AI research wiki
social-stream · 2026-05-20

2026-05-20

Summary

The day's strongest cross-slot thread is Anthropic talent flow: Karpathy publicly joining the pre-training team (morning, confirmed across Algorithmic Bridge, TechCrunch, Axios, and his own tweet) plus Soizig moving from Mistral interpretability to Anthropic SF (afternoon, via @brivael), a second Paris-to-SF interp hire in a quarter. The single-slot standout is SpectralQuant in the afternoon curated set: 5.95x KV cache compression on Mistral 7B at +7.5% perplexity overhead against TurboQuant's +22% at the same ratio, a direct shot at a baseline the wiki already has on file. Morning is the substance-heavy slot, carrying Google I/O's coordinated Co-Scientist / Gemini for Science / ERA push with three same-day Nature publications, Cursor Composer 2.5 at 63.2% on CursorBench 3.1 for $0.55 per task against Opus 4.7 Max's $11.02, and a single-source SpaceXAI S-1 rumour at $2T target valuation. Afternoon curated content is unusually research-dense (AIRA, Code-as-Agent-Harness 100-page survey, On-Policy Mix, a mech-interp claim that LLMs carry multiple non-overlapping circuits per task). Evening is thin filler with one aphorism on bias-variance versus compute-efficiency and Tesla's Hey Grok consumer rollout, otherwise skip. AI-account noise (Scobleizer business posts, @brivael French political commentary) clusters in afternoon and evening.

