OpenAI ships Codex inside the ChatGPT mobile app — phone becomes a control surface for laptop-side agents
No codebase on the phone. Files, credentials and local setup stay on the developer's machine; only diffs, terminal streams and approval prompts flow over a secure relay.
OpenAI ships Codex inside the ChatGPT mobile app — phone becomes a control surface for laptop-side agents
No codebase on the phone. Files, credentials and local setup stay on the developer's machine; only diffs, terminal streams and approval prompts flow over a secure relay.
- Codex is now embedded in the ChatGPT mobile app for iOS and Android — in preview, available on every ChatGPT plan including free.
- The phone is a control surface, not a host: developers monitor active Codex sessions running on laptops / Mac minis / managed devboxes.
- Capabilities: start tasks, inspect threads, approve commands, switch models, add context, follow terminal output / diffs / test results in real time.
- Security model: code, credentials and local config stay on the host machine; only real-time updates stream to the phone over a secure relay.
Implications: OpenAI is staking out "agent oversight, anywhere" as a product category rather than just a feature. Anthropic, Cursor and Windsurf all need a credible answer for "what does an agent on a long-running task look like when you're away from your laptop?"
PwC × Anthropic expands — 30,000 PwC consultants get certified on Claude, new Office of the CFO business unit
Claude Code + Cowork roll out to PwC's US team first, then global; Office of the CFO becomes PwC's first business unit explicitly anchored on Anthropic.
PwC × Anthropic expands — 30,000 PwC consultants get certified on Claude, new Office of the CFO business unit
Claude Code + Cowork roll out to PwC's US team first, then global; Office of the CFO becomes PwC's first business unit explicitly anchored on Anthropic.
- Joint Center of Excellence + a certification program for 30,000 PwC professionals on Claude.
- Three focus areas: agentic technology build, AI-native deal-making, and reinvention of the enterprise function.
- New Office of the CFO business unit at PwC is the first standalone group anchored on Claude.
- Already in production: professional sports ops, insurance underwriting, mainframe modernization, HR transformation, cybersecurity — PwC claims up to 70% delivery-time reduction.
Implications: Following the KPMG × Claude rollout to 276K staff earlier this month, every Big-4 firm is now either officially "with Anthropic" or "with someone else." Watch Deloitte and EY for their next-quarter announcements.
Karpathy's Anthropic remit comes into focus: Claude-using-Claude as the next pretraining engine
Follow-up to last week's hire — Anthropic also ships dreaming, multi-agent orchestration, outcomes, and webhooks for Claude Managed Agents.
Karpathy's Anthropic remit comes into focus: Claude-using-Claude as the next pretraining engine
Follow-up to last week's hire — Anthropic also ships dreaming, multi-agent orchestration, outcomes, and webhooks for Claude Managed Agents.
- Karpathy's new sub-team uses Claude to automate parts of the pretraining research loop — concretely, dataset selection, ablation design, and run analysis.
- Same week, Anthropic shipped higher Claude Code and Opus API limits.
- New Managed Agents primitives: dreaming (passive context build-up), multi-agent orchestration, outcomes, webhooks.
- Plus self-hosted sandboxes and MCP tunnels (from the prior week) for air-gap-friendly enterprise deployment.
Implications: "AI to accelerate AI training" stops being a research vibe and starts being a product line. If Anthropic's pretraining cycle compresses meaningfully, the gap between Claude Opus N and Opus N+1 shortens — and everyone else has to match.
Pope Leo XIV releases "Magnifica Humanitas" — the first papal encyclical on AI
235-page text presented today at the Vatican's Synod Hall alongside Chris Olah (Anthropic) — Pope Leo XIV is the first pontiff to personally present an encyclical, and AI is the chosen subject.
Pope Leo XIV releases "Magnifica Humanitas" — the first papal encyclical on AI
235-page text presented today at the Vatican's Synod Hall alongside Chris Olah (Anthropic) — Pope Leo XIV is the first pontiff to personally present an encyclical, and AI is the chosen subject.
- Released May 25 at 11:30 AM at the Synod Hall; signed by Pope Leo XIV on May 15, 2026.
- Core thesis: control of AI must not remain "in the hands of a few"; use in warfare must be subject to "the most rigorous ethical constraints."
