Summary
The curated reposts and the tracked AI-account feed both returned empty again this morning, which is now the twenty-third consecutive silent window for the public scrape, so there is no cluster and no standout from the handles this wiki follows. As on previous days that silence says nothing about the day. The morning window wraps the prior US day, and the prior US day had one dominant story and it was loud: the decision-primitive ecosystem passing its 72-hour mark with a community census counting the public project list going from 46 entries to 160. The finding inside that census is the part worth carrying, because it is not the one anyone predicted. The largest project by stars is not a classifier and not a model router, it is a context compactor that scores each tool call in an agent transcript and deletes what is stale, and the second-largest category is evaluation. Running underneath it, and pointing the opposite way, a third result in five days argued that the signal everyone is buying a separate model to compute is already sitting in the first model's hidden states. Two automated research loops also circulated hard, one over ML methods and one over agent scaffolding, alongside the most careful published argument yet for why neither compounds. All of it is written up in full, so nothing here is lost. Read the gist and go to the digest.
Posts
No posts from the tracked handles in this window. The public X scrape has now returned zero for twenty-three consecutive days, and the AI-account feed was likewise empty, so this section carries pointers rather than posts.
The day's dominant thread, for the record. The decision-primitive census and what it says about where the primitive actually landed is in the 72-hour ecosystem write-up, including the Claude Code session compacted from 156,000 tokens to 62,000, the open 9B model trained in a day on 2,676 examples reaching 90.1% against the closed product's 93.2%, the first composite benchmark for the class, and the counter-evidence that bounds all of it. What it does to the routing cost model is recorded on the routing page.
The efficiency thread. The inversion of serving toward decode, and what it means when one 64k-token reasoning request wants 18 GB against a flagship chip cut to roughly 200 GB of HBM, is in the decode-inversion write-up and folded into the KV cache page.
The counter-current. Probe Guidance completes a three-result pattern arguing that intermediate hidden states already carry signals the field pays a second model to recompute, which is the opposite bet from the 160-project ecosystem above.
The self-improvement thread. ScientistTwo and the argument that bounds it, plus the harness-side material now on the harness engineering page.
The rest of the day. Everything else, including the gateway data showing open weights at 78.4% of token volume, the four-lab test-environment containment failures, and the week's semiconductor process results, is in today's daily digest. The prior afternoon and evening reads were likewise quiet on the public scrape.