AI research wiki
social-stream · 2026-06-18

2026-06-18-morning

Summary

The morning slot is carried by two threads. The first is talent: Noam Shazeer, the "Attention Is All You Need" co-author Google paid a reported $2.7B to bring back and lead Gemini, announced he is leaving for OpenAI, and the timeline lit up with a cluster of seven reactions (@eliebakouch, @ns123abc, @MillionInt), Sam Altman's "divine benevolence" quip included. The second is the open-weight flood: GLM-5.2 (MIT weights, 1M context, pitched as Opus-4.6 / GPT-5.4 tier) and DeepSeek V4 Vision both landed within hours. The sharpest policy story is the Anthropic export-control fallout, a four-post cluster plus WIRED reporting that the White House ordered SK Telecom's Mythos access revoked over alleged China ties, with JPMorgan separately cutting its Hong Kong staff off from Claude. On industry infrastructure, AWS Summit NY dumped a full agentic stack in an eight-post @mattsgarman thread and Cursor shipped cloud subagents, while Midjourney's pivot into an ultrasound body-scanner drew outsized hype. The substantive research item is Google's AMIE longitudinal-care paper in Nature. The rest is heavy noise: a large block of French and US political content from @brivael and others, Kilo Code star-count promo, and two Intel-logo meme images. Signal-to-noise is poor this slot, but the talent, open-weights, and export-control threads are all real.

Posts

  • Noam Shazeer leaves Google for OpenAI (cluster of 7: @eliebakouch · @ns123abc · @MillionInt). The Transformer co-inventor, brought back to Google to lead Gemini for a reported $2.7B, is moving to OpenAI after under two years. Shazeer's own note calls it "a difficult decision"; Sam Altman replied with a "we attribute his arrival, as all else, to divine benevolence" quip, and @eliebakouch riffed by begging Altman for GPT-6 access. The framing across the cluster is the talent war as Anthropic-Karpathy versus OpenAI-Noam. Also surfaced in RSS via The Decoder.
  • GLM-5.2 open weights (@ns123abc · HF weights). Z.ai's new MIT-licensed model, with a usable 1M-token context, two reasoning-effort levels (max for peak, high for token efficiency), and GLM-5.1 pricing. @ns123abc rates it "easily Opus 4.6 / GPT 5.4 tier" on coding and agentic tasks. The docs pitch 1M context as stable for project-scale work. See GLM-5.2 summary.
  • DeepSeek V4 Vision (@ns123abc). DeepSeek added vision to V4, now live on web and app, the second open-weight frontier multimodal drop in the same window as GLM-5.2. See DeepSeek V4 architecture.
  • Anthropic export-control fallout (cluster of 4: @ns123abc on SK Telecom · WIRED · @ns123abc on JPMorgan). WIRED reports the chain: days before Anthropic took its most advanced models offline, the White House ordered it to revoke SK Telecom's access to Claude Mythos over alleged China ties (SK Telecom: "no ties to China"); Anthropic complied the same day, then Amazon flagged Fable 5 guardrail vulnerabilities, producing the export-control letter that revoked all foreign nationals' access. Separately, JPMorgan cut its Hong Kong staff off from Claude, following Goldman Sachs last month. The mechanistic backdrop is today's SAE-interventions paper (a safeguard blocks the visible route while the capability stays reachable). See Anthropic Glasswing/Mythos vulnerabilities.
  • G7 AI summit positions (@ns123abc). At Évian-les-Bains, Altman reportedly pitched a multilateral standards forum that includes China ("we develop the technology, citizens of the free world make the rules"), while Dario Amodei pitched a US-led bloc with structured frontier-model access for allies. The two lab CEOs split on whether China is inside or outside the governance tent, directly relevant to the export-control thread above.
  • AWS Summit NY agentic stack (cluster of 8: @mattsgarman · recap). AWS shipped AgentCore (Harness, Policies, Web Search, Managed Knowledge Base), AWS Context (a knowledge-graph service that maps organizational data relationships for agents at runtime), AWS Continuum (autonomous vulnerability validation and remediation, explicitly motivated by frontier cyber models like Claude Mythos finding vulns at machine speed), Amazon Quick (natural-language agent builder), and Kiro for iOS. The throughline is governed prototype-to-production agent deployment at enterprise scale.
  • Cursor cloud subagents (cluster of 4: @cursor_ai · changelog). /in-cloud spins a subagent in its own cloud VM and branch for long-running or parallel work (fixing CI, exploring a codebase) while the local workspace stays clean; environments save as reusable snapshots and agents are drivable from a phone. Same parallel-cloud-agents pattern as AWS Kiro.
  • Google AMIE longitudinal care (@GoogleResearch · blog). New Nature paper with DeepMind: AMIE uses Gemini's long context to ground clinical reasoning over multi-visit patient management, scoring physician-level on longitudinal management and significantly higher on plan preciseness and guideline alignment. The substantive research item of the slot, and a concrete long-context application.
  • Midjourney Medical (cluster of 4: @Scobleizer · @_sholtodouglas · @ns123abc). David Holz launched a new division with an ultrasound "Midjourney Scanner" body-scan device, framed as preventive cancer screening. Reaction was effusive (Anthropic's Sholto Douglas predicted it saves 100x Midjourney's profit to date), but all of it is hype, not technical detail.
  • Cursor 1.5T-param model claim (@eliebakouch). Eliebakouch pushed back on the claim (attributed to Cursor's CEO) that their new 1.5T-param model, pre-trained on 100k+ GPUs, is "as big as Opus," noting it would imply crazy closed-model margins versus open-source serving, and that the comparison ignores scale economies in batched inference. Posted alongside blended input/output price plots. Cost-realism signal feeding the routing economics thread.
  • Kilo cost levers and cheap-build demos (cluster of 6: @kilocode spend levers · SpaceX-bought-Cursor blog). Mostly star-count giveaway promo, but with real signal folded in: the same coding task costs $15 on Opus versus $1.70 on a cheaper model ("most of what an agent does does not need a frontier model"), a MiniMax M3 vision build for $0.53, a Step 3.7 Flash free demo, and an argument that vertically integrated stacks (the $60B SpaceX-Cursor deal) are the wrong bet on model choice. The production case under today's SciOrch routing paper.
  • Kilo + Inceptron EU sovereign inference (@kilocode · blog). Kilo Gateway now routes to Sweden-hosted Inceptron with lockable EU-only data residency, offering Kimi K2.6, GLM 5.1, and MiniMax M2.5 ($0.15/1M input) under GDPR/ISO compliance. Routing-by-jurisdiction as a product, the commercial mirror of the export-control thread.
  • NVIDIA VivaTech / France AI (@nvidia · blog). Mistral's 44MW Bruyères-le-Châtel data center is operational with 18,000 GB200 Blackwell systems, en route to 200MW across Europe by 2027. France's AI factories moving from announcement to production. Hardware-side context for the open-weight push.
  • Two Intel meme images (@magicsilicon). Image-only posts showing classic "Intel Inside" and "intel delivers" logos, no text or technical content. Skip.
  • Off-topic political noise (cluster of ~18: @brivael French/UK politics · @AustinJustice · @DoWCTO · @WHFraudTF · @spencerpratt · @SeanParnellASW). Almost entirely non-AI politics, crime, and military content that slipped the keyword filter. Skip.