About a year ago I wrote a long, drunk post saying local models would get properly good and this would all go further than people expected. I’ll dig it out. Some of it holds up.

I’m a developer. Six months ago writing code was the job. Now it isn’t. I design the system and review what gets built, and things that would’ve taken a team months take hours. I don’t think people outside the industry get how fast that changed.

It’s not just me. Anthropic just wrote up how they replaced git internally: www.anthropic.com/research/redesigning-source-control

Merged code changes went from ~2,500 a day in May to over 30,000 by September. Writing code isn’t their bottleneck any more. Git, CI and code review were built for humans working at human speed, and they can’t keep up. So a team of seven rebuilt source control in five months, with Claude doing a lot of the work.

My job doesnt even work at human speed anymore. IN LESS THAN 6 MONTHS.

The model is building the tooling that lets more model-written code ship. It’s not recursive self-improvement YET. Humans still set the direction, and the article says they regularly had to tell Claude no. Writing infrastructure also isn’t the same as coming up with the research ideas that make the next model better.

But the loop is getting tighter. My guess: within one year we'll reach RSI. 2 Max.

As soon as AI can improve AI better than we can - which GENUINELY IS SO CLOSE and I don’t think people realise how close it is. ALL BETS ARE OFF.

When that loop closes, I don’t think anyone, including the people building it, can tell you what happens next. That’s what should worry people. Not that it’ll be bad, just that nobody will see it coming.