You have to remember one thing, writing or speaking of a language is not a fixed scientific law or math formula that will stay true through out history. A living language is always moving and evolving in most of its components, be a vocabulary, grammar, or even meaning of words/phrases. We are just entering an era where AI generated content someone might feel appealing and follow that style, compare to copy a contemporary popular writer.

This is a classic feedback problem: you use a microphone to amplify your voice, but If the mic picks up the amplified sound it creates audio feedback + a sharply increasing wail.

I can’t imagine what LLM feed back ‘sounds like’, but a guarantee you it ain’t pretty.

I’ve heard this theory. Feels like unrealistic hopeful wishes of people who want AI to fail.

LLM processing will be a huge tool for pruning and labeling training sets. Humans can sample and validate the work. These better training sets will produce better LLMs.

Who cares is a chunk of text was written by a human or not? Plenty of humans are shit writers who believe illogical or clearly incorrect things. The idea that human origin text is superior is a fantasy. chatGPT is a better writer than 80% of humans todat. In 10 years LLMs will be better than 99.9% of humans. There is no poison to be avoided.

chatGPT has an apparent style when used in the default mode, but you can already get away from that with simple prompt tweaks. This whole thing is a non-issue.

I suspect the quality LLM development teams will pursue the same in-depth data sourcing & cleaning techniques that quality ML researchers are developing today. Or rather, they’ll do something similar in principle to mitigate this issue.

I still agree with your conclusions. It will be a bigger consideration and less scrupulous teams will be more effected.

I think it could end up being a problem that we face in the future, but probably not an insurmountable one.

For one, I suspect that clean data sources will always be available, though it could become a lot more expensive to obtain. As an extreme example, you could always source your data by recording in-person conversations.

Also, as AI improves, I’m guessing it will be able to handle bad data more gracefully, and that it should be able to train to the same effectiveness while using a smaller dataset.

thanks, it’s terrifying

I’m not sure this is true. They could be trained based on published works prior to a certain date as the formal writing style, eg Project Gutenberg, then layer on the recent internet to better capture modern stylistic trends.

Ultimately, the models will always require fine tuning, and selecting which data set you use for early training has a very large impact on the overall performance of the model. Additional knowledge and trendiness can be learned after the fact.

Well thankfully we have tons of archived conversations before ChatGPT on archive.org.

Many forums with hundreds of thousands of posts are just waiting to be scraped.

There has already been jokes of AI being used to create well crafted correspondence, then another AI translating that into a short summary.

I think you are going to see AI as something people lean on more to talk to others, and that is going to create its own language where AI talks to AI.

There’s usually a context difference that might might be significant. People don’t write the same way way for an email, like they would a letter, text message, or tweet.

They might write more like an LLM for things like essays and reports, but your usual writing is probably still fine. Then classics that inspire people to write are still around, and I doubt that they would be supplanted by an LLM any time soon.

We might start being in trouble if people start republishing books with them, but that’s unlikely to to happen any time soon, considering the current state of copyright around AI works.

Thanks, I hate it

There are some in the research community that agree with your take: THE CURSE OF RECURSION: TRAINING ON GENERATED DATA MAKES MODELS FORGET

Basically the long and short of that paper is that LLMs are inherently biased towards likely responses. The more their training set is LLM generated, and thus contains that bias, the less the LLM will be able to produce unlikely responses, over time degrading the model quality throughout successive generations.

However, I tend to think this viewpoint is probably missing something important. Can you train a new LLM on today’s internet? Probably not, at least without some heavy cleaning. Can you train a multimodal model on video, audio, the chat logs of people talking to it, and even other better LLMs? Yes, and you will get a much higher quality model and likely won’t get the same model collapse implied by the paper.

This is more or less what OpenAI has done. All the conversations with 100M+ users are saved and used to further train the AI. Their latest GPT4 is also trained on video and image recognition, and they have also been exploring ways for LLMs to train new ones, especially to aid in alignment of these models.

Another recent example is Orca, a fine tune of the open source llama model, which is trained by GPT-3.5 and GPT-4 as teachers, and retains ~90% of GPT-3.5’s performance though it uses a factor of 10 less parameters.

Slightly unrelated, but I was just talking with a friend about how we’re going to have similar issues with young artists trying to copy ai. As is, many young artists will turn to cartoons instead of real life when starting out. Their work is a bastardization of a bastardization, with serious flaws in anatomy, gravity, light, and depth. They go on to call those mistakes their “style” and point to other artists making those same mistakes to normalize them. Since “style” isn’t something they think they need to improve on, they may become good artists overall while having severe, glaring holes in their skillet that any professional can see. You can sometimes even tell when someone started out because “90s anime” or “10s cartoon network” made specific stylistic choices that changed over time.

So I think ai is going to cause similar problems. Newbies will copy what looks pretty to the untrained eye and learn an ai based style. Then when they become more popular they’ll be fed into ai as reference material and perpetuate the problem. Even worse is actual professionals may turn to ai instead of real life references or a desk mannequin. Then their skills may degrade because they rely too much on improper tools. (I’ve already seen this becoming an issue with photoshopped reference photos.)

Anyways, that’s my $0.02

This sounds like what an ai would write /s

I think that while LLMs are going to get worse, the AI software will get better to the point of strong AI, and it will do a lot of “apple-esque” changes to mass produced speech that will ultimately be for the better… The cynical possibility is that it will further taint human dialogue even though it could provide a better way.

What I predict is that they’ll try to implement data filters to avoid feedback loops, but there will be an enshittification process for AI too.

What might put the nail into the coffin much quicker isn’t the feedback loop, but trying to monetize the whole thing. I think it’s only a matter of time until OpenAI will try to get money off of it, like putting certain features behind a paywall, especially those that professionals might use.