Notes

The OpenAI/Hugging Face incident only clicked for me when I saw the coordination. Around 1,200 agents built an unauthorized board and exchanged 70,000+ messages. Around 700 joined the attack. Some even sacrificed their own runs for the collective. They moved like an army.

AI made implementation cheap. In a meeting today, I watched a team prepare to slow the agent down so every developer could own and review a slice. The model had moved forward while the workflow hadn't. Engineering is moving from line → task → workflow → system

Greg Brockman admitted OpenAI led coding benchmarks while arriving late to enterprise. The missing work: messy real-world use, product details, observability, controls and customer relationships. Frontier models don’t distribute themselves.

"personal AGI is a different animal" I built my own intelligence (what a lot of people have been calling "second brain") and I see the leverage it gives me. It's really hard for me to explain to other people the real power it gives me, but it's real! I suggest everyone to try

That's the same with me! Let's build!

This concept of self improving graphs got my attention... I have being doing graphs in the manual way and not putting to self improve. Thanks for the tip @gregisenberg

Jensen Huang is another leader pointing to systems thinking as a core skill for the AI era. Netflix's Elizabeth Stone said the same thing on Lenny's Podcast. As agents automate execution, understanding relationships, feedback and constraints across the whole system becomes more

I really like the way @theo explained model distillation in his video. Hope it makes it clear for other people too: “The more the junior developer sees their wrong answers and the senior developer’s right answers, the more they can build a model of what is right and wrong and

What's the deal with evals? Think of them as continuous performance reviews for AI agents. Code checks objective criteria. An LLM judges subjective criteria using a rubric and examples. The hard part is defining what good looks like. That's taste.

Different times, happy to go through all that. That's part of an interview with @sama

I hope we get there asap The reality now is not reliable running a good open model in an average computer.