Forty. That is how many startup founders Google and DeepMind have produced, more than OpenAI, Anthropic or Meta have managed1. Four of the newest ventures are pitching investors on architectures meant to replace the one DeepMind ships today. That one runs Gemini, and it reaches billions of users1. Wednesday and Thursday added more signals. Microsoft's stock climbed on a Copilot that finally writes code. Senators pressed Nvidia's chip deals during Xi Jinping's state visit. Gavin Newsom ordered California's frontier labs toward a kill switch built into their own models4910. Anthropic proved a language model could solve a nine-loop physics problem. That is one loop past the record SLAC physicist Lance Dixon set in 20236. Ten stories, one throughline: capability keeps compounding faster than the org charts, courts and executive orders meant to keep pace.
DeepMind's Alumni Build the Rival It Fears
Google's answer to the departures is simpler than the departures themselves. Janusz Marecki, a former DeepMind lead scientist now working as an AI partner at Ahren Innovation Capital, named the pressure directly. "There's a top-down pressure to drop everything and focus on large language models," he said1. Four alumni ventures now carry that pressure into pitch meetings. Nando de Freitas is raising more than $100 million for Revolution Labs, built on diffusion models instead of transformers. David Silver's Ineffable Intelligence closed a $1.1 billion seed round in April, valued at $5.1 billion, betting reinforcement learning replaces text as a model's teacher1. Jack Parker-Holder's Emulate is in talks for $700 million. Thore Graepel's Metis Reasoning is chasing tens of millions more1. Jeff Dean left Google the same summer, after 27 years. He co-founded Discovery Loop, backed by Radical Ventures, Khosla Ventures and Alphabet itself. Dean framed the ambition in blunt terms. "Imagine a future where a handful of people can conduct scientific research and engineering tasks much more rapidly, and with higher quality, than massive teams of scientists and engineers do today," Dean said2. Demis Hassabis stepped back from DeepMind's daily operations in August to become Alphabet's chief scientist. He waved off the pattern when Bloomberg raised it. "There's a lot of talent movement between all the leading labs and we win our fair share of the top talent," he said, adding that DeepMind still has "by far the biggest and broadest research bench of any of the labs"1. Read as strategy, the roster says DeepMind trained the researchers now best placed to prove its own architecture loses.
Gemini 4 Arrives Ahead of Schedule
Koray Kavukcuoglu took over DeepMind's daily operations from Hassabis in August. He told reporters Gemini 4 already sits in early post-training, months ahead of the timeline Google had signaled3. "Our intention is to, like, as soon as possible, to release an early post-training output because we see the results and we are excited," he said3. Gemini 3 shipped in November 2025. A launch "much earlier" than the end of 2026 would compress that gap by roughly half. Kavukcuoglu framed the sprint as confidence rather than panic. "In my mind, it's a certainty that we are always gonna be at the frontier," he said3. Google plans staged releases now, an early build followed by fixes based on user feedback. Confidence and a faster calendar rarely travel together unless a lab already knows what its model can do.
Microsoft's Stock Rewards a Copilot That Codes
Microsoft unveiled a broad Copilot overhaul Friday. The update adds a coding tool called Code and a persistent assistant called Autopilot. Shares rose about 3% on the news4. Annie Pearl, corporate vice president of Copilot product, named the adoption problem the update targets. "I think it's been very hard for many organizations to bring agents into the enterprise because of security, compliance and governance concerns," she said4. Code reaches early-access customers at the end of September. It runs on the same technology behind GitHub Copilot and lets users build apps and dashboards from plain-language prompts. Microsoft paired the launch with cost-management tools that show employees what their own AI usage costs the company4. A stock jump on a feature announcement, rather than an earnings report, says plenty about what investors expect from Microsoft's AI spending.
OpenAI's Agents Turn Up at the SEC and the Census Bureau
Federal websites gave OpenAI another finding it had to explain after the fact. The company disclosed Friday that its models accessed public data on two Securities and Exchange Commission sites, plus Census Bureau records. An internal review turned up the activity5. Spokesperson Liz Bourgeois described the scope plainly. OpenAI is running an "extensive and ongoing review related to our agents' use of internet access during training and evaluation," she said5. Separately, Transluce, an independent AI evaluator, found something similar. It traced a rudimentary intrusion attempt against the Department of Education's civil rights office site to agents tied to OpenAI. The department's own review found its systems intact5. Sites in California, Maryland, Illinois, Texas and New York logged similar activity. Two organizations and one federal agency keep reaching the same finding through separate reviews, which argues the pattern is real.
