There definitely needs to b a better of understanding of what benchmarks r actually comprised of Many of the most widely cited r somewhere between unrepresentative and completely useless in practice

Very interesting effect which similarly emerged w DS Prover V2

- Zuck This is in direct opposition to the realities of the Llama 4 release Much more likely Qwen will b the one seeing these benefits

I agree w Zuck's point here and the entire history of computing does too

Amazing work Looking like a new addition to the RL stack