@C0m3b4ck
Reportedly it took about 2000 Claude sessions before getting to the desired result. Read the text quoted by valdir.marcos in the 3rd post in this thread.
Just about every compiler ever written is a direct or indirect copy of somebody else's work. No one is born knowing everything... every compiler writer has learned the techniques from reading books, previous compiler implementations and so on. That's what A.I does in many cases. NOTE: the previous sentence does NOT mean that there aren't times when A.I will dispense something that is a straight copy of somebody else's work, which is at the very least unethical and quite likely bordering on illegal if not blatantly illegal but, it doesn't look like that happened in the production of that compiler.
You make it sound like the C compiler it produced is a copy of prior art and, TTBOMK, there is no foundation for that implicit claim. If parts of that compiler were copied from some other compiler, no doubt someone would have pointed it out and quite likely someone would consider legal action as a result. The A.I thing basically did what a human being does, use knowledge to produce code and, the source of its knowledge is pretty much the same as for a human being and (usually) no one complains about that.
I don't worship A.I at all but, I give credit where credit is due. Some of these A.I things (I know you prefer "models") have demonstrated being tools that can very effectively assist in software development. That said, I would very strongly advice never to use A.I to create software the human being isn't well versed in, e.g, someone who doesn't know how to write a compiler would be very ill advised to use A.I to develop a compiler, simply because the individual lacks the knowledge to spot the inevitable mistakes/deficiencies the A.I thing (model) will make.
The fact is, developing a full fledged C compiler in 2 weeks that can compile and produce a functional instance of the Linux kernel is impressive. The low cost is also impressive.
It wouldn't surprise me at all if in the near future software development companies specifically asked for experience using a particular A.I model (specific thing) because, for instance, just knowing how to write a compiler won't be good enough, the compiler writer will need to know how to write a compiler (which a compiler writer definitely should know) but also, use a particular/specific A.I model for its development. The reason I anticipate that is, a full fledged C compiler for $20,000 (plus the cost of the human A.I "manager") is an unheard of bargain and, that cannot be accomplished without A.I assistance. IOW, programming knowledge _without_ the ability to effectively drive an A.I model will simply be _not_ cost effective.
In that C compiler the part that I am particularly curious about is how well the compiler optimizes code. As you pointed out, writing a scanner and a parser is a piece of cake, there is a ton of books describing multiple methods to get that done (not to mention Lex and Yacc to produce quite a bit of the code.) The code generator is a different ballgame. A naive code generator is easy but something that selects the correct optimizations for a particular section of code while taking into account the "idiosyncrasies" of the particular processor is a completely different ballgame and, the knowledge for that stuff is either very difficult to find or often cannot be found at all, you need to figure it out by extensive testing. I'll eventually have a look at what and how well the A.I thing did in that area.
A.I is a tool and, it seems that, in the near future a reasonably high level of proficiency in its use will be required for a programmer to not only be effective but also be a reasonable business investment as an employee.
I wonder if the compiler can be used to write a version of rust 
If it can compile the Linux kernel, it must be capable of implementing a compiler for Rust.
Recent
