AI research wiki
social-stream · 2026-10-28

2026-10-28

Summary

Compute supply ran through the whole day. Google's TPU going merchant showed up in both the morning and evening slots: Oracle will host TPUs from 2027, OpenAI plans to buy them, and Google claims up to 80% better performance per dollar for serving. It sat next to a memory squeeze (SK hynix at an 86% gross margin, with over half of 2027-28 capacity pre-sold) and GB300 rents near $14/hr. The best single post was the evening thread on DeepSeek's open-sourced DeepInfer, which serves KV cache hits from SSDs and gets a 128K-token hit to first token in 500ms instead of 13s. The morning's agent-evaluation cluster was the strongest research block: where to put the checker in a harness, the memory tax on agent memory, and cheap verifiers that approve too much, and it fits what the reader has been saving. Anthropic's IPO discourse, the "coding interview is dead" takes and the 50-agents virality were mostly noise. The afternoon slot was empty.

Posts

  • Google's TPU goes merchant (cluster of 6) (@rohanpaul_ai · @StockSavvyShay · @dylan522p · @rohanpaul_ai · The Information) [morning + evening]. Oracle runs TPUs in its own data centers from 2027 and OpenAI plans to buy them. TorchTPU running PyTorch, vLLM and SGLang without rewrites is what actually lowers switching costs. Summary.
  • SK hynix: memory sold out to 2028 (cluster of 3) (@StockSavvyShay · @StockSavvyShay) [morning]. 86% gross margin, DRAM prices up 28% quarter on quarter, and 5-year deals covering over half of 2027-28 capacity. HBM is eating capacity that would otherwise make regular DRAM.
  • DeepSeek DeepInfer: KV cache served from SSD (@_avichawla) [evening]. MLA shrinks the cache enough that reading it from distributed SSDs beats recomputing prefill, and a cache hit costs 50x less than a miss. See KV cache and DeepSeek KV cache economics.
  • Cache engineering for agents (AKV) (@TheTuringPost) [evening]. New "drop" and "repos" ops let agents prune tool outputs without recomputing the cache: 27.7% lower cost. Already covered in AKV.
  • GPU rents and infrastructure money (cluster of 3) (@StockSavvyShay · IREN · Vertiv) [evening]. GB300 rents run near $14/hr against $3.50 for an H100, and IREN borrowed $363M against its GPUs. Supports rising GPU rents.
  • Where the checker goes in a coding-agent harness (@omarsar0 · paper) [morning]. Of 17 verification designs, an independent critic gives the best accuracy per dollar: +5.9 to 9.5 points on Terminal-Bench at 1.13-1.25x cost. Summary.
  • The memory tax (@omarsar0 · paper) [morning]. Many agent-memory systems lose to plain retrieval over the raw history and cost more. Reranking raw history together with memory cuts the loss. Summary.
  • A 187x cheaper RAG judge that approves too much (@rohanpaul_ai · paper) [morning]. Accuracy within 0.6 points of the frontier judge, but it lets through 23.2% of unsupported claims, and the confidence cascade saved no more than random routing. Summary.
  • Decision models vs rerankers (cluster of 2) (@omarsar0 · paper · @omarsar0) [morning + evening]. A plain reranker fine-tune beats a decision fine-tune by 5 to 29 nDCG@10 points up to 9B, and the two tie at 27B.
  • MANDATE failure taxonomy (@dair_ai · paper) [morning]. 23 categories of delegated agent work going wrong, mined from 163K Reddit posts. The most severe failures were also rated the hardest to notice.
  • Models building decision models (@TheTuringPost) [morning]. Pointer heads, fast readouts and agent-caught labeling errors (one fix took a Qwen student from 72.6% to 78.5%).
  • Agent tool routing (cluster of 2) (@rohanpaul_ai · @rohanpaul_ai) [morning + evening]. Jev Router puts 10K+ tools behind one MCP server, and a second post argues tool choice now matters more than which model writes the answer. Vendor numbers are unverified. See tool calling.
  • Axiom formalizes the largest proof in mathematics (cluster of 3) (@rohanpaul_ai · @omarsar0 · @rohanpaul_ai) [morning + evening]. Agent swarms put the 15,000-page proof classifying finite simple groups into Lean in 24 days, catching 2,000+ errors. Summary.
  • Anthropic's day (cluster of 3) (@AnthropicAI · @AnthropicAI · @omarsar0) [morning]. Drug-discovery partnerships, 40% off Claude for US government, and a study in which analysts followed their agent's hidden lean.
  • SkillSpector (@alex_verem) [morning]. NVIDIA scanned 31K Claude Code skills: 26% had security holes and 5% looked deliberately malicious. The scanner is open-source.
  • Physical Intelligence near $20B (cluster of 2) (@rohanpaul_ai · @rohanpaul_ai) [morning + evening]. It is raising $1B+ at nearly double its last valuation. Separately, Nvidia, AMD and Broadcom have shown nonbinding interest in up to $5B of Meta stock (@rohanpaul_ai).
  • Genie 4 (cluster of 3) (@demishassabis · @dioscuri · @rohanpaul_ai) [morning + evening]. DeepMind's real-time world model with native audio. No efficiency angle.
  • Liquid AI LFM3 (@rohanpaul_ai) [evening]. New open-weight family. The 1.7B model beats its 2.6B predecessor, which makes it worth a look for small-model efficiency.
  • Local always-on agents (@rohanpaul_ai · @omarsar0) [morning]. Sentient OS builds an on-device profile overnight. The case made is that proactive agents are the best unclaimed use of cheap local inference.
  • Agent infrastructure notes (@harjtaggar · @ycombinator · @TheTuringPost · @garrytan) [morning + evening]. Mintlify's docs traffic is now increasingly from agents, a YC company launched a payment router for agents, Oracle added visual agent memory, and GStack is cutting its token use.
  • Photonic GPU disaggregation at 40km (@Ronald_vanLoon) [evening]. IOWN claims a 1.9% penalty at 40km. It is tagged #AD, so treat it as vendor framing.
  • Truncated reposts, click through (@TheTuringPost · @thegautamkamath · @rohanpaul_ai) [evening]. Papers of the week, how research teams use coding agents (likely the best of these), and Pichai on agents.
  • Chasing the Karpathy peak (@IntuitMachine · essay) [evening]. Substack essay posted with no commentary. Skim only.
  • The coding interview is dead (cluster of 2) (@omarsar0 · @VaibhavSisinty) [evening]. Anthropic dropped coding puzzles because Claude beats candidates at them. Opinion, no data.
  • Anthropic IPO discourse (@GaryMarcus · @rohanpaul_ai · @GaryMarcus) [evening]. Criticism of the S-1's $40T market sizing and a debate about a unilateral pause. Heat, little new information.
  • Viral and noise posts (50-agents post, "prompting is dead" clip, Claude-degree listicle, a bias thread, EMNLP event posts, tickers, hashtag spam) [morning + evening]. Skip.

See the daily digest for the full synthesis.