Summary
One save carried the whole day, and it appeared in two slots rather than one, which makes it the only thing resembling a cross-slot cluster. A screenshot of Sparse Reward Subsystem in Large Language Models landed in the morning window and the evening slot pointed back at it for want of anything newer, so the day's social signal is a single item read twice. It is a good item: reward-related information in a model's hidden states turns out to be concentrated in a sparse set of individual neurons rather than smeared across the vector, splitting into value neurons that predict whether continuing from a given point yields a correct answer and dopamine neurons that encode step-level surprise, and the second type works as a process reward model, the scorer that grades each intermediate reasoning step. That is a cost result in interpretability clothes, since a process reward model normally means serving a second model alongside the policy and this suggests it may be a gather over known indices instead. The noise is worth naming too, because the post that surfaced the paper got the attribution, the headline number and the recency all wrong, and the same account produced the previous day's save with the same distortion pattern, so the source is worth reading and the post is not. Everything else was silence: the afternoon and evening windows captured nothing, no night slot ran, and the public Nitter scrape passed its eleventh straight day without a reachable instance, which is a pipeline fact rather than a quiet news day.
Posts
Sparse reward neurons inside LLMs, saved as a screenshot of the paper's first page (@HowToPrompt__ · arXiv 2602.00986) [morning + evening] (cluster of 2). Guowei Xu of Tsinghua with Mert Yuksekgonul and James Zou of Stanford probe at the neuron level rather than fitting black-box probes on the whole hidden-state vector, and find the reward signal concentrated in a sparse subset that splits into value neurons and dopamine neurons. Value neurons are reported as robust and transferable across datasets and models with causal evidence; dopamine neurons guide inference-time search as a process reward model, which is where the serving cost comes out. → Wiki summary
The framing around that save is wrong in three specific ways (@HowToPrompt__) [morning]. The post credits Stanford and drops the Tsinghua first author, quotes a 50 percent math-reasoning collapse from ablating value neurons that does not appear in the abstract, and presents a paper HuggingFace surfaced on 2026-05-11 as new. It also adds a "nobody programmed this" biological-inevitability story to a paper that predicts the finding in advance from a maximum-entropy reinforcement learning argument.
Nothing from the public timeline in any slot [morning + afternoon + evening] (cluster of 3). Zero curated reposts and zero posts from the tracked AI handles across all three windows. The general scrape has produced no item since 08-27, an eleventh consecutive day with no reachable Nitter instance.
No night slot ran [night]. The window is absent from the day rather than empty.
Where the day's substance actually sits, flagged from the evening slot [evening]. ContextPipe plans an agent's prompt the way a database plans a query and cuts the token bill 31 percent while its cache-hit ratio gets worse, the first clean counterexample to the append-only prefix-protection rule the wiki has carried since 08-14. See the 09-06 daily digest and the 09-06 Media Zone.
Most recent prior saves, referenced from the afternoon slot [afternoon]. The 09-04 pair still stands as the last social material before today: an explainer separating the four caches routinely conflated in deployment, where the per-request KV cache, server-side prefix caching and provider-billed prompt caching all match on exact tokens and cannot change an answer while semantic caching matches on similarity and can return a wrong one, plus a repost arguing an agent's context capacity scales with compute while a human's does not. On the wiki as four cache layers.