Learning an optimal policy. YC F26
Joined March 2021
- Tweets4
- Following198
- Followers51
- Likes64
Asher retweeted
In MIT CSAIL's recent "Sparks of Weight-Space Programming", GPT-6 Astra fixed Qwen's thinking loops by editing its weights directly.
I replaced Astra's trial-and-error tuning with linear programming, constructing a weight edit that matched its results while taking 99% less time (minutes vs 8 hours) and changing 8x fewer weights.
Blog post and repo are linked below
Such a great experience! Will share more about what we built soon!
Excited to share that @asherdfeldman and I won first place overall at the Vibecheck formal methods hackathon yesterday! Had a great time meeting super cool people and learning from everyone there. We built Sound and Complete, a verified tool to help guitarists practice call-and-response blues.
Thanks to @maxvonhippel, @qd_forall and everyone else who helped put on this fantastic event!
There are some large winds and great people moving FM forward, excited for what is to come!
