ASM -> C ≠ C -> LLMs

same reason why the highest reliable language decompilers (ghidra, ida, etc) can reconstruct from arbitrary binaries is…pseudo-c. predictable enough that you can reverse it

That is fascinating!

honestly I recommend any dev to try reverse engineering their own (small) binaries at least once! try writing a really simple program in C, and then do it again in rust. drop it in ghidra, use a model / mcp / whatever to churn away at it, and watch how different (or not!) the decompilation is from the original source code. you'll be surprised how it looks...especially rust (lol)

Content like this is why I open this piece of shit twitter thing.

A major point of C was to create a layer of abstraction to enable portability across many architectures. If we lived in a monoarchitecture world where all hardware was identical, I'd imagine there would still be a hold out assembly programmers who hold quality above all else.

The contract for C code to ASM is "the ASM will produce the same outputs for the same inputs". You can specify contracts of that type for LLM code.

That is not the contract for C code to ASM. But in any case, if the contracts are not computable, then they cannot be trusted by themselves.

I can’t believe we have to have this totally retarded discussion but here we are 🤦

yeah the thing is LLMs are non deterministic while C compilers are supposed to be deterministic. I can exactly predict what it will output from a compiler but not the same from a LLM

It takes intelligence to design and craft a flute. Once created, the instrument reliably produces a musical tone when played correctly. A compiler is like a flute (tool), and a program is like music.

Very interesting video, probably. I'm sure the content is very deep. I just couldn't hear a word over how good that beard looks.

Like all analogies, it breaks down the closer you get to the source.

I mean if you ask an LLM , “solve XYZ” and it just solves XYZ, 99% of the tlime, even it can be super predictable. You just don’t like the fact that it solves XYZ, you would rather be doing that problem solving. Unfortunately the time of human problem solving is over

Embarrassing response & instantly trivially disproven

off topic but what camera are you using? looks reallly clean and detailed

PDP-11 heritage that we take for granted today through Intel and other ISAs...is C maps best to ISAs that used a heavily stack based abstraction with native hw stack frames call graphs. Recursion, nested calls, local variables, and function pointers etc. came easier.

You cut off right at the point it was getting interesting, the real question for me is: how can you make these tools as realiable as a compiler, these are supposedly stochastic