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

2026-05-26

Summary

A day with no single paper reposted across slots, but two coherent thematic threads. The strongest is responsible-AI / governance: morning surfaced the EU agentic-AI compliance paper via Luiza Jarovsky's Paper Club pick, evening delivered Chris Olah's Vatican remarks at the presentation of Pope Leo XIV's AI encyclical, where a frontier-lab co-founder named the structural-incentive problem from inside and pointed at interpretability findings that mirror human neuroscience, introspection, and functional emotional states. The second thread is Grok Build's product push: morning had the CLI Beta opening to all SuperGrok and X Premium+ users, evening had Kilo Code's benchmark of Grok Build 0.1 one-shotting a polished 3D OS demo in under two minutes for seven cents. The day's single strongest standalone signal is evening's Sholto Douglas tweet confirming Mythos has independently solved the Erdős unit distance problem with a "cute, simple proof" days after OpenAI 5.5 cracked it; two top-lab models closing the same open problem in a week is real model-as-mathematician evidence, not hype. Morning also carried Rohan Paul's ZEDA writeup (50 percent of MoE expert FLOPs removed with near-zero accuracy loss by letting easy tokens skip experts), the cleanest efficiency result of the slot. Outside those threads the day was noise: Ferrari Luce mockery dominated both slots, plus the usual Spencer Pratt LA mayoral content and brivael French political rants.

