Summary
No curated reposts or saved posts landed in the morning slot, so the signal comes from the morning capture of the Following feed, which covers the late US Monday. The strongest signal is a cluster of three posts on Anthropic's two Monday research posts: distillation defenses that Anthropic says cut harvested reasoning data by 90-99%, and a Frontier Red Team study showing small open models are close to self-replicating across cheap IoT devices. The second cluster (four posts) is infrastructure economics: Lisa Su's raised CPU forecast, Zuckerberg on H100s holding value, a Navitas 800V power board for NVIDIA racks, and UBS on Cerebras. A third cluster covers Meta's Muse Wonder launch, now with a Vals Index #5 placement. Standout single items: NVIDIA's LoGRA paper on running full-weight RL in less memory than LoRA, Alibaba's QwenGyre on idle GPUs during agent rollouts, Microsoft's DUET on co-training graders, and LMArena's finding that July's best agent is now merely average at up to 96% lower cost. Decision models keep trending: a Jev-in-a-harness test and a note on SubQ Snap's linear latency scaling.
Posts
Anthropic on distillation and self-replication (cluster of 3) (@AnthropicAI via post, @ns123abc, @ChrisGPT, self-replication post). Distillation attacks harvest a model's reasoning traces through fake-account networks to train a cheaper copy. Anthropic's new defenses: "preserved thinking" stops new API accounts from editing prior context, reasoning is readable only by the requesting organization, and a two-stage classifier with under 0.1% false positives blocks extraction attempts, which are now billed. Summarized reasoning alone dropped a distilled model from 80.4% to 71.8% on SWE-bench Verified. The red-team post ran Gemma4-26B across 15 $55 Android TV boxes at 0.39 tokens per second; a 9B model set up inference on 8 of 9 unfamiliar hardware types; GLM-5.2-class models completed end-to-end replication in simplified tests. The viral thread's framing ("unkillable compute", "digital slime mold") goes beyond the post, which assumed rather than chained the exploit steps. Ed Zitron's reply asked how much Anthropic revenue came from distillers. Wiki: Anthropic self-replication and distillation defenses.
Compute economics (cluster of 4) (@StockSavvyShay on Su, on Zuckerberg, on Navitas, on Cerebras). Lisa Su says server CPU demand could grow ~50% a year to $200B+ by 2030, her third raise in nine months, driven by agent workloads. Zuckerberg says an H100 is worth more now than a year ago, which supports longer GPU depreciation for neoclouds. Navitas and NVIDIA's 24 kW board converts 800V directly to 12V; 1.5 points of efficiency frees ~15 MW per GW. UBS sees Cerebras revenue ~4x to $3.3B in 2027 with ~$25B in contracted backlog, arguing on-chip memory matters more as agents make dozens of calls per task. Analyst framing throughout. Wiki: OCP 2026 and Maia 300.
Muse Wonder (cluster of 3) (@alexandr_wang, @AIatMeta, @rohanpaul_ai). Wang says training was steered by an internal Personal Superintelligence Index (PSI) built to measure usefulness for personal agents. Vals AI placed Muse Wonder Full (Max) at #5 on its index, above every OpenAI model. Zuckerberg named "memory and persistence" as next priorities for weeks-long tasks. Wiki: Meta Muse Wonder.
LoGRA: RL in less memory than LoRA (@omarsar0, paper). NVIDIA compresses each weight gradient with a random rank-256 projection as microbatches accumulate, so the full gradient is never stored, and rebuilds the update from two small matrices. A predicted-KL controller scales each step. 7.2 GiB vs 13.2 GiB for rank-256 LoRA at 1.5B with equal Pass@4; 45.7% less memory at 7B; a 27B model trains on one 8-GPU node. Wiki: efficiency shorts.
QwenGyre: idle GPUs during agent RL (@VaibhavSisinty, arXiv 2609.33848). Agent rollouts of 700K+ tokens leave GPUs waiting on tools. An elastic scheduler lends them to other training and reclaims them; shared trajectory prefixes are not recomputed; partial progress after timeouts is scored. Up to 1.85x faster than colocated training.
DUET: co-evolve the grader (@dair_ai, DAIR). Alternate updates to a solver agent and its LLM grader. Finance Agent Benchmark 96.5% vs GEPA 94.4%; GDPval grader agreement 68.4% to 78.5%. Updating the grader with the solver frozen barely helps.
Harness learning, transferable (@Xudong07452910, arXiv 2609.35738). A 4B "proposer" trained with RL to revise a solver's harness (model calls, tools, information flow) beats a 35B teacher on 21 unseen Reasoning Gym tasks (0.62 vs 0.56 average revision score) and keeps improving over 10 rounds at test time. Covered in the 10-04 and 10-12 digests; this is a Chinese-language explainer.
Agent Arena: the cost of "good enough" collapses (@arena). Claude Fable 5 led on 1 July; the top-20 average matched it after 12 weeks. Kimi K3 matched it in 3 weeks at 89% lower cost per task, DeepSeek V4.1 Flash at 96% lower. Top five now confirm 94% of tasks done, up from 80%.
Decision models in practice (cluster of 3) (@akshay_pachaar, @rohanpaul_ai, @huggingface RT of mrm8488). Jev inside an open agent harness matched Opus on 40 decisions at ~150x lower decision cost, routing low-confidence answers to review. A note on SubQ Snap: 660 ms for 100K tokens and ~2.5 s for 384K, roughly linear, as expected from subquadratic attention. Manuel Romero's Typed Decision on EmbeddingGemma 2 returns calibrated typed answers from an encoder. Wiki: decision-model price war.
Prompt optimization for a 17x cost cut (@artemg314). Collect agent trajectories, build a golden train/test set, run GEPA so a cheaper model matches a larger one. The eval set was the slow part.
Perplexity's agent revenue (@rohanpaul_ai). Srinivas says revenue grows 5x YoY because agents need search, a browser and multi-model switching; sovereign funds asked to run it entirely on their own hardware.
EmbeddingGemma 2 (@TheTuringPost). A 740M on-device multimodal embedding model; Matryoshka truncation to the first 256 of 768 dims keeps ~95% retrieval quality at 1/3 storage.
TencentDB Agent Memory (@Divyyanshishrma). Open-source layered agent memory (conversation, atom, scenario, persona) in readable markdown on SQLite; claims 61% fewer tokens and PersonaMem 48% to 76%. The "killed the vector database industry" framing is hype.
Skip: job ads, an airship funding congratulation, sponsored enterprise-AI graphics, a synthetic-actor ad thread, and politics.