I've built 4 pretty novel features where the LLM didn't know how to solve the problem, but with design/decision input from me was able to implement it. Between the two of us it was done, I would estimate, 10 - 20 times faster in terms of man hours than the same implementation would have taken without the use of LLMs. Like it's team level productivity from a single developer. In my case it was done infinity times faster because I just wouldn't have done these things, it would have been cost-prohibitive.

> But if you are doing things that are not well represented in publicly available code, or in unusual languages, then LLMs are not great currently.

Have you tried since November 2025?

For what it's worth, engineers at proprietary trading firms are finding that they can get substantial productivity boosts by using LLMs and it'd be hard to find a domain that's less represented in the training data.

Does definitely require more steering than if you're churning out slop apps/websites though.