Hacker News Reader

Comment by choeger | Hacker News Reader

I tend to agree with the "bullshitting" part. But that gives us a very useful tool to estimate future applications of LLMS.

Will a LLM ever program on its own? Certainly not. It gets semantics wrong so often, it's not even funny. Syntax on the other hand is surprisingly correct. So it could help with refactorings (but be aware of subtle conflicts). Of course, code search will be a killer application.

In other cases it helps to ask "will someone really notice if the LLM puts out bullshit?" What will happen if your manager and your client both believe what the model generates? Will people die? Or will the negative effects be untraceable? In the latter case, why object?

View on HN

I think it's pretty far from certain that an LLM won't ever "program on its own". I don't have much confidence that we won't have LLMs widely programming within 12 months. As soon as you give it access to a lot more context (your whole codebase), a precise prompt, access to edit and run code, access to a debugger, and then put the whole thing in a loop...you're pretty close at that point. Maybe it just won't be able to cross the gap, but I don't find it hard at all to believe that we could soon have AI agents taking in issue tickets and opening PRs to address them. And if the issue tickets are too vague, they'll ask for clarification. That's an insane thing if you think about it. Six months ago I thought this was a decade or two away, if we could do it at all.

Impossible. In order to fix a bug, you need to do abstract interpretation or something equivalent. Language models simply can't do that. And even the best bug report will not tell you what you need to do to fix it.

Of course, there are other change requests: Change a color here, store this information alongside that record, modify that output, etc. But even those changes can very quickly spiral out of control. Color models, database schema compatibility, pretty printing ...

No, I think the idea that LLMs take over programmers jobs is born out of a complete misunderstanding of both. Of course that doesn't mean that no one will try to do it. I bet people will even report success in doing it, simply because their bonus depends on it, but I don't think it will ever really work.

What might happen though is that LLMs replace many of the programs we write nowadays. After all, who needs a web browser when a LLM can give you a summary in natural language?

It’s pretty funny that you think this is impossible based on some theoretical idea of what LLMs are. Playing with LLMs today quickly demonstrates how things that shouldn’t be possible in theory are trivial in practice. In a few years, who knows where we’ll be.

You have obviously no clue what "theory" means in that context. Playing with LLMs has so far demonstrated very little. Yes, these language models are subjectively impressive, but that's exactly what they have been designed to be. Usefulness is a completely different metric and it will be much harder to demonstrate convincingly.

From a pure programming point of view, the situation is simple: Even if a LLM could translate human language into working code, it just becomes another compiler and the input just another, admittedly novel, kind of programming language.

Your hyperbolic dismissal in the face of the existing usefulness of these models is kinda sad, if understandable. I mean, I get it, it’s kinda terrifying, and that can easily manifest as the kind of overly confident dismissal that you’re demonstrating here. But it just doesn’t match reality. I’ve been building software for 25 years, across multiple platforms and every kind of organization under the sun, and I’m using GPT4 to extreme usefulness right now, as are many others. And we’re just scratching the surface.

But by all means, continue to emphatically insist that LLMs have not demonstrated usefulness, or that they’ll never be able to “program on their own”.

ChatGPT plugins are in alpha as far as I know. 12 months seems like a reasonable timeframe.

I can totally see a set of plugins that write, debug and test code — all steps a human must now do — all automatically in the next 12 months. That, plus the GPT4-25k token model and you've got something that can replace high-level, high-boilerplate jobs or at least allow one person to 10x their output.

If it can reason code from the garbage programmers get handed, it’s AGI rather than LLM.

Reply on news.ycombinator.com