Shmuel Berman

@ShmuelBerman

PVL Lab @ Princeton | Memory and Perception | Anthropic Fellow | thebestworstcase.substack.com

New York, USA shmuelberman.com Joined February 2020

211 Following 246 Followers 83 Posts

Latest posts 5

Pinned post

Shmuel Berman @ShmuelBerman Sep 2

Researchers are actively improving memory for LLMs/VLMs. But are we measuring it right?

A system can score 100% by rereading the input for each question. If a human did that, we’d call it terrible memory!
We test beyond accuracy with ECC: Efficiency, Compression, Calibration. 🧵

Shmuel Berman @ShmuelBerman 6d

Had a great time working on this project. Especially rewarding to get a better sense of when ensembling, often overlooked for LLMs, is worth the compute. Turns out it can be. It was wonderful working with the amazing @EthanBaron78553!

Andrew Gordon Wilson Verified account @andrewgwils Nov 2

New paper on compute-optimal scaling prescriptions for data constrained pre-training. Our prescriptions include a return of ensembling in compute-optimal training, and uncover surprising pathologies of past approaches! Led by @EthanBaron78553.
arxiv.org/abs/2610.37842
View quoted post on X

Shmuel Berman @ShmuelBerman Jul 15

I am excited to share my work as an @AnthropicAI Fellow!

Can language models control robots? ⬇️

Shmuel Berman @ShmuelBerman Jun 30

LLMs make human behavior more predictable. Prediction markets incentivize people to change outcomes they have control over in order to make a profit.

I'm worried that these two technologies might have dangerous synergies that could be disastrous for a stable society. 1/3

Shmuel Berman @ShmuelBerman Mar 19

10x is crazy.. glad to have helped on this :)

Samip Verified account @industriaalist Mar 19

Announcing 10x data efficiency on NanoGPT Slowrun!

There are two macro trends worth highlighting:
- pretraining is nowhere close to done, and
- 100x looks feasible.

Writeup on all the core ideas: qlabs.sh/10x

View quoted post on X