The water from sewage might end up there after it's extracted and sanitized, but all the solids have to be disposed of too. Those solids, plus the leftover chemicals used to extract and sanitize the water, go to landfill.
Essentially LLMs are recontextualizing their training data. So on one hand, one might argue that training is like a human reading books and then inference is like writing something novel, (partially) based on the reading experience. But the contract between humans considers it plagiarism when we recite some studied text and then claim it as your own. So for example, books attribute citations with footnotes.
With source code we used to either re-used a library as-is, in which case the license terms would apply OR write our own implementation from scratch. While this LLM recontextualization purports to be like the latter, it is sometimes evident that the original license or at least some attribution, comment or footnote should apply. If only to help with future legibility maintenance.
It's a zero sum game. AI cannot innovate, it can only predictively generate code based on what it's already seen. If we get to a point where new code is mostly or only written by AI, nothing new emerges. No new libraries, no new techniques, no new approaches. Fewer and fewer real developers means less and less new code.
Nonsense indeed. The model knowledge is the current state of the art. Any computation it does, advances it. It re-ingests work of prior agents every time you run it on your codebase, so even though the model initializes the same way (until they update the model), upon repeated calls it ingests more and more novel information, inching the state of the art ever forwards.
I've seen terrible things where it would overcomplicate and duplicate. But I've also seen it write really good code. I've been trying to get it to do the latter consistently. Detailed specs and heavy use of agents really helps with the code quality. The next step is editing the system prompts, to trim away any of the fat that's polluting the context.
Nonsense. LLMs can easily build novel solutions based on my descriptions. Even in languages and with (proprietary) frameworks they have not been trained on, given a tiny bit of example code and the reference docs.
That's not novel, it's still applying techniques it's already seen, just in a different platform. Moreover it has no way of knowing if it's approach is anywhere near idiomatic in that new platform.
I didn't say the platform was the novel aspect. And I'm getting pretty idiomatic code actually, just based on a bit of example code that shows it how. It's rather good at extrapolating.
Practical Common Lisp by Peter Seibel, and then The C Book by Mike Banahan, Declan Brady and Mark Doran. No clue why those books in that order, but they both proved to be decent choices.
Then I had a couple of jobs where I was given access to data and opportunities to go beyond my expected duties by doing things with that data, i.e. automation and reporting.
LLMs generate the most likely code given the problem they're presented and everything they've been trained on, they don't actually understand how (or even if) it works. I only ever get away with that when I'm writing a parser.
It matters if AGI is the goal. If it remains a tool to make workers more productive, then it doesn't need to truly understand, since the humans using the tools understand. I'm of the opinion AI should have stood for Augmented (Human) Intelligence outside of science fiction. I believe that's what early pioneers like Douglas Engalbert thought. Clearly that's what Steve Jobs and Alan Kay thought computing was for.
AGI is such a meaningless concept. We can’t even fully design what human intelligence is (and when a human fails it meaning they lack human intelligence). It’s just philosophy.
If it empirically works, then sure. If instead every single solution it provides beyond a few trivial lines falls somewhere between "just a little bit off" and "relies entirely on core library functionality that doesn't actually exist" then I'd say it does matter and it's only slightly better than an opaque box that spouts random nonsense (which will soon include ads).
Late 2025 models very rarely hallucinate nonexistent core library functionality - and they run inside coding agent harnesses so if they DO they notice that the code doesn't work and fix it.
This sounds like you're copy-pasting code from ChatGPT's web interface, which is very 2024.
Agentic LLMs will notice if something is crap and won't compile and will retry, use the tools they have available to figure out what's the correct way, edit and retry again.
reply