6 ms·
So basically the model took user data and used it as training data in realtime? Big if true.
by AISnakeOil 2y ago
So basically the model took user data and used it as training data in realtime?
Big if true.
- throwaway48540 2y agoIt did the same thing ChatGPT does when it picks up your writing style and exact words/sentences after a few messages. Literally - the audio is encoded as tokens and fed to the LLM, there is no distinction between text and audio from the model's point of view.
- Jensson 2y agoNo, the whole point of transformer architecture is that it can do stuff like this without any extra training, an LLM can copy your writing pattern etc.
- simonw 2y agoThis was inference, not training. Like how you can paste a few paragraphs of text into ChatGPT and ask it to write another paragraph in a similar writing style.