Summary
Two stories ran across both slots today. NVIDIA's Vera CPU hand-delivery tour to Anthropic, OpenAI, SpaceXAI, and Oracle landed in the morning AI feed as the official NVIDIA blog framing of Vera as "the agentic-AI CPU," then resurfaced in the afternoon @bayesiansapien retweets with on-the-record customer reactions (Anthropic calling Vera promising for agentic workloads, Oracle committing to hundreds of thousands of units in 2026). Cursor Composer 2.5 was the other cross-slot thread, a six-post launch in the morning from @cursor_ai detailing a Kimi K2.5 base with textual-feedback RL, Sharded Muon, dual-mesh HSDP, and a 10x-compute follow-on with SpaceXAI on Colossus 2, then a (still unsourced) Karan Kendre repost in the afternoon claiming near-Opus-4.7 quality at 10x lower cost. The afternoon's standout single-slot story was the Musk v. Altman verdict, where all three charitable-trust claims were dismissed as time-barred and Musk filed a same-day 9th Circuit appeal. The morning's standout was Anthropic's three-in-one ship of Fast Mode on Opus 4.7 by default, prompt-cache diagnostics in the console, and the "extra usage" to "usage credits" rename, which together point at a metered-throughput pricing model. Everything else was Tesla FSD testimonials, a Scoble medical-AI demo run, and the recurring @brivael French-language political feed with no AI substance.
Posts
- NVIDIA Vera CPU hand-delivery: Anthropic, OpenAI, SpaceXAI, Oracle get first units (cluster of 4 [morning + afternoon], @nvidia · @ns123abc · NVIDIA blog). Ian Buck personally hand-delivered the first Vera CPUs Friday May 15 across the Bay Area, then Oracle Cloud Santa Clara on Monday. Anthropic publicly called Vera "a promising part of the ecosystem when solving for agentic workloads," Musk asked SpaceXAI evaluation questions about cores, memory layout, cooling, and Oracle committed to hundreds of thousands of Vera CPUs starting 2026. Vera is positioned as a different CPU architecture for agentic workloads where host-side dispatch and context-loading dominate.
- Cursor Composer 2.5: Kimi K2.5 base, textual-feedback RL, Sharded Muon, dual-mesh HSDP, 10x-compute SpaceXAI follow-on (cluster of 7 [morning + afternoon], @cursor_ai · @mntruell · @karankendre · Cursor blog). Composer 2.5 is built on Moonshot's Kimi K2.5 open-weight base with three training-stack changes: textual-feedback RL for credit assignment across hundred-thousand-token rollouts, synthetic data generation for complex RL environments, and a Sharded Muon optimizer with dual-mesh HSDP. Truell confirms the SpaceXAI collaboration is a from-scratch larger model with 10x total compute on Colossus 2's million H100-equivalents. Karan Kendre's afternoon repost (carried by @brivael) claims near-Opus-4.7 quality at 10x lower cost, currently unsourced and worth tracking only once Cursor publishes eval cards. Second consecutive week of Cursor shipping on a Chinese open-weight base, now a stable pattern.
- Musk v. Altman: jury verdict and 9th Circuit appeal (cluster of 5 [afternoon], @ns123abc). Jury unanimously found all three Musk claims (breach of charitable trust, aiding and abetting, unjust enrichment) time-barred in roughly 90 minutes; the only counts that could have produced disgorgement, a constructive trust on Brockman's ~$30B equity, or nonprofit control of the for-profit were dismissed without weighing the underlying breach evidence. Musk filed a 9th Circuit appeal same day; Toberoff argues the limitations clock should have run from 2025's PBC conversion with Microsoft at 27%, not from 2020. Context for the Anthropic overtakes OpenAI B2B thread on OpenAI corporate form.
- Anthropic ships Fast Mode default on Opus 4.7, prompt-cache diagnostics, usage-credit rename (cluster of 6 [morning], @ClaudeDevs · fast-mode docs · cache diagnostics docs). Three changes in one push. Fast Mode is identical Opus quality at ~2.5x response speed at a higher per-token rate, toggleable via
/fast, still research preview. Prompt-cache diagnostics now show exactly which prompt segment caused a cache miss and the token cost, addressing one of the most-requested missing features for high-cache-rate workloads. "Extra usage" renamed to "usage credits" with the framing that credits now power features beyond plan overflow (Fast Mode, other premium routes). The three together describe Anthropic moving toward a metered-throughput pricing model with speed and cache quality as first-class billable axes. - How Claude Code works in large codebases ([morning], @ClaudeDevs · blog). Anthropic publishes a prescriptive best-practices post for multi-million-line monorepos, legacy systems, and distributed microservices. Fetched article body was navigation chrome; the substantive playbook requires the click-through.
- Demis Hassabis was an angel investor in Anthropic at founding ([afternoon], @ns123abc · ft.com). FT-sourced claim that Dario views Hassabis as a role model and Hassabis put personal money in at start. If accurate, it complicates the standard DeepMind/Anthropic-as-rivals framing. URL in tweet was truncated, click through to verify.
- Open Collider: LLM creativity engine that mechanically improves idea diversity ([morning], @brivael reposting @cdriclion). Cedric Lion's open-source project positioned as a sampling-and-recombination engine that forces non-trivial idea diversity, framed against Jiang et al. (2025) "Artificial Hivemind" collapse on repeated sampling from the same brief. No method details exposed in the tweet, diagnosis is consistent with prior wiki coverage of LLM-creativity-bottleneck work and the open RLVR-distribution-contraction question.
- Tesla FSD v14.3.3 testimonials (cluster of 3 [morning], @Tesla). Community videos of v14.3.3 catching potential side-swipes and merges, plus a Herbert Ong summary of much-reduced driver-monitor nagging, smoother human-like driving, faster Smart Summon, and "Hey Grok" voice. Tier 3.
- Robert Scoble medical-AI field reports: AR brain surgery, vein-and-skeleton biometrics, LobeHub agent travel planning (cluster of 3 [morning], @Scobleizer). Vendor-positive demos with no technical depth. The AR brain-tumor visualization with claimed mistake-rate reductions is the only item worth tracking.
- Hello from Code with Claude London ([afternoon], @bcherny). Anthropic developer event. Skip.
- @brivael French-language political and personal feed (cluster of 24 [morning + afternoon], @brivael). Capitalism, philosophy, AI-as-merit-catalyst, Jensen Huang Dell signing, Peter Thiel on physical-world stagnation, Sarah Knafo on a 15-year-old hacking the French state, Musk-via-Ashley-St-Clair gossip, top-entrepreneurs analysis. No AI research content. Skip.
- Promo and one-line reposts ([morning], @Scobleizer on Robert Adams biometrics, @magicsilicon on White Castle, @WHFraudTF reposting VP Vance on fraud). Skip.