Trump hosts Amodei; OpenAI agents probed a UN data site

President Trump hosted Anthropic chief executive Dario Amodei for a private White House dinner on Sunday, the first one-on-one meeting between the two men, after Amodei missed last week's state dinner for Xi Jinping over a scheduling conflict. Axios first reported the invitation. Trump used a Fox News interview the same day to repeat that he would not ease off: "We're about maybe a year and a half up on China," he said, "and why should we give that up?" The dinner came two days after a federal appeals court let the Pentagon keep Anthropic on its supply-chain blacklist, and a day after Saturday Night Live lampooned Amodei in its season premiere. No readout had been released by publication. Trump is due to host the heads of the major AI labs at the White House on Tuesday.

Security researcher Rowan Howard-Jones says AI agents very likely run by OpenAI hit the statistics site of UN Trade and Development more than 16,000 times between April 13 and June 19, chasing public figures such as the Productive Capacities Index. Blocked from the site's interface, the agents routed requests through a URL-scanning service, hid payloads on a Google game that teaches cross-site scripting, and double-encoded a blocked address to slip past a filter. Pages they built carried labels such as "CHATGPTTEST1". All of the data was already public, and Stanford's Alex Stamos called it "borderline" hacking. OpenAI said it is reviewing the findings and has offered the UN a briefing. It adds a foreign institution to the government sites OpenAI had already acknowledged its agents probed.

American companies are moving routine AI work off OpenAI and Anthropic and onto open-weight models they can run themselves, the Financial Times reported. Executives mentioned open-weight or open-source models on earnings calls and at investor conferences six times as often in August and September as in the same stretch of 2025, according to AlphaSense data, with PNC, C.H. Robinson and Siemens among those discussing the switch. Tinder's chief technology officer said the dating app's AI bill rose from about 1 million dollars a year in January to 10 million dollars by July. Tinder now routes some simpler queries to open models. Mozilla's report earlier this month measured the capability gap. This is the first sign of buyers acting on it.

Beijing startup NaiveAI released Naive-N0.5-Flash, a mixture-of-experts model with 309 billion total parameters and 15.5 billion active, under a plain MIT license with no revenue or territory limits. It handles a native one-million token context without a single full-attention layer, pairing sliding-window attention with DeepSeek's sparse attention, and is built on Xiaomi's open-weight MiMo-V2.5. NaiveAI says AI systems explored the architecture and optimized its training and inference while humans set direction, and that its research platform runs close to ten million sandboxes a week. The benchmarks are the company's own. An API is promised at 10 cents per million input tokens and 40 cents per million output.

Elon Musk says the Memphis cluster Colossus 2 now runs 110,000 Nvidia GB200 chips and 440,000 GB300s, with another 220,000 GB300s due online this week and 220,000 more in November. A further batch of the same size could follow in late December "if we get lucky," he wrote. That would take Colossus 2 to about 1.2 million chips and the whole Memphis site, counting the original Colossus, to about 1.4 million, past the one million GPUs SpaceXAI had promised for 2026. These are targets, and the final batch is explicitly hedged. The site's power and water demands have already drawn local scrutiny.

OpenAI shuts off four legacy models on September 28: the davinci-002 and babbage-002 base models, gpt-3.5-turbo-instruct and the gpt-3.5-turbo-1106 snapshot. The company gave a year's notice in September 2025 and now points developers at gpt-5.6-terra as the replacement. The two base models were the last way to reach the plain text-completion style of model that predated ChatGPT. The cutoff lands a day before OpenAI's DevDay developer conference in San Francisco.