Three threads run through today’s issue. Most of the news is companies reporting on themselves, and almost none of it has been checked by anyone else.
Anthropic wrote about Anthropic four times: an equity pledge, its own shrinking junior hiring, Claude working out where some famous encryption numbers came from, and its Claude Code team’s account of automating its own code review. DeepSeek’s scorecard is DeepSeek’s too.
The second thread is measurement. A Mercor map says about 90% of American working hours have little or no AI test coverage. That tells us what nobody has checked, and says nothing about what AI can or cannot do.
The third is where tools get found and bought: Replit’s new default, a plugin inside ChatGPT, a web key from DeepSeek, and a chip boss’s phone wish list.
Anthropic on Anthropic: A pledge, a hiring slump and a codebreaking trick
Three of today’s stories come straight from the company’s own pages. Each one tells you something, and each one is graded by the person who wrote it.
- Anthropic pledges 5% of itself to a new public fund, if shareholders agree. The stake would grow as the company’s value does, but the extra share is the board’s to grant. Anthropic would own the fund outright, and the rules for running it are not out yet.
- Anthropic’s junior technical hires fell from 27% to 9%, its own data shows. Junior hires dropped from 27% to 9%, while junior employees overall slid from 21% to 14%, which is a different count. Anthropic admits the fixes it lists do not touch the problem.
- Mystery numbers in popular encryption came from a weak generator and the clock. Anthropic says Claude traced the unexplained numbers the NSA supplied for a common digital signature standard back to a 1990s random number routine fed the clock. So far, the only person who has checked that is Anthropic.
Who Checked: Gaps in the scorecards
The numbers keep coming, but they are mostly written by the people being measured. Here is what each one does and does not show.
- Most American work has barely been tested on AI, a Mercor map finds. Mercor says about 90% of US working hours fall on tasks that its 29 chosen AI tests barely cover. The figure counts what has not been measured, so it is no verdict on what AI can do.
- DeepSeek’s free model wins at getting jobs done, still loses at hard thinking. On DeepSeek’s own scorecard, a model anyone can download beats Opus 5 on four job-style tests and trails it on seven, including the toughest reasoning exam. Nobody outside the company has rerun it yet.
- Anthropic’s Claude Code team has machines read most of its code changes. Claude writes about 90 percent of the product’s code, so review piled up. The team now sorts each change onto a risk ladder that sets how many people must sign off, and the results are its own.
Shelf Space: Where AI tools get found and bought
A default setting, a chatbot plugin and an API key can each decide which tool wins. Today’s stories show all three moving.
- Replit swaps Anthropic for OpenAI as its main AI, then softens the claim. One of Anthropic’s top five customers made OpenAI the default behind its coding product. Its CEO first said the Anthropic relationship was over, then said customers can still choose Claude.
- In OpenAI’s own telling, a ChatGPT plugin is how Stack Overflow got found. A customer story on openai.com says signups tripled after launch, with no starting number and no outside check. Even so, developers increasingly find tools inside the chatbot.
- DeepSeek lets your AI read the live web with the key you already have. The docs describe a search tool and a page reader, each $0.75 per 1,000 successful calls. They say little about how the page reader copes with sites that fight back.
Today’s Quick Hits
- Three unannounced OpenAI model names briefly surfaced in its API. TestingCatalog caught the names, then saw them in blind image tests, and guesses a successor to GPT Image. Its report cites no comment from OpenAI.
- Trusting AI-written code is mostly about fixing the codebase. One engineer ranks fixing the code itself first and human review last, with prompts and rules files near the bottom. It is her judgment and has not been tested.
- A chip boss says AI firms want phones that run huge models all day. Qualcomm’s chief says unnamed AI companies want phones running 100-billion-parameter models by 2028. He offers no plan for the memory or power, and Qualcomm sells the chips.