Davis Yoshida

[Davis Yoshida @davis_yoshida]

@yoavartzi @dlwh You can always use things as functions (params, inputs, state) -> (inputs, state) so it's generally easy to interact across framework boundaries.

I used a mix of both haiku and flax models for a couple project and I didn't run into any problems.

[Davis Yoshida @davis_yoshida]

@suchenzang I tried asking it to tell me what happened to Trump in July 2024 without searching, and it did list the assassination attempt with really low latency

[Davis Yoshida @davis_yoshida]

Moved to SF for work! If a mutual (you perhaps) wants to get coffee they should DM me.

[Davis Yoshida @davis_yoshida]

@pli_cachete @himbodhisattva For soft verifying the point is the same. Usually the thing you want to do is produce good things, not just identify them.

[Davis Yoshida @davis_yoshida]

@pli_cachete @himbodhisattva Because (e.g.) the Lean checker doesn't make new proofs for you, and you want cool new proofs.

[Davis Yoshida @davis_yoshida]

@pli_cachete @himbodhisattva If you use the filter at inference time, maybe you generate 1000 candidates, and one passes (or none). If you train the model to pass the filter, then you can just run it once at inference time.

[Davis Yoshida @davis_yoshida]

@srush_nlp We had an ACL paper arguing that searching for high-likelihood outputs _should_ give you garbage, under mild assumptions: https://t.co/In0GcGWitE

IMO this is a big reason why sampling beat out search over the last 5-6 years.

[Davis Yoshida @davis_yoshida]

@boydgraber I swung by the poster, sad to hear you couldn't make it!

[Davis Yoshida @davis_yoshida]

@srush_nlp 2. The response to "but this task looks hard even for humans" was "okay but people say this is AGI so it should be better than humans".

I didn't see anything which would reject LLMs being (e.g.) 25% as good at reasoning as humans.

[Davis Yoshida @davis_yoshida]

@srush_nlp Not to agree too closely with the aggressive question/comment which was made, but I felt like it was arguing against a bit of a strawman. I consider it an open question still to what extent LLMs are reasoning, since:
1. "Cannot do X reasoning task" != "Cannot reason in general"

[Davis Yoshida @davis_yoshida]

@typedfemale I had a bug where I had two components, and my model only worked if both or neither were torch.compile'd. If just one was I got NaNs.

[Davis Yoshida @davis_yoshida]

@srush_nlp (As for why mypy complains, no idea)

[Davis Yoshida @davis_yoshida]

@srush_nlp From the python docs:

> Using a generic class without specifying type parameters assumes Any for each position

[Davis Yoshida @davis_yoshida]

Suffix scaling will hit a wall: https://t.co/BvEOZj7T3r

[Aran Komatsuzaki @arankomatsuzaki]

GPT series still hasn't shown any signs of saturation 😲

[Davis Yoshida @davis_yoshida]

@cgarciae88 There's no _with_path ones in there last time I checked though

[Davis Yoshida @davis_yoshida]

@finbarrtimbers DM'd you a preprint

[Davis Yoshida @davis_yoshida]

@nadienadianadie @MasterTimBlais Ah good point. Without forced captures, it must always be a stalemate right?

[Davis Yoshida @davis_yoshida]

Just writing some markdown in nvim and copilot attacks me out of nowhere