Summary
The morning slot is empty on the public timeline side. Zero curated reposts and zero original posts from the tracked AI accounts came through in the 24-hour lookback window, so there is no cluster to name and no standout item, and the public X feed has now produced nothing on three of the last four days. The one piece of signal that did arrive came through the saved-reading channel rather than the timeline: a single new bookmark, the first in five days after a four-day quiet stretch, and it is a long-form X article by Avi Chawla separating the four things the serving stack calls caching. That is a topic pivot worth recording on its own. The saved trail before it ran thirteen items deep on harness and loop engineering, meaning how to drive an agent loop well; this one is about what the loop costs at the memory layer, which is the same transition the research feed made yesterday when the first harness paper published a serving-side cost-per-success number instead of an accuracy number. With the timeline silent, the day's social signal is one post, and it happens to be squarely on the routing and KV cache axis. It is treated in full in today's Media Zone, which is where saved reading is synthesized.
Posts
Public timeline: nothing captured. The
@bayesiansapiencurated repost section and the tracked AI-handle feed both returned zero items for the 24-hour window ending 2026-08-29 11:13 IST. This is an absence of captured content rather than a judgment that nothing was posted, and it should not be read as a quiet day across X generally.The four caching layers, and the one that can be wrong (@_avichawla, long-form X article; public mirror at Daily Dose of Data Science). The single saved item of the day, and the substance is worth carrying even though it arrived through the private channel. The argument is that four different mechanisms in an LLM serving stack are all called caching, and they sit at four layers keyed on four different things. The KV cache is the per-request store of key and value vectors so tokens already processed do not get recomputed at every decode step; it lives in GPU memory, dies with the request, and runs to roughly 40 GB for a 70B model at BF16 with a 128K context. Prefix caching persists those same tensors server side across requests: vLLM chunks the sequence into fixed 16-token blocks, hashes each block over its parent block's hash plus its own token IDs, and the scheduler walks incoming blocks in order and stops at the first miss, which is why one changed byte near the front of a prompt invalidates everything after it. Prompt caching is the provider's billed version of that same lookup, roughly 1.25 times the base input rate to write a cache entry and 0.1 times to read one. Semantic caching is a different object entirely: it stores finished response strings keyed by embedding similarity at the application layer, so it is the only one of the four that is fuzzy-match rather than exact-match, and it is therefore the only one that can hand back a different question's answer with an HTTP 200 success code. The three operational details that matter most and are hardest to find elsewhere: Anthropic's prompt cache walks backward through at most 20 blocks on a read, so more than 20 blocks of conversation between two calls pushes the last write out of range; cache writes only happen at a breakpoint you explicitly placed; and cache entries are keyed to a specific model, which means routing a conversation to a cheaper model part-way through still pays a full cold prefill on the entire accumulated history. The article closes on four production failure modes, all of them prefix-shaped rather than capacity-shaped: a timestamp or user ID at the front of the prompt, tool schemas reordered, feature toggles rendered into the prompt text, and summarizing conversation history. Its operative advice is to truncate tool outputs in place rather than summarize, because that keeps the prefix byte-identical and the cache alive. Written up in full at the four cache layers and folded into the KV cache concept page.