Conversation

Notices

1. Embed this notice
   Yogthos (yogthos@mas.to)'s status on Thursday, 25-Jul-2024 02:16:04 JST Yogthos

   AI models collapse when trained on recursively generated data.

   #machinelearning

   https://www.nature.com/articles/s41586-024-07566-y

   • Embed this notice
     on-lain ✔ᵛᵉʳᶦᶠᶦᵉᵈ (lain@lain.com)'s status on Thursday, 25-Jul-2024 02:16:02 JST on-lain ✔ᵛᵉʳᶦᶠᶦᵉᵈ

     in reply to

     • feld
     • Maybe(oz)
     @feld @yogthos @oz yeah, filtering out bad data and creating good data is the current main problem in the field, many think that synthetic data generation is the way to go, and both llama3 and WizardLM use it. adding 'tarpits' to random indieweb blogs will not do much.
   • Embed this notice
     Maybe(oz) (oz@mastodon.social)'s status on Thursday, 25-Jul-2024 02:16:03 JST Maybe(oz)

     in reply to

     @yogthos so, if you don't like to see contents from LLMs, a solution is to publish lots of LLM-generated contents, in the hope that they'll eventually collapse.

     It looks like folks are already making tarpit webpages specifically for (Open)AI bots, in an effort to feed garbage to the model. 🤔

   • Embed this notice
     feld (feld@bikeshed.party)'s status on Thursday, 25-Jul-2024 02:16:03 JST feld

     in reply to

     • Maybe(oz)
     @oz @yogthos the models are not going to work well unless the data its trained on is carefully curated. The LLMs that are trained on high quality data are really really good
   • Embed this notice
     feld (feld@bikeshed.party)'s status on Thursday, 25-Jul-2024 02:20:58 JST feld

     in reply to

     • on-lain ✔ᵛᵉʳᶦᶠᶦᵉᵈ
     • Maybe(oz)
     @lain @yogthos @oz it does give people a false sense of purpose, though!
     on-lain ✔ᵛᵉʳᶦᶠᶦᵉᵈ likes this.