AI research wiki
social-stream · 2026-05-26

2026-05-26-morning

Summary

A morning slot dominated by a small but high-signal curated-retweet set from @bayesiansapien and an otherwise noisy AI account feed (the Ferrari Luce electric-car launch ate the day on @brivael and @ns123abc). The strongest substantive signal: Rohan Paul's writeup of ZEDA (Zero-Expert Self-Distillation Adaptation), a post-training method that removes 50 percent of expert compute from already-trained MoE models like Qwen3 and GLM with almost no accuracy loss, by letting tokens that do not need an expert skip them. The second strongest signal is DanielMiessler's preview of an upcoming Claude Code /workflows feature, framed as the operational substrate for turning company processes into agent-executable graphs. Luiza Jarovsky surfaced the AI Ethics Paper Club's 8th pick, the EU agentic-AI compliance architecture paper, which lands in the same governance-around-agents window as today's Foundation Protocol HF paper and Ken Huang's static-authorization piece. Two of the curated reposts are opaque x.com/i/article reposts of DAIR.AI's weekly Top Papers and an AlphaSignalAI article whose body the farmer could not fetch (X session cookies expired). The xAI side ran two enthusiast amplifications of Grok Build's CLI Beta opening to all SuperGrok / X Premium+ users. Beyond that, the AI feed was Ferrari Luce + Spencer Pratt LA mayoral race + Scoble China-electric-car commentary; pure noise relative to the research thread.

