Hacker News Reader

Comment by disgruntledphd2 | Hacker News Reader

I think that this is a great article on all of this: https://substack.com/home/post/p-224216856

He distinguishes between weak recursion (using sotfware to build more software), and strong recursion (handing over model development to AI which does everything), and makes a pretty strong case for the fact that only weak recursion is happening.

This maps to my experience with AI models, they reduce the cost of experimentation massively, but they can't (currently) design experiments themselves.

Even then, training RL is very data hungry, and most profitable tasks don't have good data so even if we got superintelligence (whatever that means), it's entirely possible that progress would still be bottlenecked by lack of high quality training data.

View on HN

Thats a great distinction.

The hope was that weak recursion could produce strong recursion by automating the iteration of ML pipelines.

Theres also the idea of LLMs producing synthetic data to contribute to new models. Perhaps that is not as effective as hoped.

> the idea of LLMs producing synthetic data to contribute to new models

This never made sense to me. Seems analogous to inbreeding in biological organisms.

It works because artificial neural nets take a lot of examples of something to learn it (way more than natural neural nets). So at the simplest level it's just copying items in the training data enough that they become salient.

More complex synthetic data generates things that could be in the training data but aren't for some reason. For example, machine translation training corpuses often contain a lot of formal language (because formal documents are the ones that get translated), so you might generate a bunch of casual conversation to improve the translation model's ability to handle that.

> Theres also the idea of LLMs producing synthetic data to contribute to new models. Perhaps that is not as effective as hoped.

Wasn't model collapse/model decay a big issue since the beginning? i know that RL kinda mitigate the effect, but if this was close to not be an issue anymore, wouldn't we have a paper on that already? All that i've seen on the subject was about mitigation, never about solving the issue.

Sorry if my question is dumb, I need to do some research about the difference between image generator and text generator, i feel like my understanding of LLM is skewed by my usage and understanding of Z-turbo, but i only have that much time in my week.

Reply on news.ycombinator.com