The Archive

Sunday 20 September 2026

What the deck led with that day: 7 stories, each with the note as it read at the time. The record is not rewritten; corrections are logged separately.

Commentary2d 1h ago19 Sep, 1:21 PM CDT
NO PHOTOGRAPHPUBLISHED WITH THIS STORY

Microsoft director: AI scraping 'the largest theft of labor in human history'

A quote story, so the first question is always what surrounds the quote. It is carried here because the underlying dispute is the one that governs whether sites like this one can exist, and because an executive at a model company saying it out loud changes the shape of that argument regardless of the context.

Hacker News

Reported2d 1h ago19 Sep, 1:21 PM CDT

Microsoft director called AI scraping ‘the largest theft of labor in human history,’ while OpenAI head brands ChatGPT an ‘existential threat’ to publishers — revelations come from legal briefs filed in NYT lawsuit

A quote story, so the first question is always what surrounds the quote. It is carried here because the underlying dispute is the one that governs whether sites like this one can exist, and because an executive at a model company saying it out loud changes the shape of that argument regardless of the context.

Tom's Hardware · carried by 2 sources

Reported2d 10h ago19 Sep, 5:00 AM CDT

Mathematicians Hate AI. They Can’t Quit It

The interesting part is not the friction, it is who is reporting it. Mathematics is the field with the least tolerance for an answer that cannot be checked, so when its practitioners say a tool is both unbearable and indispensable, that tension is data about the tools rather than about the mathematicians. Worth reading alongside anything claiming the argument is settled in either direction.

Wired

Reported3d 5h ago18 Sep, 10:00 AM CDT

If the AI Industry Followed Its Own Research, It Might Have Paused Already

A lab's own chief executive is the source here, which makes this a primary claim wearing a reported headline. That is the interesting shape: the gap named is between what a company's safety research finds and what the same company ships. Hold the specifics loosely until the underlying research is read directly, and note that the strongest version of this argument is always the one made from inside.

Wired

Commentary3d 20h ago17 Sep, 6:37 PM CDT
NO PHOTOGRAPHPUBLISHED WITH THIS STORY

How to Write with an LLM

Willison writing about process rather than product, which is the rarer and more useful thing. Note where this landed: his own site first, then the top of Hacker News, with no outlet in between. The people who build with these tools are still learning how to write with them in public, and the practitioner posts are running well ahead of the coverage.

Hacker News · Simon Willison · carried by 2 sources

Reported4d 1h ago17 Sep, 1:33 PM CDT

LLMs respond differently to harmful prompts when AI watermarking is used

Read this one carefully, because the finding runs opposite to the instinct. Watermarking is sold as a safety measure, and the claim here is that adding it changed how a model answered prompts it would otherwise have refused. If that holds, a mitigation designed to make output traceable also moved the behaviour it was meant to leave alone, which is the kind of result that matters far more than whichever model topped a benchmark this week.

Ars Technica

Primary4d 2h ago17 Sep, 12:58 PM CDT
NO PHOTOGRAPHPUBLISHED WITH THIS STORY

An Empirical Study of Harness Design for Coding Agents

The only Tier 1 item in today's leads, and it is a paper rather than an announcement. Worth the attention precisely because harness design is the part of coding agents nobody markets: the scaffolding around the model, not the model. When a measurement lands on the unglamorous layer, that is usually where the next year of gains comes from.

arXiv · carried by 3 sources