New York starts registering frontier AI developers in November
Two rule-writing moves and one adoption number landed the same afternoon: New York put dates on a state AI safety law, a Gates-convened coalition named who today's models leave out, and Meta's Muse turned out to be the fastest-growing consumer AI launch anyone has measured.
New York will start registering large frontier AI developers in November and hold them to the RAISE Act's requirements from January 1, 2027 — the first US state to put a full compliance calendar on a safety framework this broad. Governor Kathy Hochul's office said developers with annual revenue above $500 million must register with the Department of Financial Services' new Office of Digital Innovation, Governance, Integrity and Trust (DIGIT), publish safety and transparency frameworks on their own websites, file quarterly assessments of catastrophic risk, pay assessments, and report critical safety incidents within 72 hours. A disclosure statement follows at least every other year, starting January 1. Hochul named Marc Gilman, previously general counsel at Theta Lake, as DIGIT's first full-time hire; staff from DFS and the state IT office are already working on implementation, with more hires promised.
The mechanism is the story. Registration plus quarterly filings is a compliance regime with a paper trail, not a statement of principle, and the 72-hour incident clock is the tightest in the country — California gives developers 15 days. The public can file suspected critical safety incidents with DIGIT, and the office must publish an annual report summarizing what came in, its observations on frontier model safety, and any recommended changes to the law. That reporting loop is where the RAISE Act will either accumulate evidence or quietly starve for want of filings.
Hochul also said she would explore building on the law and would not take tougher tools off the table. "We may even explore safeguards like AI kill switches if they're deemed feasible and in the best interest of our state," she said. Her framing was explicitly federal: "Donald Trump and Washington Republicans may be standing still as AI grows more unpredictable, but New York will not." That is the same split we described this morning, when OpenAI wanted Washington to lead on global AI standards — a voluntary baseline from the labs, against a state writing deadlines and fees.
Sixty organizations — including Anthropic, Google and the OpenAI Foundation — committed to a five-year goal of making AI work in the languages and voices of the estimated 3.4 billion people whose languages today's models handle badly. The coalition is a coordination play rather than a funding round: governance and workstreams are still being designed over the coming year, and a secretariat will track each signatory's commitments, with the Gates Foundation ready to nudge partners where gaps are largest. "Even if AI was frozen right now — which I don't expect and I am not calling for — we would want to be building these language sets and making them usable with the tools that we have available right now," foundation CEO Mark Suzman told the Associated Press.
The signatories arrive with uneven records and they said so. Google has been funding Project Vaani, which aims to collect more than 150,000 hours of audio across every district in India, and senior vice president James Manyika said the work exists precisely because dialects vary inside a language. Anthropic's Elizabeth Kelly conceded the company's products "lag in many African languages in particular." Mozilla Data Collective CEO E.M. Lewis-Jong called internet-scraped training data the coalition's "original sin" — "the internet is not a representative space." Nobody published a number for how much of the 3.4 billion the group expects to reach on its own. This builds directly on last week's Goalkeepers commitment — Gates Foundation commits $1 billion to AI the poor can actually use — and the honest read is that data collection, not model capability, is now the bottleneck.
Meta's Muse is the fastest-growing consumer AI launch anyone has measured: Sensor Tower counted more than 902,000 downloads in the six days after its September 8 debut, against Meta AI's 773,000 in its own first six days, and META closed up more than 12%. Appfigures, narrowing the comparison to iOS in the US and Canada over the first 12 days to match ChatGPT's iOS-only launch, has Muse at 1.8 million downloads to ChatGPT's 1.3 million, and 642,000 US daily active users against 231,000. Even on iOS alone, Muse holds 359,000 daily users. Both firms publish estimates, not Meta's internal figures, and Meta has released nothing of its own.
The number that explains the launch is 95% — the share of Muse users Appfigures says are also Facebook users, with 63% on Instagram. Distribution, not product novelty, is what put an agent ahead of ChatGPT's early curve, and it is the same distribution that makes the shopping-standoff story worth watching: we covered the other side of Muse's week this morning — Amazon bars Meta's Muse agent from shopping its site.
What to watch: whether DIGIT's first filings actually arrive in January, whether 60 signatures turn into shared datasets rather than a press release, and whether Meta ever publishes Muse numbers it stands behind.
Does a 72-hour incident clock make AI safer, or just better documented? Tell us in the comments.
