3 ms·

subjective opinion, since LLMs can be constructed in multiple layers (raw output, enhance with X or Y, remove mentions of Z,...), we should have multiple purpos

subjective opinion, since LLMs can be constructed in multiple layers (raw output, enhance with X or Y, remove mentions of Z,...), we should have multiple purpose built LLMs: - uncensored LLM - LLM which censors political speech - LLM which censors race related topics - LLM which enhances accuracy - ... Like a Dockerfile, you can extend model/base image, then put layers on top of it, so each layer is independent from other layers, transforms/enhances or censors the response.

evilduck 3y ago
You've just proposed LoRAs I think.
wongarsu 3y ago
As we get better with miniaturizing LLMs this might become a good approach. Right now LLMs with enough world knowledge and language understanding to do these tasks are still so big that stacking models like this leads to significant latency. That's acceptable for some use cases, but a major problem for most use cases. Of course it becomes more viable if each "layer" is not a whole LLM with its own input and output but a modification you can slot into the original LLM. That's basically what LoRAs are.