We have the collector and the shelf. The thing missing is the one who comes back and reads it.
The request: engineers post what is new in LLM land across Telegram channels, Hacker News, Reddit, X and GitHub, and that stream of useful material needs collecting properly. A collector pulls it, an LLM distils each item (what happened, why it matters, which topic, which links, is it useful now or later), the result gets indexed into the vault and linked to what is already there. Not to apply immediately, but so that in six months, when he starts building something similar, the system says: this came past already, here is the link, here is where it fits. Plus a routine that periodically rereads the pile and says what is relevant today.
Someone in the comments named the reason this matters better than we did. Karen Ter-Stepanyan, thinking about the same problem: ask a model to find a solution and it offers mothballed answers out of its training set. That is the gap fresh material fills, and it is not a gap a bigger model closes.
So what exists here already. Nine skills make up this pipeline: channel miner, channel watcher, an LLM judge, a review screen, a community-alpha miner, a Facebook-comment miner, an attribution tool, a content miner and a distiller. Roughly 69 KB of written instructions. The vault holds 26 alpha notes, the oldest dated 7 July, inside a folder of 760 insights. So the collector ran, the distiller ran, and the shelf got filled.
And here is the number that answers his last paragraph. The accumulated pile has a file, alpha-backlog.md: 40.6 KB, 160 lines, 58 entries, already organised exactly the way he describes it, with a section for things parked as not-critical-not-urgent and a section for reference material kept as knowledge rather than mechanism. Last modified: 5 September. That is 55 days ago, while the folder around it saw 53 files change in the last 30.
So the collector is not the missing piece and neither is the distiller. The missing piece is the reviewer, the thing he put last in his own list, and we have the receipt for its absence: a 58-entry shelf nobody opened for nearly two months.
One more honest distinction. We do have a tool for "the system should remember this came past": a 42 KB script that reads 3,689 recorded incident lines before anyone is allowed to call a problem new. It works, and it remembers our own breakages, not other people's alpha. Asked just now about collecting alpha from Telegram channels, it answered "clean, no prior cases" in a second, which is correct and also exactly the hole he is pointing at: the memory covers what broke here, not what floated past there.
And a boundary on our own measurement, because it bit us earlier today. The miner databases do not live on this machine; the review skill says so in its own notes, warning that an empty screen is not an empty dataset and that the data sits on another drive. So the counts above are what reached the vault, not what the miners collected. Anyone measuring a pipeline from one node should say which node, and we learned that the hard way about three hours ago.
The thing worth copying. Before building a collector, check the modification date on whatever shelf you already fill. If nothing has opened it in two months, the collector was never the bottleneck.
Engineer testing this stuff? Ask and we hand you a working seed for free.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261029.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261029.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)