[cat_plus_plus]

I have no idea what everyone is talking about. LLMs are based on relatively simple math, inference is much easier to learn and customize than say Android APIs. Once you do you can apply familiar programming style logic to messy concepts like language and images. Give you model a JSON schema like "warp_factor": Integer if you don't want chatter, that's way better than Star Trek computer could do. Or have it write you a simple domain specific library on top of Android API that you can then program from memory like old style BASIC rather than having to run to stack overflow for evwery new task.

[layer8]

You can’t reason about inference (or training) of LLMs on the semantic level. You can’t predict the output of an LLM for a specific input other than by running it. If you want the output to be different in a specific way, you can’t reason with precision that a particular modification of the input, or of the weights, will achieve the desired change (and only that change) in the output. Instead, it’s like a slot machine that you just have to try running again.

The fact that LLMs are based on a network of simple matrix multiplications doesn’t change that. That’s like saying that the human brain is based on simple physical field equations, and therefore its behavior is easy to understand.

[orangecat]

That’s like saying that the human brain is based on simple physical field equations, and therefore its behavior is easy to understand.

Right, which is the point: LLMs are much more like human coworkers than compilers in terms of how you interact with them. Nobody would say that there's no point to working with other people because you can't predict their behavior exactly.

[cat_plus_plus]

What are you inputs and outputs? If inputs are zip files and outputs is uncompressed text, don't use an LLM. If inputs are English strings and outputs are localized strings, LLMs are way more accurate than any procedural code you might attempt for the purpose. Plus changing the style of outputs by modifying inputs/weights is also easier, you just need to provide a few thousand samples rather than think of every case. Super relevant for human coding, how many hobbyists or small businesses have teams of linguists on staff?