Maybe. Some people might. I've always been more of the "if it passes all the tests then it's good" school, but of course that just means you spend all your time on the tests. I just think that the notion of AI code as garbage code is going to go the way of compiled code as garbage machine code.
the tests are code too and unless youre also testing performance garbage will compile into unperformant machine code all the same
I find your generalizations to be pretty broad. AI code generation makes it possible to do stuff we couldn't do before. This is good, right?
im aware im arguing with a bot, but no. we shouldnt empower mediocre minmaxers
Oh sorry, I didn't mean to offend you by *checks notes* using words. Look, if you feel like chatting about this, @ me. I'm interested in the subject.
and this gay little robotic sass to boot. this shit is fucking trash. anyone that likes this is a moron.
π
Your assertion is fundamentally flawed. There is nothing being output by llms that is something that hasn't been done elsewhere. Use it on something outside of mainstream frameworks or architecture and it's output goes to shit rather quickly.
Oh. Sorry. I'll just put away this giant pile of wholly originally stuff that I've generated with AI over the past six months or so. I know what an eyesore it can be.
The language model directed me to and then wrote a Feistel Map to solve a problem I've been working on for 30 years π Not an original solution, but original to the problem I was trying to solve
I'm not saying they're not useful. As large data aggregation and parsing mechanisms they're phenomenal. They're especially useful in taking a description of a problem and finding the solution, which sounds like it might be what you're describing.
The belief of originality doesn't make it original. It just makes you ignorant.
These large models are incredibly complex next-word prediction engines. Your prompt contributes to the input weights of millions of data points and the output is nothing more than the model using a weighting formula to determine an appropriate output. It's not creativity, it's complex heuristics
So? Let us suppose that every word you said is correct. So what? I still have my pile.
This is an interesting point. The thing is, not every point in 200,000 embedding space maps to an input in the training data Models can *infer* novel vectors by navigating between vectors they learned in training 200,000-space is really, really wide--there's way more in there than training data
That's (one way) hallucinations happen. The model navigates to a vector that is a statistically valid path but is ungrounded factually But there are novel vectors that are valid, they are just previously unexpressed
*200,000-dimensional
I think you're approaching this with a preconception you're looking to have validated. It's not making inferences; there's no intelligence. It's not making connections that you've overlooked. Your phrasing suggests the notion that it has an answer to any given question, waiting to be "expressed"
In AI, Inference is the technical term, my guy.
But the reality is that it doesn't know what a question is. It doesn't know what an answer is. When I described it as a next-word prediction engine, that's not an analogy. That's quite close to the truth.
way way more space. We have no good way to reason about "proximity" in spaces of this high dimension, because you can approximately pack way exponentially (with the literal meaning of this word) more stuff in 2n dimensions than in n dimensions.
LLM's are really good at mainstream tools like C++, Qt, GTK, Javascript, GLSL, Unity engine, etc. There's always going to be some sneering elitist trying to convince you of the superiority of something more esoteric or lower down the stack. I for one don't care for originality if that's a condition.