Summary
No curated reposts or saved posts landed this morning. This synthesis comes from the morning capture of the Following feed, which covers the US Monday daytime and evening. The strongest signal is a cluster of 7 posts arguing over whether there are too many GPUs or too few. Business Insider reports Amazon has warehouses of GPUs it cannot install, while Nebius B200 spot prices reportedly tripled and Gartner and Morgan Stanley published trillion-dollar spend forecasts. A second cluster of 4 covers Meta's infrastructure day: the Prometheus supercluster write-up, its reliability numbers, and the Muse Glimmer 30B open-weight report. Inference efficiency had a strong showing outside the clusters: FlashMap (a new speculative decoding drafter, up to 6.1x), vLLM Speculators v0.9, and an llm-d poster on tracking where offloaded KV blocks live. Agent papers dominated research reposts: DeepMind's Busy Beaver discovery agent, a CoT-controllability result with a monitoring twist, a Salesforce study of agents that never check each other, and Meta's ProPerAgent harness. Funding news centered on Greptile's $50M Series B.
Posts
GPU glut or GPU shortage? (cluster of 7) (@edzitron, @edzitron, @GaryMarcus, @StockSavvyShay, @StockSavvyShay, @rohanpaul_ai, @rohanpaul_ai). Business Insider reports Amazon has warehouses of GPUs it cannot yet install. Its "early rack receipt" program takes about $3B of GPUs a month with no clear forecast path to installation. Critics, including Ed Zitron and Gary Marcus, read this as Nvidia channel stuffing and proof of a bubble. On the other side, Nebius B200 spot pricing reportedly more than tripled in seven days. IREN's CEO says 2026 capacity is sold out, contracts are priced at more than double last year's, and bringing GPUs online is the bottleneck. Gartner forecasts $6T of AI infrastructure spend in 2027-2030 with over $3.5T at risk of write-downs. Morgan Stanley puts 2027 spend at $1.58T once leases count as capex, with Nvidia taking about 44%. Read together, both sides fit one story: powered floor space, not chips, is the limit. See the industry roundup.
Meta's infrastructure day (cluster of 4) (@AIatMeta, @AIatMeta, @StockSavvyShay, @StockSavvyShay). Meta says Prometheus, its gigawatt cluster in Ohio, trained the Muse models with over 85% goodput (useful training time) and about 47K GPU-hours between interruptions. It runs hundreds of thousands of GB200 and GB300 GPUs today, heading for about 1 million with Vera Rubin next. Meta also published a tech report on Muse Glimmer, a 30B open-weight model for agent tasks that runs at 4-bit on one consumer GPU and leads similar-size open models on tool use and planning. See Prometheus.
FlashMap: speculative decoding with flow maps (@jeff_shen_; paper). Speculative decoding has a small drafter propose a block of tokens that the big model verifies in one pass. Block-diffusion drafters like DFlash write the block in one pass but predict each token independently. FlashMap conditions a flow-map language model on the target's hidden states, so it writes the whole block in one pass while modeling how the tokens depend on each other. Up to 6.1x lossless on Qwen3-8B and 7% more average speedup than a matched DFlash. See speculative drafting.
vLLM Speculators v0.9 (@RedHat_AI; release). The library that trains drafters for vLLM got a fused loss kernel (up to 2x faster forward passes on ~1B drafters) and an FP8 hidden-state backend that halves storage with no loss in drafter quality.
Where offloaded KV blocks live (@PyTorch). When GPU memory fills, vLLM offloads KV cache to CPU or storage, and llm-d tracks where every block lives and routes each request to its cache. A PyTorch Conference poster from Moreh will show when CPU offload is enough and when shared storage pays. Same theme as AWS's KV transfer posts.
llm-d Semantic Classifier (@RedHat_AI). An X Article on adding a difficulty signal at the gateway so a simple lookup and a multi-step proof go to different models. The article body was not captured; click through to read.
CompileIQ + Helion (@PyTorch). NVIDIA's CompileIQ exposes ptxas compiler decisions (register allocation, fine scheduling) as control files, and PyTorch Helion's autotuner now searches them together with tile shapes, warps and pipeline stages. The pitch is extra speed on kernels already tuned to the limit.
