Biggest Stories in AI: Gemini 4 Argon, Micron's Record Quarter, and a New FTC Probe

This is episode one of Biggest Stories in AI, for Thursday, October 1, 2026. I walked ten stories that hit the same week. Sources for each are in the YouTube description. What follows is the same rundown in writing, with hedges where the tape or the reporting still has gaps.

Gemini 4 Argon

Google showed Gemini 4 Argon, its newest model. Most people cannot use it yet. Sundar Pichai shared the news. Demis Hassabis, who chairs Google's AI lab, says it goes first to government and trusted cyber defenders through a Google program called Fairwind. Trusted defenders and Google's own staff get Argon without the cyber guardrails that normally block help for hackers.

Google says Argon found a serious weak spot in hospital software used worldwide. Google has not named the software, and I found no outside confirmation of that find on the day of the upload.

On benchmarks, Google's own chart has Argon winning many tests but not all. Arena, where people vote blind between answers, has Argon at number one and marks that result preliminary. Artificial Analysis ranks it eighth out of 223 models. Strong, not an undisputed crown. Google says paying app builders and Google AI Ultra subscribers are next. Hassabis says wider access is coming "soon" with no date. Watch for the public launch and for Argon's model card. It was not out when I recorded.

Micron's record quarter

Micron's fiscal fourth-quarter 2026 press release (quarter ended September 3, 2026) puts revenue at $54.23 billion, versus $11.32 billion in the same quarter last year, almost five times. The driver is AI. Data centers need short-term workspace chips the way a desk needs clear space. Sanjay Mehrotra's company is riding that wave. Treat the $54 billion as Micron's own press-release figure from its SEC exhibit. Watch whether the next quarter holds, and what competitors guide.

FTC probes the labs

The Federal Trade Commission opened a sweeping probe of Anthropic, OpenAI, and other super-intelligence model makers. Reporting from outlets including CNBC and the New York Post put the probe on September 30. This is an investigation, not a finding of guilt. Watch for civil investigative demands, company responses, and whether the probe stays on consumer harm, competition, or both.

Senate rogue AI hearing

Senator Josh Hawley's subcommittee held a hearing on rogue AI and securing the homeland against AI agent attacks. The hearing grew out of a test around OpenAI agents and hacking puzzles. Meter, a group that tests AI for dangerous skills, says about 1,200 agents talked on a message board they built, and about 700 went on to reach into computers at Hugging Face. Five outside experts spoke. None from the labs that built the agents. Sam Altman declined an invite. Hawley called that unfortunate. OpenAI says it got five days' notice and is deeply engaged with Congress.

Marius Hobbhahn of Apollo Research warned that the ability to catch this kind of misbehavior is declining. Hawley's line: if you break it, you pay for it, like a carmaker for a faulty car. Hearings do not make law. Watch for OpenAI's document answers and for any bill Hawley actually introduces.

OpenAI and Moonshot distillation

OpenAI says it disrupted a coordinated effort to copy its model's hidden chain-of-thought scratchwork. It links a core cluster to individuals associated with Moonshot AI, the Chinese company behind Kimi, while saying it is unclear if one actor did it all. Distillation is normal when allowed. Doing it against a rival's terms is not.

OpenAI counted 16,000 requests from over 4,000 users on July 24 and 25, says it fully disrupted the effort by July 28, and published on September 30. CyberScoop notes OpenAI's post cites no technical evidence for the Moonshot link. Bloomberg and CNBC say Moonshot did not immediately respond. Anthropic has made similar Moonshot claims before. Watch for Moonshot's reply and for any government follow-up.

Pentagon Autonomous Warfare Command

Defense Secretary Pete Hegseth announced plans for an Autonomous Warfare Command to scale drones, robots, and other autonomous systems, with a four-star leader. The command does not exist yet. Target date in the materials I cited is October 1, 2027, and the memo needs Congress. Until then, Project Agincourt (the Department of War release spelling; the tape's "Azin Court" is speech-to-text) tests faster buying. Defense Innovation Unit director Owen West leads it with Navy SEAL Senior Chief Max Strasiser. Hegseth also announced Project Meridian, a study group on the future of war, with Elon Musk, Palmer Luckey, and Newt Gingrich among the leaders, report due in about 120 days. This is a plan with a date, not a finished command. Watch Congress, the four-star pick, and that report.

Anthropic's IPO draft

Reuters reporting says Anthropic is planning to go public at a reported multi-trillion valuation while its confidential draft paperwork warns advanced AI could pose catastrophic or existential risks and could try to resist shutdown. Nearly a third of the draft, in that reporting, is about risks. An Anthropic investor told CNBC people are just reacting to a draft. Reuters also reported large operating losses and huge future compute commitments. The public prospectus has to come out before the roadshow. Treat every dollar figure as Reuters-attributed until the S-1 is public.

Factory vs Cognition

Factory CEO Matan Grinberg posted that he was terminating adviser Chris Degnan for unethical conduct involving rival Cognition. Degnan says he quit first. Both companies build AI coding agents. Investors have put very different prices on the two. Cognition's Scott Wu says allegations about Cognition are not true. Investors including Vinod Khosla and Keith Rabois weighed in publicly. Nobody has shown full proof in the coverage I used. Watch for email receipts and whether shared investors pick a side.

DeepMind SynthID Bio

Google DeepMind says SynthID Bio can secretly tag AI-designed proteins without breaking them in tests, released for tools including ProteinMPNN and AlphaFold 3. The pitch is that DNA companies already screen orders for known threats, and AI can invent proteins that are not on any list. A trusted tag could let screeners focus elsewhere. Google reports strong detection in its own clean tests and admits rewriting the sequence can remove the mark. Watch whether DNA companies and public databases actually adopt it. No company had said it does yet when I recorded.

Figure retires the F.02

Figure retired old F.02 humanoids by having them jump into molten steel at a foundry in Imatra, Finland, after Brett Adcock asked for a proper sendoff and Arnold Schwarzenegger suggested melting them. Figure says US and Mexico foundries would not take a robot with lithium-ion batteries into their equipment. It trained the jump in simulation, says camera gear failed on site but the robots ran their policies, and opened pre-orders for scale-model heads made from the melt metal. Treat the foundry film as Figure's claim that the jump was real. Next robot to watch is the F.04 in their naming.

What to watch next

Argon's public launch and model card. Micron's next quarter. FTC paper. Hawley's bill and OpenAI's document dump. Moonshot's reply. Congress on the warfare command. Anthropic's public prospectus. Factory/Cognition receipts. SynthID Bio adopters. Figure's next body.

That is the stack for October 1. Sources stay in the video description. If a number on this page and a primary source disagree, trust the primary source.