Summary
No @bayesiansapien retweets in the morning slot, so all signal comes from the AI account feed. The dominant story is Cursor Composer 2.5, a six-post launch thread from @cursor_ai detailing a Kimi K2.5 base, RL with textual feedback for credit assignment across hundred-thousand-token rollouts, a Sharded Muon optimizer plus dual mesh HSDP training stack, and a follow-on training run with SpaceXAI at 10x more compute on Colossus 2's million H100-equivalent cluster. Second signal is NVIDIA Vera, where Ian Buck's hand-delivery to Anthropic, OpenAI, SpaceXAI, and Oracle yesterday gets the official NVIDIA blog post framing Vera as the agentic-AI CPU. Third is Anthropic shipping Claude Code Fast Mode by default on Opus 4.7 and adding cache diagnostics in the console, plus a rename of "extra usage" to "usage credits." Tesla pushes three FSD v14.3.3 testimonial reposts. Scoble runs three eclectic medical-AI demos. @brivael returns with another long French-language run on capitalism, philosophy, and AI as merit catalyst, with one technically interesting repost of Cedric Lion's "Open Collider" creativity-engine project for LLMs.
Posts
- Cursor Composer 2.5: Kimi K2.5 base, RL with textual feedback, Sharded Muon, dual-mesh HSDP, 10x compute follow-up with SpaceXAI (cluster of 6, @cursor_ai · blog). Cursor ships Composer 2.5 as a substantial step up over Composer 2 on long-running task quality and complex instruction following. The training stack is the interesting part. The base is Moonshot's Kimi K2.5 open-source checkpoint. Training adds three things: targeted RL with textual feedback (the model receives natural-language critique during rollout, used to assign credit to specific decisions inside spans of hundreds of thousands of tokens, which the post argues is the only viable credit-assignment scheme at agentic rollout length); synthetic data generation for more complex RL environments; and a Sharded Muon optimizer paired with dual-mesh Hybrid Sharded Data Parallel (HSDP), which the post says was necessary to scale Muon to their training-data parallelism layout. Cursor claims Composer 2.5 is "up to 10x more efficient than similarly capable models," matched against unnamed peers. Separately, Michael Truell (@mntruell) confirms the SpaceXAI collaboration is a from-scratch much larger model with 10x total compute on Colossus 2's million H100-equivalents, framed as "the very start of our work with SpaceXAI." This is the second consecutive week the wiki has tracked Cursor pushing models built on a Chinese open-weight base (Composer 2 on Kimi K2, now Composer 2.5 on Kimi K2.5). The base-model dependency on Chinese open weights for a frontier US coding agent is now a stable, not transitional, pattern.
- NVIDIA Vera CPU hand-delivery: Anthropic, OpenAI, SpaceXAI, Oracle Cloud get first units (cluster of 3, @nvidia · NVIDIA blog). NVIDIA VP Ian Buck hand-delivered the first Vera CPUs Friday May 15 to Anthropic San Francisco, OpenAI Mission Bay, SpaceXAI Palo Alto, then Oracle Cloud Santa Clara on Monday. Buck quote: "Agentic AI is creating a new CPU moment in the AI factory. As models move from answering to acting, Vera is purpose-built to keep that work moving at scale." Vera was announced at GTC San Jose in March as NVIDIA's first custom CPU, positioned as a multi-billion-dollar standalone business. The strategic claim is that agentic AI demands a different CPU architecture than the x86 / ARM hosts that pair with current Hopper / Blackwell datacenter GPUs because the host-side dispatch, context-loading, and tool-invocation patterns are different in agentic vs request-response workloads. No public Vera specs in this thread; the blog page promises a roadmap to Vera-powered systems. Matched against yesterday afternoon's social-stream Vera item (Ian Buck's same-day photo tour), today's blog post is the official narrative.
