🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Today’s report on AI Governance in CA builds on the urgent conversations around AI governance we began in the Legislature last year. I thank Fei-Fei Li, Jennifer Chayes, and Tino Cuellar for the hard work and keen insight they provide in this urgent report.

When will AI systems be able to carry out long projects independently? In new research, we find a kind of “Moore’s Law for AI agents”: the length of tasks that AIs can do is doubling about every 7 months.

AI models are *not* solving problems the way we think using Docent, we find that Claude solves *broken* eval tasks - memorizing answers & hallucinating them! details in 🧵 we really need to look at our data harder, and it's time to rethink how we do evals...

NEW on Transformer: Where's Gemini 2.5 Pro's system card? Google's previously promised governments that it will publish such information — but with the latest model, it's nowhere to be found.

Notable also that OpenAI's Deep Research system card came out over two weeks after the model dropped. Good rule of thumb: if your AI lab is releasing models too fast to even keep up with your own voluntary transparency commitments, then you are moving too fast period.

Even the most basic AI transparency is falling to the wayside in the competitive crush. OpenAI puts out a model card weeks late for Deep Research; Google doesn’t publish any at all for Gemini 2 or 2.5; now Meta puts out the shortest vaguest one I’ve ever seen with Llama 4. Sad!

"Especially to the extent AI developers continue to stumble in these commitments, it will be incumbent on lawmakers to develop and enforce clear transparency requirements that the companies can’t shirk." -- Kevin Bankston Agree! The time for these requirements is now.

I remember talking about competitive pressures and race conditions with the OpenAI’s safety team in 2018 when I was an intern. It was part of a larger conversation about the company charter. It is sad to see OpenAI’s founding principles cave to pressures we predicted long ago.

In MIT Technology Review, I wrote about the crisis in AI evaluations — and why a new focus on validity could be the best way forward

🚨 New paper alert! 🚨 Are human baselines rigorous enough to support claims about "superhuman" performance? Spoiler alert: often not! Patricia Paskov and I will be presenting our spotlight paper at ICML next week on the state of human baselines + how to improve them!

Think the upshot here is that we should certainly leverage chain of thought as low-hanging fruit for safety alignment but also continue to invest in other more direct methods of interpretability since CoT can be unreliable. Yes and, not either or.