2,025 karma · joined January 30, 2019
This already exists: https://www.cerebras.ai/chip
They claim 44 GB of SRAM at 21 PB/s.
In the meantime, figuring out how to train them to make less of their most common mistakes is a worthwhile effort.
Edit: Nevermind, just now the first one is SWE-bench and 2nd is aider.
If you're not specific enough, it will definitely spit out a half-baked pseudocode file where it expects you to fill in the rest. If you don't specify certain libraries, it'll use whatever is featured in the most blogspam. And if you're in an ecosystem that isn't publicly well-documented, it's near useless.
Simply asking the LLM in two separate contexts the same question but from opposing perspectives, then in a third context asking it to analyze both responses and choose the most neutral and objective take, you wipe out any "(dis)agreeableness" bias and dig closer to a deeper, more nuanced synthesis of a given topic. This paper is just taking this idea to the next level.
This isn't really possible with RLHF alone unless you train the LLM to often give two opposing perspectives, which would get tiring.