5 ms·

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 trai

by Imnimo 4y ago

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 learn to imitate their own errors?

berkle4455 4y ago
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.
- TillE 4y ago
  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.
  - whimsicalism 4y ago
    Corpora is likely too small. It would just be an "LM"
- whimsicalism 4y ago
  Probably from multiple modalities as well as extending the sequence lookback length further and further. They have low perplexity now, but the perplexity possible when predicting the next word on page 365 of a book where you can attend over the last 364 pages will allow even more complexity to emerge.
whimsicalism 4y ago
But Bard isn't a foundation model? Clearly this data has value as some sort of RLHF finetuning dataset. Honestly they probably used it for negative examples.
sdenton4 4y ago
Things like ShareGPT or PromptHero give vast repositories of human-curated ML outputs, which make them fantastic for at least incremental improvement on the base model. In the grand scheme of things, these will be just another style, mixed in with all the other crap in the training set, so I don't imagine it's too harmful... eg, 'paint starry night in the style of midjourney 5'
pessimizer 4y ago
The internet is an easy, convenient way to train LLMs, but I'm pretty sure you could train them with microphones. One cloud surveillance company, like maybe for networked security monitoring, or maybe just Alexa/Siri etc. could dip into as many and as varied communications per hour than all the books ever written.

by Imnimo 4y ago

berkle4455 4y ago

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.

TillE 4y ago
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.
- whimsicalism 4y ago
  Corpora is likely too small. It would just be an "LM"
whimsicalism 4y ago
Probably from multiple modalities as well as extending the sequence lookback length further and further. They have low perplexity now, but the perplexity possible when predicting the next word on page 365 of a book where you can attend over the last 364 pages will allow even more complexity to emerge.

whimsicalism 4y ago

But Bard isn't a foundation model? Clearly this data has value as some sort of RLHF finetuning dataset. Honestly they probably used it for negative examples.

sdenton4 4y ago

Things like ShareGPT or PromptHero give vast repositories of human-curated ML outputs, which make them fantastic for at least incremental improvement on the base model. In the grand scheme of things, these will be just another style, mixed in with all the other crap in the training set, so I don't imagine it's too harmful... eg, 'paint starry night in the style of midjourney 5'

pessimizer 4y ago

The internet is an easy, convenient way to train LLMs, but I'm pretty sure you could train them with microphones. One cloud surveillance company, like maybe for networked security monitoring, or maybe just Alexa/Siri etc. could dip into as many and as varied communications per hour than all the books ever written.