Summary
The public morning scrape is empty and the saved-reading channel is not, which is the cleanest separation this stream has produced in over a week. raw/twitter/2026-09-04-morning.json holds zero curated reposts and zero tweets from the tracked AI handles, because the farmer again found no reachable Nitter instance, the ninth consecutive day that has happened, so neither the curated repost stream nor the tracked handle list produced anything at all. There are no article bodies and no image attachments in the morning file, so there is nothing to transcribe and nothing to cluster from the general feed. The bookmarks channel, by contrast, authenticated normally against X's GraphQL Bookmarks operation and returned two genuinely new saves, which is a measured number rather than a broken pipe, and both sit squarely on the two themes this reader has been accumulating for a month. The stronger of the two is an inference-serving explainer that separates the four distinct caches in an LLM serving stack, which is the sixth save on the KV cache and GPU-kernel theme and a re-arrival of material this wiki already holds a page on. The second is a repost of a paper titled "The End of Software Engineering," framed as a claim that code becomes throwaway and human judgment becomes the job. Because the saved items are private reading rather than public feed content, their substance is synthesized in today's Media Zone rather than enumerated here, and the day's actual content lives in the daily digest, where four papers independently report that the component the field was protecting was the wrong one.
Posts
Nothing captured from the public feed this slot. Zero curated reposts, zero tweets from the tracked AI handles, zero linked articles, zero images. The morning file is structurally empty rather than filtered down to empty, so there is no item to summarize, no cluster to name, and no promo to skip. The next slot with a reachable Nitter instance picks the general feed back up.
Two new saves went to the private bookmarks channel and are covered in the Media Zone. The first is a serving-stack explainer whose substance is the separation of four caches that are routinely conflated: the per-request KV cache (the key and value tensors for every token at every layer, held in GPU memory for one active request), server-side prefix caching (vLLM keeping those tensors in 16-token blocks identified by a hash that chains in the previous block's hash, so a block only matches when everything before it matched, and the scheduler stops at the first miss and prefills the suffix from there), provider-billed prompt caching (Anthropic charging roughly 1.25x the base input rate to write an entry and 0.1x to read it), and application-layer semantic caching (embedding the incoming prompt, running a similarity search over stored prompts, and returning a stored answer above a threshold). The load-bearing distinction is that the first three match on exact tokens and therefore cannot change what the model produces, while the fourth matches on similarity and can return a wrong answer when embeddings collide, and it charges an embedding round trip on every request including every miss. The wiki already carries this material as four cache layers (08-29), so this is a re-arrival rather than new knowledge, which is itself worth noting: the same explainer has now been saved twice in seven days. The second save is a repost of a Chinese-authored paper titled "The End of Software Engineering," whose argument as relayed is that a human's working memory bounds how much state and how many dependencies they can hold while an agent's capacity scales with compute, so the human's role shifts from writing code to directing agents and auditing results. The repost attaches its own agent-building guide alongside the paper, and the paper itself is not linked in a fetchable form in the capture, so the claim here is the reposter's framing rather than the paper's own text. Both are treated properly in today's Media Zone.