Summary
The morning slot captured no reposts and no tweets from tracked handles, so this synthesis reads the morning Following-feed capture, which covers the US Sunday afternoon and evening. The strongest technical signal is a cluster of 4 harness posts arguing for less automation, not more: Elvis Saravia's "boring deterministic steps" thread and its follow-up, Lauren Tan's four-question triage for agent mistakes, and the Northwestern and Google self-improving harness paper. The best paper in the slot is a Google DeepMind and Columbia system that lets agents debug chip simulations by turning 10 GB waveform dumps into a SQL database. Leaderboard news came from Agent Arena: Claude Fable 5.2 (Max) took first place with a record margin and Qwen4-Max debuted as the top Chinese model. The Nvidia story produced a cluster of 4 posts, led by Ed Zitron's claim that Microsoft cancelled some GPU orders and a PageRank analysis of how the press coverage spread. Samsung's $1.9B Austin expansion for Starlink chips was the slot's one semiconductor item. No new saved posts were captured this run.
Posts
Harnesses need boring deterministic parts (cluster of 4) (@omarsar0, @omarsar0 follow-up, @0x_kaize on Lauren Tan, @rohanpaul_ai). Elvis Saravia says the best agent systems he has built or seen "involve lots of boring deterministic steps." Adding more LLM agents compounds errors, so you need classifiers, verifiers and evaluators in the loop. In a follow-up he says most current harnesses try to automate everything and invite more failures during long runs. Lauren Tan, who ships over 3,000 pull requests a month with agents, applies the same idea to mistakes. Each one goes through four questions: can the code be redesigned so the mistake is impossible, can a lint rule or type catch it, can the agent catch it by running its own checks, and does it really need a written rule. A new AGENTS.md line is the last resort, not the default. Rohan Paul's repost of the Northwestern and Google paper (agent harnesses that improve on new tasks without answer keys, by testing each proposed edit on real runs) is the research version of the same discipline. See agent harness engineering.
Agents that debug chips through SQL (@rohanpaul_ai, paper). Chip simulations log every signal change, and the waveform dumps can exceed 10 GB of text that engineers still read by hand. The paper notes about 74.6% of LLM-for-EDA (electronic design automation) work targets RTL code generation, not this verification step. BTTF converts the dumps into a compact SQLite database, and a team of agents turns English questions into SQL, checks design rules and audits each answer. It gets 95.33% of 150 expert-labelled questions right with Gemini 2.5 Pro. The multi-agent split scores 17.3% higher on a quality audit than one agent at 1.96x the latency. The takeaway generalizes: when the data will never fit in context, give the agent a query interface. See the BTTF summary.
Agent Arena: Fable 5.2 first, Qwen4-Max the top Chinese model (cluster of 3) (@arena, @arena on Qwen4-Max, @Alibaba_Qwen via @arena). Claude Fable 5.2 (Max) took first place on Agent Arena with what the leaderboard calls an unprecedented lead, a +30.4% net improvement. Qwen4-Max debuts at #14 overall with +7.5% net over Qwen3.8-Max, making Alibaba the top Chinese lab and fifth overall, behind Anthropic, Google, OpenAI and Meta.
Antigravity doubles context (@_mohansolo via @demishassabis). Google's Antigravity coding environment is doubling the context window for its Argon, Carbon and Hydrogen models. No numbers on cost or latency in the post.
Nvidia after Huang (cluster of 4) (@edzitron, @edzitron on capacity, @varun_mathur, @rohanpaul_ai). Ed Zitron posts what he says is an email from Microsoft CFO Amy Hood cancelling some GPU orders, and argues Huang had feared overbuying TSMC and memory capacity back in August, the same month Nvidia guided about 70% growth for FY2028. Neither claim is independently confirmed. Varun Mathur ran a swarm of agents over about 1,660 stories from 40 outlets (27 Oct to 14 Nov) and applied PageRank to their citations. He finds the New York Times the most-cited origin, with Bloomberg (140 stories) and the WSJ (99) driving volume. Nvidia has lost over $1 trillion in market value since the first NYT story. Rohan Paul reports the reinstatement campaign "gaining fast." Nvidia reports earnings on 11-17.
Samsung $1.9B for Starlink chips (@SawyerMerritt via @elonmusk). Samsung will expand its Austin, Texas fab with a $1.9B investment to make semiconductors for SpaceX's Starlink.
Sinkhorn balancing, a 1963 result still everywhere (@docmilanfar). Peyman Milanfar on Sinkhorn's algorithm: alternately normalize a non-negative matrix's rows and columns until it is doubly stochastic. It solves entropic optimal transport and shows up in PageRank, SLAM, multi-omics alignment and market design. It is also the routine behind balanced token-to-expert assignment in some MoE (mixture-of-experts) routers.
e-means, second day (@HowToPrompt__). The Meta clustering result from the 11-15 slot is recirculating with hype ("killed a 70-year-old algorithm"). The substance: an EMA (exponential moving average) update of cluster centers from small batches, with scaling laws that set batch size and step size automatically.
OpenAI's science blog launches (@OpenAI, post). Physicist Nathaniel Craig describes working with GPT-6 Astra to develop a candidate non-perturbative lattice definition of the Standard Model and turn it into a detailed proof.
Zep Agent Skills (@TheTuringPost). Zep introduced a procedural memory system that turns agent experience into reusable skills. The capture holds only the repost stub, so no details.
Models versus systems (cluster of 3) (@GaryMarcus, @fchollet, @ChrSzegedy). Gary Marcus argues a system that combines neural networks with symbols (Lean checking an LLM's proofs, or Claude Code running tools) is by definition neurosymbolic, and that you cannot understand current AI without separating the model from the system. François Chollet says the substrate does not matter: transformers trained to be inductive (reasoning models, a 2024 idea) gained real fluid intelligence. Christian Szegedy traces that training method to STaR (self-taught reasoner, 2022).
Skip. The "GPU-like minds" education post, the Dwarkesh-critique thread, the "lock yourself in for 15 minutes" agents link, the "hire a crew of Claude agents with 20 repos" list, Cohere's North 2 certification post, GitHub's Copilot security extension, and Elon Musk's off-topic posts.