really is. So it was nice to see it killed violently! Dwarkesh took the fall for all of us and I think he did so knowingly.

[Henry Bigelow @hrbigelow]

@suchenzang 😀 Tim & Eric!

[Henry Bigelow @hrbigelow]

@charles_irl @modal 😁

[Henry Bigelow @hrbigelow]

@fchollet I agree with all of this. But how could one train an ML model to do this, since we don't have access to the "thought traces" ultimately responsible for creating the available text.

[Henry Bigelow @hrbigelow]

@algekalipso the logic of it checks out ✅

[Henry Bigelow @hrbigelow]

@suchenzang very nice together with cool clean simple Equinox

[Henry Bigelow @hrbigelow]

@DimitrisPapail @ChrSzegedy yes we can! I believe he's saying that AI will ingest a human written, human comprehensible proof, translate it to Lean, and thus be able to verify the correctness of the human proof. And the human doesn't need to look at the Lean translation.

[Henry Bigelow @hrbigelow]

One unexpected and rewarding side effect of studying ML is that it forces one to scrutinize how human intelligence works. This is one of these moments.

[Oleg Mürk @oleg_murk]

OpenAI's reasoning system just scored at the gold-medal level at this year's IOI online competition — ranking #6 when measured against human competitors and #1 among all AI submissions. With @SherylHsu02 @alexwei_
@bminaiev
@ahelkky

My personal notes:
https://t.co/yZQWQIPbXz

[Henry Bigelow @hrbigelow]

@tunguz Dinosaur with a paunch lol

[Henry Bigelow @hrbigelow]

@dzhulgakov What was the bug?

[Henry Bigelow @hrbigelow]

Good question. I think of "world model" as anything that allows an agent to predict how its environment will respond to its actions, which could include predicting observations and rewards. So it's a functional definition. Of course to achieve this, a world model has to include a dynamics model. But you could have a dynamics model that doesn't involve any agent, just some dynamical system. So to me "dynamics model" is less specific.

[TheVariational @TheVariational]

Vector-quantization is taking over! @BytedanceTalk @keyutian @pess_r @robrombach @OriolVinyalsML @koraykv The details of VQ methods are highlighted, including the VAR @NeurIPSConf paper of the year. check out the following PDF https://t.co/FE2WH4Kmqc

[Henry Bigelow @hrbigelow]

@suzannegildert I do not know about consciousness but as for free will, I don't see how it is mechanistically possible. Either the brain process is deterministic or it is not. If so, will is not free. If not, then choice arises from randomness outside of will's control.