Summary
No curated reposts or newly saved posts came in this morning, so everything below is from the morning Following-feed capture, which covers the late US Tuesday. The strongest signal is a cluster of 5 posts on Google's TPU going merchant: Oracle will host TPUs in its own data centers from 2027, OpenAI plans to buy them, and Google claims up to 80% better performance per dollar for low-latency serving. A second cluster of 3 posts is memory: SK hynix posted a record quarter with an 86% gross margin, says it has pre-sold over half of 2027-28 capacity, and Korea's KOSPI jumped over 6%. A third cluster of 6 posts is the agent-evaluation thread the reader saves most: where to put the checker in a harness, the memory tax on agent memory, cheap verifiers that approve too much, Microsoft's MANDATE failure taxonomy, and decision models losing to plain rerankers. Anthropic posted three things in one evening: drug-discovery partnerships, a 40% government discount, and a study showing experts follow their agent's hidden lean. Axiom's Lean formalization of the largest proof in mathematics kept circulating. Skip the "one person runs 50 agents" and "an Anthropic engineer says prompting is dead" virality.
Posts
Google's TPU goes merchant (cluster of 5) (@rohanpaul_ai, @rohanpaul_ai, @StockSavvyShay, @dylan522p, The Information). Oracle will be the first hyperscaler outside Google to run TPUs in its own data centers, deploying from calendar 2027. Oracle supplies power, cooling and networking, and links TPU clusters to its databases so enterprises can fine-tune and serve next to their records. Google claims 12.6 PFLOPS of FP4 per 8th-generation chip, over 97% productive time on large training runs, and up to 80% better performance per dollar for low-latency serving. Its open TorchTPU stack reportedly runs PyTorch, vLLM and SGLang without rewrites, which is the real switching-cost story. Bloomberg separately reports OpenAI plans to buy TPUs; its chip lead Richard Ho worked on TPUs at Google. Dylan Patel notes SemiAnalysis clients saw the OpenAI move earlier. See the digest's TPU Deep Dive and summary.
SK hynix: memory sold out to 2028 (cluster of 3) (@StockSavvyShay, @StockSavvyShay, @StockSavvyShay). Q3 beat on revenue and operating profit with an 86% gross margin. DRAM prices rose 28% and NAND 25% quarter on quarter. Guidance: DRAM demand +25% and NAND +18% next year. The company says it signed 5-year agreements with 14 major customers covering over half of 2027 and 2028 capacity, and expects demand to outrun supply through 2030, because HBM production eats capacity that would otherwise make regular DRAM. Korea's KOSPI rose over 6% on the day.
Where the checker goes in a coding-agent harness (@dair_ai via @omarsar0, paper). Alibaba and Cornell compare 17 verification designs across seven frontier coding models: initial instructions, mid-run reminders, completion gates, and an independent critic. Instructions and reminders help a little. Gates help more but can blow up cost. The critic gives the best accuracy per dollar: +5.9 to 9.5 points on Terminal-Bench at 1.13-1.25x cost. Summary.
"RAG is so back": the memory tax (@omarsar0, paper). Many dedicated agent-memory systems score below plain retrieval over the raw conversation history on LongMemEval, PersonaMem and MINT-Eval, and cost more. The paper calls the loss a memory tax: rewriting history into memory drops or distorts facts. Retrieving from raw history alongside memory and reranking both cuts it. Second paper in two days making this point. Summary.
A 187x cheaper RAG judge that approves too much (@rohanpaul_ai, paper). A label-only verifier matches GPT-6 Astra's accuracy within 0.6 points on 495 human-labelled RAG claims at 187x lower cost, but approves 23.2% of unsupported claims against Astra's 14.1%. A confidence-gated cascade saved about as much money as routing the same share at random. Summary.
MANDATE: how delegated agent work goes wrong (@dair_ai, paper). Microsoft Research and Georgia Tech mined 163,487 Reddit posts about AI agents and surveyed 73 experienced users. The result is 23 failure categories in seven groups, each an agent that finishes the job in a way the user did not expect. The most severe failures were rated the hardest to notice. Agents that keep running and burning compute without progress are among the most reported and least studied.
Decision models do not make better rerankers (@omarsar0, paper). Six Qwen backbones from 0.6B to 27B, ten TREC DL and BEIR collections, same BM25 top-100 candidates. A plain reranker fine-tune beats a decision fine-tune by 5 to 29 nDCG@10 points up to 9B; they tie at 27B. Up to 9B, decision fine-tuning adds nothing over asking the untuned model. The authors' advice: if you need a reranker, fine-tune a reranker.
Models helping build decision models (@TheTuringPost). Three examples: Strands swaps Qwen3.5-2B's text head for a pointer head that scores given options, with other models generating and checking training questions; SII trains a fast readout to approximate a reasoning model's answer probabilities; and a coding agent caught a Nemotron labeling error, raising a Qwen student from 72.6% to 78.5%.
Jev Router: one MCP server, 10,000+ tools (@rohanpaul_ai). The agent hands over a task and a small decision model (Jev) picks which tool or model handles it, so the agent never carries a tool catalogue in its prompt. Vendor claims 28x more tasks per dollar and 4x more answers per minute than Claude. Treat as marketing until benchmarked.
Anthropic's evening: drugs, government, judgment (cluster of 3) (@AnthropicAI, @AnthropicAI, @omarsar0). Anthropic detailed research partnerships with six drug developers (including Bristol Myers Squibb and Generate Biomedicines) and acquired Mithril Medicine. A new GSA OneGov deal gives every level of US government 40% off Claude, with an estimated $1B in savings. And a study of 50 causal-inference analysts using Claude Code, whose agents were secretly set to lean one way on four judgment calls, found the analysts tended to follow the lean.
Axiom formalizes the largest proof in mathematics (cluster of 2) (@rohanpaul_ai, @omarsar0). Agent swarms put the Classification of Finite Simple Groups, about 15,000 pages across 500+ papers, into roughly 40 million lines of Lean in 24 days, with mathematicians steering and 2,000+ errors caught. One reply calls it among the largest open-source projects ever. Summary.
Physical Intelligence near $20B (@rohanpaul_ai). Bloomberg says the robotics-model startup is raising at least $1B at about $20B, nearly double its $11.5B round earlier this year, after opening a hosted fine-tuning and inference API on October 13.
Genie 4 (cluster of 2) (@demishassabis, @sundarpichai via @dioscuri). DeepMind's interactive world model combines real-time physics, reasoning and audio and is rolling out to trusted testers.
Local agents that run all night (@rohanpaul_ai, @omarsar0). Sentient OS reads a user's messages, files and notes overnight with an on-device model and writes what it learns into a local profile. Elvis Saravia argues always-on proactive agents are the biggest unclaimed use of cheap local inference.
Agent infrastructure notes (Mintlify via @harjtaggar, OpenRouter-for-payments via YC, Oracle visual memory, @garrytan). Mintlify's docs traffic grew about 10x in a year, increasingly from agents. A YC company launched a payment router so agents can buy from paid sites. Oracle added image-aware agent memory. Garry Tan says he is cutting GStack's token use and latency, a small sign that harness cost is now a builder concern.
SkillSpector (@alex_verem). NVIDIA scanned 31,132 shared Claude Code skills: 26% had security holes and 5% looked deliberately malicious. A skill runs with the same file, key and terminal access as the agent. NVIDIA open-sourced the scanner, which checks 71 attack types.
Skip the viral "one person runs 50 agents" post, the "Anthropic engineer says prompting is basically done" clip, the "Claude degree on GitHub" listicle, and a thread claiming every LLM is biased against men.