Nix also needs the build output to be deterministic to calculate the hash. It also has the problems of timestamps etc. The build environment tries to be hermetic by setting the time to be epoch among other things.
These seem very reasonable, the workarounds used are natural, and overall the article is not at all congruous with the conclusion in the (clickbait?) title?
Reading this, I think low level engineering is actually more dependent on specific environments. Hardware also has its own points of change. Usually, when you think at a high level, environmental changes are less significant than you might expect. But low level thinking tends to be tied to specific environments, which is what makes it difficult. The reason low level is hard is that even if the code itself is short, the hidden assumptions inside it are difficult and place a heavy cognitive load on the programmer. For example, even a short snippet in C like
`int value = (int)buffer`
requires a lot of implicit knowledge about the 4 byte alignment of the buffer, or whether int is exactly 32 bits. LLMs do not seem to be very good at knowing these things. Rather, they are strong at high level wrapping, but at the low level, they seem surprisingly difficult and somewhat useless. Hardware has CPU generation changes, and in the case of PLCs, where I mainly work, the protocol differences between vendors are far too severe. There does not seem to be any technology with a very long lifecycle.
Time and date are... tolerable. There's SOURCE_DATE_EPOCH which should always be set to whack it into submission when used. ASLR of the _compiler being invoked_ resulting in a difference in the _program being compiled_ is nuts and would break any self-hosting compiler with consistency checks.
If Clang generated non-deterministic output due to pointer addresses then that's a bug (happens regularly) that should be fixed. The most common way this happens if it some code path is iterating over a DenseMap which is non-deterministic. Sometimes that's fine and sometimes that's not depending on how that map is used. The common way to fix that is to switch to a MapVector which pays some additional runtime/memory cost to guarantee deterministic iteration order.
I'll try and make a minimal reproduction case and file a bug. Do you know if any tooling that can take a binary and fuzz it down to a minimal reproduction set?
> lol you'd think, but no, it's not. In theory it is (and for small scale compilers it definitely is), but in practice compilers are strange and complicated beasts containing multitudes that no mere mortal can fully comprehend on their own.
Umm there are like 60,000 lit tests checked into LLVM that verify that the output is absolutely a deterministic function of the input of at least that compiler.
> Even though the source code had the same bytes, the output of the compiler was wildly different.
This is the goofiest I've seen written unironically in quite a long - the C preprocessor is not part of the compiler. The pre in preprocessor should probably give it away.
Just a tip: you should probably actually understand something before you decide you hate it.
> This is the goofiest I've seen written unironically in quite a long - the C preprocessor is not part of the compiler. The pre in preprocessor should probably give it away.
This is true but doesn't seem relevant; does replacing the word "compiler" with "build chain" change anything? Because that seems like the clear meaning.
On the off chance that you’re serious, that would result in disastrously bad output. The difference between “jmp $+15” and “jmp $+16” is inscrutable and the LLM would not be able to pick the right one without tooling.
That tooling is a compiler. The higher level, the better chance the LLM can be steered to good output. Machine code is hopeless, don’t bother.
> I don't see why that's the case. LLM trained on binary would totally see it, not?
It would not. You find the correct version by counting the number of bytes to the destination. LLMs are famously bad at this kind of problem (counting).
> Also the tool can also be running the test and a debugger.
The test needs to provide a good amount of signal. That’s too hard if you are throwing machine code at the wall.
In order for debuggers to work, you need some kind of model that describes what the code should do and what state the computer should be in after each instruction. That model is high-level code.
I can understand the intuitive appeal of training LLMs with machine code, but all of my experience with LLMs suggest that they are incredibly ill-suited to the task, and we just don’t have the capacity to train them to make useful machine code.
No, I don’t think so. LLMs are good at a lot of simple tasks, but bad at certain simple tasks. Moravec’s paradox in a new iteration.
It applies to humans too. Calculus is “simple” but it takes something like sixteen years to train a human to do it, if all goes well. Meanwhile, most humans think that inverse kinematics is, like, the easiest thing in the world (it’s a super complicated task).
I would phrase it as "LLMs are good at big picture stuff and bad at fine detail", or to put it another way, they're accurate, but imprecise and with low reproducibility.
That compiler does wonders with languages that have UB on their specs, especially when having optimizations passes with heuristics.
Also there are dynamic compilers were the shape of machine code changes as the code executes, and each single execution will certainly generate different sequences, depending on the program execution and where it is running.
Deterministic JIT compiler code generation, at least on optimising ones, is not a solved problem.
Why? I mean this is all emergent, right? And it’s not like humans ever work at this level. It would be very interesting to see what sort of outputs and abstractions an LLM comes up with.
Generative algorithms have been studied for decades now and while they have led to some interesting results they're a bad fit for LLMs because there's no such thing as a "plausible" binary: a small perturbation yields an unusable result.