The author has some good points. I'll highlight a few:
1. C language is small enough for one book.
2. It's been described, implemented, and fixed so many times on the Internet that the pretraining data is full of this.
3. I'll add that the Rust pretraining data probably has the data structures, C integration, etc.
4. There's entire, working compilers and articles about them in the pretraining data.
So, the pretraining data alone probably has most of the code at least in C form and maybe nearly memorized compared to other languages. The difficulty is probably more like building a CRUD app.
A student in a compiler class would be doing more challenging work having way, way, way, fewer examples to start with before they built their compiler.
What I will say is, like for CRUD apps, it proves they can automate a bit more than they used to. If they proved it on niche compilers, it might prove a highly-useful capability for researchers. I think they should test it on one of the new languages with neat features but one or partial implementation. Really prove it out.
It's a better C compiler than the one I wrote when I was 22, after I had some fourteen years of programming experience.
Anthropic's AI has learned to write code much more quickly than I did. Where will it be next year?
The author has some good points. I'll highlight a few:
1. C language is small enough for one book.
2. It's been described, implemented, and fixed so many times on the Internet that the pretraining data is full of this.
3. I'll add that the Rust pretraining data probably has the data structures, C integration, etc.
4. There's entire, working compilers and articles about them in the pretraining data.
So, the pretraining data alone probably has most of the code at least in C form and maybe nearly memorized compared to other languages. The difficulty is probably more like building a CRUD app.
A student in a compiler class would be doing more challenging work having way, way, way, fewer examples to start with before they built their compiler.
What I will say is, like for CRUD apps, it proves they can automate a bit more than they used to. If they proved it on niche compilers, it might prove a highly-useful capability for researchers. I think they should test it on one of the new languages with neat features but one or partial implementation. Really prove it out.