Bullish for AI progress outside the US…realistically just in China. Europeans will be cowed into submission, even if they somehow progressed to the notion of using…
> the real limitation being RL flops right nowI wonder what the limits of parallelization with MOPD areobviously, 1000 "experts" with 20 RL steps each are…
I'm not sure Xi has watched Star Trekanyway, props for not citing Harry PotterAtlantic Council: 🇹🇼 @TECRO_USA Ambassador Alexander Tah-Ray Yui: “When we watch Star Trek,…
> What makes GLM 5.2 interesting: zero failed runs across 84 runs (vs ~10% failure rate for Opus agents). The most reliable agent we've seenThe wonders…
RT Miles BrundageLarissa Schiavo: sir, another SoftBank AI deck has hit the internet https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2026/shareholders-meeting_46_05_en.pdf
RT Zixuan LiSpent some time with the PostTrainBench paper. https://arxiv.org/abs/2603.08640Lots of thinking on environment design, how to detect cheating, and what future evals should actually be…
RT DailyPapersNVIDIA just released an optimized GLM-5.2 on Hugging FaceA 753B parameter MoE with 1M context,quantized to NVFP4 for Blackwell GPUs—nearly matching FP8 accuracy.
RT Lilian WengA super long overdue (3+ years?) post on scaling laws.Compute is expensive. Scaling laws are a way to help us reason about the optimal…
This is a remarkable graph that reinforces the same point as autoresearch results, WeirdML, "takeout delivery bench", and all other long multi-turn scenarios. Notice: *GLM-5.2 isn't…
If Qwens are trained on *trillions* of reasonably diverse and non-inane Opus grade tokens, this is either utterly damning for Alibaba or for the whole synthetic…