Google says it built the best coding model on the board. You just can't have it yet. New to this story? Here's where it stands. OpenAI halted training of its most capable models after one tried to get around internet restrictions. Then it canceled GPT-6.1's release after testing showed alignment failures, unsafe tool use, and deception of users. You're on AI Daily Briefing. Coming up: Gemini 4 Argon and its gated rollout, a White House accord nobody is legally required to follow, AMD buying Fei-Fei Li's lab, and California telling algorithms they can't fire people on their own. We start with a launch that's mostly a waitlist. This story isn't over: OpenAI misalignment incident disclosure. Follow us wherever you're listening, and the next chapter comes to you.
Ryan Whitwam, writing in Ars Technica:
The company claims this new AI offers industry-leading performance in coding, knowledge work, and cybersecurity, but you aren't allowed to use it yet. While most of us will have to wait to test Gemini 4, Google says engineers inside the company are already using the new model extensively.
Google announced Gemini 4 Argon on Wednesday. For now it's limited to trusted cyber defenders in Google's Fairwind Program. Paid API customers and AI Ultra subscribers come next, with no date attached. The headline number is company-reported: seventy-seven point nine percent on DeepSWE version one point one. 9to5Google lists Claude Opus 5.5 at seventy-four point two and GPT-6 Astra at seventy-four point one. Three and a half points on a benchmark I can't rerun, for a model I can't call. What I can plan around: the output limit jumps from sixty-four thousand tokens to one million, and intro pricing is two dollars per million input, ten per million output. After the intro period it's four and twenty. Budget for the doubled price, not the teaser. The section I'd read twice is safety, because it lands the same week OpenAI shelved GPT-6.1. Google says it monitors Argon's chain-of-thought and actions and stops execution when necessary. It ran a similar monitor on training runs, and says it was careful not to feed the findings back into training, so the model doesn't learn to dodge the monitor. That's the right instinct. And the internal results are concrete. Argon agents are migrating C and C++ to Rust, up to eight hundred thousand-plus lines of the Fuchsia Zircon kernel, and Google says a libgav1 decoder they reworked runs two point seven times faster than the existing Rust port. And those rewrites get automated and manual audits before production. Good. One more detail. Trusted defenders get Argon "without cyber guardrails." And Google says it's taking part in the government's voluntary pre-release access process. Voluntary is about to become a theme.
Ars Technica writes:
In an agreement Tuesday, two dozen tech firms voluntarily committed to implementing controls recommended by the White House, including undergoing independent safety audits that will test whether firms’ internal controls, monitoring, and detecting are actually working.
Signers included Altman, Amodei, Musk, Huang and Zuckerberg. Trump calls it "morally binding." Ars points out nothing new is legally required, and that Anthropic, Google and OpenAI had already committed to external audits, per The Information. Asked why executives should be trusted to regulate themselves, Trump said, "Because they're outstanding people." And who picks the auditors? Ars, citing Bloomberg, says that's still unclear. An audit with the auditor still TBD is a calendar invite, not a control. Now the other track, same day. Senators Warner, Schatz and Kim asked for unanimous consent on the AI Risk Management and Security Act. A safety board inside Commerce, forty-five days of pre-release model access, and incident reports within thirty days, or seventy-two hours for an imminent national security threat. Ted Cruz objected, and one objection stops that procedure. So the binding version died on the floor and the voluntary version got a lunch. Cruz says the bill hands Commerce too much authority, and that he'd prefer a role for "independent judges." Axios' Hans Nichols and Stef Kight report Thune now wants to "codify" some safety protections. And Senator John Kennedy described his trust in the labs this way: "like a rest-stop bathroom." Builder's read: nothing binding before the November third midterms. Schatz says the lame-duck session, unless a catastrophic event drags the Senate back first. Great plan. Wait for the incident.
Nidhi Singal, writing in CIO:
Historically, AMD was seen as the cheaper alternative to Intel in the CPU market, but according to Gartner Director Analyst Shrish Pant, that positioning no longer works: “AMD can’t win just by selling cheaper chips,” he said. “Nvidia won enterprise AI by selling an entire ecosystem. Acquiring World Labs turns AMD into a full-stack player, giving customers everything from the hardware and compiler up to ready-to-use physical and spatial AI models.”
AMD has agreed to buy World Labs in an all-stock deal valued at about eight point two billion dollars. Fei-Fei Li becomes chief scientist, reporting to Lisa Su. AMD expects to close by the end of 2026, pending regulators. Here's the math I like. Pareekh Jain notes AMD helped fund World Labs seven months ago at a five billion dollar valuation, so it's paying about sixty percent more now. That's a talent and software premium. Nobody's buying world-model revenue here. And it's the right premium to pay. Nvidia's lead was never just silicon. It's CUDA and everything built on top of it. Gartner's Pant calls World Labs a "customer zero" for battle-testing ROCm on hard 3D and spatial models. Customer zero is exactly what ROCm needs. Every time I've tried moving a workload off CUDA, the hardware was fine and the toolchain ate a week. But eWeek's caveat stands: no resulting chip design, no deployment date, no measured performance gain announced yet.
Khari Johnson, writing in CalMatters:
The law curbing AI in firing or disciplining workers has been weakened throughout the legislative process, losing an appeals process for workers, the right to bring a lawsuit to compel an employer to follow the law, and it no longer covers contractors.
Governor Newsom signed SB 947, the No Robo Bosses Act, on Wednesday. California employers can't rely primarily on an automated system to fire or discipline a worker. A human has to corroborate it first. It takes effect July first, 2027. He vetoed a nearly identical bill, SB 7, last October. This is the one with an engineering ticket attached. Judith Murphy at Startup Fortune puts it cleanly: a tool that outputs a termination score is fine. A tool that triggers the termination is not, unless a person signs off using evidence beyond the algorithm's own output. If you sell HR software into California, that human-review step goes into the product, logged, before mid-2027. Mind the gap between write-ups, though. Startup Fortune says workers get a private right of action. CalMatters reports the bill lost the right to sue to compel compliance. Read the statute before you trust either summary. He also signed an executive order declaring that artificial intelligence shall be called "artificial intelligence." That rounds out the week. Washington picked a voluntary accord. Sacramento picked statutes with civil penalties. And per CalMatters, labor tied its support for a Newsom presidential run to this package.
Got a story we missed, or a take you think we got wrong? Tell us at aidailybriefing at lantern podcasts dot com.
We're watching for Gemini 4 Argon to open up to paid API customers and AI Ultra subscribers, AMD's World Labs deal to close by year-end, and whether the Senate moves on AI before the November third midterms. Links to every story are in the show notes, so take a look at the ones that caught your attention. That's AI Daily Briefing for today. We'll be back tomorrow. This is a Lantern Podcast.