Hacker News Reader

Comment by harshreality | Hacker News Reader

No, but large established copyright interests will go to the ends of the Earth arguing that it is.

https://www.youtube.com/watch?v=5pIVVpoz5zk

It is the same as humans learning from art to imitate one or multiple other artists' styles. The differences are in scale, and flexibility in mixing and matching different styles (a very primitive, not-very-accurate model or description would be a vector of [0,1] values, one for each artist[1]), and which vector an AI art program happens to use as a default style for any prompt that doesn't specify something akin to "in the style of <x>"

[1] I'm not claiming these neural nets are learning models of each artist and linearly combining them, but if you squint hard enough it approaches that.

View on HN

Everyone knows that the term machine learning doesn't actually have anything to do with human learning, right? It's just a set of words that were catchier than gradient descent or curve fitting.

Asserting legal rights based on a naming quirk seems pretty unfounded.

Let's try that statement again: 'It is the same as executing gradient descent over multiple artists works' That sounds a heck of a lot closer to running a Photoshop filter than learning anything now doesn't it.

I fail to see how that last sentence makes it sound like a photoshop filter. Do you mean if you can express it as an algorithm it cannot be compared to human learning?

What if we can reduce human learning to making predictions based on training data? Then it's not human learning anymore but more like a photoshop filter?

"Everyone knows that the term machine learning doesn't actually have anything to do with human learning, right?" Do you have sources for that? They're not equal but I see a lot of parallels.

When someone successfully reduces human learning to predictions on training data, I will celebrate the Nobel prizes they are awarded for medicine, philosophy, and possibly physics and economics.

Until then maybe we shouldn't just assert that it's true.

Until then, we shouldn't just assert that it's false either.

GPT-3 is coming quite close to being able to write novel convincing research papers. With GPT-4 and GPT-5 on the horizon I don't see why language models couldn't be significantly involved in winning a Nobel prize in the future. At least I wouldn't categorically exclude that possibility.

> Everyone knows that the term machine learning doesn't actually have anything to do with human learning, right

Unfortunately the answer is definitely no. People have heard the terms “learning” and “train” in this context and have anthropomorphized weak AI.

The level of ignorance on display is sad to witness.

Deep learning is definitely inspired by biological learning and also shows surprising parallels in terms of for example how language rules are learned by children.

"Machine learning doesn't have anything to do with human learning" is just not true. There are many parallels but sure they are not equal.

Just happened across this paper:

Theory of Mind May Have Spontaneously Emerged in Large Language Models

Theory of mind (ToM), or the ability to impute unobservable mental states to others, is central to human social interactions, communication, empathy, self-consciousness, and morality. We administer classic false-belief tasks, widely used to test ToM in humans, to several language models, without any examples or pre-training. Our results show that models published before 2022 show virtually no ability to solve ToM tasks. Yet, the January 2022 version of GPT-3 (davinci-002) solved 70% of ToM tasks, a performance comparable with that of seven-year-old children. Moreover, its November 2022 version (davinci-003), solved 93% of ToM tasks, a performance comparable with that of nine-year-old children. These findings suggest that ToM-like ability (thus far considered to be uniquely human) may have spontaneously emerged as a byproduct of language models' improving language skills.

Actually I think I use gradient descent in a lot of situations.

Consider the act of adding salt to chicken broth for example but I can quite easily think of dozens of other things.

Absolutely. Gradient descent is a fantastic technique for a lot of problems. This and many other algorithms can be performed by both humans and computers and give useful results.

Notice neither your or my comments connect any of this to learning.

After numerous times making chicken broth, something will change about your brain/body and we will say 'you learned how to season broth.' We have basically no idea what this process is, or whether it relies on gradient descent or anything else.

Hmm okay I see your point now.

Machine learning is not human learning in any meaningful or legal way because laws are for humans and crucially machine is not human.

It will be the case as long as that machine lacks free will, human rights and all that important stuff. If it gets to a level where a machine is granted human rights this could be reassessed. Until then, whoever operates the machine is to whom laws apply, and in this case they violate copyright by capitalizing on others’ creative work and intellectual property through creating derivative works.

The large interests in question are guys like OpenAI/Microsoft, and it is in their interest that we continue to see AI as some human-like enigma and don’t think about the ethics of what they do. If you ask actual small artists, most do not in fact share this enthusiasm.

> No, but large established copyright interests will go to the ends of the Earth arguing that it is.

Small independent copyright interests trying to eke out a living from their art will also argue this.

I love how both sides in this debate immediately claim the other side is huge, faceless and corporate.

In truth both sides are a mix of both individuals and monied interests but this rhetorical tactic seems to be irresistible.

This is also a few big firms in a way harvesting the public domain - not sure we need new forms of rent seeking via AI.

Scale matters. And there is no reason we have to have the same rules for people and for software tools.

If you're learning from existing works at scale, aren't you less infringing than someone who only learns from a handful of existing works?

Reply on news.ycombinator.com