Comment by fragmede

> I wonder if Nicholas Carlini (the author of the Anthropic post) would have had more success using Claude Code alongside his own abilities for significantly cheaper.

You're thinking like an individual, not a corporation. $20,000 is a lot of money for me to go and pay the bill for as an individual. That's a car for most of America! However, if I'm earning $20,000/year at my job, that's peanuts. Thus Mr. Carlini (whom surely makes vastly more than $20,000/year) being able to do, what previously would have taken a team of people to do, is nothing short of astounding. I don't know how well the compiler stacks up against, say clang or gcc, the real question is how much did it cost Intel to make v0.1 of icc.

> For example, someone might feel like they're more productive, but their output is roughly the same as what it was pre-LLM tooling.

There is just no comparison. It's not about how much faster it is, it's about could I have attempted this project before? Yes. Would I have attempted it? Probably not! The start up cost for a project was just so high that I've a list of things that I'd love to attempt but never made the time for. With AI, I'm slowly knocking things off that list (most of them don't actually go anywhere, but there's an itch to scratch, as a hobby).

> not nearly as good as some would have you believe.

Hallucinations from LLMs are interesting as a concept, but they can hardly be blamed for it as they learned to ability from humans. (Some) humans love to blow smoke up your ass in pursuit of the all mighty dollar. LLMs have their limitations. There's some prognostication about the future, but I'm interested in what they can do today.

Thank you for the thoughtful response!