Summary
The evening window closed with nothing in it: no curated reposts, no posts from the tracked AI handles, no new saves, no linked articles or images. There is no cluster to count and no standout to name, and inventing a theme out of an empty slot would be worse than saying it in one line. The day's real social signal all arrived this morning and it was a single item, a screenshot save of Sparse Reward Subsystem in Large Language Models, which argues that the step-level scorer you normally pay a second forward pass for on every search step is already sitting in a sparse set of neurons inside the policy you are running. That one save is worth more than a busy timeline would have been, and it is already written up. Everything else on 09-06 came through the research feed and Gmail rather than social, led by ContextPipe cutting an agent's token bill 31 percent while its cache-hit ratio got worse.
Posts
- Nothing captured in this window. No item to rank, cluster or skip from either the curated reposts or the tracked AI accounts, and no newly saved posts.
- The day's one save, from earlier: sparse reward neurons inside LLMs (@HowToPrompt__ · arXiv 2602.00986). Reward-related signal in a model's hidden states turns out to be concentrated in a small set of individual neurons rather than smeared across the vector, splitting into value neurons that predict whether continuing from here yields a correct answer and dopamine neurons that encode step-level surprise. The cost angle is the second one: a process reward model normally means serving a separate scorer alongside the policy, and this says it may be a gather over known indices instead. The viral framing of the paper gets the attribution, the headline number and the recency wrong, so read the source and not the post. → Wiki summary
- Where today's substance actually sits. ContextPipe plans an agent's prompt the way a database plans a query and lands 31 percent fewer tokens with a lower cache-hit ratio, which is the first clean counterexample to the append-only prefix-protection rule the wiki has carried since 08-14. See the ContextPipe summary, the 09-06 daily digest and the 09-06 Media Zone.