Hacker News new | ask | show | jobs
by runarberg 914 days ago
Doesn’t that just mean it was trained with inputs and outputs from simple programs?
2 comments

No, it's actually executing that Python code. This is what allows an LLM (or an 'llm based system', I guess) to do something like "reverse <some uuid that has never been observed before>" - it can't just memorize the output and map it to the input because the output has literally never been observed. Instead, if it knows the algorithm for reversing a string, it can just use that and offload the execution to Python.
That is even less impressive. I was thinking—like normal linear models—it would be capable of interpolation.
It is. It doesn’t even need an existing language. You can define your own psuedo language in the prompt and have ChatGPT “execute” it (works best with 4 nonturbo).

You can even combine your pseudo language with natural language. See the OP’s custom GPT and the comments here: https://news.ycombinator.com/item?id=38594521

That looks totally different. In the case of the Python code it is literally executing it by calling out to CPython.
…got a source for that claim?
Why is it less impressive?
I would say creating a model which is able to interpolate from training data in a way which produces an accurate output of a new input is a little impressive (if only as a neat party trick), however anybody can run a python interpreter on a server somewhere.

I’m sure there are use cases for this. But in the end it is only a simple feature added onto a—sometimes—marginally related service.

Hm, I don't think of it that way I guess. What the LLM is doing is generalizing a problem based on previous problems it has seen and then offloading the execution of that problem to a machine with some defined, specific semantics.

This is a lot more than a party trick. The model is able to describe the program it wants to execute and now it can accurately execute that - that it 'offloads' the work to a specialized program seems fine to me.

It's way more than a simple feature, this is enabling it to overcome one of the biggest limitations and criticisms of LLMs - it can answer questions it has never seen before.

So it only works if the code is 100% correct. Which according to my experience doesn't happen often.
Sure, don’t expect perfection, but it still works great.
Most of the time it's just doing really simple stuff like this.