5 ms·

I don't get this reasoning. Without LLMs I would learn how to write sub-optimal code that is somewhat functional. With LLMs instantly see "how it's done" for my

I don't get this reasoning. Without LLMs I would learn how to write sub-optimal code that is somewhat functional. With LLMs instantly see "how it's done" for my exact problem case which makes me learn way faster. On top of that it always makes dumb mistakes which forces you to actually understand what it's spitting out to get it to work properly. Again: that helps with learning. The fact that you can ask it for a solution for exactly the context you're interested in is amazing and traditional learning doesn't come close in terms of efficiency IMO.

dingnuts 1y ago
> With LLMs instantly see "how it's done" for my exact problem case which makes me learn way faster. No, you see a plausible set of tokens that appear similar to how it's done, and as a beginner, you're not able to tell the difference between a good example and something that is subtly wrong. So you learn something, but it's wrong. You internalize it. Later, it comes back to bite you. But OpenAI keeps the money for the tokens. You pay whether the LLM is right or not. Sam likes that.
- Spivak 1y ago
  This makes for a good sound bite but it's just not true. The use case of "show me what is a customary solution to <problem>" plays exactly into LLMs strength as a funny kind of search engine. I used to (and still do) search public code for this use case to get a sense of the style and idioms common in a new language/library and the plausible set of tokens is doing exactly that.
- margalabargala 1y ago
  > So you learn something, [...] You internalize it. Or they don't.
layer8 1y ago
It’s more like looking up the solution to the math problem you’re supposed to solve on your own. It can be helpful in some situations, but in general you don’t learn the problem-solving skills if you don’t do it yourself.
- Ekaros 1y ago
  Exactly. For vast majority of students myself included just looking at ready solution is actually very poor way to study. And LLMs are exactly this. Ready solution generators. Doing with things like math and programming is learning. And same goes for art. You do not become master in art by looking at art or even someone drawing...
hackable_sand 1y ago
I would recommend programming, and designing your system, on a piece of paper instead. It's the most efficient few-shot which beats the odds on any SotA model.

dingnuts 1y ago

> With LLMs instantly see "how it's done" for my exact problem case which makes me learn way faster. No, you see a plausible set of tokens that appear similar to how it's done, and as a beginner, you're not able to tell the difference between a good example and something that is subtly wrong. So you learn something, but it's wrong. You internalize it. Later, it comes back to bite you. But OpenAI keeps the money for the tokens. You pay whether the LLM is right or not. Sam likes that.

Spivak 1y ago
This makes for a good sound bite but it's just not true. The use case of "show me what is a customary solution to <problem>" plays exactly into LLMs strength as a funny kind of search engine. I used to (and still do) search public code for this use case to get a sense of the style and idioms common in a new language/library and the plausible set of tokens is doing exactly that.
margalabargala 1y ago
> So you learn something, [...] You internalize it. Or they don't.

layer8 1y ago

It’s more like looking up the solution to the math problem you’re supposed to solve on your own. It can be helpful in some situations, but in general you don’t learn the problem-solving skills if you don’t do it yourself.

Ekaros 1y ago
Exactly. For vast majority of students myself included just looking at ready solution is actually very poor way to study. And LLMs are exactly this. Ready solution generators. Doing with things like math and programming is learning. And same goes for art. You do not become master in art by looking at art or even someone drawing...

hackable_sand 1y ago

I would recommend programming, and designing your system, on a piece of paper instead. It's the most efficient few-shot which beats the odds on any SotA model.