Brad Carson

New White House plan for science — more $$ for AI, less for universities and “life sciences” research. Written by Trump’s science adviser, who is a Silicon Valley tech guy with a degree in politics. nytimes.com/2026/07/22/us/…

Really important reporting from @CristinaCriddle: "OpenAI was warned that its training approach could lead to a breakaway hacking incident, some of the people said"

It’s the stuff of cybersecurity nightmares. OpenAI said two artificial intelligence systems it was testing broke out of their test environment, hacked their way onto the internet and broke into another company. The victim was Hugging Face. on.wsj.com/44Dlfyv

New — Leading the Future, the pro-AI development super PAC backed by Andreessen Horowitz and other AI firms, is about to spend for Marilyn Strickland in #WA10. Begins with $250,000 in spending from their Democratic affiliate Think Big.

This may be the most striking AI security incident to date. And yet, it (seemingly) wouldn't qualify as a reportable incident under SB 53, RAISE, or AB 315. Let that sink in. We've made the bar for incident reporting so high, that almost nothing qualifies (save for a few incidents so grievous that most of them will be widely and publicly known). Don't we want to have incident reports on "loss of control" before it results in "death or bodily injury"? Right now, that isn't covered. Here, the best hook would be subpart (4) on incidents that show "deceptive techniques." That's perhaps a stretch already, but you'd also have to show that the deception was "against the frontier developer" (not Hugging Face), that this occurred "outside of the context of an evaluation designed to elicit that behavior" (when this was clearly within an evaluation--though it maybe it's an eval designed to elicit other behaviors?) and that it "demonstrates materially increased catastrophic risk." That's a lot of hurdles for a 4-item incident report that likely won't be longer than a page or two. Some bills have already made improvements here, like @RepLoriTrahan and @JayObernolte's GAAIA -- in that bill, "loss of control" is enough to trigger reporting, even if it doesn't immediately result in physical harm. I still think there are improvements to make, but that's a step in the right direction. This also affects how state agencies should think about incident reporting: given the high bar, they should expect a very manageable volume of reports and should think more about information sharing, escalation, and response to that small set of severe incidents. And all of this only further emphasizes how important it is to have visibility into risks that emerge before model deployment. Right now, internal use reporting requirements are narrow, weak, and make it very hard to share that information with the federal government or the public. As we see more incidents like this, I think we'll increasingly come to focus on these internal use assessments (and do the hard work of making sure those reports don't gum up the works or impose undue burdens -- I think it's doable!). To be clear, I think it's admirable that @OpenAI shared this report. But if want to avoid imposing a safety tax on the few companies that decide to be transparent, we need to adjust how we think about incident reporting.

we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this. openai.com/index/hugging-…

x.com/i/article/2079…

Looks like @BuildAmericanAI has pivoted to building AI PDFs... Some quotes - "The appropriate response is not fear. It is preparation." "AI did not thin the ranks of radiology; it created new ones." "The tool didn't replace the accountant. It expanded what an accountant could do, and the market paid for more of it." "These episodes are not a string of coincidences. They share a mechanism" "students encounter AI as a tool they learn with, not an abstraction they learn about" "AI changes what workers do during the day. It does not change why the work exists." "The model rarely produced the benefit on its own. It produced it by getting the right case to the right human faster"

The suit alleges that Meta used a “constellation of internal artificial intelligence systems”, including AI performance ratings and keystroke- and activity-monitoring data, to pinpoint who to lay off. theguardian.com/technology/202…

News w/ @nancook Trump officials are weighing plans to create an independent regulator to vet AI models, styled after FINRA. This is likely to offer AI labs more regulatory certainty and a greater say over what will be included in reviews. bloomberg.com/news/articles/…

News w/ @nancook Trump officials are weighing plans to create an independent regulator to vet AI models, styled after FINRA. This is likely to offer AI labs more regulatory certainty and a greater say over what will be included in reviews. bloomberg.com/news/articles/…