Original ·2060008768671301747· vitalik.eth @VitalikButerin · · en XEn respuesta a ·2060004717196153142· tmo @tmophoto · · en XThis is really terrible advise. NEVER run a 2 or 4 bit quant if you can run a different model that is at least a Q6.
The amount of stupid mistakes, empty files it creates, and shortcuts the highly quantized models make is unreal and stacks up if you use them for agents over@tmophoto The long-term and even short-term solution for this is safety-first workflows.
At the very least, make it generate lots of test cases for everything and make sure they pass. Ideally write things in lean and prove theorems.
(And obviously, only run serious AI inside sandboxes)
RT @VitalikButerin: @tmophoto The long-term and even short-term solution for this is safety-first workflows.
At the very least, make it ge…