Uploaded September 2026 | Updated November 2026, 2 days ago
Mercury 2 is the first reasoning model built on a diffusion architecture and the fastest model on the Artificial Analysis index. Modern diffusion LLMs demand a rebuild of the post-training stack, from algorithms to systems.
At Ray Summit 2026, Aditya Grover, co-founder and CTO of Inception Labs, covers why standard policy-gradient methods break when sequence log-probabilities are intractable, how efficient estimators make GRPO-style RL viable, and the levers unique to this paradigm, plus how diffusion LLMs herald a new era of efficient on-policy training with a well-optimized inference engine.
You'll leave with a view of the applications powered by Mercury 2 and the future of diffusion LLMs.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
Mercury 2 is the first reasoning model built on a diffusion architecture and the fastest model on the Artificial Analysis index. Modern diffusion LLMs demand a rebuild of the post-training stack, from algorithms to systems.
At Ray Summit 2026, Aditya Grover, co-founder and CTO of Inception Labs, covers why standard policy-gradient methods break when sequence log-probabilities are intractable, how efficient estimators make GRPO-style RL viable, and the levers unique to this paradigm, plus how diffusion LLMs herald a new era of efficient on-policy training with a well-optimized inference engine.
You'll leave with a view of the applications powered by Mercury 2 and the future of diffusion LLMs.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com










