Anyone who argues that LLMs are "just stochastic parrots" fundamentally doesn't understand what neural networks do. The power does not come from sampling from a distributions over words but from the multi-dimensional representation. And it is that that enables LLMs to be more than just mechanisms that produce copies of material previously seen in training.
The specific AI technology being demonstrated isn't an LLM, it's a different kind of model that renders Quake from memory. It's very much trained to copy Quake.
John wasn't really talking about the Quake demo where each frame is rendered by generative AI. He started by saying:
> I think you are misunderstanding what this tech demo actually is, but I will engage with what I think your gripe is — AI tooling trivializing the skillsets of programmers, artists, and designers.
By the time he got to "power tool" he was talking about something else entirely.
It can copy/simulate anything with a stream of pixels that change in response to labels equivalent to button presses.
Put an iPhone on the dashboard of a car and simultaneously record steering wheel, pedal, and gear positions, and you've got a photorealistic driving simulator of wherever you were driving that car.
Strap it to a bodycam and pair with annotated recordings the wearer's motion as if it were control inputs, and you have a photrealistic first-person simulation of whatever they were doing, be that paintball, skating, US Army urban combat training, or birdwatching.
Feed it sports TV annotated with player motion, golf, tenis, cricket, soccer, etc.
You're still just saying that while stealing a lot from one person is wrong, stealing only a little but from many people would be OK.
And the people who are stolen from don't agree with you.
Is an artist finding inspiration from another's work stealing? Creativity is very rarely building something completely new from nothing at all - it's the end result of inspiration and a merging of ideas
Which seems to be exactly what these models are doing
Satisfactory the game didn't "steal" from Factorio - they took the idea of an automated factory game from them, brought it into a 3d world instead of 2d, and pulled in some other ideas, perhaps from other games. This is a hell of a lot closer to stealing than an LLM taking 0.001% each from 100000 images to create something new.. and yet no-one would ever call it such
It's disturbing because it's automated on an insane scale, not because of a solid legal ground
That's not what's happening. Using your analogy, you're reading the book, remembering it, and then making N more books (maybe for commercial purposes) by using what you remembered to create something very similar in terms of style, prose, plot, and so on. As a result, the person who you learned from can't get a job writing anymore because their work has been commoditized. Also, they're feeling lost because the work they devoted their lives to is now awash in a sea of similar work.
if I quote it from memory without attribution, it is arguably stealng. Certainly in bad taste to pretend that was my idea. If I type it out from memory on my blog page without permission, it is 100% stealing.
These aren't novel situations. We have centuries of case studies, precedent, and cultural osmosis giving us legal and de facto means of what we feel and say is "stealing intellectual property".
What bunkum. Stealing is only recent propaganda, and not successful at that. If anything, the act of sharing something is cultural and innate for millennia even [0].
"recent" we in 300 or so years? Yes. And guess why we made copyright to begin with? It became very easy for large companies to copy and mass produce ideas from people who made them. History repeats.
It's not like tribes and families also didn't keep secrets to maintain some power or as matter of trade. We call them "trade secrets" now but those go back before recorded history.
In addition, consider thar back the we had actual communities, unlike now. Bring those back and we can talk about tearing down copyright.
Stable Diffusion and the like is unable to perfectly reproduce an image from the training data, outside of rare exceptions, e.g. it was overrepresented in the training set.
Anyone who argues that LLMs are "just stochastic parrots" fundamentally doesn't understand what neural networks do. The power does not come from sampling from a distributions over words but from the multi-dimensional representation. And it is that that enables LLMs to be more than just mechanisms that produce copies of material previously seen in training.