In giant language mannequin collapse, there are usually three sources of errors: The mannequin itself, the best way the mannequin is skilled and the info — or lack thereof — that the mannequin is skilled on.
Andriy Onufriyenko/Getty Photos
cover caption
toggle caption
Andriy Onufriyenko/Getty Photos
In giant language mannequin collapse, there are usually three sources of errors: The mannequin itself, the best way the mannequin is skilled and the info — or lack thereof — that the mannequin is skilled on.
Andriy Onufriyenko/Getty Photos
Requested ChatGPT something currently? Talked with a customer support chatbot? Learn the outcomes of Google’s “AI Overviews” abstract function?
In case you’ve used the Web currently, chances are high, you’ve got been consuming content material created by a big language mannequin.
Massive language fashions, like DeepSeek-R1 or OpenAI’s ChatGPT, are form of just like the predictive textual content function in your cellphone on steroids. To ensure that them to “study” how you can write, these modesl are skilled on hundreds of thousands of examples of human-written textual content.
Up to now, this coaching often concerned having the fashions learn the entire Web. However these days — thanks partially to those giant language fashions themselves — a lot of content on the Web is written by generative AI.
That implies that AI fashions skilled now could devour their very own artificial content material — and undergo the results.
View the AI-generated images talked about on this episode.
Have one other subject in synthetic intelligence you need us to cowl? Tell us my emailing shortwave@npr.org!
Take heed to Brief Wave on Spotify and Apple Podcasts.
Pay attention to each episode of Brief Wave sponsor-free and help our work at NPR by signing up for Brief Wave+ at plus.npr.org/shortwave.
This episode was produced by Hannah Chinn. It was edited by our showrunner, Rebecca Ramirez. The audio engineer was Jimmy Keeley.
