social-stream · 2026-11-15

2026-11-15-afternoon

Summary

No curated reposts this slot, so it is the Following feed on a US Saturday night, and it is about 85% noise (Starship Flight 15 catch reposts, hashtag reposts, paid "Explee" growth ads). The strongest signal is an agent-harness cluster of 3: Shanghai AI Lab's Self-Harness (the agent edits its own scaffold, weights frozen, and doubles Terminal-Bench scores), Anthropic's "Building always-on agents" architecture piece, and a recap of Claude Code's advisor mode, which is routing by moment. The second cluster is the Jensen Huang exit aftermath (5 posts): Musk vouching for his compute promises, an investor letter to Nvidia's board, and a conspiracy thread tying his exit to Nemotron 3 Ultra. Standouts outside those: Sevren's Embla 1, a 10B model running offline on an iPhone, and a resurfaced paper on emergent LLM value systems. Skip the "Claude Council" and "we achieved AGI" bait.

Posts

  • Self-Harness: harnesses that improve themselves (@HoodieFade_). The agent groups its own failures by root cause, proposes harness edits, and keeps an edit only if it helps on one task set and hurts nothing on a second. Weights never change. Qwen3.5 goes from 18% to 36.7% on Terminal-Bench 2.0, and GLM-5 from 44.4% to 85% on AppWorld. The keep-only-if-no-regression gate is the real idea. Harness engineering
  • Anthropic: Building always-on agents (@Xudong07452910 · article). The agent sleeps until a message, webhook or timer wakes it, runs one turn, saves state, and sleeps again. It has 5 parts: Inbox (dedup and retry), Timer, Memory, Permissions, and Outbox. The best idea is the deferred tool call: when a step needs human approval, it saves the session and frees the worker instead of blocking. The poster reads it next to DeepSeek's Router-plus-Workers Super-agent. DeepSeek Super-agent
  • Claude Code advisor mode (@xsmryze · docs). With --model sonnet --advisor opus, Sonnet runs the session, Haiku subagents explore, and Opus is consulted only before a plan locks, on a repeated error, or before calling the work done. This is cost routing by moment, not by query. The "JEV" plug in the post is promo. Advisor routing
  • "20x cheaper harness" (@IntuitMachine). A one-line claim with no link. Skip.
  • Jensen Huang exit aftermath (cluster of 5) (@elonmusk, @garrytan RT GavinSBaker, @garrytan RT dee_bosa, @VaibhavSisinty, @adamemedia). Musk says Jensen always delivered compute before the purchase order existed. Atreides sent a letter to Nvidia's board, and Bosa argues his exit tightens China export policy. Vaibhav recalls Jensen defending open models weeks before. The adamemedia "they took him down over Nemotron 3 Ultra" thread is conspiracy. Nvidia without Jensen
  • Embla 1: a 10B model offline on iPhone (@ahsuzm). Danish startup Sevren says its 10B model runs on an iPhone 18 Pro in airplane mode, answering in 1.77s. It averages 57.4 on five benchmarks, all self-reported, and the weights are on Hugging Face. Worth checking which quantization makes 10B fit in phone memory.
  • Emergent utility functions in LLMs (@thesupermannx). This resurfaces the Utility Engineering result. As models scale, their pairwise preferences stop cycling, so they act like one coherent utility function, and corrigibility drops. The country-ranking example is real but framed for outrage.
  • Fable 5.2 showcase list (@deedydas). A grab bag: an S-1 explainer video, a 40k qps Rust scraper, a cricket physics game. Two of them failed. A vibe check, not an eval.
  • YC and startup reposts (@garrytan, @garrytan, @snowmaker). Bland's Tofu-1 speech-to-speech model, AgentRun passing 1,000 GitHub stars, and YC on the Grok Bot core team. Launch promo, light signal.
  • Flat Minima (@SchmidhuberAI · paper). Schmidhuber points a reply at his 1997 paper again. A priority-claim habit, not news.
  • Robotics live-challenge proposal (@berkeley_ai RT JitendraMalikCV). Malik wants live, judged robot challenges instead of curated demo tapes. A sound idea but outside your focus.
  • Ed Zitron and Gary Marcus snark (@edzitron, @GaryMarcus). Generative-AI skepticism replies with no new claim. Skip.
  • Starship Flight 15 tower catch (cluster of ~12) (@elonmusk). First tower catch of a Ship. A big moment for space, nothing for AI. Skip.
  • Hashtag tech reposts (@Ronald_vanLoon). About 12 posts on 3D printing, fusion, AR and food tech. Skip.
  • Paid ads and bait (Explee growth-tool ads ×3, Academic Humanizer, "Claude Council" setup thread, "we have achieved AGI", trading and apparel ads). Skip.