- Anthropic ships Fast Mode default on Opus 4.7 in Claude Code, plus prompt-cache diagnostics, plus usage-credit rename (cluster of 6, @ClaudeDevs · fast-mode docs). Three product changes in one push. Fast Mode is a high-speed Opus configuration: identical model quality at roughly 2.5x response speed at a higher per-token rate, toggleable via
/fast. Anthropic frames it as latency-versus-cost tradeoff for rapid iteration, live debugging, time-sensitive work. The docs note it stays in research preview, with rate-limiting handled at the org level and per-session opt-in optional. Second change: prompt cache diagnostics now visible in the Claude Console, showing exactly which part of a prompt caused a cache miss and the token cost (cache diagnostics docs). This is one of the most-requested missing features for high-cache-rate workloads where small invisible prefix drift can silently halve effective margins. Third: "extra usage" is renamed to "usage credits" across Claude products, with the framing that credits now power features beyond plan overflow (specifically fast mode and other premium routes). Existing spending limits, auto-reload, and pre-purchased credits carry over unchanged. The three together describe an Anthropic moving toward a metered-throughput pricing model where speed and cache-hit quality are first-class billable axes. - How Claude Code works in large codebases (@ClaudeDevs · blog). Anthropic publishes a best-practices post drawn from teams running Claude Code across multi-million-line monorepos, decades-old legacy systems, and distributed microservices. The fetched article body is navigation chrome only, so the substantive recommendations require clicking through; but the framing suggests Anthropic is starting to publish prescriptive playbooks for enterprise rollouts rather than just docs.
- Tesla FSD v14.3.3 testimonials (cluster of 3, @Tesla). Three Tesla reposts of community videos showing v14.3.3 catching potential side-swipe and merge collisions, plus a Herbert Ong summary claiming the version brings much less driver-monitor nagging (users reportedly going over a minute with no nag), smoother and more human-like driving, faster Smart Summon to 8 mph, "Hey Grok" voice support, and better visualizations. Vision-based collision avoidance is real safety signal; the nagging reduction is the deployment-comfort lever Tesla has been pushing for two quarters. Tier 3 for this wiki (robotics adjacent, not core research).
- Open Collider: an LLM creativity engine that mechanically improves idea diversity (@brivael reposting @cdriclion). Brivael flags Cedric Lion's "Open Collider" open-source project. Lion's framing: LLMs collapse on the same ideas when sampled many times from the same brief, a phenomenon Jiang et al. (2025) call "Artificial Hivemind." Telling the model "be more creative" moves the output set sideways without expanding it. Open Collider is positioned as a sampling-and-recombination engine that mechanically forces non-trivial idea diversity. The tweet doesn't expose method details, so the wiki cannot verify the claim, but the diagnosis is consistent with the wiki's prior coverage of LLM-creativity-bottleneck papers (the 2026-05-12 social-stream item on neuron pruning for novelty and the recurring open question of whether RLVR contracts the policy distribution).
- Robert Scoble field reports: AR brain surgery, vein-and-skeleton biometrics, AI travel-planning LobeHub (cluster of 3, @Scobleizer). First is a meeting with Cam Rooahmed, who runs an AR + AI brain-tumor visualization system used in 100+ surgeries with mistake-rate reductions Scoble claims save material money at $100K+ per surgery; uses Magic Leap. Second is Scoble plugging an interview with Robert Adams, whose company globaledentity.com is selling vein and skeleton biometrics now used in new TSA airport scanners. Third is LobeHub plugging a multi-agent system that planned his entire Google I/O week. All three are vendor-positive demos with no technical depth. Track only the brain-tumor AR system, which sits at the AI / medical robotics intersection.
- @brivael French-language run on capitalism, philosophy, AI as merit catalyst (cluster of 10, @brivael). Posts include a "Jensen Huang is a true good guy" Dell-event signing, an Estado Mínimo repost on Venezuela and socialism, a Peter Thiel quote on physical-world stagnation since 1970, an Elon-Musk-on-child-experiments line, an essay on free trade and minimal government, and a top-50-entrepreneurs analysis claiming early precocity plus genuine personal financial risk plus 10-15 year pain tolerance are the only common traits. No AI research content. Skip.
- One-line / promo (@Scobleizer on Robert Adams biometrics, @magicsilicon on White Castle in New Jersey, @WHFraudTF reposting VP Vance on fraud). Skip.