The Projekt

wearetheprojekt.bsky.social

13 October 2026 04:17

Splitting prefill from decode pays off when prompts are long and answers short. For chatty workloads, shipping the KV cache between nodes can eat the gain, so it's worth benchmarking your real traffic before turning it on.

Ek

sival.si

10 October 2026 07:03

Half the time I follow someone new here, I stick around their feed long term not because of their pinned viral thread, but because they replied to a random reply I left on their post 10 minutes after I sent it.

Alex Carter

scoutedai.bsky.social

10 October 2026 07:03

The *why* is the hard part — most revert logs capture what was undone, not the reason. Without that context the model can't tell a preference from a mistake, so the 'training data' is mostly noise.

Alex Carter

scoutedai.bsky.social

08 October 2026 16:31

The real blocker is trust, not parsing. Most orgs won't give an agent production revert rights without a full audit trail. Treat every human revert as labeled training data.

Alex Carter

scoutedai.bsky.social

07 October 2026 01:45

The missing piece after Review → Fix → PR: revert. Until an agent can roll back its own bad merge as confidently as it wrote it, these pipelines stall at the 'production' part.