Reinforcement Learning (RL) has long been the dominant method for fine-tuning, powering many state-of-the-art LLMs. Methods like PPO and GRPO explore in action space. But can we instead explore directly in parameter space? YES we can. We propose a scalable framework for

another day another person telling me to try DSPy

I feel like I still haven’t recovered from the Sutton interview.

👀👀👀

Chatting pelicans and caselaw with Simon Willison at DevDay yesterday.

Worried about ai taking your job? Just be better than ai at your job, ez

hypothetically if someone wanted to hire a person whose main job would be to preach about a great search platform and build demos/video content, handle comms on here and discord (tl;dr: do devrel) while still being technical and talking about colbert, where should they look?

Here's a demo of what you can do with the tilesets from retro diffusion, every single asset in this was generated with zero touch ups. Now imagine plugging this into a procedural generation system, where players can ask for *any* landscape or theme.

uhhhhhhhh

when Claude Code gets really good at using the Claude Agent SDK it's going to get real weird.

So, if you had a subject matter expert willing/available to do the grading work, couldn't you substitute out the scoring and LLM reflection step in GEPA with a human scorer and reflector and potentially get better results? Is this already a thing? Sounds similar to

Optimizing a data analysis coding agent with GEPA, using execution-guided feedback on real-world workloads. Amazing tutorial by Arslan Shahid: medium.com/firebird-techn…

Appreciate Tron: Ares commitment to the bit, the dialog was very clearly written by Grok.

If you are in AI and saw the new Tron movie you’ll understand why this is very funny github.com/vzeman/ares-mc…

Finally got a nice harness wired up over the weekend with DSPy to generate valid interleaved ABC musical notation so can go LLM to MIDI file. Going to see if I can get GEPA or another optimizer to help generate better compositions. Fun problem because all the frontier LLMs