@computerender 6/ Built with Claude Cowork- it did all the heavy liftings (thanks for creating such an amazing tool @bcherny).

@computerender 5/ This guide also covers trajectories, outcome vs process rewards, credit assignment, verifiable vs learned rewards, and more reward hacks. Code on GitHub.

@computerender 4/ Best part: watching agents hack the rewards. Rewarded "not losing HP"—agent learned fleeing has 100% success and zero damage. Ran from every fight. Zero XP, maximum reward. It did exactly what I asked. I asked for the wrong thing.

@computerender 3/ An RL environment isn't just "the game." It's observation space, action space, grader, state transitions, termination conditions, and sandboxing—all working together. The mental model: the environment is the teacher. When an agent isn't learning, it's usually an environment

2/ Left agent keeps hitting the same wall. Right agent remembers what's blocked and finds the gap. Hat tip to @computerender's Pokemon RL work, which inspired this whole thing.

1/ You make @claudeai work hard. I let it play Pokemon Red for 10 hours. Along the way, @karthikabinav & I wrote the field guide to RL environments:

Kimi K2.5 has arrived! 🥝 Here are 2 things to know: Aesthetic Coding x Agent Swarm.

Set up @moltbot on AWS and talk to it on WA. This thing is absolutely _wild_. It’s great to see a glimpse of future. So much to build!

📢 Confession: I ship code I never read. Here's my 2025 workflow.

Claude in Excel is now available on Pro plans. Claude now accepts multiple files via drag and drop, avoids overwriting your existing cells, and handles longer sessions with auto compaction. Get started: https://t.co/cAMDXM1h7r

Be your own AI’s biggest critic

Very insightful!

Today we are launching @openwork_ai, an open-source (MIT-licensed) computer-use agent that’s fast, cheap, and more secure. @openwork_ai  is the result of a short two-day hackathon our team decided to hack, which brings together some of our favorite open source AI modules into

6/6 AI coding agents are genuinely incredible. Undoubtedly it is the biggest unlock for software dev in a decade (maybe even longer). But if your success metric rewards volume, you’ll get volume. Don’t reward hack your productivity. Focus on the true objective worth pushing.

5/6 The real magic of AI coding isn’t that I write more. It’s that I can try 5 different approaches in the time it used to take me to commit to one. I prototype faster. I refactor without fear. I explore ideas I would’ve killed in my head. That’s the story worth telling.

4/6 Here’s what I’d actually want to measure: ∙Time from idea to working feature ∙Bugs per release ∙Confidence when touching unfamiliar code That’s where AI coding shines. Not raw output. None of this shows up in “lines changed.”

3/6 Don’t get me wrong—I’m extremely bullish on AI coding. Use it every day for all my code. It’s already completely changed how I work. What the data actually shows: AI makes it stupid easy to generate code. Which… yeah, obviously. That’s literally what it does. But “more

2/6 The numbers look incredible on paper: ∙Lines of code per dev up 76% ∙Medium teams nearly doubled output (+89%) ∙PRs are 33% bigger Sounds amazing right? Here’s the problem—lines of code has always been a garbage metric. AI just made it louder.

1/6 We finally have data on AI coding. I’m surprised that this report wasn’t getting enough attention. Super interesting stats by the @greptile team. Fantastic work! As I was reading this report, I’m worried that most people (especially execs!) might be drawing the wrong