At worse, newer models will get worse and you can just stick to older models.
You could also argue that proprietary models gated by an API are better than anything you can run locally, and yeah maybe those will get worse with time.
They're not going to get any worse than what you can run locally though. If they do, open models will overtake them, and then we'd be in a better position overall.
You can't run an up to date model locally. When I ask Googles models they have knowledge from stuff just a week ago, without using search. You wont get that from a giant local model.
An optimistic take would be that LLMs will make curation so much more valuable, that it will be done much better. If the wider world gets to use this curation to limit spam, and highlight good work, it would be amazing for the world.
Won't that also mean a return to a world of gatekeepers?
So I'm not nearly as pessimistic.
It's very unlikely for 100% of all intelligent curators to become dishonest either so it doesn't seem like a serious roadblock.
Open source AI needs to get a lot of investment for this to be mitigated. Relying on market incentives to drive development without the possibility of forking is too dangerous given the high stakes.
It’s just really clear that a giant text averaging machine can only go so far, and while we do see some higher level emergent properties, we’re going to have to move beyond the current state of the art in the next decade, and such future systems may be much less affected by the internet’s bullshit. Even without the bullshit singularity I think such measures would be a necessity and we would see them developed soon.
It's not really a text averaging machine, it's a pattern matching machine.
Right now the "depth" of the patterns it can match can only go so far, but in a few years with more advances in chips and memory the depth is going to increase and the patterns it can match will fan out accordingly.
Since all deployed models produce some kind of effect and feedback from the world there is an opportunity there to collect data targeted on the current level of the model, the most useful kind of data. That's why I think AI will be ok even with the proliferation of bots online. It's not 100% pure synthetic data in a loop, it is a agent-environment loop.
tl;dr Models learn better from their own experiences, not ours.
I just finished rereading Do Androids Dream for the first time in 20 years or so, and was astonished at how similar his andys really are to LLMs in the polluted / destroyed reality there. How confusing and corrupting they are to organic life. PKD describes the androif brains as neural networks with thousands of layered pathways and trillions of weighted parameters, and it's as if he was able to accurately conceive of what linguistic and "emotional" strengths and weaknesses those constructs would actually have, decades before LLMs existed. And there's this one amazing line where Deckard calls them "Life thieves". What else should we call what Sam Altman and others are building - but theft of human ingenuity, creativity, and basic reason for living, and the utter annihilation and suppression of people like this 16 year old kid who dare to hope they can contribute something more original in life than being a servant of a tech company building this shit, or a tiktok influencer who writes prompts?
What should that kid hope: That their work becomes noticeable enough to be immediately stolen and their name turned into a prompt?
Life thieves.
Just as a further aside, I had dinner tonight with a friend who's a fairly famous animator in the commercial realm, and I brought up this post. He's just sure the kid's screwed and the genie is out and creative is basically over. He's turning to building wooden clocks.
But his reaction, and the reactions I see here every time this comes up, remind me of something else. They remind me of how people react when someone is robbed. Everyone has some reason why it's bad but not that bad, it was inevitable, it'll be okay, etc. Or they go around wondering what they're going to do now. Or they paper it over with optimism. Surprisingly few people get robbed and are willing to realize they were robbed, and become wildly pissed off about it. Most people have some sort of flight reaction, as evidenced by e.g. the promotion of "prompt writing" or learning to use a paid API to do what was formerly your own creative job that now spits your own work back at you.
I say sue the shit out of all these content thieves.
If the AI was good at detecting it, it wouldn't matter if the AI-generated content sucked, yes?
Even a low probability of detection would help. Let's say our algorithm is 50% likely to detect AI junk. That means that half the junk data won't make it in to model. Even 20% would probably be worthwhile, especially if it also threw away human-generated junk (and let's be realistic here: there is, and always has been, no shortage of terrible and/or wrong human-generated content).
Let's say those crappy filler paragraphs that get stuck between pictures on meme clickbait pages... I suspect most of those people have already been replaced, but that prose was horrible long before the current LLM boom.
I suspect there is a lot of effort being expended right now on ways to ensure the training data (whether human or AI generated) isn't shite.
Then how can humans often tell that something is generated by AI?
Besides, I wasn't suggesting AI detection as such. I was suggesting crap detection, regardless of how the crap was produced.
You probably don't want crappy human-generated content going into your training set, either.
If we can only dicuss ideas which have never bee shared we can close this website right now
I'd add that majority of LLM-generated stuff put onto the internet isn't pure junk. It's endowed with human-created context and curation. If someone submits some LLM-created code to Github, it's because it works. A small fraction is pure noise (e.g. state-actors spamming Twitter) but that should remain a minority.
In the last internet revolution (web search), results started high quality because the inputs were high quality - bloggers and others just wanted to document and share knowledge. But over time, many interests (largely commercial) figured out how to game the system with SEO, and quality of search results has decreased as search's incentive structure led to lower quality data being indexed.
We're at the start of the LLM revolution now - models are trained on high quality inputs (which may be as rare as "low-background steel" in the future). But the models allow the mass production of lower quality outputs with errors and hallucinations; once those get fed back into new models, are we doomed to decreasing effectiveness of LLMs, just as we've seen with search? Will there be LLMO (LLM optimization) to try to get your commercial interests reflected in the next generation models?
I think we've got a few golden years of high quality LLMs before that negative feedback loop really starts to hurt like it did in search.