Busy Friday. California went after frontier models with an executive order, OpenAI started letting brands run their own chatbots inside ChatGPT ads, and Mozilla quietly picked a French company to power its browser assistant instead of the obvious American options. Three very different bets on where this all goes.
California Is Studying an AI Kill Switch, Not Requiring One
Governor Gavin Newsom signed an executive order Friday that convenes a group of outside experts and state officials and gives them until November 16 to recommend changes to California law. That is the whole timeline. Two months.
The headline item is a kill switch, an emergency shutoff for frontier models. Read the order and it does not actually require anyone to build one. It asks the working group to recommend how you would require it. Same deal with the other proposals on the table: having independent third parties write safety plans for frontier AI companies instead of letting the companies write their own, requiring outside verification of the safety frameworks those companies already publish, and expanding the list of safety incidents that have to get reported to the state to cover loss-of-control events.
Context matters here. Newsom already signed SB 53 in 2025, the Transparency in Frontier Artificial Intelligence Act, which was the first state AI safety law in the country. It makes frontier developers publish their safety frameworks, report critical safety incidents to the state, and protect whistleblowers. This order builds on that and speeds up the implementation. His quote was that the federal government's failure to create meaningful AI oversight should alarm every American.
Here is where I land on it. The timing is not subtle. Newsom is a presidential hopeful, an Anthropic researcher resigned earlier this month with a public warning about where the technology is headed, and Washington is still sitting on its hands. But there is a real problem underneath the politics, which is that the president signed an executive order back in December 2025 blocking states from enforcing their own AI rules. So California can convene all the experts it wants and still get preempted before a single kill switch gets built. What I want to know is whether anybody has a plan for that fight, because right now this reads like a well-researched memo headed straight into a wall.
Source: Office of the Governor of California
OpenAI Is Letting Brands Put Their Own AI Inside Your Chat
OpenAI started testing something called Sponsored Agents this week. You see an ad in ChatGPT, you click it, and instead of getting dumped on a landing page you start a labeled conversation with a bot the advertiser is paying for. You can ask it questions, tell it what you actually care about, and follow a link to the company's site when you are ready. OpenAI is stressing that the sponsored conversation is separate from your real ChatGPT thread and separate from ChatGPT's own answers.
It is live with a handful of US advertisers right now. Alongside it OpenAI shipped tools so businesses can write ad prompts inside ChatGPT Work, added AI creative tools to Ads Manager, and built integrations that put ChatGPT inside HubSpot and Shopify.
The number that explains all of this: OpenAI said its ChatGPT Ads business hit a one billion dollar annualized revenue run rate by August 31, with tens of thousands of advertisers. That is from a business that only started testing ads in free and Go tiers earlier this year.
What I keep coming back to is the labeling. OpenAI is drawing a hard line between the sponsored agent and the assistant, and today that line is real. The question is what it looks like in eighteen months when a billion dollars has turned into five. Every ad-supported product in history has started with a clean separation and then slowly sanded it down. I am not saying they will. I am saying the incentive is now permanently pointed that direction, and good intentions have a poor track record against a revenue line that big.
Source: OpenAI
Mozilla Picked Mistral, Not OpenAI, for Firefox
Mozilla and Mistral announced a partnership on September 16 putting Mistral Small 4 behind Smart Window, the AI browsing beta Firefox shipped in August. Smart Window summarizes complicated searches with sources attached, helps you dig up a page you clicked away from, and can work off the tabs you already have open.
The model is live for users in the US and Canada, and the feature itself is launching in France with French language support. The UK and Germany are next, later this year. Mozilla says it does not keep chat transcripts and does not use your conversations to train anything, and Mistral is matching that with no data retention on its side.
I think this is the most interesting of the three and it got the least attention. Mozilla could have picked OpenAI or Google and nobody would have blinked. Instead they went with a European lab whose whole pitch is privacy and multilingual performance, and they put it in a browser whose only remaining argument against Chrome is that it does not spy on you. That is a coherent product decision, which is not something I get to say about Mozilla very often. Whether Mistral Small 4 is actually good enough to keep anybody from switching back is the part nobody has answered yet, and that is the only part that matters.
Source: The Mozilla Blog