pvr
Casey NewtonI wrote about some of the early lessons from unleashing generative AI on the on the internet. Chatbot text is spreading quickly — and it may be corroding the web https://www.platformer.news/p/the-ai-is-eating-itself
Liroy Leshed
Joe ❌👑@caseynewton They will need some newer data, so the chatbot isn't permanently living in 2021. Perhaps they could take it only from Wikipedia and whatever reliable news media they can get permission for. Even those could have some contamination.
Andy@not2b @caseynewton Wikipedia assuredly has GPT text. It just makes sense for someone to decide they’re doing the world a favor and having GPT write articles.
@pierogipowered @caseynewton There is probably some, but as a proportion it is going to be a lot less than what an unfiltered 2024 snapshot of the web will be like.
Fabien Niñoles@caseynewton interesting article. However, there is already models trained (or teaches like they said) by other models. And that create better models faster. I agree that we need better validation, but I doubt that the internet was a valid source for that. You need better curation if you're looking for truth.
Earthy Tonez@caseynewton Definitely feels like the web is at a transition point... Tipping point? Google talking about moving away from the 10 blue links, twitter and reddit nodding off.
Seems like in a few years no one will recognize the web any more. Also seems like a backlash against non-mainstream users?
Richard Schneeman@caseynewton reminds me of how “pre (nuclear) bomb” steel is needed for research involving radioactivity
chris@Schneems what a great analogy
Dustin Mitchell@caseynewton a further observation: much of the commercial AI out there right now is for generating text. The next largest category is AI that reads (summarizes, categorizes, analyzes, indexes, etc.) text. In aggregate, we are building a very expensive, lossy way to encode simple ideas into verbose "content" that no human will consume.
Craig Hockenberry@caseynewton The AI feedback loop is basically attacking itself and I’m guessing it will increase exponentially. https://en.wikipedia.org/wiki/Adversarial_machine_learning
Jonathan Joelson@chockenberry @caseynewton I forget where I read this, but this is basically the opposite of the Singularity.
@caseynewton
Clearest value prop for ‘AI’ is to steer scrutiny away from VCs, who have burned more cash in last decade than bankers up to the ‘08 financial crisis.
LLMs grab human written text from the internet and squish it into a million or so human written templates (the ‘transformer’).
Glorified Lorem Ipsum.