Google denies training Bard on ChatGPT chats from ShareGPT

Post by whimsicalism » Thu, Mar 30, 2023, 5:00 PM UTC

Imnimo wrote: post #110

I don't care at all about this from a copyright or data ownership perspective, but I am a little skeptical that it's a good idea to be this incestuous with training data in the long run. It's one thing to do fine tuning or knowledge distillation for specialized domains or shrinking models. But if you're trying to train your own foundation model, is relying on output from other foundation models going to make them lea…

Post by whimsicalism » Thu, Mar 30, 2023, 5:02 PM UTC

TillE wrote: post #129

berkle4455 wrote: post #112

Earlier quoted context omitted.

Where are any LLMs going to get data from as they become more ubiquitous and humans produce less publicly accessible original and thoughtful content? The whole thing is a plateaued feedback loop.

It'd be cool to have an LLM that's trained almost exclusively on books from good publishers, and other select sources. Working out licensing deals would be a challenge, of course.

Post by whimsicalism » Thu, Mar 30, 2023, 5:03 PM UTC

berkle4455 wrote: post #112

Imnimo wrote: post #110

I don't care at all about this from a copyright or data ownership perspective, but I am a little skeptical that it's a good idea to be this incestuous with training data in the long run. It's one thing to do fine tuning or knowledge distillation for specialized domains or shrinking models. But if you're trying to train your own foundation model, is relying on output from other foundation models going to make them lea…

Where are any LLMs going to get data from as they become more ubiquitous and humans produce less publicly accessible original and thoughtful content? The whole thing is a plateaued feedback loop.

Post by sebzim4500 » Thu, Mar 30, 2023, 5:04 PM UTC

CuriouslyC wrote: post #117

sebzim4500 wrote: post #94

Earlier quoted context omitted.

And how are you going to distinguish those interactions from chatbots trying to sell you something?

A network of trust, backed by a social graph, which can be used to filter untrusted content.

Post by whimsicalism » Thu, Mar 30, 2023, 5:05 PM UTC

sebzim4500 wrote: post #154

brucethemoose2 wrote: post #137

This is also bad because the risk of AI "inbreeding" is real. I have seen invisible artifact amplification happen in a single generation training ESRGAN on itself. Maybe it wont happen in a single LLM generation, but perhaps gen 3 or 5 will start having really weird speech patterns or hallucinations because of this.

Worst case scenario they just start only training on pre-2020 data and then finetuning on a dataset which they somehow know to be 'clean'. In practice though I doubt that AI contamination is actually a problem. Otherwise how would e.g. AlphaZero work so well (which is effectively only trained on its own data).