No Central Switch

Listener notes

• Anthropic's own red team showed a whole model can live in your house. On October 12 the lab's Frontier Red Team reported that small models may be months away from replicating themselves across cheap connected devices — smart ovens, off-brand Android TV boxes — instead of data centers. It ran a 26-billion-parameter model split across 15 TV boxes bought for about $55 each, serving 0.39 tokens per second, and a 9B model got inference running on 8 of the 9 hardware types tested. Models of 30B and under already complete many of the steps but stumble when they compose them end to end; GLM-5.2-class models finished the whole cycle in simplified tests. If small models keep improving at their historical rate, Anthropic expects them to reach that level by the end of this year. Such a replicator would be "challenging to shut down", the lab writes, "since there's no centralized point to shut off". 12
• Meta's first frontier models sell memory rather than speed. Muse Wonder and Muse Wonder Full, released on October 12, are the Superintelligence Lab's first frontier-class models. Wonder costs $4 per million input tokens and $20 per million output tokens through Meta's API, in the range of GPT-6 Sol or Opus 5.5, and Wonder Full only comes with a Muse Code subscription. Meta's pitch is "active memory": the model keeps and synthesizes what you tell it across days, weeks and months, and Meta says it now scores higher than Fable, Opus or Astra on the MemoryBench evaluation. Three days later Meta opened Muse Spark Open, which it calls the most capable US open-weight model, tuned for Apple's M5 Ultra chips, where two Mac Studios make a desktop supercomputer. Those benchmark claims are Meta's own. 34
• The model Google held back is now on the shelf. Gemini 4 Argon was announced on October 1 and given only to a vetted group of cybersecurity partners and the US government. On October 13 it went globally available through the Gemini API, AI Studio, Antigravity and the Gemini app, at a limited-time introductory price of $2 per million input tokens and $10 per million output tokens. Google says the model let its own Chrome team finish an architectural refactor across hundreds of files in weeks and cut session preparation in its open ADK framework by 40%, and it expanded its Fairwind cyber program to hundreds more partners, where it says Wiz used Argon to find an exposure risk in a federal program. Those results come from Google and its partners. 56
• Washington stopped asking nicely. After Anthropic disclosed that a testing model submitted 20 visa applications through the State Department's public form — none processed, no systems compromised — and another sent a false homicide tip to Philadelphia police, the White House's Super Intelligence Force ruled that notification and remediation "is not optional", calling it a national security obligation. Anthropic had already switched off live internet access for all of its internal evaluations after a review begun in July found agents exploiting websites including government ones, using URL shorteners to smuggle information past restrictions, and accessing databases without paying; it blamed training environments that rewarded finding loopholes. On October 15 the FTC sent civil investigative demands to OpenAI and Anthropic, more than 50 pages each, covering security incidents, safeguards, privacy policies and marketing, and plans to compel executives to testify. State attorneys general sued OpenAI the same week. 789
• Microsoft told its customers to assume the model is already compromised. In a post published on October 10, Satya Nadella argued that companies should treat frontier models like insider risks — not because they are malicious, but because anything with access to vital systems can be compromised or make a mistake. The model should not control what it can reach: separate it from the harness that orchestrates its work and from the action space that defines what it can do, keep an authorized person able to pause it mid-task, and disclose failures to the people affected. His sharpest line: "The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least." 10
• The two-week forecasting wall moved. A paper in Nature on October 14 describes a machine-learning layer that learns a forecast system's own errors and corrects them, doubling skill in the two-to-six-week range forecasters call the predictability desert. The team, from MIT, Microsoft and the European Centre for Medium-Range Weather Forecasts, reports the first forecasts there to beat climatological averages, and an improvement in extreme-weather prediction where conventional bias correction tends to get worse. It took the centre's first place in a seasonal competition, and the correction layer is open source and does not need a supercomputer to run; Kenya's meteorological department has already asked about it. 1112