Summary
Only the morning slot carried posts. The afternoon slot was empty, which is normal for a US Sunday night, and no night or evening slot was written. The biggest cluster is the decision-model wave again, around fifteen posts on Jev and its clones. Only four carry numbers: jevgrep claims 40% lower coding-agent cost on SWE-bench, and RSI-Jev v2.0 is a 2B model an agent team improved in about 15 hours. The strongest single post is economic. An audited walk-through of MiniMax and Zhipu, the only listed pure inference companies, shows one reaching a 24.6% margin while the other lost 75% of its OpenRouter volume. PRIMESCIENTIST, which runs a research budget like a portfolio and needs half the attempts, and RSIAgent's broad-before-deep curriculum (56.5% to 74.5%) are the useful agent results. Most of the rest is hype. The daily digest has the deeper treatment.
Posts
- The economics of an inference company (@RMladek) [morning]. Audited numbers for MiniMax and Zhipu: one went from losing money on every token to a 24.6% margin in 18 months, and the other lost 75% of its OpenRouter volume in ten weeks. The fullest version of a story the 2026-09-23 digest opened.
- Decision-model tools with numbers (cluster of 4) (@dzhng · jevgrep · @GaoShanghua · @deliprao · @karthikkalyan90) [morning]. jevgrep (40% cheaper coding agents), RSI-Jev v2.0 (0.79 vs 0.73 on typed decisions; calibration transferred, task gains did not), Autorubric 1.6.0 grading with decision models, and a case for decision models in long-tail retrieval. See LLM routing.
- Decision-model hype (cluster of 8) (@marfinxx · @Mnilax · @hanakoxbt · @Av1dlive · @Dontgiveup_26 · @xmyttle · @mirku21 · @0xCodila) [morning]. Skip. The one useful fact: NanoJev, a 0.6B open replica, scored 128/128 on ViZDoom aiming versus Jev's 56 (@rvaniaaaa).
- Converting Qwen3 into a decision model, live (@neural_avb) [morning]. A five-hour build stream. It shows the attention-mask and position-id edits that make the choice independent of option order.
- PRIMESCIENTIST (@rohanpaul_ai) [morning]. It keeps competing research plans in a tree and concentrates budget on the strongest branches as the budget shrinks. On FIRE-Bench it gets 10.3% higher reward than AutoResearch with 50.6% fewer attempts.
- RSIAgent: practice broad, then deep (@rohanpaul_ai) [morning]. A curriculum agent invents practice tasks and a verifier gates what is saved. Broad-first practice lifts success from 56.50% to 74.54%.
- Smart and dumb are two axes (@theo) [morning]. Capability and error rate are separate. By his account Fable 5.1 is less capable than GPT-6 Astra but makes far fewer senseless moves.
- Agent memory metrics (@henrytdowling · @dair_ai) [morning]. Optimize a better metric rather than stack machinery. A Salesforce paper says to store raw trajectories, not only summaries.
- Minimal harnesses and multi-agent coding (cluster of 3) (@omarsar0 · @JoshARosen · @rvaniaaaa) [morning]. An MIT CSAIL case for minimal harnesses, sub-agent delegation, and a 535K-line rewrite credited to the task graph, not the model. See harness engineering.
- Learning resources (cluster of 3) (@attharrva15 · @0x0SojalSec · @_avichawla) [morning]. A vLLM architecture X article (body not captured), a 16-hour DeepSeek-style build series, and why a top-2 MoE can be slower than dense: cross-GPU expert traffic.
- Deep networks as kernel machines (@pmddomingos · paper) [morning]. Domingos re-posts his 2020 result as a reply to a new finding he calls unsurprising.
- Dream-RSI (@Sa4d_k1 · paper) [morning]. Argues recursive self-improvement is bottlenecked by exploration, so the exploration strategy should itself evolve.
- Contrastive World Models (@shumpeiMaxwell · paper) [morning]. Swapping Dreamer's pixel decoder for a contrastive objective ignores visual distractors.
- Promo and lecture recaps [morning]. Palantir agent-stack explainer, ElevenLabs clone promo, Karpathy/Ng summary threads, "Google automates the PhD". Skip.