8217
Substack spent two years telling writers to post more. Then it shipped an AI scanner and sent the verification bill to readers and writers.
Attackers are starting to embed text meant to make defensive AI refuse to work. Hugging Face investigated its own breach on a downloaded open-weight model.
One year of Tech Trenches, run through the same method the publication uses on everyone else: what was promised, what happened, and who paid.
Anthropic watermarks everything Claude touches, from a full draft to one fixed comma, and bills you by the token for it. The mark can't say what Claude did.
OpenAI changed its Microsoft deal on April 27. Brockman welcomed the AGI era on September 3. The dates are clear. What did the launch prove?
Washington spent years trying to choke China's AI with chip controls. It failed. Now it has called a summit, and calling it is the concession.
One vendor's misconfiguration reached four AI labs. Nine intrusions, weak passwords and leaked credentials, and a boundary that existed only in the prompt.
TypeSafe’s Jev returns decisions, not text. Paired with LLM executors inside a harness, it targets the decision gap exposed by bankrupt AI food trucks
Every few months the AI industry announces the end of how we work, then forgets it and sells a course on the next one. Word, mandate, metric, syllabus, repeat.
A sanction that switches off when fuel gets expensive is a price list. Trump’s Russian diesel deal, Hormuz, and the cost of an American promise.





