Andrej Karpathy, former Director of AI @ Tesla and founding team @ OpenAI, explores a vision where frontier LLMs spawn teams of autonomous agents to tackle complex research. He also shares how he’s using LLMs to build personal knowledge bases — indexing research documents into wikis that let him ask complex questions against ~100 articles and 400K words of material, with the LLM auto-maintaining indexes and reading related data efficiently.
Box CEO Aaron Levie outlines the key lesson from building AI agents: be brutally unsentimental about architecture. As models improve, previously necessary scaffolding becomes a constraint. The loop involves building systems around the LLM, then removing them as capabilities emerge, then building again for harder problems. Many mitigations put in place (like chunking for context limits) eventually hurt results as models get better.
Replit CEO Amjad Masad announced a new sales office in Salt Lake City and shared an enterprise-grade auth solution for developers.
Cursor Designer Ryo Lu makes the case for “glass” over “black box” AI tools. The terminal was a black box where you learned to think like the machine. AI kept that pattern — you type a wish, something comes out, you accept or reject the whole thing. Glass breaks this: the PM writes a plan and watches it become real, the engineer sees every diff, the new programmer reads and learns. As AI gets more powerful, glass becomes more important — not because you need to watch every move, but because the best work happens when you know you can.
OpenClaw’s Peter Steinberger highlights that AI-generated security reports have become a significant burden on open-source maintainers — reports jumped from 2-3 per week two years ago to 5-10 per day now, with most being correct but requiring new maintainers to handle the volume.
Anthropic’s Claude announces computer use in Claude Cowork and Claude Code Desktop is now available on Windows.
Y Combinator President & CEO Garry Tan calls Perplexity Computer “quite special” and shares thoughts on loving your work.
OpenAI CEO Sam Altman calls TBPN his favorite tech show and expects them to “not go any easier on us.”
Training Data — “How Autonomous Labs Will Transform Scientific Research: Ginkgo Bioworks’ Jason Kelly”
Jason Kelly founded Ginkgo Bioworks in 2008 with the goal of making biology programmable — a vision that has taken on entirely new meaning in the AI era. The previous revolutions in tech (internet, social media) were “totally meaningless to biotechnology and biopharma” — just back-office IT improvements. But AI is different: it’s actually going to change the fundamentals of how we do science, and big science industries like biopharma are going to get disrupted.
Ginkgo bootstrapped for six years before raising capital in 2014 — something biotech VCs rarely tolerate. The company is now building “autonomous labs” that combine AI with robotics to run experiments at massive scale, essentially creating a “ ChatGPT for biology” that can design and execute experiments without human intervention in the loop. The vision: instead of humans hypothesize → experiment → learn, you have AI hypothesize → robot executes → AI learns → repeat, running millions of experiments instead of dozens.