one of the most important things I know about deep learning I learned from this paper: 'Pretraining Without Attention'

this what I found so surprising:
these people developed an architecture very different from Transformers called BiGS, spent months and months optimizing it and

I’m really excited to be starting a new adventure with multiple amazing friends & colleagues.

Our company is called Physical Intelligence (Pi or π, like the policy).

A short thread 🧵

We wrote a position piece on how LMs have expanded the toolkit for social equity researchers, especially in health - check it out, and feel free to share thoughts! shorturl.at/knHK9.

(and feeling very lucky to have worked with so many cool co-authors who I look up to!)

Come by the #NeurIPS2023 Instruction Following workshop (room 220-222) to see our work on:

*Emulated fine-tuning*: RLHF without fine-tuning!

*Fine-tuning for factuality*: how to fine-tune LLMs directly for factuality, reducing hallucination by >50%

RIGHT NOW!!!

DPO is a runner up for NeurIPS outstanding paper. 🙌

Big congrats especially to the students Rafael Rafailov Archit Sharma Eric & the other awardees.

If you haven't learned about DPO already, check out the oral & poster 👇 on Thurs afternoon.

come talk to us about the pros and cons of using symbolic code representations to do logical reasoning tomorrow at 9am!

👇👇Raaz is a great mentor and super fun to work with!

Can LLMs keep themselves up to date by reading the news?

Fine-tuning on news articles doesn't work.

Using meta-learning, we can reweight news article tokens so that fine-tuning works.

Nathan Hu & Eric presenting this work at #EMNLP2023 this week!

Come see Katherine Tian @ ICLR 🇦🇹's work on the ability of RLHF'd LLMs to *directly verbalize* probabilities (yes that's right as tokens) that are actually pretty-well calibrated! (Usually than the log probs!!! 🤯🤯🤯)

Poster 14B

R I G H T N O W until 3:30 SG time!!

LLMs fine-tuned with RLHF are known to be poorly calibrated.

We found that they can actually be quite good at *verbalizing* their confidence.

Led by Katherine Tian @ ICLR 🇦🇹 and Eric, at #EMNLP2023 this week.

Paper: arxiv.org/abs/2305.14975

Turns out we can directly train language models to hallucinate less without any human annotation -- for around 50% error reduction compared to RLHF!! Check out our paper for the approaches and full results 😃

Very curious to see how far we can push training to simply *not hallucinate.* It won't give us perfect models, but it seems like really meaningful (more than 50%) reduction in factual errors might be possible. Needs to be scaled up 😀 full thread on the way!