Nvidia's CEO falls in a spy scandal. Can anyone track where AI chips go?

Remember Sunday's token about a US-approved Nvidia chip buyer in China with an empty office? The chip question now gets sharper. MIT Technology Review reports that Nvidia representatives told Rep. Bill Foster it would be technically possible to track chips, as the Chip Security Act asks, to fight smuggling.

At the same time, an industry group publicly called those measures "potentially infeasible." So, possible in private, maybe impossible in public. Separately, a post titled "Chinese frontier AI models carry the fingerprint of export-controlled GPUs" got 296 points on Hacker News. Export-controlled means the US forbids selling those chips to China. Our sources do not give the method, so treat it as a claim for now.

With Huang gone and Congress investigating Nvidia's China sales, these questions will likely come up on Wednesday.

Yesterday we only had a headline: Anthropic reported that a key safety alarm was triggered, days before its IPO. Now we know more. In its amended filing on Tuesday, Anthropic said it met the "research acceleration threshold" of its Responsible Scaling Policy, its own set of safety rules.

This is the point where AI help is expected to roughly double the pace of AI progress. Anthropic stresses the judgment is about expectations for future models and the overall pace, not proven abilities of today's systems. It says it is still looking at the data and will publish more reports.

Meanwhile, the IPO website is live. It got 931 points and 866 comments on Hacker News. An IPO is when a company sells shares to the public on the stock market.

Two days ago, Gemini 4 Carbon opened to everyone after a careful start. Now Google adds Gemini 4 Hydrogen, the cheapest member of the Gemini 4 family. It powers Search's AI Mode and the Gemini app starting today. Google says it handles complex questions better: many parts, typos, very specific requests.

Ars Technica notes Hydrogen is not getting the slow, security-conscious release that Carbon and Argon got. Simon Willison finds the pricing confusing: Hydrogen is not far from its bigger siblings, and Carbon and Argon double in price on December 1, from $2 to $4 per million input tokens. A token is a small piece of text. Google also launched an Orchestrator in its Antigravity coding tool, and a Decisions API that answers up to 1,000 yes/no questions per call.

Not everyone wants more Gemini. Ars reports families are calling for AI moratoriums in schools after a Gemini pilot, and a teen's lawsuit calls Gemini a "suicide coach." Chicago schools told Google they would rather lose Search than be forced to accept AI.

This week we covered Claude Code leaking secrets through its 'Don't Ask Again' option. Now a user reports the opposite problem: asking was on, and it still did not ask. They filed it on GitHub last Friday. This is one person's report, not confirmed.

They used Claude Code on Windows in manual approval mode, with a rule that PowerShell commands must stop and ask. They say they never saw a prompt. The agent ran 9 PowerShell commands that moved 11GB of files, and the log shows approvals they say they never gave.

The same day, a forum post claiming "Humble Bundle just killed most of their security through vibecoding" got 382 points on Hacker News. Vibecoding means letting AI write code with little checking. Our sources give no details. United Airlines now cites cybersecurity risks to block AI agents from booking flights. And Anthropic's new Claude Meetings joins Google Meet calls to take notes, on by default for Team accounts.

Yesterday we told you about a blog post claiming AI agents found an unmapped Maya site, with a big question mark. Now we have the answer. The Verge reports the site was already discovered and published in an academic journal last year. The claim, by Google researcher Pietro Miotti, drew criticism from archaeologists online.

A second study shows the same trust problem in daily life. IP protection firm Podqi asked ChatGPT, Gemini and Perplexity shopping questions. The Verge reports the chatbots sometimes pointed to imposter and other risky websites.

AI can search fast. Checking is still a human job.

Local AI means the model runs on your own computer, not in a company's cloud. Your data stays at home. Two projects push it further today.

Edge0-145B is an open-source model with 145 billion parameters, built to run on your own computer. Parameters are the numbers that make up a model. It got 248 points on Hacker News. Smolmaxxing, a free GitHub project, tries to make LLMs work on small and old devices. It got 365 points.

This matters more this week. With Nvidia's open-source support now in question, a Reddit user advises backing up your favorite models and datasets, just in case. My old laptop just looked at me with hope.