+1 to @TacoCohen on differentiability. Moreover it is important to remember that most RL for LMs involves only one step of RL where the state (question, instruction) is provided by the environment and where RL must generate a single sequence of tokens (1 action). This results in
@nandodf
-
AI’s Potential to Generate Revolutionary Knowledge Beyond Competition
By
–
We are insignificant in our universe, but we could write its most significant story. Catching up, leading LM Arena, the AI Race, competing with China or USA are such little uninspiring stories. Who wants to get out of bed for this? Our AIs will create new knowledge that will
-
Future of AI Training: Beyond Supervised Learning and SFT
By
–
Maybe one day supervised pretraining and SFT will be gone but it’s not clear how. RL has its issues too, and maybe one day we’ll have something smarter.
-
Reinforcement Learning Enables Causal Knowledge Beyond Supervised Fine-Tuning
By
–
Thanks for these thoughts. Very helpful. RL does indeed enable us to generalise better. I fully agree. In principle, RL results in causal knowledge whereas SFT only on association. But I feel the question is still relevant because we still rely heavily on (self-)supervised
-
Flow Matching for Video Generation: Data Scale and Generalization Benefits
By
–
I loved this research paper on Flow Matching, the most popular approach for video gen. TLDR: More data means harder to fit any specific data (say image), better generalisation, greater coverage of learned concepts. Pretraining: Use Billions of data to learn many things and none
-
New Math and Reasoning Benchmark Released for AI Evaluation
By
–
New math and reasoning benchmark. Highly needed because most of the current ones are nearly or completely saturated.
-
Call for Elite Engineers to Build AI-Powered Products
By
–
If you’re eager to learn all about AI, with complete intellectual freedom, or build the best products to empower users (education, healthcare, productivity,…) and believe you’re one of the best engineers in the world- or even better dream of this one day – apply.
-
MAI voice team showcases exceptional audio capabilities
By
–
This btw is another taste of what the MAI voice team can do. They’re absolutely the best I’ve seen (heard )
-
Microsoft AI releases MAI 1 preview and MAI Voice technology
By
–
This was an amazing week at @MicrosoftAI !! We released MAI 1 preview and a taste of MAI Voice. I’m super happy with this team – only about 100 people and already shipping in @lmarena_ai in less than a year. Strong support. More soon. Thanks for feedback! pic.twitter.com/bZzDmPsfI5
— Nando de Freitas (@NandoDF) 30 août 2025This was an amazing week at
@MicrosoftAI
!! We released MAI 1 preview and a taste of MAI Voice. I’m super happy with this team – only about 100 people and already shipping in
@lmarena_ai
in less than a year. Strong support. More soon. Thanks for feedback! -
Microsoft’s Mai Voice leads AI speech generation frontier
By
–
https://
copilot.microsoft.com/labs/audio-exp
ression
… Mai Voice – Leading the frontier of AI speech generation Bravo team