Trainium joins PyTorch natively (cluster of 3) (@PyTorch, @PyTorch, @StockSavvyShay). TorchNeuron registers Trainium as a native PyTorch device, built on the same extension points other accelerators use. Switching from CUDA becomes one setting. AWS is bringing several sessions to PyTorch Conference.
Marvell: memory 50 meters from the GPU (@rohanpaul_ai). A photonic module moves HBM and DDR traffic over optics so the memory can sit up to 50 m from the GPU instead of millimeters, with latency in the hundreds of nanoseconds. If it holds, memory and compute scale separately.
Nvidia Vera CPU wins Modal (@StockSavvyShay). Modal, which used AMD and Intel CPUs, reportedly moves to Nvidia Vera after testing showed 2x capacity for agent workloads. Agents run code and call tools constantly, so CPUs matter more. Nvidia targets about $20B of CPU revenue this fiscal year.
Looped Transformers reading list (@TheTuringPost). A list of 10 new looped-transformer papers (models that rerun the same layers to think deeper) was one of the most-shared AI posts of the window. It matches today's HF papers LEV and Looped Models Done Right.
DeepMind's Busy Beaver agent (@omarsar0; paper page). A Gemini-based agent searched the BB(6) frontier of the Busy Beaver problem, picked its own Turing machines, found patterns in their execution traces, and proposed compact rules that summarize many steps. Each rule was proven in Lean, and only proven rules were kept. Eleven previously unresolved machines now have exact descriptions, though whether they halt is still open.
Reasoning models can control their CoT after all (@dair_ai; paper). Earlier tests put a constraint on the chain of thought in the user prompt and found models rarely followed it. Placed as a self-instruction inside the chain of thought, the constraint is followed far more often, and a generic reminder helps too. The catch: the authors show this control can be used to evade a CoT monitor. See oversight shorts.
Agents for different parties never verify each other (@rohanpaul_ai). Salesforce ran over 2,500 agent-to-agent sessions across 488 scenarios using Claude Code, Codex and Cursor in separate sandboxes and found 20 failure modes. No agent re-fetched a document to check a peer's claim. One $24,000 dispute settlement passed all 10 outcome checks while the vendor agent leaked confidential details. The advice: score each agent against the party it serves.
ProPerAgent: a harness for proactive assistants (@dair_ai). Meta's harness decides when a wearable assistant should speak. It spots objects in view, retrieves personal memories from a graph, and runs a separate LLM check on whether the memories justify interrupting. Accuracy rises from at most 15% with standard memory to 38% with no weight changes. Removing the filter drops it to 24.0%, and unneeded interruptions fall to 15.3% vs 77-84% for standard RAG.
Harness² and self-tuning agents (cluster of 4) (@openmys, @wohsj110, @RoundtableSpace, @_guillecasaus). Google Research's open-sourced Harness² lets an agent rewrite its own prompts, skills and memory by running the same task repeatedly and comparing results, with no weight updates. A JEV team says plugging it in improved vision benchmarks round after round; their repo is not out yet.
Human-centered coding agents (@math_rachel; paper). A position paper argues the bottleneck for coding agents is no longer task-solving but how people communicate with, supervise and trust them. It names four dimensions: task alignment, verifiability, steerability and adaptability.
Mercury 3, a diffusion LLM (@deedydas). Inception's Mercury 3 generates text by diffusion instead of token by token. It claims 4-8x the speed of comparable small models at the same price, and OpenRouter data shows 2-4x effective.
DeepSeek ban in US civilian agencies (@rohanpaul_ai). DHS ordered civilian agencies to remove DeepSeek and every product built on it, including US-hosted weights, citing data flows to China for the hosted service and possible vulnerabilities for weights. Federal software vendors now have to show which open-weight models sit inside their products.
Greptile's $50M Series B (@ycombinator; blog). Led by Menlo Ventures at a $750M valuation. Greptile reviews and tests over 1% of all GitHub pull requests daily, acting as an independent validation layer whatever coding agent wrote the code.
Claude Code model-mixing hype (cluster of 2) (@mirku21, @Sprytixl). Viral threads repackage Claude Code's documented subagent model setting (Haiku for search, Sonnet for the main session, Opus for review) as a secret trick. The underlying idea, routing roles to cheaper models, is real and matches Cohere's 10-19 result; the framing is engagement bait. Skip.
Skip: AI-infra career listicles, procurement-AI promos, RAG explainer graphics, and partisan GPU rants without new facts.