Moonfarm 🇸🇪

1/4 This year AI and specifically AI Agents have taken a huge leap forward with @openclaw, @anthropics coworker and now @cursor launching their own "always on agent". 🔥 I have fully embraced "Agentic engineering" as @steipete so gracefully put it, and have had a blast working on an agent-first saas, I actually think I need to start using all my time to build agent-first saas. 🫣 So! Here are the new goals for this year 🧵👇

@geggleto @TukiFromKL I know, that's the only thing ive been tweeting about the last hour 😂

@moonfarm_dev @TukiFromKL x.com/fynnso/status/…

🚨 Cursor just dropped Composer 2.. their own AI model.. not Claude.. not GPT.. their own.. and it beats Claude Opus on coding benchmarks.. at a fraction of the cost.. a code editor with 50 people just outperformed a $30 billion AI lab.. at coding.. which is supposed to be their whole thing.. the vibe coding era just got an upgrade..

@geggleto @itsolelehmann It kinda helps make a generic skill more specialized, but of course, if the verification the agent tests it against is poor, the skill will not become better

@moonfarm_dev @itsolelehmann na this is only 1/2 the battle. what its really doing is limiting context to a specific solution

i built a skill that 10x's all your other claude skills on autopilot (using karpathy's autoresearch method) the problem: most of your claude skills are quietly broken and you have no idea i didn't either. i ran my landing page copy skill through the autoresearch scoring loop and it was failing its own quality checks 44% of the time(!!) karpathy (co-founded openai, coined "vibe coding") built autoresearch to improve machine learning code autonomously. turns out it works on skills too here's how: you give it 3-6 yes/no questions that define what "good" means for that skill. ex: "does the headline include a specific number?" or "is the copy free of words like revolutionary and game-changing?" then it loops: 1. runs your skill 2. scores the output against your questions 3. tweaks one thing in the prompt 4. runs it again 5. score went up? keep the tweak. score went down? undo it 6. repeat i walked away and came back to a landing page skill that passes 92% of its checks. 4 rounds of changes, zero manual editing works on anything where you can define what good looks like. outreach sequences, newsletter drafts, any prompt you reuse, etc note: method credit to karpathy, i just adapted it for claude skills full skill below.

@geggleto Haha 😂

@moonfarm_dev absolutely wild. I was thinking about spinning up a K2.5 instance this weekend, guess I dont need to now lol

@T_Zahil But is it okay to brainstorm distribution ideas? 🤔

❌ Stop brainstorming product ideas ✅ The only ideas worth your time are distribution ideas.

@leerob Can i just say that it's great that you decided to be transparent a it this, kudos for not just leaving it in the dark 👏

Yep, Composer 2 started from an open-source base! We will do full pretraining in the future. Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training. This is why evals are very different. And yes, we are following the license through our inference partner terms.

@MythThrazz Haha 😂👏 working on it!

@moonfarm_dev Whatever, justs give the bots their own emails so they can spread! :D

I don't know if this is any good for a new domain, but I guess it's progress 🥳

@OskarWieckowicz 🙌🥳

@moonfarm_dev I am sitting at 6 with 39 backlinks. So you are good.

@sown0x @amanrsanger This is hilarious

@amanrsanger 😅

Composer 2 marks the one-year anniversary of our large model training efforts. Since then, we've built an exceptionally talent-dense team of ~40 people with some of the best researchers and engineers from the labs, academia, industry, and more heterogeneous backgrounds. And we are exclusively focused on coding. We don't care about models that can respond to emails, do your tax returns, or be your friend. Every FLOP, token, parameter, and researcher is entirely dedicated to software engineering.

Oh so that's how you "release your own model" these days 😂

@MartinTale Yeah, my openclaw is just a mess of fixing issues, ill probably lean into whatever this turns out to be haha

@moonfarm_dev Hmm, I want this now 😊

Starting to feel big openclaw vibes 😄

@suida_ajdini That's awesome, congrats! 🥳

Finally 1K verified followers 🥳

@JaredC50767 Yeah, was hoping this would be compatible with the remote control feature

@moonfarm_dev Yeah, the interesting split is convenience vs control. Messaging Claude Code from your phone is nice, but the durable layer is still whether you own the runtime, local files, browser auth, and the recovery path when the workflow gets weird.

@jordienr Damn, you solved it!

yall spending thousands on GEO tools and this shit so easy

Next video in progress 🔥 Got some great feedback on the last one, need to slow down the pace and instead of focusing on doing something cool, make it understandable as well! x.com/moonfarm_dev/s…

@shredandship I couldnt agree more, i love how AI enables us to do so much more with the same resources

Some developers fell in love with writing code. I fell in love with building things. That’s why AI doesn’t feel like a threat to me. The how has changed a dozen times over my career. The why never has.