This Meta paper just gave an empirical proof of why RL updates in LLMs are so sparse! They proposed this "Three-Gate Theory” & showed that pretrained models have highly structured optimization landscapes, so geometry-aware RL is much better than heuristic based SFT methods
RESEARCH
-
LLMs Continue Advancing Beyond Predicted Performance Walls
By
–
Absolutely. I've had the same experience. So much for LLMs hitting a wall. An amazing time to be alive.
-

AI Partnership with DOE Advances American Energy Dominance
By
–
We're proud to partner with @ENERGY and the Trump Administration on the Genesis Mission. By combining DOE's unmatched scientific assets with our frontier AI capabilities, we'll support American energy dominance as well as advance and accelerate scientific productivity.
-
Gemini 3 Launch: Google DeepMind CTO Discusses AI Progress
By
–
A LOT happened last week. ICYMI, @koraykv (CTO of @GoogleDeepMind and Chief AI Architect of @Google) and @LoganKilpatrick discuss Gemini 3, the state of AI, and where we are on the path to AGI.
— Google AI (@GoogleAI) 24 novembre 2025
Chapters:
0:00 – Intro
2:00 – Gemini 3 launch reception
4:16 – Continuous progress… pic.twitter.com/AudpzVmFpjA LOT happened last week. ICYMI, @koraykv (CTO of @GoogleDeepMind and Chief AI Architect of @Google
) and @LoganKilpatrick discuss Gemini 3, the state of AI, and where we are on the path to AGI. Chapters:
0:00 – Intro
2:00 – Gemini 3 launch reception
4:16 – Continuous progress -

Cerebras Co-hosts 8th Neural Scaling Workshop at NeurIPS 2025
By
–
Heading to @NeurIPSConf 2025? Cerebras is co-hosting the 8th Neural Scaling Workshop at NeurIPS 2025 with @mbzuai
, @Mila_Quebec and CERC-AAI Lab. Join us: https://
luma.com/gy51vuqd -

Claude Opus 4.5 outperforms GPT-5.1-Pro in efficiency
By
–
Claude Opus 4.5 – multi-turn (8 edits). A really nice result, I wouldn't say it is beating GPT-5.1-Pro, but it didn't take anywhere near as much time. I haven't shared Gemini 3 Pro properly because I could never get anything that I liked, so Opus is definitely winning here! https://t.co/dCFOyDu0Sx pic.twitter.com/LRLt2FNk9a
— Peter Gostev (@petergostev) 24 novembre 2025Claude Opus 4.5 – multi-turn (8 edits). A really nice result, I wouldn't say it is beating GPT-5.1-Pro, but it didn't take anywhere near as much time. I haven't shared Gemini 3 Pro properly because I could never get anything that I liked, so Opus is definitely winning here!
-

Why GPT 5.1 Underperforms on ARC-AGI-2 Benchmark
By
–
ARC-AGI-2 is being solved incredibly fast, but I still don’t understand why GPT 5.1 scores so low even though it’s the newest model
-

Anthropic releases Claude Opus 4.5 scoring 80.9% on SWE bench
By
–

BREAKING : Anthropic released Claude Opus 4.5 and it scores 80.9% on SWE bench verified! Where is the wall?
-
NeurIPS San Diego meetup invitation for AI discussions
By
–
Hey! I'm going to be at NeurIPS in San Diego next Weds-Fri! Ping me if you want to grab coffee or lunch and talk about AI (particularly juicy things nobody knows about yet)! I'm rmetz17@bloomberg.net, or rachelmetz.11 on signal!
-
Meta and ConservationX Launch SA-FARI Animal Dataset for Global Conservation
By
–
We partnered with @ConservationX to build the SA-FARI dataset with 10,000+ annotated videos including over 100 species of animals. We’re sharing this dataset to help with conservation efforts around the globe. Find it here:
