Summary
No curated reposts or saved posts landed in the morning slot, so this synthesis is built from the morning capture of the Following feed, which covers the US Wednesday. The strongest signal is a cluster of six posts on agents improving their own training loop: Microsoft's AutoEnvScaling (agents build their own RL environments), Google's Harness² (harness edits checked without an answer key), DeepMind's Evo² resurfacing, Tev-Gym's 1,170 dense-reward environments, and a DAIR.AI pick on self-rewarding for SWE agents. The second cluster (five posts) is compute money: CoreWeave's bonds falling below 85 cents, an investor's case for Cerebras' on-chip memory, Microsoft's Maia becoming its primary inference chip, Epoch's estimate that China is 2.7x as exposed to chip shortages as the US, and a photonics disaggregation demo. A third cluster covers Claude Code: two viral posts about running Opus as planner over Haiku and Sonnet workers (pointing to the real dynamic-workflows and advisor docs), plus agent step tracing and a session sidebar update. Standout single items: Databricks fine-tuning Qwen3.8-27B 6x faster at a reported 98% GPU utilization, the System One decision papers JevGPT and JevEval, and Meta's Muse Wonder landing at #12 in Agent Arena.
Posts
Agents that build their own training environments and improve their own harness (cluster of 6) (@omarsar0 on AutoEnvScaling, @Xudong07452910 on Harness², @omarsar0 on Evo², @gordonhu608 on Tev, @dair_ai on Behavioral Self-Rewarding, @bibryam). AutoEnvScaling turns environment design into a terminal task: a proposer agent reads the solver's rollouts and writes new RL environments, with up to 25x more valid environments than prompt-only generation at up to 75% lower cost each, and one Qwen3.6-35B-A3B in both roles gains 12.2 points on Terminal-Bench 2.1. Harness² (the Chinese-language thread explains it well) proposes a harness edit, runs old and new versions under the same conditions, keeps edits whose behavior is better, and logs how each edit was made so future edits improve; up to +11.8 points with no ground-truth verifier. Evo² (covered 10-14) trains one policy as both harness editor and reviewer, reaching 48.3% on SWE-bench Verified. Tev-Gym builds 1,170 environments from hidden "latent world" specs so progress can be rewarded per step. Behavioral Self-Rewarding adds small bounded pairwise preferences when every rollout in a GRPO group passes or fails, so the group still teaches something (Qwen3-14B 32.1% to 34.7% on SWE-bench Verified). @bibryam pushes back on harness taxonomies that fold sandboxing and persistence into the harness: those are infrastructure. Wiki: harness and environment self-improvement.
Compute money and chip exposure (cluster of 5) (@rohanpaul_ai, @StockSavvyShay, @rohanpaul_ai, @rohanpaul_ai, @Ronald_vanLoon). CoreWeave's $1.25B of 9.625% 2032 notes, sold in June, fell below 85 cents (yield near 15%). Its Q2 net interest expense ($640M) exceeded its net loss ($626M), and Oracle bonds, Nebius and Dell fell alongside. An investor thesis argues Cerebras wins agent inference because its on-chip SRAM avoids scarce HBM (high-bandwidth memory) and agents make dozens of calls per task; OpenAI already buys Cerebras inference. A truncated repost says Satya Nadella confirmed Maia as Microsoft's primary inference system (consistent with the 10-13 OCP page). Epoch AI: chips are $15.2 of every $1,000 Chinese spending versus $5.7 in the US (2.7x), rising to 4.7x for Taiwan-made chips. IOWN Global Forum showed ~40 km of separation between data and GPUs at ~1.9% penalty using photonics. Wiki: TSMC Q3 and neocloud credit.
Claude Code: plan in a script, run on cheap models (cluster of 4) (@thedelost linking the dynamic workflows docs, @mirku21 linking the advisor docs, @VaibhavSisinty, @aakashgupta). The two viral posts use hype framing ("pure goldmine") but point at real docs. Dynamic workflows are JavaScript scripts Claude writes and a runtime executes: the script holds the loop, branches and intermediate results, so dozens to hundreds of subagents can run while the main model's context holds only the final answer. The docs list quality patterns such as independent agents adversarially reviewing each other's findings. The first post also promotes "JEV engineering" for millisecond micro-decisions, which reads as decision-model promotion. Separately, Claude agent tracing is in public beta (latency, tokens and tool calls per step, with an opt-in detailed mode), and a Claude Code update sorts sessions by whether they need you and gives a recap card on return. @aakashgupta ties that to Gloria Mark's finding that an interruption costs about 23 minutes to recover from.
Databricks: 6x faster fine-tune at "98% utilization" (@FeitengLi). A Qwen3.8-27B fine-tune already showing 98% GPU utilization went from 7h44m to 1h17m by using fused kernels matched to the model's linear-attention layers (removing many small kernel launches) and repacking samples to remove padding. The poster's comment: routine work. The lesson is that utilization measures busy, not useful. Wiki: efficiency shorts.
HobbyLM: a 1B sparse MoE in ~300 GPU-hours (@NammaCoimbatore). A two-person team in Coimbatore trained a 1B-parameter sparse mixture-of-experts from scratch on 4 H200s in 76 hours, compared against ~123K GPU-hours for Sarvam-1. No paper or benchmarks in the post. Treat it as a cost anecdote.
System One decision models on the research side (cluster of 2) (@_reachsumit on JevGPT, @_vmlops on JevEval). JevGPT runs fast System One judges in parallel to check sources and return original passages, under a GPT planner that sets evidence goals. DeepEval's JevEval scores model outputs with explicit questions and calibrated probabilities rather than a generated score. Both sit next to the Jeva architecture paper. Wiki: Jeva and the System One cluster.
Muse Wonder lands at #12 in Agent Arena (@arena, @AIatMeta). After 5K+ real agent sessions: +6.4% net improvement at $1.13 median per task (2.1x Muse Spark 1.3), within 1 point of GPT-5.6 Sol at 25% lower cost. Meta's own post claims state-of-the-art results on agentic tasks; the arena number is more modest.
Anthropic deals and positioning (cluster of 2) (@rohanpaul_ai, @rohanpaul_ai). The Opaque Systems deal is a license plus hiring the CEO and CTO. Opaque runs AI jobs in hardware-sealed enclaves that produce cryptographic proof of which code touched which data. Anthropic's international chief told Bloomberg that open-weight models carry more safety risk by definition.
Hybrid Reasoning Agents, resurfacing (@omarsar0, @Jiaru_Zou, paper). Covered 10-14. The agent reasons in latent space and switches to text only to call tools or answer: +10.4% across 11 benchmarks with 35.5-54.7% fewer tokens and 2.7-3.7x lower cost. Today it gets a new pairing with the monitorability essay.
Jensen Huang on safety and GPU access (@rohanpaul_ai, @StockSavvyShay). He says he would not sell GPUs to irresponsible labs: "If the product is not safe, just don't release it."
Skip: a viral "Harvard proved universities are obsolete" thread and an invented Hassabis dinner quote (engagement bait with no source), Karpathy "keep Graph" clip-farming, travel-agent launch promos, and real-estate and trading ads.