AI roundup: Expectations come down, censorship and data harvesting move up

Money looms large in the AI world, but this week the tone has become a bit more subdued. Both OpenAI and Anthropic are busy raising capital, and both have to adjust expectations.

At the same time, questions about privacy and control are surfacing in new places: in a headband, in a Chinese chatbot and in the mailbox of a water utility. Meanwhile, AI for everyday use and research is quietly getting better. Here are this week's threads.

Dreams of value are shrinking

Anthropic, the company behind Claude, wants to go public before Thanksgiving, according to investors. The stock market is where ordinary people can buy shares in a company. Tech stocks have fallen, and the company's valuation looks set to land below the $2 trillion some had hoped for. Anthropic wants to go public in November, but must lower expectations OpenAI has closed a first part of a new round of about $11 billion at a valuation of $1.25 trillion. That is below the original target, and the whole round could end up at about $20 billion rather than up to $30 billion. OpenAI raises $11 billion, but must lower expectations These are still staggering sums, but a sign that investors will not pay just anything.

Who gets to use your data?

A meditation headband, Muse, has been used to train an AI model on users' brain data without the privacy policy mentioning it, according to a study in the journal Science. The policy has since been changed. Your headband may have trained AI on your brain waves Danish utilities and educational institutions have received an unusual number of identically worded freedom of information requests, that is, demands to see public documents. The Ministry of Education sees signs of an AI bot, and the matter has been reported to the police. Mysterious requests fill the mailboxes of water utilities and district heating companies And Meta has put new terms on hold after publishers feared they gave Meta the right to pull content from their sites. Meta puts new scraping terms on hold after publisher complaints

When AI answers in the spirit of the state

Alibaba's Qwen model has answered evasively or denied the facts about Tiananmen Square in 1989, according to several tests. Alibaba will look into the matter internally. Chinese AI model dodges questions about Tiananmen Square Nous Research has made an adapted version that, according to its own test, less often follows the Chinese line, but the figures come only from Nous itself, and no independent party has checked them. New AI model is meant to remove the Chinese line, but the figures come only from the developer

Everyday tools are moving

If you have an Android phone, it may be interesting that a writer at Android Police, after a month, names Gemini a clear favorite over Claude and Grok. It is, however, one person's experience. One-month test on Android: Gemini beat Claude and Grok as a mobile assistant Gemini can now also create and edit PDFs directly in the chat, but a test shows clear limitations. Gemini can edit PDFs in the chat, but the test runs into problems For businesses, Anthropic has made Smart Reports available in Claude Enterprise, so you can see what your team's usage goes toward and where there is waste. Claude can now show what a company's AI spending goes toward

From smartwatches to cells

Google is researching whether an ordinary smartwatch can give signs of elevated blood pressure without a cuff. It is not yet a diagnosis. Google tests whether your smartwatch can catch high blood pressure Anthropic's biology team has, with help from Claude, determined the first 3D structure of a vault particle from bacteria, but the AI also made mistakes that cost time. Claude helped map a mysterious cell particle in bacteria And the evaluation organization METR estimates that Anthropic is probably improving its models on well-defined tasks about twice as fast as before. Anthropic has, however, redacted much of the evidence, so others cannot verify it. METR: AI probably makes Anthropic's models progress twice as fast

What you can do yourself

Check who the numbers come from. Many of this week's results were measured by the developers themselves, and that is a reason to keep a cool head. Also read the privacy policy before you give a new product access to personal data, especially if it measures your body. And if you feel like experimenting, try Gemini's new PDF feature on a document task that is not sensitive.