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 Sunday. The strongest thread is a cluster of four posts on agent compute: Sakana's MASS (one model trains on its own best multi-agent runs), reposts of Meta's GitSwarm (500 agents sharing one Git repo) and Amazon's Research World Models (predict an experiment's gain before running it), and an argument that agent swarms force metered pricing. A second cluster of three is about inference money: the WSJ report that Amazon is in talks to buy Decart for about $7B to make Trainium faster, Jaya Gupta's "reverse auction" view of inference hosting, and a reported Anthropic cloud deal with IREN. A third, smaller cluster is harness talk: Pedro Domingos on Harness Learning ("the model swallows the harness"), Musk comparing it to Tesla FSD eating its C++ code, and a 48-page harness handbook. Standouts outside the clusters: Karpathy's capability funnel (about 20M professional frontier users, about 5,000 inside labs), Uber's 800-server MCP gateway design, the White House restoring UK model access for safety testing, and Meta Muse's retention numbers.
Posts
Agent compute that compounds (cluster of 4) (@SakanaAILabs, @hardmaru, @rohanpaul_ai, @rohanpaul_ai; papers 2610.12176, 2610.04862, 2610.12235). Sakana's MASS (Multi-Agent Self-Supervision, with UC Berkeley) has one shared 27B open model solve tasks as a team of virtual subagents, search for better team workflows, judge the results itself, and then train on the best team's runs. After two cycles, score per output token reached 1.2-1.6x the base model on four research benchmarks. Training only as the task solver also improved the model as optimizer and evaluator, and multi-agent traces were better training data than single-agent ones. Sakana flags the safety risk: when a model is its own evaluator, shared blind spots can be reinforced as progress. Meta's GitSwarm runs hundreds of agents over one Git repo where each commit cites what it builds on; it scored 79.4% vs 65.1% for one Codex agent on ProgramBench, and 70% of failed research commits were reused later. Amazon's Research World Models use logged experiments to predict whether a change will help before GPU time is spent; past records lifted ranking correlation from 0.506 to 0.774. See GitSwarm, Research World Models and MASS in the 10-11 harness summary.
Swarms force metered pricing (@rohanpaul_ai). In a swarm, a parent agent plans and child agents do narrow, easy-to-check jobs like searching files or running tests, which can be routed to the cheapest model that passes. So agent token volume grows faster than frontier-model revenue, and per-seat pricing, which relied on human typing speed to cap usage, gives way to metering.
Inference money (cluster of 3) (@rohanpaul_ai, @StockSavvyShay, @JayaGup10, @StockSavvyShay). The WSJ reports Amazon is in talks to buy Decart for about $7B, after Nvidia talks and a $6B Anthropic deal collapsed. Decart's DOS stack moves workloads across chips and took real-time video from 26 fps on GPUs to about 101 fps on Trainium, so Amazon is buying the software that makes its own chip cheaper per token. Jaya Gupta argues hosted inference is a "permanent reverse auction": easy capital, spare neolab compute and one-line switching keep pushing prices down. IREN reportedly signed a multiyear cloud deal with Anthropic, possibly 50 MW of B300 capacity. See the Decart summary and the industry summary.
The model swallows the harness (cluster of 3) (@pmddomingos, @elonmusk, @KirkDBorne; paper 2609.35738). Domingos points at Harness Learning, the CMU paper where a proposer model is trained to rewrite the agent's harness (the program that organizes model calls, tools and information flow) using execution feedback, framed as meta-learning over programs. His read: the learned model is absorbing what used to be hand-written scaffolding. Musk says the same happened with Tesla FSD, where the network ate the C++ code. A 48-page "Understanding Harness Engineering" handbook circulated alongside. Garry Tan pushed the other direction: replace LLM calls with deterministic code wherever possible. See harness engineering.
Karpathy's capability funnel (@karpathy, @IntuitMachine). Napkin math: about 6B people barely touch LLMs, 1-2B use free chatbots casually, about 20M (0.2%) use frontier models professionally in math and code and see weeks of work done by agents, and about 5,000 inside labs see swarms of thousands of agents running for weeks. He says the gap in shared understanding is widening into a funnel. Critics think 20M is too high.
Uber's MCP gateway (@stretchcloud). Uber's gateway fronts over 800 MCP servers and 5,000+ tools, mostly generated automatically. A control plane (registry for discovery and ownership) is split from a data plane (a proxy translating MCP calls into HTTP, gRPC and TChannel). A crawler finds new services and has an LLM write tool descriptions, but every server starts disabled until its owner enables it, and description changes need owner approval.
White House restores UK frontier-model access (@Miles_Brundage). A repost of a Tony Diver exclusive: the White House has reinstated the UK's access to frontier AI models for safety testing after ministerial lobbying. Only the headline was captured.
Agent failure diagnosis talk (@SharonYixuanLi; papers 2607.12747, 2606.26080, 2605.30834). A Simons Institute talk on three NeurIPS 2026 papers: locating which step in a failed agent trajectory caused the failure without expensive prompting pipelines, building step-level progress rewards for agents without human labels, and finding failure signals for robot (VLA) runtime monitoring.
Meta Muse retention (@deedydas). Day-7 retention above 30%, higher than ChatGPT, and 22% at day 14, just below it. Still far from WhatsApp (67%) or YouTube (59%).
Liquid open d1 for robots (@TheTuringPost). Liquid's open decision models (d1-3B, d1-omni-600M) report 35 ms per image on Jetson AGX Thor, pitched as cheap on-device checks between calls to a big planner. See decision models.
Mathematicians react to OpenAI's paper dump (@rohanpaul_ai). Some Northwestern students are asking whether to stay in mathematics after OpenAI posted hundreds of AI-generated papers on open problems; one professor says his priority is now keeping talent in the field.
Smaller items. GitHub Security Lab's Fuzzing Taskflow agent writes fuzzing harnesses, runs AFL++ and files one report per bug (@github). Tab exits stealth at a $300M valuation (@Ronald_vanLoon). A launch of fast production inference for open-weight models with $25 credits, reposted by Garry Tan and Y Combinator (@garrytan).
Skip. AWS webinar promo, a "Jev workflow cuts support costs 90%" hype thread, a fabricated-looking "Karpathy 2-hour lecture" promo, cold-email repo lists, Grok app promotion, and political reposts.