Builder
Jack Rae
Distinguished scientist, Meta
Led Gemini Thinking, co-led Gemini pre-training, and contributed to Gopher and Chinchilla language-model scaling research.
Follow Jack Rae
We’ll email you each new moment, one email a day at most. Plus the best of every big interview, Wednesday and Saturday.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
It’s free.
Meta@jack_w_rae(opens X in a new tab)Real Alpha has cut 5 moments of Jack Rae (Distinguished scientist, Meta) talking, from an interview on Unsupervised Learning published Mar 17, 2025.
What Jack Rae said
Unsupervised Learning · Noam Shazeer and Jack Rae: Scaling Test-time Compute, Reactions to Ilya & AGI
5 moments · 17 min
- 25:29
AI math progress depends on asking better questions
“Whether it's possible, I'm not sure, but I would agree maybe the key crux is can I ask actually novel questions?”
at 31:30 - 9:39
Qualitative tasks need broad reward criteria
“take in a very broad set of criteria and then apply a reward signal once we have the reward signal we can train with reinforcement learning against it”
at 10:54 - 15:57
Agent research depends on choosing the right environment
“I don't think anyone has solved the perfect environment yet”
at 18:57 - 42:03
Use one model unless specialization earns its cost
“If there's positive transfer, put it in the same model, as long as it doesn't then become too expensive to serve.”
at 43:55 - 1:06:30
Inference-heavy AI can use more distributed compute
“if it turns out that building AGI becomes mostly an inference problem, an inference problem that can be much more distributed than maybe synchronous large batch training”
at 1:06:44
- 25:29
Is this you? Add the badge to your site.
