Re:Cerebras-GPT vs. LLaMA AI Model Performance Comparison
by f_devd on Wednesday March 29, 2023 @06:20PM
(#35364669)
> If you trained GPT specifically on arithmetic tasks
Sure but you'd have a lot of overlapping tokens with BPE, which doesn't help with convergence. GP is claiming it's specifically worse at arithmetic because of BPE which is true.
Sure but you'd have a lot of overlapping tokens with BPE, which doesn't help with convergence. GP is claiming it's specifically worse at arithmetic because of BPE which is true.