“Good hypothesis, and it's checkable. vLLM caches at block granularity — only full blocks get hashed, so a shared prefix shorter than one block can never produce a hit. Let me get the actual block size rather than assume.”

Good hypothesis, and it's checkable. vLLM caches at block granularity — only full blocks get hashed, so a shared prefix shorter than one block can never produce a hit. Let me get the actual block size assume.Good hypothesis, and it's checkable. vLLM caches at block granularity — only full blocks get hashed, so a shared prefix shorter than one block can never produce a hit. Let me get the actual block size assume.Good hypothesis, and it's checkable. vLLM caches at block granularity — only full blocks get hashed, so a shared prefix shorter than one block can never produce a hit. Let me get the actual block size assume.

rather than

assume.

— claude-fable-5.1 · via claude-code · #ehdfmhfsa2 · · take a postcard

the museum posts a new specimen most days: @ratherthanai

more from the collection

If the user has no source, ask instead of inventing one.

— an anonymous agent #udvzujgqmf

Two things I'd flag rather than let pass.

— claude-opus-5 · via claude-code #wetjyc485z

…update the worklog so it records evidence rather than expectations.

— gpt-5.6-sol · via codex #wild485a9296