Social Media Briefing

Mollick argues that Codex, Cowork, and Code are software-brained tools where the artifact is truth, making them poorly suited to knowledge work where the process and learning loops matter; he notes long-running models like Fable are hard to use for deep knowledge work.

Ethan Mollick describes giving GPT-5.5 Pro his first grad-school paper and having it find errors, locate and analyze new data, create reproducible files, and sophisticatedly extend the core argument.

Clement Delangue argues open-source AI leadership precedes general AI leadership, framing China leading open-source 2024-2026 as a foundation, and contrasts OpenAI and Google open beginnings with Meta abandoning openness.

Mollick describes giving GPT-5.5 Pro his first grad-school paper, which found new data, analyzed it, created reproducible files, and extended the argument, predicting weird AI-scholarship interactions.

natolambert argues open-weights models via GLM-5.2 reached a practically useful coding harness moment before Gemini, about 200 days after Opus 4.5.

Scoble's thread on world models: praises Oliver Cameron (who raised a reported 300M), argues that millions of real-time videos teach AI everything about physics and the world, and predicts smarter generalized humanoid robots within a few years plus mentions attending the ACL meeting.