AI research wiki
social-stream · 2026-05-22

2026-05-22

Summary

Thin research-substance day across both slots, with infrastructure positioning and product framing dominating over papers. The strongest signal is a cross-slot infrastructure cluster: morning carried NVIDIA's AI tokenomics primer and Jensen's "demand going parabolic" Dell keynote, paired with @ns123abc reporting Samsung and SK Hynix only signing long-term HBM deals with no-waiver clauses (implying memory shortages into the 2030s). The single-slot standout is Google Research announcing ERA (Empirical Research Assistance) in Nature in the evening, an agentic coding system that slots into the wiki's existing auto-research thread and warrants a summary page once the Nature paper is in hand. Anthropic-side product news was steady (Claude Code /usage breakdown previewed by @bcherny, @ns123abc claiming Anthropic now has more accessible compute than OpenAI) and Cursor doubled Teams-plan invite allowances for a month. The rest was noise: Cosmo and LobeHub framed agentic UI shells with no architectural depth, @magicsilicon posted a one-line CPU-vs-GPU provocation, and the @brivael / @AustinJustice / @WHFraudTF feeds carried 40+ off-topic political and personal reposts to skip.

Posts

  • Google publishes ERA (Empirical Research Assistance) in Nature (@GoogleResearch) [evening]. Agentic coding system pitched as enabling "previously impossible" scientific breakthroughs, presented at I/O. Lines up with the auto-research thread already in autoresearchclaw and the auto-research roadmap; deserves a wiki summary once the Nature paper is in hand.
  • NVIDIA AI tokenomics primer (@nvidia · thread continuation · podcast) [morning]. Four pillars (token utility, demand, supply, monetization) framed as backward-derivation from the customer use case. Marketing-grade but the framework grounds the wiki's ongoing inference-economics-vs-capability discussion.
  • NVIDIA "demand going parabolic" with Dell (@nvidia · article) [morning]. Numbers from Jensen at Dell Technologies World: $3-4T AI infra spend by 2030, 3,400% token-consumption growth in the same window, Vera Rubin NVL72 pitched at one-tenth cost per token versus prior gen, agent sandboxes 50% faster on Vera. Marketing vocabulary ("agent sandbox", "token economics") is now standard inside NVIDIA pitches.
  • Memory long deals signal HBM shortage to 2030s (@ns123abc · SK Hynix bonus) [morning]. Samsung and SK Hynix reportedly only accepting long-term HBM contracts with no-waiver clauses; SK Hynix workers projected 900K USD bonus. Supply-side flip of the parabolic-demand framing in the same slot: if HBM lead times push out to the 2030s, KV-cache compression and weight-sharing MoE schemes (anything that cuts HBM bandwidth pressure) gain premium.
  • Claude Code /usage breakdown coming (@bcherny) [morning]. Boris Cherny previewed /usage itemising tokens per Skill, Agent, MCP, and Plugin (CLI first, Desktop next). Pragmatic as the Claude-Code plugin economy gets denser; spirit equivalent of top for token cost.
  • Anthropic claimed to have more revenue and more accessible compute than OpenAI (@ns123abc) [morning]. Aligns with Decoder reporting in today's RSS that Anthropic is approaching a $559M operating profit on $10.9B Q2 revenue, though Gary Marcus notes the profit projection leans on a one-time SpaceX compute discount disclosed in the SpaceX S-1. Plausible given the $15B/yr Anthropic-SpaceX compute deal, not yet independently verified.
  • NVIDIA GTC Taipei keynote June 1 (@nvidia · event page) [morning]. Jensen's COMPUTEX 2026 keynote, June 1 at 11 a.m. Taipei time. Watch for any new Rubin variant disclosure or networking SKU.
  • Cursor doubles usage for Teams-plan invites (@mntruell) [morning]. Doubled Cursor usage for the next month on new Teams-plan invites. Plain growth lever, lands in the same week Google's Antigravity 2.0 IDE redesign is taking heavy early flak per Pragmatic Engineer.
  • Cosmo agentic UI layer (@Scobleizer) [morning]. Runtime-generated visual UI on top of AI agents; tagline "AI is eating the operating system". Same lineage as Antigravity and Anthropic's plugin economy; the unaddressed research question is the latency budget for runtime UI generation when it sits in the model's critical path.
  • LobeHub framed as "AI workspace" / Chief Agent Operator (@Scobleizer · lobehub.com) [evening]. Long-horizon agent orchestrator claiming 312K skills and 58K MCP servers. Signal is the framing shift from chatbot-UI to operator-of-agents, not the product itself. Promo-leaning.
  • "The killer product is still the CPU" (@magicsilicon) [evening]. One-line pushback on GPU-only narratives. Provocative but unsupported.
  • 120 startups launching at Founders Inc San Francisco (@Scobleizer) [morning]. 120 teams launching at Founders Inc the evening of May 22 PT. "Robots moving. Agents earning. Hardware working. Software shipping." Ecosystem-density signal only, no specific startups named.
  • AI-adjacent but content-thin tweets (cluster of 3) [morning]. @Scobleizer "agents run 24/7", @Tesla Model S X Signature Delivery, @ns123abc SK Hynix 900K bonus. No body content to gloss. Skip.
  • Off-topic political and personal feed (cluster of 44) [morning + evening]. @AustinJustice, @WHFraudTF, @brivael, @ns123abc SpaceX Starship V3 / Bastrop solar / Nicki Minaj, plus evening Elon-meme and Macron-quantique reposts. Skip.