Jeremy Berman
@jerber888
Followers
2K
Following
1K
Statuses
80
research @ndeainc. co-founded https://t.co/aY50hNeJUD. yc w19.
NYC
Joined August 2017
I just got first place on the public ARC-AGI benchmark using Claude Sonnet 3.5 and Evolutionary Test-time Compute
2024 ARC-AGI-Pub SoTA! đź‘ľ 53.6% @jerber888 47.5% MARA(BARC) + MIT (@ellisk_kellis, @akyurekekin) 43.0% @RyanPGreenblatt
40
87
1K
Internet connected code execution is the only tool an LLM needs.
0
0
9
Returning the literal reasoning tokens at inference is important because it’s useful to manipulate the reasoning. If the reasoning tokens are part of the context, we can edit / guide the reasoning in real-time to do interesting things. Having root access to is valuable. This puts o3 at a disadvantage relative to DeepSeek R1.
2
0
10
@ElliotEvNo I’m sure we do have some preset parameters. We don’t start our RL loop from random weights
1
0
0
RT @mikeknoop: just published my full @arcprize analysis of deepseek's r1-zero and r1. link below. key points: r1-zero is more important t…
0
136
0
AIs have bad taste now because they are confined to the tastes within their training distribution, and the average taste on the internet is bad. This is temporary. We will figure out AGI: how to generalize materially outside of the training distribution — how to create new knowledge and new taste on the fly through exploration. Then, AIs will have extraordinary taste.
0
0
6
“No man ever steps in the same river twice, for it's not the same river and he's not the same man.”
0
0
0
RT @ClementBonnet16: Excited for the launch of @ndeainc 🔥 We’re making the bet that merging structured reasoning with deep learning will c…
0
10
0
RT @mikeknoop: AGI is the most important technology in the history of the world. That's why I'm going all in on a new adventure with @fchol…
0
96
0
RT @michaelxbloch: I’m trying something new: a weekly roundup of ideas and articles that made me stop and think. Here’s what stood out to m…
0
1
0
Facts and deduction are all we need to create knowledge. Intelligence is intuiting which facts to deduce from.
2
0
9
@techczech The best thing to do is to have an LLM semantically chunk the training data. But that’s super expensive
0
0
0