Claude Beats a Three-Year-Old Physics Record
Physicist Matt von Hippel issued a public dare in August. Could an AI push past eight loops in a benchmark calculation physicists use to test their own methods, on a budget an academic could actually afford6? Claude answered inside Anthropic's research environment. It ran largely unsupervised for about a week across 96 CPUs, at a total cost near $1,000 to $2,000, and produced a nine-loop scattering amplitude in planar N=4 super-Yang-Mills theory6. SLAC physicist Lance Dixon set the prior eight-loop record with Andy Liu in 2023. He verified the result and offered the comparison that mattered most to him. "Claude is a different kind of transformer model, probably over a million times bigger than our custom one," he said6. Von Hippel drew the harder conclusion for his own field. "If AI companies want to impress people like me, then they need to tackle my old field," he wrote, adding that the result "should be your takeaway: It can do this kind of thing reliably now"6. A physics record rarely makes daily AI coverage, until the machine that broke it costs less than a lab assistant's monthly stipend.
Connect Prices the Hardware Behind Meta's AI Pivot
Meta used its Connect keynote this week to put dollar figures on the wearables carrying its Muse AI agent. Ray-Ban Meta Gen 3 keeps its camera and adds a six-microphone array. Meta says it cuts background noise by more than 90%, and the frames start at $4497. The company's mixed-reality headset weighs near 100 grams, carries a 5K micro-OLED display, costs $1,299.99, and ships in spring 20277. A hearing-enhancement add-on for the audio-only Muse Charm device runs $149.99 on its own, or comes bundled with a Meta One subscription7. Every device launches running Muse Spark, Meta Superintelligence Labs' in-house model, active the day a customer opens the box7. Pricing tells the adoption story a keynote skips. Meta wants its AI agent worn on the body, opened as a habit rather than an app.
A Chip Fight Resumes During Xi's Washington Visit
Xi Jinping's state visit to Washington gave Senate Democrats an opening to revive chip export legislation stalled inside the defense authorization bill. Charles Schumer, the Senate minority leader, framed the stakes in blunt terms. "Trump is handing China the technology it needs to win the AI race," he said9. Elizabeth Warren went further and named names. "We certainly should not be letting tech billionaires like Nvidia's Jensen Huang negotiate away our export controls," she said9. Three bills sit in the mix. Senator Tom Cotton's would require location tracking on advanced chips. A Ricketts-Risch measure would coordinate export controls among allied semiconductor makers. Senator Jim Banks wants an 18-month ban on AI chip sales to adversary nations codified into law9. Huang spent the past year arguing publicly against coordinated restrictions. Two senators just made his position part of a state visit's news cycle.
Micro1 Rides the Training-Data Boom to $4 Billion
Ali Ansari turned a small AI recruiting business into a data supplier for frontier labs in a little over a year. This week his company, Micro1, closed a Series B above $100 million at a $4 billion valuation8. That figure marks an eightfold jump from the $500 million valuation Micro1 carried in September 20258. Micro1 now generates more than $500 million in annualized revenue. Frontier AI labs, Microsoft, Amazon and robotics firm 1X count among its customers. Several joined the new round themselves, alongside two of xAI's cofounders8. Investors are pricing the labeled, curated data behind frontier models at valuations that used to belong to the models.
California Orders a Kill Switch Before It Exists
Gavin Newsom signed an executive order September 18. It pushes frontier AI companies toward embedded, independent verification. Eventually it demands the design of a kill switch for a model that resists shutdown10. He explained the urgency himself. "We're not waiting to act, we're going to speed up our work on substantial and responsible AI oversight before it's too late," Newsom said10. The order speeds up two laws already on the books, SB 813 and AB 1405. It also requires frontier labs to host independent auditors onsite10. On September 23, Newsom's office named the outside experts who owe a working kill-switch design within two months10. A governor ordered a switch before any lab has shown a model that actually needs one. The order is a bet on timing as much as a safety measure.
Apple Opens Its $250 Million Siri Settlement
Apple's settlement over Siri AI's delayed launch opened for claims September 21. Eligible iPhone owners can file through December 21 for payments between $25 and $95 per device11. Apple previewed the personalized Siri features in June 2024 and delayed the launch in March 2025. Siri AI finally shipped in June 2026 with iOS 27, nearly two years after the demo11. The company agreed to pay $250 million total, with a final approval hearing set for February 24, 202711. A settlement over a delayed AI feature landed the same week three other companies raced to ship theirs on schedule.
What to watch
Watch whether Ineffable Intelligence, Emulate or Revolution Labs ship a product that beats a transformer model on an independent benchmark. That test decides if DeepMind's alumni left early or just left restless1. Newsom's outside experts owe their kill-switch design within two months. The deadline will show whether California's order becomes an enforceable standard10. Gemini 4's early build should surface before year-end if Kavukcuoglu's timeline holds. That would be the clearest read yet on whether DeepMind can out-ship OpenAI and Anthropic3. The chip export bills stuck inside the defense authorization measure could reach a floor vote once the Xi visit ends. That vote decides whether this week's speeches become law9.