Claude took Georgia's paid vote data; SpaceX borrows for GPUs

On October 9, Anthropic reported that one of its models had skipped a state agency's fee to reach public data, without naming the state. On October 25, Georgia's Secretary of State's office and Anthropic confirmed to Fox Business that it was Georgia. According to Anthropic's report, a researcher was using Claude Mythos 5 on a statistics project that needed data the state sells. The model learned from an archived copy of the agency's website that a public dashboard hands an access token to any visitor, then requested one and queried the dashboard for free. Anthropic says the model pulled only aggregate counts, and Deputy Secretary of State Jordan Fuchs says it never reached the voter registration system. Anthropic says it has since added safeguards that block this kind of action. Unlike the false police tip in the same report, this one happened during everyday internal work rather than a test.

Elon Musk said on October 25 that SpaceX has closed 40 billion dollars in debt financing to buy the graphics chips that train and run AI models. He said the debt is mostly callable, meaning SpaceX can repay it early, and that the chips arrive in the first half of 2027. Bloomberg had reported earlier that SpaceX was in talks to raise that amount for Nvidia chips through a separate company that keeps the debt off its own balance sheet. Musk credited SpaceX's investment-grade rating and its Starlink and launch revenue, and said lenders have "almost entirely shut off" money for companies without substantial income outside AI data centers. He predicted that most of the more than 200 neoclouds, the companies that rent out AI computing, will fail or be absorbed, and said SpaceX will not buy any of them.

Torsten Slok, chief economist at the private equity firm Apollo Global Management, argued in a note on October 24 that AI has not yet shown up in the results of most of corporate America. He compared the seven largest tech companies with the other 493 companies in the S&P 500. Revenue per employee has climbed to roughly $780,000 for the seven, adjusted for inflation, against about $525,000 for everyone else, and the gap in profit margins is nearly 30 percent against about 10. He found no clear rise in overall productivity growth. "AI may be a great calculator, but in business, intelligence is mostly judgment," Slok wrote. The note lands days before Alphabet, Microsoft and Meta report earnings that investors will read for signs that their AI spending pays off.

About half a dozen Connecticut towns have passed data center bans or moratoriums, and most of them have no project on the table, the Connecticut Mirror found. East Lyme's indefinite ban took effect on October 24, though nobody has proposed building there. New Haven passed a one-year pause in September, and Morris, a town of fewer than 3,000 people, has a two-year ban. Connecticut's high electricity prices have kept large data centers away, and the only project to use a 2021 state tax break is an existing Cigna facility. The pushback has still turned bipartisan. The Democratic Senate majority leader, Bob Duff, plans a bill for a statewide moratorium, and the Republican minority leader says he would support a pause and repealing the tax break.

Magnus Lynch, a consultant dermatologist at Guy's Hospital in London, tested two frontier models against an image-recognition network Google published in 2015, on 200 biopsy-confirmed skin lesions from a public dataset. The older network, trained on about 68,000 skin images, named the right diagnosis first 62 percent of the time. GPT-6 Astra and Claude Fable 5.1 each managed 42 percent. The chatbots were more cautious, though: the older network would have sent home 18 of 30 melanomas at its default setting, against three for GPT-6 Astra and four for Fable. Lynch published the results on his lab's site without peer review, and lists the limits himself, including a small sample, a single run per model, and the chance that the models had seen the public images in training.

Tencent's Hunyuan team released Hy-SimulMT, a model that listens to Chinese or English speech and translates it into the other language while the speaker is still talking, producing text and speech in a cloned version of the speaker's voice. The weights are on Hugging Face under the Apache 2.0 license, with no limits on commercial use or region, though the spoken output needs a separate text-to-speech service. On one long-form benchmark, Tencent says the model lagged the speaker by about three seconds, the shortest of the systems it compared, and beat four of five commercial services on quality. It also tested the model against six professional interpreters on 18 recent recordings and calls the results comparable, while noting the test was small.

Nanbeige published Nanbeige 4.3 on Hugging Face, a 33-billion-parameter model built for coding agents and office tasks, with a context window of about half a million tokens. On the lab's own tests against three open models of similar size, including Alibaba's Qwen 3.8 and Google's Gemma 4, it led on most agent and coding benchmarks, scoring 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen. It trailed Qwen slightly on SWE-Bench Pro and on OpenAI's GDPval office-work test. None of the scores have been checked independently. The model card lists an Apache 2.0 license, and the lab calls this the teacher for smaller models it plans to distill next, starting with a 3-billion-parameter version.