Hacker News Reader

Comment by Havoc | Hacker News Reader

This sounds promising. Smaller but custom trained/tuned models would be ideal - works for the task without the overhead

Agreed, mostly in real life we have specialized experts. You go to doctor to ask about health related stuff, you ask your colleague who is expert in ML only about ML and probably not much about mobile development.

Instead if having ML that is expert at coding in all languages probably would be better to allow to switch context.

Native mobile dev LLM? - e.g. train only on swift, objc, kotlin, java, c, c++ code

Python dev LLM? - train only on python, c, c++, rust code

Would this way final model be smaller, faster and maybe event better?

In my experience it doesn't work like that. It's like if you take an idiot and spend a bunch of time training them, they'll still perform much worse than a moderately intelligent person with much less training. And smaller models can be pretty idiotic.

Then why do Chess AI perform much better than LLMs trying to play chess.

Because they do completely different things? They literally have nothing to do with each other. Why do planes fly better than ships if ChatGPT can't do math?

Provided that the problem is suited to the strengths of an LLM at all. An example might be a small ai custom trained on documentation for libraries. You ask it a question like "how do I make the background move with parallax effect when you move the cursor". It's a little ambiguous, high-level concept, and probably not a single function.

Small ai: likely makes up a function or suggests a single function which isn't sufficient. Refuses to budge from its answer or apologies and gets confused

Large LLM: able to actually understand the question, combine several functions. If it doesn't work you can tell it why and it fixes it

Why would a language model be good at playing chess?

Why no? Chess notation is text. But the problem is that LLMs are not that good for problems which require evaluation of a search tree. Also leading chess engines such as lc0 are without search better than 90+x% of All humans

Because there’s a world of difference between a reinforcement learning trained special purpose model and asking a general purpose large language model to have a go at something.

Because they have an explicit model of chess and specific heuristics for learning chess.

An LLM could have picked up some chess patterns through osmosis, but it can not reason explicitly in the domain.

No"they" (lc0) don't have specific heuristics

Unfortunately, for transformer-based LLMs the magic starts only when they are trained by more that 10^22 TFlops (preferably 10^24) so smaller models might not cut it even for fine-tuned tasks.

At medium size (13B) Microsoft Orca demonstrated you can trade off size with larger fine-tuning dataset.

Any references on this?

https://www.youtube.com/watch?v=tVtOevLrt5U

Agreed.

Reply on news.ycombinator.com