3 ms·
There have been ambitions to do that kind of thing with LoRA - see the leaked "no moat" Google memo from a couple of years ago for one example: https://simonwil
by simonw 1y ago
There have been ambitions to do that kind of thing with LoRA - see the leaked "no moat" Google memo from a couple of years ago for one example: https://simonwillison.net/2023/May/4/no-moat/ https://simonwillison.net/2023/May/4/no-moat/
It hasn't really happened though. I suspect that's because it turns out techniques like RAG or tool calling are massively easier and more effective than trying to tech models new information through shared model weights.