Posts

  • Google I/O 2026 Co-Scientist + Gemini for Science + ERA, three same-day Nature publications (@GoogleResearch via blog · Gemini for Science) [morning]. Multi-agent Co-Scientist (hypothesis generation, debate, evolution) plus Empirical Research Assistance for expert-level scientific coding. Most coordinated science-AI push by a frontier lab the wiki has tracked. Pairs naturally with today's AutoResearchClaw, which beats AI Scientist v2 by 54.7% on ARC-Bench via a Pivot/Refine loop.
  • Karpathy publicly joins Anthropic's pre-training team (Algorithmic Bridge essay, TechCrunch, Axios via The Decoder, Karpathy tweet) [morning]. Works under Nick Joseph on using Claude to accelerate pre-training research. Algorithmic Bridge frames it explicitly as Anthropic hiring Karpathy to prepare Claude to improve itself, along the recursive-self-improvement axis Jack Clark called out 4 May. Substantive backdrop is Karpathy's autoresearch repo finding ~20 validation-loss-improving nanochat changes in ~2 days of unsupervised running.
  • Cursor ships Composer 2.5 plus a Jira-native agent surface (@cursor_ai changelog) [morning]. CursorBench 3.1: Composer 2.5 at 63.2% / $0.55 vs Opus 4.7 Max at 64.8% / $11.02 and GPT-5.5 Extra-High at 64.3% / $4.37. 20x cost-per-task delta versus Opus at parity within two points. Jira integration lets users assign issues to Cursor or @-mention it in a comment to spin up a cloud agent.
  • SpaceXAI S-1 filing rumour at $2T target valuation, $75B all-primary raise (@ns123abc) [morning]. Goldman leading, 30% retail allocation, $SPCX trading 12 June, Nasdaq-100 inclusion within 15 days, Musk retaining ~42% equity and ~79% voting control. Single-source tweet pending the actual S-1; directional signal is consistent with Composer 2.5's SpaceXAI follow-on training at 10x Colossus 2 compute.
  • Anthropic publishes "Best practices for computer and browser use with Claude" (claude.com blog) [morning]. Click accuracy, thinking effort levels, long-session context management, recording demonstrations Claude can replay. The replay primitive converts one-shot agent behaviour into reusable trajectories, the same mechanism AutoResearchClaw uses for cross-run safeguard evolution.
  • NVIDIA at Google I/O ships JAX learning path and Dynamo on GKE codelab (NVIDIA blog) [morning]. Dynamo is the inference orchestration layer for disaggregated prefill and decode; pairing it with GKE pushes the orchestrator into the production Kubernetes surface most enterprises actually run.
  • Logan Graham frames third-wave philanthropy math: roughly 0.5 Manhattan Projects per year liquid within ~2 years (@logangraham) [morning]. Anchored to Nan Ransohoff's catalogue of the OpenAI Foundation's 26% stake (~$220B at today's valuation) plus Anthropic co-founders' 80% giving pledge. Aligns with the Karpathy timing: philanthropy capital arrives roughly when frontier compute consolidates and AI-for-safety is positioned to absorb it.
  • SpectralQuant: 5.95x KV cache compression at 3x less degradation than TurboQuant (@ashwingop) [afternoon]. Mistral 7B at +7.5% perplexity vs TurboQuant's +22% at the same ratio. 15-second per-model calibration, drop-in for any HuggingFace LLM, ViT, ESM, AlphaFold Evoformer, or VideoMAE. Direct shot at the TurboQuant baseline. Tier 1 if numbers replicate. See also KV cache concept page.
  • AIRA: Meta's dual-agent system that autonomously designs sub-Llama-3.2 architectures in 24h (@omarsar0 · arxiv 2605.15871) [afternoon]. AIRA-Compose searches macro architecture (11 agents), AIRA-Design implements low-level attention (20 agents). AIRAformers / AIRAhybrids beat Llama 3.2 by 2.4 to 3.8% accuracy at 1B and scale 54 to 71% faster. Cross-source confirmed with the wiki's existing AIRA summary.
  • Code as Agent Harness, 100-page survey (@omarsar0 · arxiv 2605.18747) [afternoon]. Frames agent infrastructure around harness interface, harness mechanisms (planning, memory, tool use), and single-to-multi-agent scaling. Argues future systems need executable, inspectable, stateful, governed code. Tier 2 summary candidate.
  • On-Policy Mix: data mixing algorithm spanning pretraining, midtraining, and instruction tuning (@michahu8) [afternoon]. Pitched as solving the right-data-mix problem under distribution drift. Only thread opener captured; click for method.
  • Mech-interp claim: LLMs carry multiple non-overlapping circuits per task (@che_shr_cat) [afternoon]. If it holds, it undermines the canonical-circuit framing that a lot of interp work assumes. Relevant to recent probe-trajectory monitoring.
  • Soizig moves from Mistral interpretability to Anthropic SF (@brivael RT) [afternoon]. Second Paris-to-SF interp hire in a quarter, sitting alongside the Karpathy move at the top of the day's talent-flow signal.
  • Two opaque x.com/i/article reposts by @bayesiansapien, farmer can't fetch without cookies (@DimitrisPapail, @AmarSVS) [afternoon]. Click through to read.
  • Bias-variance vs compute efficiency aphorism (@MillionInt) [evening]. High-variance low-bias methods scale and waste compute; low-variance high-bias methods are efficient and hit walls fast. MillionInt notes one side has historically won and leaves which side as an exercise. The bitter lesson restated.
  • Hey Grok wake-word voice activation lands in Tesla vehicles (cluster of 3, @Tesla, @Tesla, @Tesla) [evening]. Enabled via Grok Settings, "Goodbye" or "dismiss" ends the session. First wide consumer deployment of an LLM tied to a wake word inside a moving vehicle.
  • Scobleizer cluster: Aligned News finances, Soma tribute, paid-post disclosure, robotics demo reposts (cluster of 5, @Scobleizer plus four others) [afternoon]. $15K month-one spend on alignednews.com/ai, $4K of that on X API fees. Skip unless tracking AI-media business models.
  • @brivael French political and personal cluster, plus Hassabis admiration RTs and a grandes écoles polemic (cluster of ~12 across afternoon and evening) [afternoon + evening]. Skip.
  • DAIR.AI Academy and taap.it / argil.ai affiliate promos [afternoon + evening]. Skip.