Obvious joke, LLMs, don't tell me this is inherent to LLMs because that is in fact the obvious joke
If LLMs could solve the "don't confidently output shit that makes no sense sometimes" problem they could really be something.
@jonny well... We can train a human to be an expert in a field and (1) they can output shit sometimes and (2) they can out reasonable and logical responses that other highly trained human experts vehemently reject.
It's not just that you can't train an AI to "not be wrong". That's not even how knowledge works.
re: Obvious joke, LLMs, don't tell me this is inherent to LLMs because that is in fact the obvious joke
@SecondUniverse
Please see the cw wherein I say "don't tell me this is inherent to LLMs because that is the obvious joke" which also includes "obviously there is no way to make something never output things that make no sense because 'making sense' is not a well defined concept"
Its just sort of amazing how shallowly this becomes a problem. Ive been using the LLM basically as a search index into rust concepts, telling it to basically just direct me to the docs for the language features I need, but it can't help itself but comment on the code, and I'm now spending just as much time trying to tell it "OK no we are not going to add another random method to patch some problem we can resolve by just doing it right, your job is to serve as a reference for the language, you are a language reference machine, that is all you do"
Dog I just started learning I should not be having to tell my cocaine-addled teacher to not touch my code and make it worse or wander off and tell me about random other shit.
@jonny you can just disable the write tool
re: Obvious joke, LLMs, don't tell me this is inherent to LLMs because that is in fact the obvious joke
@starsider
It was just pitching in the chat output
i've said this a thousand times in the past few years. i can't believe people let these things do anything important for them. I am simply trying to use this thing for reference lookup becasue it can be challenging to map concepts from between contexts because none of the words overlap but there is nevertheless a well defined mapping between them. I am now about 5-layers deep in a loop of "solving a problem it was not asked to solve in a way that makes no sense" where it is telling me to duplicate all my code into two branches with a read lock or a write lock because it can't understand the flow of the code literally one layer up from the last code that it ingested in its context window. even with the latest models trying to do an extremely constrained task that should literally be the only thing they can do - fuzzy mappings in language space - these things just spontaneously combust at the starting line
@jonny I honestly don't know how you tolerate interacting with these things. I just use the "go to definition" tool in gram, or the "hover for docs" tool until I feel satisfied.
If that doesn't work, I look up the rust forums. They seem pretty good. Usually the same 5 people answering every question, but they sure know the language.
@jonny strong Butter Robot energy
@wrstscrnnm6 it EVEN "KNOWS" THAT and imitates the butter robot in its responses by going "right. i reference rust. oh my god" and then completely fails to be a butter robot
Expensive, non-deterministic machines being implemented
to solve cheap, deterministic problems will never be not funny.
@paninid it was making some confident assertion about the behavior of some of my python code that i'm porting, and at first it turns out that it just used the wrong long numbers when it used sed to print the lines of the method so it only saw half of it, but then it told me that the method was never used because....
- the name of the class is
Epoch - usually when it's assigned as a local variable it is named
epoch - the method it was searching for uses was a python dunder method,
__sub__, which overloads the subtraction operator
so it had grepped for epoch - 1, not found anything, and concluded the method was unused.
meanwhile, i am over here with a "normal IDE" with "normal language tooling" where i can "deterministically find all the places this method is used" being like "what in the world". important to note that it has a python LSP tool and instructions to use it rather than grepping are injected into the context window at every launch but it just fails to use it when a human being would simply never fail to use the "find usages" button and accidentally grep instead
