Scientists warn of AI collapse

We’ve all become used to AI-generated art in the form of text, images, audio, and even videos. Despite its prevalence, scientists are warning that AI creativity may soon die. Why is that? What does this mean for the future of AI? And will human creativity be in demand after all? Let’s have a look.

🔗 Join this channel to get access to perks ➜

#science #technews #ai #tech #sciencenews

[jalvrus]

"The more it eats its own output, the less variety the output has"... sounds exactly like the YouTube recommendations algorithm.

[femkeligtvoet]

This reminds me of when I was playing with play-doh when I was young. You start out with many different colors, and somehow always end up with a big brown ball.

[organfairy]

My biggest problem with AI is that it needs to get the information from somewhere, and sometimes these sources can be slightly dodgy. I did an experiment where I asked ChatGPT about some very narrow subjects: The Danish organplayer Peter Erling, the trio Klyderne, and the artist Jørgen Fonemy. These are subjects that I have some knowledge about and actually have written Wikipedia articles about. I could see that most of the answer that I got from ChatGPT was based on the exact Wikipedia articles that I wrote! I have tried to write the truth in those articles, but if I didn't care if things were correct - or worse; if I deliberately wanted to mislead people, then AI would base the answers on wrong data, if there wasn't multiple sources available. The problem I see with AI is that we trust it too much. Already now there are people who believe that it is an omniscient trustworthy source of all answers and that it will always be more correct than human knowledge or just knowledge that we have googled or looked up in an oldfashioned book.

[Koperviking]

I remember as a kid I used to record my voice and play it back on my speakers, which I then proceeded to record once more. By repeating it, I could hear how it slowly degraded until it was nothing more than a weird, synth-like sound.

[alieninmybeverage]

3rd possibility: AI learns how to gaslight us, and we forget how many legs elephants have.

[realpdm]

This could lead to a .. Nightmare on LLM Street....

[Ibhenriksen]

It's kinda like Google search engine. It started out sucking, then there was a time it was pretty good to find stuff. Now it sucks again....

[andrewhall]

This actually is not that suprising, when you think about it: these AIs are basically using huge amounts of data to approximate averages of various things, and with more iterations they extract more and more core features until they just have the same set of features they are using all the time. It's like taking data scores and continually averaging them until you are left with one value.

[MichaelDembinski]

A friend in the UK is a graphic designer; he says that over the past few months, more and more clients have been saying 'NO!' to AI-generated artwork - "it's too samey". They'd rather pay more for something original. Trouble is, AI has pushed down the rates; so while designers and artists are noting an uptick in requests for proposals, the money is much worse.

[markvoelker]

Apparently in the original Matrix movies storyline, the reason why the machines needed to keep those troublesome humans around was not as an energy source (“batteries”) but as a source of creativity. But the writers thought that this idea was too complex so they substituted the battery idea instead.

[tygorton]

Glad you started this conversation. There is also the theft component of generative AI. A YouTuber like yourself will get automatically copyright struck for using 4 seconds of a clip in a 20 minute original video. Yet these generative AI companies can use entire social media platforms with content painstakingly created by individuals across decades to create their data sets. This is peak hypocrisy in which, as per usual, corporate "big money" is protected while the individual is left with no means of defending their content. Generative AI is 100% theft in my opinion.

[Rosie-ufox]

I love that this underscores how complex human intelligence really is.

[robertruffo]

As someone who used to play with photocopiers as a kid... A copy of a copy of a copy is always much worse and weirder than you might think. Small flaws amplify until you all you get is a smudged blur.

[PhilMoskowitz]

Garbage In/Garbage Out. I've been saying the same thing about both AI and Analytics for the past decade and a half. People only want to look at processes, algorithms, ease of use, speediness, raw power, TCO, design and pretty UI with both AI and Analytics. You rarely hear people talk about things like bias, data integrity and context. Those three things only come into conversation when AI and Analytics produce horribly incorrect results.

[tchaika]

AI is geared to come up with a solution using the smallest amount of computations possible. It means that ignoring diversity and details is part of its basic make-up. It can't capture interesting quirks and details and spurt them out once in a while, in some outputs but not others. It also means that if it found one way to get to an acceptable solution, it will only try to get there quicker the next time around. If you've experienced getting stuck in a rut with ChatGPT, that's why.

[arifchagla]

That’s really interesting.. because comparing it to ‘bad cinema’, most bad cinema is bad in the same way, if that makes sense. Overused tropes, predictable storylines, cliche characterisation. Is there someone that can expand on this thought?

[joanlopez]

You reminded me of Pandora, the music recommendator that provided you with music according to the 👍 and 👎 that you gave to the songs proposed. No matter if you started with Black Sabbath, Chopin or Yunchen Lahmo. Eventually, after a couple dozen songs, you always ended up in a Coldplay loop.

[tamlynburleigh]

You make a good point. Already I can usually pick the ‘style’ of AI generated images. They have a certain ‘style’ because they are in a sense too perfect, too smooth, too balanced. It is not something one could define in some cases, but the human brain is good at recognising patterns.

[drachimera]

Sabine, as a professional in the application of machine learning in medicine I would like to thank you for making this video! It’s understandable and it reaches a lot of people! There is the AI hype (which people should not believe because it comes from executives and rookies) and there is the machine learning reality that veterans understand. This technology will be useful in automating some drudgery and common simple tasks…. It’s dogshit at doing anything truly valuable. What’s most worrying is the very real threat, without laws, that this nonsense will create such a firehouse of bullshit that we can’t get through our email, find what we need on the web, tell the difference between fact and fiction, and generally think for ourselves!

[billdavis]

I think Frank Herbert might have already told us the eventual solution in Dune.