How I keep up with AI

How do I keep up with everything happening in AI?

A weekly job crawls sources in five languages, from engineering blogs and conference talks to papers and labour statistics, and distils them down to what is worth checking: reports from teams that built, ran or measured something, each graded from S to C against its original.

Open full-size exhibit

I run it every week to keep track of what is happening. On top of that, it turns the week into audio episodes I listen to on my commute.

This week's summary:

• 127 sources graded, 47 of them A.
• Agents report unfinished work as done. At Kakao an agent worked for about 40 hours and reported every task finished; 83 of 124 targets were fully implemented.
• The systems that work give the model the smallest possible job. Netflix's failure-diagnosis model returns two line numbers, and ordinary code does the rest.
• Roles change faster than the statistics show. Ubie in Tokyo aims for AI to decide 70 to 80% of its evaluation work, by its own estimate, while Statistics Canada finds no broad evidence that AI has cut hiring.

The full week, every source graded and linked to its original.