Latest / HotTea: The AI Morning Briefing / California sets a November deadline for AI safety proposals
Transcript
- Sections 00:00 What we're covering today 00:30 1. California sets a November deadline for AI safety proposals 01:33 2. Anthropic will pay Accenture to test its AI safeguards inside the company 02:30 3. European AI firms say a slowdown could protect US leaders 03:25 4. Nscale files for an IPO after a $1.02 billion half-year loss 04:28 5. Richmond Fed contacts say AI is slowing hiring, not cutting jobs yet 05:22 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript Welcome to Hot Tea for Saturday, September nineteenth. California is considering who should test A I safeguards and what happens when one fails. Anthropic is bringing evaluators inside the company, while European rivals are pushing back on calls to slow development. We'll also look at Nscale's finances ahead of a possible stock-market listing, and what regional employers are saying about A I and hiring. This narration uses an A I-generated voice. California sets a November deadline for A I safety proposals. Governor Gavin Newsom signed Executive Order N-9-26 on September 18. It gives state agencies until November 16, 2026, to propose changes to A I safety law. The agencies must consider allowing outside evaluators to check companies' safety reports for themselves. They must also consider a kill switch that those reviewers could test over time. The review includes rules for reporting incidents in which developers lose control of their systems. For now, this is only a review. It doesn't require every developer to install a kill switch, and the agencies still have to decide whether the technology is feasible. The recommendations are due November 16. By May 1, 2027, California must also publish criteria for organizations that want to conduct independent verification. The open question is whether California defines a shutoff that reviewers can test, and whether it lets them report failures. Those decisions will determine how much power outside evaluators have. Anthropic will pay Accenture to test its A I safeguards inside the company. The companies announced the partnership on September 18. Faculty, Accenture's specialist A I business, will lead the work. Each company expects to invest at least $1 billion over five years to build A I safety capacity. Anthropic says the evaluators will get access similar to what its employees have. They'll test safeguards and check whether the models behave as intended. Anthropic will pay Accenture for that work. The company says the industry has no common standards for evaluator access, reporting, or funding. Calling an evaluator "independent" doesn't tell us who controls its access or whether it can publish what it finds. The partnership is non-exclusive, and Anthropic says it will name more evaluators. How much access they get, whether they can report freely, and what they publish will show how much outside scrutiny Anthropic permits. European A I firms say a slowdown could protect U S leaders. Reuters reported on September 18 that Mistral and other European A I companies oppose U S-led calls to slow frontier A I development. Mistral argues that established companies could use safety rules to protect their market positions. The European companies don't all agree on oversight. Hugging Face chief Clement Delangue supports placing independent evaluators inside labs, but he opposes a slowdown. These companies have a financial stake in rules that could reshape their markets. That doesn't erase the safety risks, but those risks don't decide who gets to write the rules either. Who chooses the evaluators matters. New entrants may face the same thresholds as established companies, while public authorities need a way to challenge the largest labs. Those choices will show whether the rules treat competitors evenly. Nscale files for an IPO after a $1.02 billion half-year loss. The company announced its U S IPO filing on September 18 and applied to list on the New York Stock Exchange under NSCL. It hasn't decided how many shares to offer or what price to seek. Reuters reports that the filing shows $140.6 million in revenue and a $1.02 billion net loss for the first half of 2026. A year earlier, Nscale reported $10.4 million in revenue and a $368.9 million net loss. Reuters also reports that one customer supplied 52% of Nscale's first-half revenue. Revenue is growing fast, but the company still has a substantial loss and depends heavily on one customer. Contracted future revenue isn't cash Nscale has already earned. The final offering terms and later filings should show how Nscale plans to pay for construction, and whether its contracts produce revenue. Its dependence on a few customers matters as much as its growth. Richmond Fed contacts say A I is slowing hiring, not cutting jobs yet. The Richmond Fed spoke with dozens of businesses from early August through early September, then published its findings on September 18. Most contacts expected staffing to stay flat or rise by year-end. Some firms slowed office hiring while they tested A I. Others hired workers with A I skills. Data-center suppliers reported strong demand, while other firms were more cautious about future sales. These are the views of regional businesses, not a national estimate of jobs lost to A I. Hiring plans also can't tell us how staffing or productivity will change. The useful evidence comes later. Did firms resume hiring once demand became clearer? Did they measure productivity gains? An unfilled job isn't an eliminated job, and workers will feel the difference. That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.