if it starts to ingest that data it will only get more wrong over time. Unless it also ingest the replies that say "ChatGPT is full of shit here?"
At least, unlike Bing Chat, it apologizes when corrected instead of gaslighting me.
[1] https://www.reddit.com/r/tipofmytongue/comments/118id2f/tomt...
And it just reminds me of the Steve Martin line about playing a joke on young children: “Talk wrong“.
Seems to already be happening with Bing Chat, except it led to the bot threatening the user: https://twitter.com/tobyordoxford/status/1627414519784910849...
You seem to assume that it would be particularly confused about its own content
I made vxai [1] to do that for movies. Sometimes it shines sometimes not su much.
[1] https://vxai.com
This isn't a huge problem, it's barely a problem.
It’s a standard feedback loop.
The amusing thing is that controlling nonlinear multivariate feedback systems is an unsolved problem.
ChatGPT trains Bard which trains Bing which trains ChatGPT, etc.
Then the watermark should be made public. We need a way to tell ai-generated content from human-generated content without actually having to read it.
In what way can millions of lines of auto generated additions of text to the internet, be suddenly and reliable expected to have all been somehow curated by humans?
I used to think the same but after reading and learning some more, I realized not.
Playing one AI against another is an established technique to developing AI.
Furthermore, content on the internet will always vary from more reliable (well established wiki pages, Reuters) to less reliable (random blog posts, disinformation).
Whether or not an AI generated text doesn't seem to be that important - what's more important is how reliable it is, and how well humans engage with it.
AI that benefits from the kind of adversarial training that you mention are mostly _planning_ type AIs (think AlphaGo). The problem these AI systems try to solve is that you have some constraints, and as a human "can't be bothered to" work out an optimal solution, so you have the computer do it for you and it does so by starting off with a bad estimate of a solution and "improving" it by trying stuff.
LLMs, on the other hand, are more of a modelling/compression type AI --- (well, more traditionally it wouldn't even be considered AI per-se due to the lack of planning capability). The problem here is to try to represent a huge swath of data in the as efficient a way as possible, thereby forcing it to find and collapse "patterns" by adjusting the representation. Here it's generally not the case that you want to train with adversarial distributions.
An easy thought experiment is: say you take a single case (e.g. a game of Go or a paragraph of text) and massively over-represent it to both models, so much so that that single case eventually covers 99% of all your training instances at the end. For a "planning" AI this over-representation isn't a huge deal --- once it's learnt all it can from that case, seeing it again is but a waste of time*. It merely makes that particular plan more "clear" but not more "desirable". However, a modelling/compression type AI will continuously adjust to adapt to the increasing occurrence of that case. It truly "believes" that the more often it sees some pattern, the more important it is, right until it has "forgotten" everything else.
*: This is kind of an over-simplification.
1) We may be able to model a subset of chatGPT's abilities as adversarial questions. For example, can we write an AI that finds sources and generates questions that a solving AI should be able to figure out. Can we write a test framework such that AI's can challenge themselves to write optimal code for a given solution. Etc.
2) Like I mentioned, if you're scraping the internet you are inherently needing to build some kind of relevance model. E.g. highly up items answers on stack overflow have more weight. In said situation whether or not the content is written by human is largely irrelevant - if you have a reasonable ranking method than highly ranked content is important regardless of source.
Eventually someone's going to write an "AI Trap" that serves up a seemingly infinite forum or reddit-style site, but is actually just generating an endless stream of (non)consciousness from some LLM chatbot.
[0] https://en.wikipedia.org/wiki/Spider_trap
[1] https://www.gsp.com/support/virtual/web/cgi/lib/wpoison/
It actually reminds me of an early part of Greg Egan's "Orphanogensis" where the orphan has randomly stumbled into a working resource, but doesn't yet understand how to navigate properly, so it just keeps accessing that same resource, over and over, that one works, other random nonsense doesn't work, that one works, the other random picks do not.. until it finally figures out some basic navigation and then it leaves.
An AI like a human may read a few pages of your forum about Invisible Moon Donkeys or whatever, but after not long its interest is sated, the forum's location may be noted down for future interest "OK, more about Invisible Moon Donkeys here" but the AI moves on, what's this "Hacker News" forum about ?
GPT could filter out anything they themselves emitted in future trains, yeah? Because they know what their bot's said. They get the benefit of looking at a conversation, knowing reasonably well what's copy/pasted from ai.com and what's the exasperated expert trying to correct a doomed world :p
The only way it eats itself is 1. Colossal mistakes. 2. Everyone decides to get off the internet and go outside.
2 seems pretty unrealistic, we put up with a lot :D
There is no issue with AI ingesting data from itself in itself. Humans do it as well. That data might even be higher quality than human data. The scale at which humans produce data will most likely stay higher than AI data for a long time.
There is already bot data out there from lower quality AIs/bots, and chatGPT has ingested it.
LLMs are made to be good at some textual tasks, and not for what they're being used right now. They're not information stores, or Q/A. It only answers what a human is likely to answer.
But this gets at the heart of the issue - separation. If we ensure that humans and AI are trained on roughly the same data then we will stay connected and be able to understand each other. We may even end up borrowing a few gpt-isms, and that's actually totally fine.
Back in 2011, Google faced the same problem mining bi-texts from the Internet for their statistical machine translation software. The thought was that one could utilize things like multi-lingual websites to learn corresponding translations.
They quickly realized that a lot of sites were actually using Google Translate without human intervention to make multi-lingual versions of their site, so naive approaches would cause the model to get trained on its own suboptimal output.
So they came up with a whole watermarking system so that the model could recognize its own output with some statistical level of certainty, and avoid it. It wouldn't be surprising if this is being done for LLMs too. The more concerning problem is when different LLMs, who are not aware of each others' watermarks, end up potentially becoming inbred should the ratio of LLM content rise dramatically...
As long as you agree with the new facts, you’re fine. Problem solved!
“ChatGPT, a version of OpenAI’s GPT-3.5 model… gained more than 100m users in its first two months, and is now estimated to produce a volume of text every 14 days that is equivalent to all the printed works of humanity.”
— Dr Thompson, Feb/2023, cited in report by the National Bureau of Economic Research (Scholes, Bernanke, MIT)
https://www.nber.org/system/files/working_papers/w30957/w309...
This is of course also a necessary condition for ChatGPT to come up with original insights. Except perhaps where it comes to things like fiction, which probably has value in itself.
seems more like it's gonna eat its own vomit, degrading it (maybe not completely) to inbreed (?)
Maybe our view of AI is being colored by sci-fi stereotypes of robots malfunctioning when asked to compute really hard problems generating infinite recursion. I'm not so sure that LLMs will totally destabilize. We might see some interesting output, but I don't think we know yet whether the stability of the system will merely fluctuate as a whole without falling apart.
