Hacker News Reader

Hacker News

Comment by dave_mcladty | Hacker News Reader

Parent Full thread

dave_mcladty

When I see AI take an large existing legacy codebase and rewrite it perfectly then I will believe that

View on HN

You will never see this from a single human being.

Probably, within a few months, one (or more, possibly colaborating) LLMs just might.

Always fun to see "the difficult ones," proven.

You'd still have to describe to an LLM what to do, which strikes me as about equivalent complexity to simply writing the code in the first place (which is, after all, a formal description of how a program should behave, even if we perceive it as of a complexity which should not be strictly necessary). The big wins have been simply leveraging LLMs to apply common patterns to multiple codebases (albeit in a buggy and haphazard fashion), but it's still up to humans to compose these patterns into meaningful programs and validate that it's actually functioning as expected or desired. Humans still have far superior understanding of what software is, how it functions, what our intentions are using software, and how to derive good software from bad—we know what a bug is intuitively in a way that LLMs have not been able to demonstrate at all.

However, being able to rewrite a program with formally well-defined behavior (i.e. code) should be in an LLM's capability, but LLMs are a long ways away from demonstrating semantically coherent coding skills, just the ability to regurgitate common patterns (often filled with bugs and/or incoherent semantics).

An LLM is only an Internet search engine with a fancier interface, it doesn't actually reason about anything. There is no "semantic" anything about an LLM's output.

Non-RAG LLMs don't search the internet, nor even have the capability to do so.

I don't personally detect sentience, yet — but about 5% of my inputs result in some sort of interpretation and/or reasoning (sometimes I have to think for days "why did LLM.xyz make such a strange connection..?" only to realize the schizotypisms of machine aren't often wrong (just different).

People ascribe sentience and emotion to a smiley face picture, and that is just two dots and a curved line.

That's just what people do, we are hard-wired to see social cues even if there are none.

Absolutely false, at the core of every LLM is a highly compressed text corpus from an Internet search engine.

(The wonder here isn't that an LLM succeeds at text retrieval tasks, the wonder is how highly compressed the index turns out to be. But maybe we just severely overestimate our own information complexity.)

So, you're saying an LLM is a just a database that does text retrieval?

Yes, using a statistical model which is in effect a very lossy compressor.

So, what you're telling me is that every thing they say has already been said before, completely verbatim? Like, if I asked it to write a story about a dog named Jebediah surfing to planet Xbajahabvash, it would basically just find a link to someone else's story that wrote about the same dog surfing to the same planet? That sounds like an infinitely large amount of combinations. Perhaps the internet is just infinitely large, squared (or even circled).

So, like a human, then?

Reasoning is not necessary, but semantic coherence is.

It's only as semantically coherent as its training database. An LLM is, in effect, just a lossy compression of its training database. The compression is based on statistical maximum likelihood estimation, there are no mental (or any other kind) of models involved in compressing the training database.

You can claim that mental models don't actually exist and everything in the universe is just maximum likelihood, but that would be a religious/spiritual statement, outside the realm of science.

We'll have to see what Google's most advanced model with the ridiculously larger context window can do, once it's fully released into the wild. Refactoring an entire code base is presumably asking too much, but it'll be able to do small refactoring since the released model can do that, so we'll have to see where its upper limits are.

Of course, to get it better at refactoring, everyone has to write blog posts on refactoring to feed the machine.

Why blog posts? Wouldn't the codebase "self evolve" based on the proficiencies of other published code?

AI will probably just delete it.

Worked at a few of these dog shite companies and the amount of low quality shit driven by half baked initiatives from the C-level suite (ie, “we are microservice oriented now, do that”) is astounding.

This seems to me like it would be easier than having it create a greenfield project. The legacy code tells the AI the needed functionality so I would think in the near future an AI could rewrite something in a more concise way with the same style throughout, even adding useful comments and docstrings.

Reply on news.ycombinator.com