Is the Apple Macintosh charging brick the least-well-designed device ever? It blocks adjacent sockets completely. That is, until it falls out of the wall.

New Rules: processes that are compulsory ought to work. @WeWork Merchant Square London

I'm proud we are releasing LAION-fMRI, a densely sampled 7T fMRI dataset of natural images, with very broad stimulus sampling for testing countless hypotheses & deeply exploring brain representations. It is now available at laion-fmri.hebartlab.com/ What does LAION-fMRI offer? 🧵

Check out my latest article: Fast Loop: Aotearoa Agentic AI linkedin.com/pulse/fast-loop… via @LinkedIn

Do Impossible Things. AI can help.

It'd be great to see @Tesla help out some of these guys with semis, across the market. Maybe 1 month of production edition.cnn.com/2026/03/28/b… @elonmusk

I (and a reasoning model) did a calculation. A 2 earth G (fairly mild) passenger transport mass driver on the moon only needs to be 144km long to reach escape velocity in about 2 minutes. To make it follow the curvature (otherwise you need quite a tower at the end) you need a max anti-levitation force of about one lunar gravity, the levitation force that you need to keep it off the rails at the beginning anyway. With care, this could be used for landing too. @elonmusk @PeterDiamandis

Announcing The Future Vision XPRIZE. A global competition with $3.5M in prize funding challenging creators anywhere on Earth to imagine hopeful, technology-forward futures worth building toward. Not warnings. Blueprints. Futures that inspire us to go boldly. Someone in your timeline is sitting on a vision that could change the world and doesn't know this exists yet. Share this. Be the reason they find it.

Computer science will stop being taught as a useful craft; it has become an explanation of how the world works, like physics, biology and chemistry. This change is long overdue.

New term "distillation poisoning"; I expect to see it used, and papers about it, soon.

Steering and queuing in GPT-5.3-Codex Extra High in VSCode is game-changing for productivity. Now, please have @OpenAI Codex write its android and windows apps.

It might be worth trying opt-in AI-based review of one's papers, in return for partial or complete reduction of reciprocal review load. Achieving high levels of calibration with human review may not be difficult, and knowing whether it is would be a valuable research outcome. @ReviewAcl

We (@OpenAI Codex 5.3 and I) just replaced the very unreliable @Jira MCP toolset with API level tools from Von. Now it's reliable and provides tools (like linking) missing from @Atlassian MCP. This will happen a lot, as we code away barriers (reflects the recent @karpathy library replacement post).

Dear @OpenAI - please tell your internal coding agents, human or otherwise, to always simultaneously test (an use in training) bash (all recent versions), PowerShell (ibid) and any sh variants in the macoverse. Trivial things not working because of shell assumptions is bad.