Summary
Only two slot files exist for the day: morning and evening. The evening slot was empty, so all the day's signal comes from the morning capture of the Following feed (the US Wednesday). The strongest cluster has six posts on agents improving their own training loop. Microsoft's AutoEnvScaling has agents write their own RL environments, Harness² keeps a harness edit only if it beats the old version, with no answer key, and Tev-Gym and DAIR's Behavioral Self-Rewarding add denser reward signal. The second cluster is compute money and real cost signal. CoreWeave bonds fell below 85 cents, there is a bull case for Cerebras SRAM, Maia is now Microsoft's primary inference chip, and Epoch finds China is 2.7x as exposed to chip shortages as the US. The best single item is Databricks cutting a fine-tune from 7h44m to 1h17m at a reported "98% utilization". It is a clean lesson that a busy GPU is not the same as useful work. Noise was high: the Claude Code "goldmine" posts point to real docs under hype framing, and a batch of engagement bait got cut.
Posts
- Agents that build their own training environments and improve their own harness (cluster of 6) (@omarsar0 · AutoEnvScaling, @Xudong07452910 · Harness², @omarsar0 · Evo², @gordonhu608 · Tev, @dair_ai · Behavioral Self-Rewarding, @bibryam) [morning]. AutoEnvScaling gets up to 25x more valid environments and +12.2 on Terminal-Bench 2.1. Harness² reaches +11.8 with no ground-truth verifier, and @bibryam pushes back that sandboxing is infrastructure, not harness. Wiki: harness and environment self-improvement.
- Compute money and chip exposure (cluster of 5) (@rohanpaul_ai, @StockSavvyShay, @rohanpaul_ai, @rohanpaul_ai, @Ronald_vanLoon) [morning]. CoreWeave's 2032 notes yield near 15%. Other posts: a Cerebras SRAM thesis, Maia confirmed as Microsoft's primary inference chip (10-13 OCP page), Epoch's 2.7x China exposure figure, and a 40 km photonics disaggregation demo. Wiki: TSMC Q3 and neocloud credit.
- Claude Code: plan in a script, run on cheap models (cluster of 4) (@thedelost · workflows docs, @mirku21 · advisor docs, @VaibhavSisinty, @aakashgupta) [morning]. The framing is hype, but the docs are real: the script holds the loop, so the main context only sees final answers. Also new: agent step tracing (public beta) and a session sidebar sorted by which sessions need you.
- Databricks: 6x faster fine-tune at "98% utilization" (@FeitengLi) [morning]. Fused linear-attention kernels plus padding-free packing cut a Qwen3.8-27B fine-tune from 7h44m to 1h17m. Wiki: efficiency shorts.
- HobbyLM: a 1B sparse MoE in ~300 GPU-hours (@NammaCoimbatore) [morning]. Two people trained it on 4 H200s in 76 hours. No paper or benchmarks, so treat it as a cost anecdote.
- System One decision models: JevGPT and JevEval (cluster of 2) (@_reachsumit · JevGPT, @_vmlops) [morning]. JevGPT uses parallel fast judges to check sources under a planner. JevEval scores outputs with explicit questions and calibrated probabilities. Wiki: Jeva and the System One cluster.
- Muse Wonder lands at #12 in Agent Arena (@arena, @AIatMeta) [morning]. It is within 1 point of GPT-5.6 Sol at 25% lower cost per task. Meta claims state of the art, but the arena number is more modest.
- Anthropic deals and positioning (cluster of 2) (@rohanpaul_ai, @rohanpaul_ai) [morning]. Anthropic is licensing Opaque Systems' enclave tech and hiring its CEO and CTO. Its international chief says open-weight models carry more safety risk by definition.
- Hybrid Reasoning Agents, resurfacing (@omarsar0, @Jiaru_Zou · paper) [morning]. The agent reasons in latent space and switches to text only for tool calls, at 2.7-3.7x lower cost. That raises a monitorability question (essay).
- Jensen Huang on safety and GPU access (@rohanpaul_ai, @StockSavvyShay) [morning]. He says he would not sell GPUs to irresponsible labs.
- Engagement bait and promos [morning]. Skip. These were a "Harvard proved universities obsolete" thread, an invented Hassabis quote, Karpathy clip-farming, travel-agent promos, and trading ads.
- Evening slot [evening]. No posts. See the 2026-10-15 digest for the day's synthesis.