Summary
A thin day: the night slot never ran, the afternoon slot (the US overnight) was empty, and nearly all the signal sits in the morning's small Following-feed capture. There is no true cross-slot cluster, but a loose RL-efficiency thread runs from morning to evening: Tengyu Ma's claim that GRPO-style RL cannot be optimal in the morning, and the RL-Kernel library (GPU kernels for GRPO/PPO post-training, up to 163x per-op speedups) resurfacing in the evening. The strongest single item is a clean explainer of chunked prefill in vLLM, which pairs with the day's KV cache papers (CacheBack, Extender, QuantWM). Leviathan's 436-token-instead-of-107K agent search claim and a three-post "measure the agent, do not trust it" cluster are the other keepers. The rest is noise: an unverified Newton's-laws thread, career content and a recycled LeCun talk. The daily digest carries the day's real substance.
Posts
- Chunked prefill, clearly explained (@akshay_pachaar) [morning]. Splits a long prompt's prefill into chunks so other users' decode streams keep flowing, trading slightly longer time-to-first-token for smooth streaming. Context for the day's KV cache papers.
- RL efficiency: GRPO critique + RL-Kernel (cluster of 2) (@stanfordnlp resharing @tengyuma · @ContentCase · repo) [morning + evening]. Ma argues GRPO (scoring each sample against its group average) is not the end point for RL. RL-Kernel is June material on FlashInfer-based kernels for GRPO/PPO; the 163x is per-op, not end-to-end (June summary, RL for LLMs).
- Leviathan: search a million records with 436 tokens (@joshuagunnn · repo) [morning]. A static-binary full-text indexer for agents; author-reported 436 vs 107,000 tokens at 1M records with 99% hit rate. Same idea as CorpusMap (agentic search summary).
- Measure the agent, do not trust it (cluster of 3) (@businessbarista · @sermakarevich · @ashwingop) [morning]. An evals primer (a grader applied to a trace), a call for held-out test sets before claiming prompt wins, and an article on personal agents acting on stale private copies of company state.
- Choice-order invariance for decision models (@neural_avb) [morning]. Notes on training JEV-style "System One" models whose answers should not change when options are reordered. Relevant because these models are pitched as router layers (LLM routing).
- Reka Rho-1 launch (@MateuszOnAI) [morning]. A 19B omni-model across image, video and text with "one state, one loop, no tool calls." A single-network alternative to routing between specialists.
- AI rediscovers Newton's laws (@AnatoliKopadze) [morning]. A Peking University system reportedly invented mass and force from raw coordinates of 46 experiments. Single-source, no paper link, unverified.
- RL Interview Questions 2026 (@ContentCase) [evening]. X Article on RL interview prep and PhD vs industry. Skip.
- Low-signal reposts (@ylecun · @Unnati_builds24 · @abeirami) [morning]. Recycled LeCun talk, a career story and a one-line reaction. Skip.