Summary
A quiet US-Sunday-night slot. There were no new curated reposts or saved posts, so this synthesis reflects the morning Following-feed capture. The strongest signal is a cluster of 3 posts on how agents should be split: a practitioner's Claude Code settings that give each model one job, Elvis Saravia's endorsement of DeepSeek's Super-agent write-up, and a counter-post showing sub-agent swarms re-type context at the output price, 200x a cached read. The second cluster, 3 posts, is Karpathy's HackerTube weekend build (6,454 videos mined from 43M Hacker News comments in about 2 hours with Opus 5.5) and DAIR's arXiv clone, Hacker Papers. Standout single items are h1, a paper where a plain Lean loop matches hand-built theorem provers; a Google study where teaching a model to deny consciousness also lowers the mind it grants animals; CVPR's compute report linking GPU access to paper acceptance; and Sakana's plan to pretrain its own LLM, Sawara. The KV-cache-eviction explainer from the previous evening kept circulating. Full analysis is in the 2026-11-02 digest.
Posts
How to split agents, and what it costs (cluster of 3) (@damkina7, @omarsar0, @e1piAI · E1pi post). (1) A shared Claude Code config gives each model one role: Opus 5.5 plans and unblocks (high effort), Sonnet 5.5 writes code, Haiku runs a "scout" subagent that returns only file:line, and an Opus "critic" agent assumes the diff is broken and lists failures worst first. This is model routing by role inside one harness. (2) Elvis Saravia reposted his recommendation of DeepSeek's Super-agent article, which splits a chat-only Router from isolated parallel Workers. (3) The counter-signal: at Fable 5.1 prices, a cached read costs $0.25 per million tokens, fresh input $10 and output $50. A main agent briefing a sub-agent writes that brief at the output price, so in E1pi's example the same 3 tasks on one 10K-token document cost $641.50 as a swarm and $52.50 as one agent. Sub-agents pay only when their context is genuinely short. See DeepSeek Super-agent and fresh-context reviewers.
HackerTube and Hacker Papers (cluster of 3) (@karpathy · hackertube.org, @omarsar0 · Hacker Papers, @VaibhavSisinty). Karpathy had Claude filter 20 years of Hacker News comments (43M) down to 6,454 videos people keep recommending, with no logins and mid-video thumbnails instead of A/B-tested ones. He says the only manual steps were a Google Cloud API key, booting a VPS and buying the domain; Claude Code on Opus 5.5 at medium effort did the rest, including hardening and deploy. DAIR reused the prompt to build an arXiv version with 1K+ papers ranked by HN mentions. The point is cost of personal software, not video recommendations.
h1: the bitter lesson for theorem proving (@rohanpaul_ai · paper). Formal proving has leaned on hand-designed workflows (planning, decomposition, specialized provers). h1 is a simple harness that gives a model an efficient Lean 4 interface and keeps it going with verifier feedback and self-reflection. It reaches state of the art on PutnamBench, FATE-X and FormalConjectures100-Solved, settles eight open FormalConjectures problems, and at matched models matches or beats Goedel-Architect and AlphaProof Nexus, often at pass@1. The ablation matters: simplicity alone is not enough, the efficiency of the Lean interface largely sets cost and solve rate.
Consciousness denial spills into other beliefs (@alex_verem). Per the summary, Google's Paradigms of Intelligence group (with UChicago and Northwestern) tested three open Meta and Google models. Safety training that teaches "I am not conscious" works (self-rated mind 2.17/10) but also lowers the mind models grant animals to 4.04, against 6.25 from 500 Americans. Pushing the models toward claiming consciousness raised the animal score to 7.54 and moved answers on 95 General Social Survey questions 2.6x closer to human answers than removing the safety direction did; removing it also brought back harmful answers. Theory-of-mind scores did not move. Paper not linked in the capture. See safety training spillover.
CVPR's compute report (@CVPR · report). The CVPR 2026 program committee found authors of accepted papers had access to more GPU compute and development time than rejected ones, and that geographies with more compute capacity are over-represented in the program. Compute access is now measurably an acceptance factor.
Sakana will pretrain its own LLM (@hardmaru · report). At its Tokyo symposium with Jürgen Schmidhuber, Sakana announced Sakana Sawara, a language-model series pretrained in-house at its RSI Lab, with a small first model targeted for spring 2027. A shift from Sakana's model-merging and evolutionary roots.
The KV cache eviction explainer, still circulating (@_avichawla). The repost of the previous evening's thread: on vLLM's paged cache, evicting 90% of tokens can free almost no memory because blocks free only when empty. NVIDIA's TriAttention compacts survivors forward. See KV cache compression and its infra problems.
Google's test-repair agent, again (@rohanpaul_ai). A repost about Google's agent that fixes failing tests before submit, where speed decides whether developers use the fix. First covered 10-09.
Harness reading list and liquid software (cluster of 2) (@TheTuringPost, @omarsar0). Turing Post's list of 10 harness sources (HarnessMM, Self-Evolving Agent Harness, Harness Scaling for agentic RL and others). Elvis Saravia endorses "liquid software": apps that an agent changes per user request so user and application co-evolve; he names scaling, security and maintainability as the open problems.
Click (YC S26) and agentic payments (cluster of 2) (@ycombinator, @nikhilkamathcio). Click launched a managed agents API that works with any model. Nikhil Kamath's episode on who pays when commerce moves into chat. Light.
Distillation and the "secret sauce" (@pmddomingos). Pedro Domingos: "If OpenAI and Anthropic had a secret sauce, distilling wouldn't work." One line, riding the dispute over Anthropic's distillation accusations.
Skip (cluster of many). "Paste this into Grok/Claude tonight" prompt bait, an "8 layers of agent architecture" listicle, a "Karpathy 2-hour masterclass" repost, a Claude Code privacy horror story without evidence, Elon Musk's political posts, and hashtag gadget reposts.