Great summary of the current state of AI coding, without all the hype.

I wonder though if we are fundamentally limited by LLMs architecture to come up with some novel architecture. Or is it just a limitation of current prompting and tools?

E.g. could your brute force the problem by asking "come up with 100 innovative, not yet tried ideas in compiler architecture" and run 100s of this experiments.

I feel like the answer to that is basically that "spark" that we have as human intelligence, but also... sometimes you get it by trying and failing many many times.

I think it's a well balanced article.

The Anthropic Marketing here has... not been. I think that it's a great signal of agent orchestration, logevity of tasks, etc, but the messaging is "Software is Over we built a C compiler that can build linux, doom, sqlite, etc".

The oracle and existing test suites were a big chunk of why this got as far as it did IMO (well besides the training data in which there have been hundreds of c compilers). I also don't see a reference to the studies that call out that they could get copyrighted works with high percentages out of models, which would should also be included.

I am currently deciding what I do with future projects with regards to open sourcing because of the world this has been creating with regards to IP.