I don’t disagree overall (Victories continue to come closer to home), but to whackamole a specific datapoint, how much critique of the Claude C Compiler did you read? https://harshanu.space/en/tech/ccc-vs-gcc/ was on top of HN for a day and I believe Zvi reported on it so I’d imagine you have.
But to quickly summarize, CCC only implemented the “easy” parts, directly copied/transpiled the hard parts of that easier part, and most importantly is over 100,000 times slower than an actual compiler! I feel this generally smells of the types of distorted truth that you should expect to omnipresently manifest when literal trillions of capital are at stake.
I think the original Anthropic post is reasonably accurate about the limitations of the project and I don’t see how this blog post contradicts claims in the original post?
I think that this C compiler represents roughly a year of work for someone and that the issues here aren’t a crux for this. TBC, by ‘CCC only implemented the “easy” parts’ you just mean that it only build a non-optimizing compiler and didn’t build the linker or assembler? This was explicit in the original post.
(FWIW, my view is the the hardest part is optimization (while still being correct), then the compiler itself (made harder if you also have good error messages etc which CCC certainly doesn’t have), then the linker / assembler. I don’t buy that the linker/assembler are harder than a compiler that can correctly compile projects like the linux kernel.)
most importantly is over 100,000 times slower than an actual compiler
Yes, this isn’t surprising as it’s a non-optimizing compiler? (It’s also quite a bit slower even at -O0 from my understanding.)
I’m a bit confused by what your argument here is. I don’t think anyone was claiming that CCC was anywhere near a full drop in replacement for gcc/clang? TBC, I do think it would represent a much stronger result if it was a full drop in replacement for gcc/clang, like this is presumably >>10x as much human effort.
Hmm, well yes I suppose my main point is that I personally don’t believe it represents a year of work, but worry the pro-marketing slant of the Anthropic blog post pushes this perspective moreso than a cynic’s breakdown, which I wanted to make sure you were aware of. (There are ofc other analyses than the HN discussed one, see 1, 2)
But I certainly believe it replaced a month+, so as long as it’s now confirmed we are operating under similar information it becomes just a minor quibble over a singular OOM in the interpretation, which seems fine to me.
I don’t disagree overall (Victories continue to come closer to home), but to whackamole a specific datapoint, how much critique of the Claude C Compiler did you read? https://harshanu.space/en/tech/ccc-vs-gcc/ was on top of HN for a day and I believe Zvi reported on it so I’d imagine you have.
But to quickly summarize, CCC only implemented the “easy” parts, directly copied/transpiled the hard parts of that easier part, and most importantly is over 100,000 times slower than an actual compiler! I feel this generally smells of the types of distorted truth that you should expect to omnipresently manifest when literal trillions of capital are at stake.
I think the original Anthropic post is reasonably accurate about the limitations of the project and I don’t see how this blog post contradicts claims in the original post?
I think that this C compiler represents roughly a year of work for someone and that the issues here aren’t a crux for this. TBC, by ‘CCC only implemented the “easy” parts’ you just mean that it only build a non-optimizing compiler and didn’t build the linker or assembler? This was explicit in the original post.
(FWIW, my view is the the hardest part is optimization (while still being correct), then the compiler itself (made harder if you also have good error messages etc which CCC certainly doesn’t have), then the linker / assembler. I don’t buy that the linker/assembler are harder than a compiler that can correctly compile projects like the linux kernel.)
Yes, this isn’t surprising as it’s a non-optimizing compiler? (It’s also quite a bit slower even at -O0 from my understanding.)
I’m a bit confused by what your argument here is. I don’t think anyone was claiming that CCC was anywhere near a full drop in replacement for gcc/clang? TBC, I do think it would represent a much stronger result if it was a full drop in replacement for gcc/clang, like this is presumably >>10x as much human effort.
Hmm, well yes I suppose my main point is that I personally don’t believe it represents a year of work, but worry the pro-marketing slant of the Anthropic blog post pushes this perspective moreso than a cynic’s breakdown, which I wanted to make sure you were aware of. (There are ofc other analyses than the HN discussed one, see 1, 2)
But I certainly believe it replaced a month+, so as long as it’s now confirmed we are operating under similar information it becomes just a minor quibble over a singular OOM in the interpretation, which seems fine to me.
(Assuming this is true) One would wish for LW to be one of the earliest places for Truth to master pants, but this hasn’t happened quite yet...