Ashwinee Panda (@pandaashwinee) 's Twitter Profile
Ashwinee Panda

@pandaashwinee

Postdoc of @tomgoldsteincs, PhD @princeton, @Cal alum, currently working on LLMs

ID: 1224111705950547971

linkhttps://kiddyboots216.github.io/ calendar_today02-02-2020 23:26:16

2,2K Tweet

2,2K Followers

763 Following

Ashwinee Panda (@pandaashwinee) 's Twitter Profile Photo

i love it when the gemini AI overview is completely wrong on a boolean question where the google preview of the first result answers the question correctly

i love it when the gemini AI overview is completely wrong on a boolean question where the google preview of the first result answers the question correctly
Ashwinee Panda (@pandaashwinee) 's Twitter Profile Photo

important read! folks who have SFTd models like llama2/llama3 on gsm8k will recognize that most of the gain you observe is in getting the model to produce the answer after the "####", and it seems like similar things are happening in RL-land.

Ashwinee Panda (@pandaashwinee) 's Twitter Profile Photo

eric infers from "i'm a couple hundred miles from japan tonight" and "we'd be in the same time zone" that mendes is in an area within a few hundred miles of japan that is in a different time zone; that's not many places, so eric can find him. my old response, & eric's counter:

eric infers from "i'm a couple hundred miles from japan tonight" and "we'd be in the same time zone" that mendes is in an area within a few hundred miles of japan that is in a different time zone; that's not many places, so eric can find him. my old response, & eric's counter:
Ashwinee Panda (@pandaashwinee) 's Twitter Profile Photo

regardless of how you feel about "the apple paper", it should be noted that the first author was an intern, and she has clearly thought about the responses that folks have. fwiw i'm solely RT'ing for visibility here, not because i agree with the paper or anything.

Ashwinee Panda (@pandaashwinee) 's Twitter Profile Photo

i learned this from Prateek Mittal during my phd. but tbh, i prefer to allow myself to get drawn into the weeds when doing the technical work, and rely on collaborators to ask the tough questions and keep me on track. what are the pros / cons of both?

Ashwinee Panda (@pandaashwinee) 's Twitter Profile Photo

> progress is based on real-world experiments rather than raw intelligence the tech that people are cooking up now is based on the insights from deploying models. if GPT-5 can't actually deploy its creations, how is it going to figure out what is needed for GPT-6? evals?

Ashwinee Panda (@pandaashwinee) 's Twitter Profile Photo

i really like this blog because reading it feels like having a conversation with albert; specifically, statements like "I’m driven by aesthetics much more than the average person, I’d guess" i'm excited to see this new architecture that i've been hearing so much about!