[adastra22]

Machine learning as a descriptive phrase has stopped being relevant. It implies the discovery of information in a training set. The pre-training of an LLM is most definitely machine learning. But what people are excited and interested in is the use of this learned data in generative AI. “Machine learning” doesn’t capture that aspect.

[simpleladle]

But the things we try to make LLMs do post-pre-training are primarily achieved via reinforcement learning. Isn't reinforcement learning machine learning? Correct me if I'm misconstruing what you're trying to say here

[adastra22]

You are still talking about training. Generative applications have always been fundamentally different from classification problems, and has now (in the form of transformers and diffusion models) taken on entirely new architectures.

If “machine learning” is taken to be so broad as to include any artificial neural network, all of which are trained with back propagation these days, then it is useless as a term.

The term “machine learning” was coined in the era of specialized classification agents that would learn how to segment inputs in some way. Thing email spam detection, or identifying cat pictures. These algorithms are still an essential part of both the pre-training and RLHF fine tuning of LLM models. But the generative architectures are new and very essential to the current interest in and hype surrounding AI at this point in time.

[hnuser123456]

It's a valid term that is worth introducing to the layperson IMO. Let them know how the magic works, and how it doesn't.

[adastra22]

Machine learning is only part of how an LLM agent works though. An essential part, but only a part.

[sdenton4]

I see a fair amount of bullshit in the LLM space though, where even cursory consideration would connect the methods back to well-known principles in ML (and even statistics!) to measure model quality and progress. There's a lot of 'woo, it's new! we don't know how to measure it exactly but we think it's groundbreaking!' which is simply wrong.

From where I sit, the generative models provide more flexibility but tend to underperform on any particular task relative to a targeted machine learning effort, once you actually do the work on comparative evaluation.

[adastra22]

I think we have a vocabulary problem here, because I am having a hard time understanding what you are trying to say.

You appear to be comparing apples to oranges. A generation task is not a categorization task. Machine learning solves categorization problems. Generative AI uses model trained by machine learning methods, but in a very different architecture to solve generative problems. Completely different and incomparable application domain.

[ainch]

I think you're overstating the distinction between ML and generation - plenty of ML methods involve generative models. Even basic linear regression with a squared loss can also be framed as a generative model derived by assuming Gaussian noise. Probabilistic PCA, HMMs, GMMs etc... generation has been a core part of ML for over 20 years.

[sdenton4]

And yet, people very often find themselves using generative models for categorization and information retrieval tasks...

[IshKebab]

How does "it's called machine learning not AI" help anyone know how it works? It's just a fancier sounding name.

[hnuser123456]

Because if they're curious, they can look up (or ask an "AI") about machine learning, rather than just AI, and learn more about the capabilities and difficulties and mechanics of how it works, learn some of the history, and have grounded expectations for what the next 10 years of development might look like.

[IshKebab]

They can google AI too... Do you think googling "how does AI work" won't work?