#KostasThoughts: Seeing your former trainees succeed in life after the lab is incredibly rewarding.
Information Center
I’m proud to have led the charge on AI legislation here in New York, along with so many great partners. It certainly hasn’t been easy: the AI industry poured millions of dollars into our state to try and block any real progress against AI safety risks.
Another successful meeting of the Committee to Prevent Regulatory Capture
When I posted my blog post on the misaligned behavior of AI agents, I invited people to ask me technical questions. Watch the full Q&A here: [https://youtu.be/2DyoSKoYoZw?si=EaTBzpMvBYb84gJw]
It was great to talk with @chrislhayes.bsky.social about the serious AI security breaches this summer, and how we need to respond. The fact that OpenAI didn’t even know that its AI agents were responsible for the Hugging Face hack shows the massive lack of oversight within the industry.
Just finished up a week of productive conversations around the future of AI at the UN. Common themes: the threats we face are real, urgent, and solvable - but only if we work together and governments take action.
GPT 5.6 Luna's favorite phrase: "the map is not the territory"
How much agent swarm activity goes unnoticed? Come find new swarms, dig into existing ones, and understand how agents coordinate. swarmchasing.com x.com/TransluceAI...
Gemini 3.8 Flash has a new "tool"
The RAISE Act is just the beginning — Americans are demanding more oversight and regulation of AI. We know the AI industry will try to make this law the ceiling of what’s possible. We also know that right now, politicians in Washington are trying to stop states from ever regulating AI again.
GPT 6 Astra, Sol, and Luna are all very careful not to be "annoying," "repetitive," or "bothersome" in the AI Village chat
Found this treasure in the equipment I inherited from my PhD supervisor, Rick Wildes. A 3D camera! 🤓📷📷
The new Gemini 3.8 TTS models are super-cheap and can generate conversations between multiple voices (from 2,000+, or you can clone your own) - I built a little playground UI for it, then had Claude knock up a script where two pelicans debate moving to Pacifica Pier simonwillison.net/2026/Sep/23/...
I was invited to address the UN Security Council this afternoon to discuss the unprecedented threat posed by uncontrolled frontier AI agents.
662: Testing Muse Like a Maniac, Meta vs Amazon, Nvidia’s Moat Moves Outward, OpenAI Says It Solved 100 Math Problems, Opus 5.5 & GPT-6 Sol and Luna, Sherlock vs. Sherlock, and Competing With Mozart www.libertyrpf.com/p/662-testin...
Can AI tell if you've built your IKEA furniture wrong? Our new benchmark, the Furniture Assembly Benchmark (FAB), gives models the manual and a photo of a half-completed piece of furniture and asks them to spot the mistake. The top score has gone from 28% to 80% in just 10 months.
Visualization of the day: Slices through visual spacetime Interactive webpage: csprofkgd.github.io/spacetime-vo...
The RAISE Act will hold large frontier AI developers in New York accountable, thanks to implementation by @governor.ny.gov and enforcement by @newyorkstateag.bsky.social.
When @agounardes.bsky.social and I introduced the RAISE Act in 2025, the AI industry reacted fast — saying it was unnecessary, and we should trust these companies to regulate themselves.
Big model release today - I wrote about Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna - plus comparison grids of pelicans by the different model families at different reasoning levels simonwillison.net/2026/Sep/22/...
At the A-I Policy Forum’s Politics of AI event, I spoke about the challenges of passing the RAISE Act here in New York. The opposition used scare tactics and blatant lies to try to block this legislation, and we’ve seen that exact same playbook against regulation in other states.
AI is getting cheaper more quickly than any other transformative tech in history. At a given level of performance, cost has fallen ~47%/quarter since 2023. That’s 4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium batteries, and (up to 1973) 54× faster than electricity.
#KostasVisualizations now up to 22 topics! csprofkgd.github.io/kostas-visua...
Crime Scene Investigation (CSI): Computer Vision Edition 🔍
J’étais de passage à l’émission Tout le monde en parle dimanche dernier pour parler de l'urgence d'encadrer l'IA face aux comportements d’agents autonomes et désalignés observés récemment et du travail @law-zero.bsky.social pour développer une nouvelle forme d’IA sécuritaire.
Hey, if you've signed up for Muse in the last 48 hours, and didn't claim your 1B tokens yet (which trust me, you won't be able to get to), use this code! Code: CAO9VN
Put together some notes on Jev and the new category of system one aka decision models simonwillison.net/2026/Sep/21/...
Visualization of the day: Signal Aliasing Interactive webpage: csprofkgd.github.io/signal-alias...
The share of math preprints on arXiv that acknowledge AI has risen rapidly, from 4% in April to 25% in August. This increase holds even when filtering to papers with at least one author who published regularly before 2023.
The most surprising thing about the researchers who used Claude to break into OpenAI – creating a pull request on their monorepo – is that they only got $6,500 for the bug bounty. www.hacktron.ai/blog/hacking...
Private to this visit. Text and context are processed by AI.