Hacker News Reader

I don't know. Like, only eight months ago or so, I was explaining my colleagues how using AI for coding can actually work, in some cases. Now, I don't know any engineers who don't use AI tools, at least at work (our own pet projects are still fun to do without AI, but at work, AI-assisted productivity is now expected).

If that's "plateauing"...

View on HN

It doesn't feel like LLM-assisted coding has gotten materially better in the last 6 months, though. The tooling, sure, but I'm still wrestling with the exact same problems I was at the beginning of the year. That doesn't feel like the exponential improvement people have been prophesizing about.

I am fine with constant linear improvement too. What model do you use? SOTA is Gemini 2.5-pro. We definitely didn't have anything comparable in 2024.

I wasn’t talking about their adoption. I’m talking about their abilities.

In eight months, we got SOTA from 50% to 83% on Aider's Polyglot benchmark (225 Exercism problems). That increase matches my own experience.

Reply on news.ycombinator.com