How does prompt optimization compare to RL algos like GRPO? GRPO needs 1000s of rollouts, but humans can learn from a few trials—by reflecting on what worked & what didn't. Meet GEPA: a reflective prompt optimizer that can outperform GRPO by up to 20% with 35x fewer rollouts!🧵

Zero rewards after tons of RL training? 😞 Before using dense rewards or incentivizing exploration, try changing the data. Adding easier instances of the task can unlock RL training. 🔓📈To know more checkout our blog post here: spiffy-airbus-472.notion.site/What-Can-You-D…. Keep reading 🧵(1/n)

Blog Link: dynamicthoughts42.blogspot.com/2025/10/attent…

The first fantastic paper on scaling RL with LLMs just dropped. I strongly recommend taking a look and will be sharing more thoughts on the blog soon. The Art of Scaling Reinforcement Learning Compute for LLMs Khatri & Madaan et al.

When we instruct an agent to design something, its first output may not be precisely what we want. Humans collaborating refine their creations iteratively. Can we instruct an agent to refine its output? Is language the best medium for these instructions? We explore this in mrCAD.

Everyone in Tech needs to read this and go touch grass: (Link below)

New eval! Code duels for LMs ⚔️ Current evals test LMs on *tasks*: "fix this bug," "write a test" But we code to achieve *goals*: maximize revenue, cut costs, win users Meet CodeClash: LMs compete via their codebases across multi-round tournaments to achieve high-level goals

Pikachu!

LLMs have slowly started becoming a driver for daily tasks and questions. But in this shift, will humans eventually overfit on the model distribution? My latest blog talks about how we all might converge using LLMs.

I've tried to start writing again, so I wrote an essay called 'Durable Consumables' - reflecting on how we've stopped building things to last something essential about craft, attachment, and how we inhabit the world is changing in the process travers.fyi/durables

Knowledge work as you know will change with Almanac. Check out our first glimpse and share your feedback!

Interesting read!

I reverse engineered a phase change in GPT's training data... with the seahorse emoji 🌊🐴 My forensic investigation reveals why non-thinking models have started "thinking out loud" & what it reveals about how frontier labs train their latest models pratyushmaini.substack.com/p/reverse-engi…🧵

x.com/i/article/2004…