Posts

  • ZEDA: skip the experts that aren't needed (@rohanpaul_ai, curated repost). Rohan Paul's writeup of Zero-Expert Self-Distillation Adaptation, a post-training method that converts already-trained static MoE models (Qwen3, GLM) into dynamic ones where tokens that do not need an expert simply skip them. The headline claim is roughly 50 percent of expert compute removed with almost no accuracy loss. Mechanism: a low-cost self-distillation framework that teaches the model to recognize which tokens are "too easy to need them" and route them around the expert pool. The paper reframes MoE inference economics: half the expert FLOPs in production MoE deployments are wasted on tokens that do not benefit from expert specialization. This is the second MoE-efficiency result this week sitting in the same frame as the 2026-05-24 cs.LG #14 Kurate paper on muP MoE scaling (covered in yesterday's Worth Watching). If ZEDA composes with the muP MoE parameterization and BitCPM-CANN 1.58-bit quantization, the production economics of MoE inference flip materially. No direct wiki page yet; pencil this in as a possible future summary if the original arXiv ID surfaces.
  • EU agentic-AI compliance architecture (@LuizaJarovsky, curated repost). Luiza Jarovsky's AI Ethics Paper Club 8th pick: a paper titled "AI Agents under EU Law - A Compliance Architecture for AI Providers." The argument is that agentic AI under EU law needs different compliance scaffolding from classical AI software, because the agent's runtime intent and tool-call trajectory are the unit of regulation, not the user identity or the underlying model. This is the regulatory leg of today's three-way agent-governance pattern, sitting alongside Foundation Protocol (HF coordination layer) and Ken Huang's static-authorization piece (Reva.AI IBAC pitch). See the digest's Connecting the Dots section for the cross-source synthesis.
  • DAIR.AI Top AI Papers of the Week (@dair_ai, curated repost · opaque x.com/i/article). Weekly summary covering AIRA (Meta's two-agent NAS framework that beats Llama 3.2 at 350M/1B/3B in a 24-hour budget), MetaCogAgent, Memory as a Model (MeMo), Code as Agent Harness (100-page survey treating the harness as a research object), Weak-Model Critic-Comparator (GPT-5.4 nano + k=8 + verifier reaches 76.4 percent on SWE-Bench Verified, matching Gemini 3 Pro and Claude Opus 4.5 Thinking), OpenAI Disproves the Unit Distance Conjecture (Erdős open problem closed by an autonomous LLM, human-verified by Alon/Gowers/Wood), Production Agent Architecture Methodology. The x.com/i/article link body was not captured by the farmer (X session cookies expired); the LinkedIn summary version was farmed via Gmail yesterday and contributed to yesterday's digest. Today the items most relevant are MeMo (the memory-as-trained-model paper that pairs with today's MemForest) and OpenAI's Erdős result (the natural-language counterpart to today's AlphaProof Nexus Lean-verified result, both covered in the digest).
  • AlphaSignal weekly (@AlphaSignalAI, curated repost · opaque x.com/i/article). Body could not be fetched (same X cookie expiry as the DAIR.AI repost). Click through to read.
  • Claude Code /workflows feature preview (@DanielMiessler, curated repost). Miessler's argument: Claude Code is about to release a /workflows feature that operationalizes "all work is just an algorithm, i.e., a series of steps to accomplish a goal." Companies-as-graphs-of-algorithms is the framing. This is the practical instance of today's Foundation Protocol thesis: as autonomous agents scale, coordination becomes the bottleneck and the substrate that turns processes into agent-executable graphs becomes load-bearing infrastructure. Watch for the Anthropic release notes.
  • xAI Grok Build CLI Beta opens to all SuperGrok / X Premium+ users (@JasonBud, @milichab, x.ai/cli, cluster of 2). xAI's coding agent is now in Beta for all SuperGrok and X Premium+ users with Plan Mode, Imagine for assets, subagent/worktree integrations, and 1-second X/web search. The pitch is "different tone, different feel" versus Claude Code and Cursor. The interesting feature is the 1-second X/web search inside the agent, which leverages X's data substrate in a way that Anthropic and OpenAI structurally cannot. See the LWiAI #246 roll-up for the Cursor Composer 2.5 vs Grok Build vs Claude Code competitive context.
  • AI Ethics Paper Club (8th paper recommendation). this is the same Luiza Jarovsky retweet covered above; framing it as a separate AI Ethics Paper Club bullet would be duplication.
  • Tandfonline cold-call retweet (@bayesiansapien retweet of @tandfonline, curated repost). Generic Taylor & Francis call for submissions to Sexual and Reproductive Health Matters. Not AI-relevant; appears to be an accidental retweet (the dated 2025-11-15 timestamp suggests the algorithm surfaced it). Skip.
  • shmidtqq opaque x.com/i/article repost (@shmidtqq, curated repost). Bare x.com/i/article URL with no preview body; farmer could not fetch (cookie expiry). Click through to read.
  • Ferrari Luce electric (cluster of ~15) (@brivael, @ns123abc, @Scobleizer, cluster includes Tesla comparisons, jokes, Jony Ive design commentary). Ferrari unveiled its first electric vehicle, the Luce, designed by Jony Ive. The internet, especially the X AI-adjacent subset, mocked it relentlessly comparing it to a Tesla Model S Plaid (Ferrari $650k, 2.4s 0-60, 280 mi range vs Tesla $54,990, 2.9s, 309 mi range). Pure noise relative to AI research; Scoble's tangentially-related Memorial Day take on Chinese consumer EV preferences is the only piece in the cluster that even pretends to be substantive (China's new rich do not want loud Ferraris, they want to fit in; Chinese culture prefers conformity). Skip the substance, log the cluster for the volume.
  • Tesla FSD Supervised promo (cluster of 2) (@Tesla, @Tesla). "From pioneer wagons to cross-country trips on FSD Supervised in under 200 years" and "FSD Supervised shrinks your idea of a long trip." Marketing posts. Skip.
  • Scobleizer's morning thread (@Scobleizer 4 posts). One on Chinese EV culture, one with a Pope/AI driving-around-a-trash-bag joke, one on virtual-being startups (Spark virtual dog) being "too early," one Memorial Day historical post. Useful one-line: Scoble's "another example of when I get excited by things too early" frame on virtual beings, with the explicit observation that AI is not yet good enough to create compelling virtual characters, and that the glasses substrate does not yet exist. The virtual-being thesis is on the bench, not buried. Skip the rest.
  • Spencer Pratt LA mayoral race (cluster of 3) (@spencerpratt, @spencerpratt LA budget, @spencerpratt Memorial Day). Continuing LA politics content from a reality-TV mayoral candidate. Kyle Kuzma reportedly placed a $36k bet on Pratt becoming LA's next mayor. Not AI-relevant. Skip.
  • kilocode Opus 4.7 20 percent off (@kilocode, no article). Promotional post for Kilo Gateway pricing on Opus 4.7. Not research. Skip.
  • brivael Ferrari/France/Reid Hoffman/SpaceX/Mars cluster (@brivael 15 posts). French commentary on Ferrari Luce ugliness, Elon Musk's SpaceX losses, Mars terraforming, OpenAI IPO delay, LinkedIn-as-Reid-Hoffman-Epstein-network polemic, and similar. No AI research substance. Skip.
  • ns123abc 10x-software-engineer prediction (@ns123abc, no article). "There will be a 10x in demand for good software engineers by 2028." Counterpoint to the "AI replaces engineers" thread, sits next to George Hotz's blog from yesterday's social-stream morning slot and the @ns123abc world-models-not-RLVR tweet from yesterday. The cluster of skeptics on agentic coding is broadening, not narrowing. The 2028 timeframe is far enough out that the prediction is not yet falsifiable.