Posts

  • Mythos independently solves Erdős unit distance problem (@_sholtodouglas, referencing @alpoge, OpenAI proof, DeepMind AlphaProof Nexus) [evening]. Sholto Douglas (Anthropic) credits OpenAI's 5.5 for cracking it first, confirms Mythos has reproduced the result with a "cute, simple proof," and calls it his go-to example of models pulling disparate ideas into a new discovery with "serious overhang." Two top-lab models closing the same Erdős open problem in a week is the strongest single piece of model-as-mathematician evidence to date.
  • Chris Olah at the Vatican on Magnifica humanitas (@bcherny · Anthropic post, responsible-ai concept page) [evening]. Olah's full remarks at the presentation of Pope Leo XIV's May 25 AI encyclical. Two passages worth reading: a frontier-lab co-founder naming the structural-incentive problem from inside, and interpretability work finding things "mysterious, even unsettling," including structures mirroring human neuroscience, evidence of introspection, and internal states that functionally mirror joy, satisfaction, fear, grief, and unease. Substantive responsible-AI signal of the day.
  • ZEDA: skip the experts that aren't needed (@rohanpaul_ai) [morning]. Rohan Paul's writeup of Zero-Expert Self-Distillation Adaptation, a post-training method that converts already-trained MoE models (Qwen3, GLM) into dynamic ones where tokens that do not need an expert skip them. Roughly 50 percent of expert compute removed with almost no accuracy loss. Frames half of production MoE expert FLOPs as wasted on tokens that do not benefit from specialization. Pairs with the 2026-05-24 cs.LG #14 Kurate paper on muP MoE scaling.
  • EU agentic-AI compliance architecture (@LuizaJarovsky) [morning]. AI Ethics Paper Club 8th pick: "AI Agents under EU Law - A Compliance Architecture for AI Providers." Argument: agentic AI under EU law needs different compliance scaffolding because the agent's runtime intent and tool-call trajectory are the regulatory unit, not the user identity or underlying model. Regulatory leg of today's three-way agent-governance pattern alongside Foundation Protocol and Ken Huang's static-authorization piece. See the daily digest's Connecting the Dots.
  • Claude Code /workflows feature preview (@DanielMiessler) [morning]. Miessler's argument: Claude Code is about to release a /workflows feature that operationalizes "all work is just an algorithm, a series of steps to accomplish a goal." Companies-as-graphs-of-algorithms framing. 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.
  • Grok Build product push (@JasonBud, @milichab, x.ai/cli, @kilocode cluster of 7, kilo.ai/install) [morning + evening]. Morning: Grok Build CLI Beta opens to all SuperGrok and X Premium+ users with Plan Mode, Imagine for assets, subagent/worktree integrations, and 1-second X/web search inside the agent. Evening: Kilo Code benchmarks Grok Build 0.1 one-shotting a full-screen 3D OS demo (Three.js, Tailwind, parallax, reflections, responsive multi-window) in under two minutes for seven cents, using less than a fifth of context. Grok Build now sits in the same product slot as Cursor Composer 2.5 and Claude Code; see the LWiAI #246 roll-up for competitive context.
  • RAM is a three-supplier oligopoly and compute is king (@ns123abc, follow-up, SemiAnalysis 800VDC piece) [evening]. Samsung, Hynix, and Micron own over 90 percent of global RAM production; CXMT is under one percent even at full ramp. Anthropic is reportedly paying SpaceX a billion a month for compute. Companion tweet claims xAI is training 1T and 10T models on GB300s (Grok 4 was 500B). RAM-oligopoly claim is straight market structure; xAI training numbers are speculative.
  • DEEP Robotics IPO reframes humanoid economics (@Scobleizer · Hello China Tech) [evening]. Prospectus shows 79.33 percent of 2025 revenue from industrial applications (power grids, fire departments, security, logistics) against Unitree's nine percent industrial. Two Chinese embodied-AI IPOs, same exchange, completely different underlying businesses. Pattern: humanoid revenue today is research hardware sold to labs, not deployed labor. Worth tracking as the empirical anchor on humanoid economics.
  • AI regulation becomes a moral issue (@Scobleizer · Unaligned) [evening]. Scoble and Cronin's free Unaligned newsletter argues AI regulation has moved past "is it powerful or profitable" to "does it protect people and respect human dignity," with surveillance, misinformation, replacement of human judgment, and power concentration as load-bearing concerns. Op-ed positioning, not new research; barometer for where mainstream-tech commentary is settling the same week as the Pope's encyclical.
  • ns123abc software-engineer demand prediction (@ns123abc) [morning]. "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 morning slot and yesterday's @ns123abc world-models-not-RLVR tweet. The 2028 timeframe is far enough out that the prediction is not yet falsifiable.
  • Scoble virtual-being thesis on the bench (@Scobleizer) [morning]. One useful line in an otherwise skippable thread: Scoble's "another example of when I get excited by things too early" frame on Spark virtual dog and virtual-being startups generally, with the explicit observation that AI is not yet good enough to create compelling virtual characters and the glasses substrate does not exist. Thesis is on the bench, not buried.
  • Opaque x.com/i/article reposts (@dair_ai DAIR.AI Top Papers of the Week, @AlphaSignalAI weekly, @shmidtqq) [morning]. Three curated reposts whose article bodies the farmer could not fetch (X session cookies expired). DAIR.AI's items were independently captured via Gmail yesterday and contributed to yesterday's digest; the other two need a manual click-through.
  • Ferrari Luce electric car mockery cluster (cluster of ~30 across @brivael, @ns123abc, @Scobleizer) [morning + evening]. Ferrari unveiled its first EV (Luce, designed by Jony Ive); the X AI-adjacent subset mocked it relentlessly against the Tesla Model S Plaid ($650k vs $54,990, 280 mi vs 309 mi range). Scoble's tangential Memorial Day take on Chinese consumer EV preferences is the only piece pretending to be substantive. Skip.
  • Spencer Pratt LA mayoral race (@spencerpratt, cluster of 5 across morning + evening) [morning + evening]. Continuing LA politics content from a reality-TV mayoral candidate. Skip.
  • Tesla FSD Supervised promo (@Tesla, cluster of 2) [morning]. Marketing posts. Skip.
  • Kilo Code Opus 4.7 pricing promo (@kilocode) [morning]. Kilo Gateway 20 percent off promo. Skip.
  • Tandfonline accidental retweet (@bayesiansapien retweet of @tandfonline) [morning]. Taylor & Francis call for submissions to Sexual and Reproductive Health Matters. Not AI-relevant, dated 2025-11-15 timestamp suggests the algorithm surfaced it accidentally. Skip.
  • AustinJustice Austin crime posts (@AustinJustice, cluster of 2) [evening]. Local Austin DA crime reporting. Not AI-relevant. Skip.
  • brivael French political and shitposting cluster (@brivael, cluster of ~25 across both slots) [morning + evening]. French startup-fiscality rants, LinkedIn-is-dead takes, Reid Hoffman / Epstein polemic, Mars terraforming, OpenAI IPO delay, Matrix-quote shitposting. No AI research substance. Skip.