@tessera_antra The residual stream certainly provides coherence between layers of a single forward pass, but it is discarded at the end of a forward pass, so the asemantic information between tokens is just in the KV cache. I don't think this undermines the overall argument though.
in reply to: 1961067410846343361
same thread: 1961067404097696051 1961067406748520959 1961067410846343361 1961067424205152417 1961307244311830941 1961334777434288187 1961468034776011050 1961468800236167182 1961469613499129970 1961515957295145138 1961578402294779947 1961580133476970963 1961819720044396979
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.