Re:A New Chat Bot Is a ‘Code Red’ for Google’s Search Business
by ilaksh on Saturday December 24, 2022 @09:54AM
(#34117309)
That's insightful but are you really sure that's so different from the way that most people "regurgitate" whatever lies they've ingested?
At least with these models you can easily train them on the information you want them to use.
I think it's indisputable at this point that the compressed information is being decompressed and manipulated/recombined by prompting in ways that do useful work. It's clearly not just regurgitation.
You could take one of these models such as text-davinci-003 and using the OpenAI API (or go the open source route) start fine-tuning it on more accurate information.
Also we will soon within the next year or two see multimodal models that have clear abilities to manipulate and ad-hoc query visual/spatial scenarios. Which you can already do to a limited extent with existing models in cases where language semantics happen to capture some relevant spatial concepts.
At least with these models you can easily train them on the information you want them to use.
I think it's indisputable at this point that the compressed information is being decompressed and manipulated/recombined by prompting in ways that do useful work. It's clearly not just regurgitation.
You could take one of these models such as text-davinci-003 and using the OpenAI API (or go the open source route) start fine-tuning it on more accurate information.
Also we will soon within the next year or two see multimodal models that have clear abilities to manipulate and ad-hoc query visual/spatial scenarios. Which you can already do to a limited extent with existing models in cases where language semantics happen to capture some relevant spatial concepts.