Benjamin Trom (@BenjaminTrom) header
Benjamin Trom (@BenjaminTrom) profile picture
Paris
Joined February 2022
Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)Photo by Benjamin Trom (@BenjaminTrom)
when will we have a jev but for next token prediction?!
1
57
Congrats to the team, amazing release!
Millennium Prize Problems a making headlines, but your wheat 🌾 won't care ! Thrilled to introduce BOTANIC-1, our new plant DNA FM family. Outperform every model we tested on our benchmark. Weights: huggingface.co/collections/l… Paper: biorxiv.org/content/10.64898…
1
5
1,005
Benjamin Trom retweeted
Millennium Prize Problems a making headlines, but your wheat 🌾 won't care ! Thrilled to introduce BOTANIC-1, our new plant DNA FM family. Outperform every model we tested on our benchmark. Weights: huggingface.co/collections/l… Paper: biorxiv.org/content/10.64898…
5
14
4
63
88,703
Mistral was an exceptional place to learn. I worked with some of the best people in the field at a pace I didn't think was possible, and I'm truly grateful.
1
7
431
Today Magic published its first result since I joined. Pretraining is the most closely guarded capability in AI, and the one frontier labs are measured against. Magic just showed it can do it more efficiently than anyone else and is the smallest team in the world training 1T llms
2
1
7
268
If you want high impact on frontier LLMs in a small team that works from first principles, I don't know a better place to be right now. And you'd get to work with me! Roles are in SF: magic.dev/careers magic.dev/blog/pretraining
1
205
Some personal news, a bit late. Sorry for the delay: I've been too busy cooking. Last October I left Mistral AI and joined Magic in SF to do frontier AI research. magic.dev/blog/pretraining
2
2
67
4,607
Benjamin Trom retweeted
Frontier pretraining is said to be a big-lab-only game. We don’t have 100k chips yet, so there’s only one way: algorithmic efficiency. Our new recipe matches DeepSeek V4 Pro’s pretrain using 50x less compute – that’s roughly half the FLOPs used for GPT3, or ~$0.5M on GB200. magic.dev/blog/pretraining
80
177
118
2,125
472,545
"generated with claude" is becoming the new "sent from my iPhone". technically true. completely useless. so i made the slop scale
1
1
7
227
so: self-assess the slop level. then put the audit result proudly in the pr body. I made a skill for that. side effect: doing the audit may force you to actually understand your own PR. github.com/limiteinductive/a…
1
2
102
i'm pretty sure that a claude codes outages has a measurable impact on birth rates
2
172
is reverse acquire VCs reverse moat 🙃
1
467
only in SF: my Uber driver casually mentions he’s just here for the weekend... because he works in private equity and "just here to check on some houses to buy"
3
595