wayneb

Meta just beat human-engineered Claude Code scaffolding by 37%. Not by changing the model. By letting an agent rewrite its own scaffolding after reading every prior failure. The key insight the field has been missing: don't summarise the history. Keep all of it. 82 files on average. The dependencies are too long to compress. By iteration 7 it was making decisions a senior engineer would make. Co-authored by KRAFTON. The company that makes PUBG. The gaming industry's AI moment is closer than most people in it think.

@joaorcsilva @heyrimsha yes

@wayneb @heyrimsha you understand there’s ways of running models and 100% locally on your hardware and that is NOT equivalent of ollama cloud version.. right?

🚨BREAKING: A dev just open-sourced the #1 ranked OCR model on Earth. It's called GLM-OCR and it just hit 94.62 on OmniDocBench V1.5, beating every OCR model in existence. Only 0.9B parameters. One pip install. Handles documents no other model could touch. 100% Open Source.

Hermes Agent (NousResearch) has a built-in learning loop. After complex sessions it extracts skills, patches failures, saves lessons. I wanted the same thing in OpenClaw, so I built it The architecture is pretty straightforward... A trigger script checks session depth every 15 minutes. If a threshold is crossed a Flash-Lite subagent reviews the last 10 turns and asks: is anything here worth remembering? Skills get written to files. Lessons appended. Failures trigger a Sonnet patch agent. If nothing stands out it exits silently. Most runs do. The compounding is the point. Not today. In 6 months

I turned OpenClaw into a self-improving agent last night. Every 15 minutes it reviews our conversation. Extracts skills. Updates lessons. Patches failures. Cost: $0.10/month. The trigger is simple. 10 turns or 5+ tool calls in one response. If the threshold is hit it reads the session and asks one question: is anything here worth remembering permanently? 80% of runs say no. That's the point. High signal, low noise. Hermes Agent (NousResearch) gave me the reference architecture. OpenClaw handled the plumbing. Building toward an agent that genuinely learns.

I turned OpenClaw into a self-improving agent. And I used Hermes as a reference Every 15 minutes it now checks our conversation. Did I reveal a preference? Did it solve something non-trivial? If yes it writes a skill file or updates a lesson. If nothing's worth saving it exits silently. No noise. The review agent costs about $0.10 a month to run. Flash-Lite. Simple classifier. Does nothing 80% of the time. The part that actually required thought was the failure path. When the agent gets something wrong I want it to patch the specific skill that caused the failure not just log an error. Sonnet handles those. Flash handles the routine checks. It's not magic. It's just a trigger script, a state file and a prompt. But compounding over months is interesting.

@JoeTegtmeyer @TeslaPodcast @Tesla @elonmusk @farzyness

I’m really hoping Elon allows retail Tesla investors in the UK/EU some kind of allocation for SpaceX IPO

@MatthewBerman Send it my man!

Looking for some agent-addicted people to test a new project I've been working on. Comment below and I'll send you access.

@anvisha Moda

We raised $7.5M to kill AI slop. Introducing Moda: the world's first design agent with taste. RT+ comment “Moda” and we’ll design your brand for FREE.

My OpenClaw memory went from 65% --> 93%. Fully local. $0 a month. OpenClaw ships with both key memory features OFF. Most people never check. This is what I've found to be the best Memory architecture to get as close to 100% fidelity as possible. Five layers, each one catches what the others miss > Auto extraction after every conversation > Three parallel search agents > Compressed history stays searchable > Git auto commits every hour > Agent must write before it replies The model was never the bottleneck, the memory architecture was. Follow and DM me for the config files 👇

@aiDotEngineer @OpenAI Opening anymore spots? Waitlist seems like it could be infinite, so no chance to attend

We are excited to welcome @OpenAI to the AIE Expo for the first time as Platinum sponsors for AIE EU! OAI has shipped SO much for AI Engineers this year alone, and this is the best place to catch up: - Meet the team at the Ask OpenAI lounge (bring your hardest tasks and best questions!) - Hear keynotes from @steipete and @lopopolo, and - get hands on with in-depth Codex workshops from @kagigz and @reach_vb! See you April 8-10 in London! AI Engineers💙@OpenAIDevs !

I've followed Elon's companies for years. Reusable rockets. EVs. Neural interfaces. Humanoid robots. I invested in Tesla many years ago. So yes I think Elon is our modern day genius! But...I think Saturday night's TERAFAB announcement is the moment it all clicked into one picture for me. And honestly, I'm still processing it. Here's what I think most of the coverage is completely missing: Terafab isn't really about chips. It's about closing the last bottleneck. SpaceX can get things to orbit cheaper than anyone. Tesla builds autonomous vehicles and robots. xAI trains frontier AI models. But all of them are capped by one thing: chips. And the entire planet's output covers only 2% of what these projects need at scale. So Musk builds his own fab. But here's where it gets genuinely wild: >80% of the output is going to SPACE >SpaceX filed with the FCC for 1 million orbital AI satellites >Solar in orbit = 5x stronger than Earth's surface. Cooling in vacuum = free. >Musk says orbital AI will be CHEAPER than terrestrial AI within 2-3 years. And the robots? 1-10 billion Optimus units per year. Millions of them will build and operate the factory that makes their own chips. He said: "We're starting a galactic civilisation." I know that sounds like science fiction. So did landing rockets on drone ships. Now it's just Tuesday. Whether this fully materialises or not, I believe 10 years from now we'll look back at this presentation as the moment everything changed. What's the part that blows your mind most?

GitHub: github.com/msitarzewski/a…

Holy cow! this GitHub repo blew past 10k+ stars in 7 days It lays out an entire AI agency with engineers, designers, growth marketers and product managers all structured as distinct roles Super easy to follow, even if you are new to this space check first comment👇

@levelsio I run into these issues sometimes too. Drives me nuts

Claude Code with Opus 4.6 was so dumb today I finally had to write my own code again A sad state of affairs 🥹

@zach_yadegari Congrats man! I guess you didn't need that top ranking on the App Store after all ;-)

Cal AI has been acquired by MyFitnessPal 🚨 Henry and I started Cal AI as 17-year old high school students with one mission: make calorie tracking easier with AI. In just 18 months, we’ve helped millions of people lose millions of pounds. And we broke $50m in ARR along the way. We are at an incredible inflection point in history where ANYBODY can build a product that can improve lives and make millions. As founders, we get a lot of praise. The truth is that this would not have been possible without our incredible 30+ person team. We are so proud of what this team has accomplished, and are thankful to everyone that has been instrumental in Cal AI’s development and success. Cal AI will continue as a separate app from MyFitnessPal. The combined team will share resources to continue helping people achieve their fitness goals!

I built an SEO audit tool this month that doesn't just score your google ranking it scores how visible you are to AI search engines too chatgpt, perplexity, and google AI all pull from the same signals most sites score under 4 out of 10 this tool will help you hit 8+/10 DM if you want to join the waitlist

built 6 sites this year every single one needed SEO fixes after launch even with the best prompts even with AI building the whole thing and that's before GEO ranking on chatgpt, perplexity, AI overviews so I just built the fix waitlist open 👇