Three stories worth your coffee this morning. California just signed a pile of AI bills into law, Google shipped a model it will only hand to cyber defenders for now, and Tencent found a perfectly legal way to rent the chips America won't sell it.
01 of 03
California Signs 13 AI Bills in One Afternoon
Governor Gavin Newsom signed 13 AI bills on September 30. Here is the whole list, because you should know what your state is up to:
- AB 1331 (Sade Elhawary): restricts workplace surveillance in bathrooms
- AB 1864 (Marc Berman): safety guidelines for gene synthesis companies
- AB 1883 (Isaac Bryan): bans employer use of AI workplace surveillance tools
- AB 1979 (Mia Bonta): keeps doctors in charge of their own judgment when AI is in the exam room
- AB 2392 (Mike Fong): AI training and procurement standards for public universities
- AB 2713 (Buffy Wicks): stronger transparency rules for AI-generated content
- SB 503 (Akilah Weber Pierson): AI use in healthcare services
- SB 574 (Thomas Umberg): stops lawyers from handing the practice of law to AI
- SB 947 (Jerry McNerney): a human has to review it before AI fires somebody
- SB 951 (Eloise Gómez Reyes): disclosure when AI causes mass layoffs
- SB 1000 (Josh Becker): updates to the AI Transparency Act
- SB 1111 (Angelique Ashby): protection from deepfakes and unauthorized digital replicas
- SB 1159 (Christopher Cabaldon): AI transparency and governance frameworks
The one I keep coming back to is SB 574. Starting January 1, 2027, California lawyers have to verify every case and statutory citation an AI hands them, can't feed confidential client info into a tool unless access is locked down to the attorney and people bound by confidentiality, and have to disclose AI use on documents they file with a court. Arbitrators are flat out barred from delegating any part of a decision to AI. Break the citation rule and you are looking at sanctions, penalties payable to the court, and the other side's attorney's fees, and the firm is on the hook alongside the lawyer.
Newsom's line was "AI should expand opportunity, not come at the expense of workers and families."
Here is where I land. SB 574 is the easiest call on the list. If you bill a client $500 an hour and file a brief full of cases that don't exist, that's not an AI problem, that's malpractice, and now it's written down. SB 947 is the one that will actually change how companies operate, because "the algorithm let you go" just stopped being a legal sentence in the biggest state economy in the country. What I want to know is how "delegate the practice of law" gets defined, because the statute doesn't say, and some poor lawyer is going to be the test case.
Source: Office of Governor Newsom, Holland & Knight on SB 574
02 of 03
Google's Gemini 4 Argon Goes to the Cyber Folks First
Google released Gemini 4 Argon on September 30, and Koray Kavukcuoglu, an SVP at Google DeepMind and Google's Chief AI Architect, wrote the announcement himself. Google is calling it its most powerful model yet, and it's pitched hard at coding and cybersecurity. The headline claim is that it can find, validate, and patch critical software vulnerabilities on its own.
The numbers Google published: 77.9 percent on DeepSWE v1.1 for software engineering, 68 percent on CWE-bench v1 for vulnerability fixes (tied for first), 51.3 percent on AutomationBench (first place), and 91.7 percent on LVBench for long video understanding. Google also says it leads the Vals Index and Gray Swan's prompt injection robustness test, and TechCrunch reports Google claims it beats OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models. Output limit jumps from 64K tokens to 1 million.
You can't just go use it, though. It's rolling out only to "trusted cyber defenders" through Google's Fairwind Program, with Wiz already running it through its Scan for Good initiative. Intro pricing is $2 per million input tokens and $10 per million output, going to $4 and $20 after the intro period.
I like this rollout more than I like the benchmarks. A model that can find and patch holes on its own can also find holes and not patch them, and Google clearly knows it, so the defenders get first crack. That's the right order. But every lab grades its own homework on launch day, so I'll believe "most powerful" once somebody outside Mountain View gets their hands on it. And a million-token output limit is wild. That's not a chatbot, that's a contractor.
03 of 03
Tencent Rents 100,000 Chips From Oracle, and It's Legal
The Financial Times reported October 1 that Tencent signed a five-year, roughly $7 billion deal to lease about 100,000 advanced AI chips across multiple Oracle data centers in Southeast Asia, with about 30 percent paid up front. These are chips Tencent can't buy for use inside China. Neither company has commented publicly, and the reports don't name the exact Nvidia chip or the specific countries.
Here's the loophole. U.S. export rules stop Chinese firms from buying top end Nvidia hardware. They don't stop them from renting compute sitting in somebody else's building overseas. Tencent already pulled this move with a Japanese data center arrangement for Nvidia Blackwell chips back in December 2025, and ByteDance got at Nvidia B200s through Nscale's facility in Norway by way of a Singapore subsidiary. Wall Street liked it fine. Oracle stock was up 1.86 percent premarket to $139.85, though it's still down about 30 percent on the year.
So what are we doing here? Washington built a fence around chips and left the gate wide open on cloud rentals. I don't blame Oracle for taking $7 billion that's legal to take, and I don't blame Tencent for walking through an open door. I blame whoever wrote export controls that care where a chip ships but not who logs into it. Until someone in Congress closes the rental gap, the chip ban is mostly a shipping rule.