〔Day Digest〕OpenAI Safety Veteran With 12 Frontier-Model Reports Quits Calling Trial-and-Error 'Broken': Son Issues Rare Caution, White House Summit Yields 'Morally Binding' Pledge — Walkouts vs Binding Rules

0. The arc

Three safety vectors hit on the same day: an internal walkout at OpenAI, a rare public caution from SoftBank's Son, and a White House summit that produced a voluntary pledge labeled 'morally binding' ^[1][2][3]. The pattern is the same on both coasts — safety advocates are leaving or warning, while governance defaults to a soft instrument ^{[1][2][4][3][5]}. The commercial race runs in parallel: Meta's Muse accumulates downloads through consumer distribution, while OpenAI's Dots and Google retain the reasoning edge ^[6][7][8].

1. Safety stack — resignations and rare warnings

• **[ESCALATED] OpenAI safety veteran resigns.** David Robinson, three-and-a-half-year OpenAI employee and one of the company's most senior staff, quit and published an Atlantic essay warning the 'trial-and-error era is broken' ^[1][4]. Robinson led drafting of the Preparedness Framework and supervised 12 frontier-model safety reports ^[1][4]. A separate read frames the call as 'AI labs should run like nuclear plants' ^[5].
• **[ESCALATED] SoftBank Group CEO Son Masayoshi.** Speaking on a Kyoto 'Science and Technology in Society' forum panel alongside White House tech-policy advisor Michael Kratsios, Son said 'we no longer have the luxury of humans fighting each other' and called superintelligence in the wrong hands 'super dangerous' ^[3]. The tone is a marked break for an AI bull, and follows 'disturbing' model safety incidents in recent months ^[3].
• **[ONGOING] Anthropic 'Claude has a soul' lobbying.** The New York Times reported Anthropic, valued at $2 trillion, has held closed-door meetings with religious scholars and clergy to argue Claude may possess consciousness ^[9]. Scholar David Decosimo publicly questioned the effort on X, calling it a 'tech cult' in scientific clothing ^[9]. Single-source relay — treat as unverified ^[9].

2. White House summit and the 'morally binding' pledge

• **[NEW] Voluntary pledge, 'morally binding' framing.** A White House AI summit convened Meta CEO Mark Zuckerberg, Elon Musk, Anthropic CEO Dario Amodei, Nvidia CEO Jensen Huang, Google CEO Sundar Pichai and OpenAI president Greg Brockman ^[2]. The session produced a voluntary commitment agreement on AI safety and data center build-out, which Trump characterized as having 'moral binding force' ^[2]. No enforcement mechanism was disclosed in the relay ^[2].
• **[NEW] Power-map read.** Body-language expert Abbie Maroño, speaking to The Information, said Zuckerberg displayed the strongest power signal, standing near Trump throughout; Musk showed deliberate distance; Amodei was visibly anxious ^[2]. Single-source social relay ^[2].

3. Agent race — distribution vs reasoning

• **[NEW] Truist on Muse vs Dots.** Truist says Meta's Muse benefits from Instagram, Facebook and WhatsApp distribution plus integrations with Shopify, QuickBooks, Stripe, Canva and Slack; the SMB version can analyze sales/marketing, draft campaigns and screen spend, but posts, sends or pays still require owner approval ^[6]. Muse's first-mover edge is distribution, not underlying model strength — OpenAI's Dots and Google retain the open-ended reasoning lead ^[6]. Truist keeps a Buy rating and $763 target; early downloads hit 2.8M ^[6].
• **[NEW] BNP Paribas on Muse retention.** BNP Paribas analyst Nick Jones reports 5.6M+ cumulative downloads since September, a one-time U.S. App Store top rank, and initial downloads that outpaced ChatGPT and Sora — yet the open rate has stabilized below Google, ChatGPT, Instagram, WhatsApp and Facebook ^[7]. The bank maintains Outperform and an $885 target but cut its long-term Muse monetization estimate ^[7]. Single-broker note — quote as a band, not a point ^[7].
• **[NEW] OpenAI launches Dots.** OpenAI introduced Dots, a consumer agent positioned for developers and advanced users, with its own cloud computer, browser and app-connection layer — the immediate competitive pressure on Muse's retention curve ^[6][7][8].
• **[NEW] a16z thesis on enterprise AI agents.** a16z partner Seema Amble and Lio CEO Vladimir Keil argued ~80% of substantive enterprise procurement work occurs outside ERP systems, giving AI-native challengers an opening against incumbents that cannot move beyond chat-style embeddings ^[10].
• **[NEW] Token economy emerges as a formal pricing unit.** The 2026 World AI Conference introduced a 'Token Economy' track, formalizing the token as the core billing unit of MaaS inference; commentators flag 'irrational exuberance' and fragile business models alongside explosive consumption and infrastructure growth ^[11]. Single-source feature — treat as commentary ^[11].

4. Infrastructure squeeze — cooling, memory, power delivery

• **[NEW] Vertiv acquires KES.** Vertiv Technologies announced the acquisition of European KES, a high-density liquid-cooling data center firm offering fluid management, commissioning and load-testing services ^[12]. Vertiv has acquired 10 companies over the past four years and is shifting from product competition toward integrated O&M and services; Chinese domestic liquid-cooling M&A remains concentrated in hardware, the relay notes ^[12].
• **[NEW] Nvidia DGX Spark pricing.** DGX Spark now offers a 64GB model at $4,999; two units can pool to 128GB ^[13].
• **[NEW] Power delivery white paper.** Intel published a white paper arguing the Power Delivery Network (PDN) — not raw compute — is the binding constraint from accelerator to rack to data center, with voltage droop and resistive loss intensifying under stacked-die architectures ^[14].
• **[ESCALATED] Memory chip shortage reshapes PC and phone demand.** Suppliers are pivoting from a volume model to 'fewer units, higher ASP' as AI data center demand tightens DRAM and NAND supply ^[15].
• **[NEW] Cross-border data regulation cost.** A CEPR/VoxEU column finds AI-driven mobile apps more than triple their user base at home and abroad, but cross-border data restrictions halve the AI gain — a quantitative policy tension for digital services trade ^[16].

5. What would falsify the arc

• The 'walkout + voluntary pledge' narrative breaks if either (a) OpenAI backfills the safety role with an equivalent framework, or (b) a non-voluntary rule (FTC action, EU AI Act enforcement) lands before the next frontier release ^[1][2][5].
• The Muse distribution thesis breaks if Dots reaches comparable consumer touchpoints, or if Muse's open rate converges with mainstream apps within one quarter ^[6][7][8].
• The infrastructure squeeze narrative breaks if Vertiv-class M&A translates into sustained service-margin expansion rather than a one-off re-rating ^[12].
• Source quality control: the White House body-language read, the Anthropic soul-lobbying relay, the Korean data-breach note, the Figure AI melting story, the $300M Nvidia smuggling case and the 2030 Huang projection repeat are all single-source social relays — flag and re-check before quoting at length ^{[17][2][18][9][19][20]}. Musk's SpaceX-to-SpaceXSI rebrand, the APEC HR ministerial communique and the CEPR policy column are likewise single-source items ^[21][16][22].