- Presented alongside Cardinal Fernández, Cardinal Czerny, Cardinal Parolin, theologian Anna Rowlands, and Chris Olah (Anthropic head of interpretability).
- Frames AI as the "social question" of this era — explicit parallel to Rerum Novarum (1891) on industrial-era workers' rights.
"Control of artificial intelligence must not remain in the hands of a few... the use of AI in warfare must be subject to the most rigorous ethical constraints."
— Pope Leo XIV, Magnifica Humanitas
Implications: Anthropic gets a global moral platform that compounds with its safety-first commercial story — a counterweight to the Pentagon "supply-chain risk" designation from earlier this month. Watch which AI labs publicly engage with the encyclical's framing on AI-in-warfare vs. which stay silent.
Anthropic + Gates Foundation commit $200M over four years — global health, education, agriculture
$100M grant funding from Gates, $100M Claude credits + technical support from Anthropic; biggest line item is health outcomes in low-income countries.
Anthropic + Gates Foundation commit $200M over four years — global health, education, agriculture
$100M grant funding from Gates, $100M Claude credits + technical support from Anthropic; biggest line item is health outcomes in low-income countries.
- Four-year, $200M partnership announced May 14, 2026.
- Health focus: extends Gates' work on overlooked diseases — polio, HPV, eclampsia/preeclampsia — and on health access for 4.6B people without essential services.
- Education: Claude-powered evidence-based tutoring for K-12 and career guidance; foundational literacy / numeracy in sub-Saharan Africa and India.
- Agriculture: Claude improvements for local crops + agriculture-specific benchmarks.
Implications: Anthropic deliberately stacks public-good wins — Gates, Vatican, KPMG/PwC — into a single "responsible frontier" brand. It's both moral positioning and an enterprise sales aid; safety-conscious buyers prefer the partner the Pope hosts.
Microsoft Agent 365 hits general availability — $15/user enterprise control plane for AI agents
Observe, govern, secure. Microsoft's bet that the next enterprise budget line item is agent management, not agent creation.
Microsoft Agent 365 hits general availability — $15/user enterprise control plane for AI agents
Observe, govern, secure. Microsoft's bet that the next enterprise budget line item is agent management, not agent creation.
- GA on May 1, 2026; standalone at $15/user/month, also bundled in the new Microsoft 365 E7 at $99/user/month.
- Three pillars: observe (visibility into agent fleet), govern (consistent controls), secure (identity + data + threat defense).
- Extends Microsoft's enterprise identity / data / threat-defense capabilities across the AI agent ecosystem.
- Sits in the same procurement conversation as Claude Managed Agents and Google Antigravity — but Microsoft sells it as the control plane, not the agent runtime.
Implications: Enterprise IT teams now have a budget category for "agent governance" — separate from the agents themselves. Whoever owns this control plane becomes structurally hard to displace, the same way Active Directory locked in identity for two decades.
X1 demo: humanoid carries a morphing drone-rover hybrid on its back — first multi-form rescue team
A modified Unitree G1 walks heavy gear into the field, then launches an M4 robot that switches between drone, wheeled rover and crawler depending on terrain.
X1 demo: humanoid carries a morphing drone-rover hybrid on its back — first multi-form rescue team
A modified Unitree G1 walks heavy gear into the field, then launches an M4 robot that switches between drone, wheeled rover and crawler depending on terrain.
- X1 is a tandem rescue rig: humanoid base (Unitree G1, bipedal locomotion) carries a morphing payload robot (M4).
- M4 has three modes — quadrotor drone, four-wheeled rover, and a crawler — and switches mid-mission.
- Demo scenario: G1 walks the assembly to the edge of an open area, M4 launches in drone mode, lands, folds rotors into wheels, drives to a pond, then flies again.
- Built by Caltech's research group; targets disaster response and inspection where one form factor isn't enough.
Implications: Multi-modal robot teams are leaving the lab. The Pentagon's counter-UAS contracts (Perennial Autonomy's Merops / Bumblebee / Hornet) are already paying for variants of this thinking on the offensive side; rescue / inspection is the dual-use parallel.
Humanoid market state: pilot → platform transition is now measurable
Agility Digit (7+ at Toyota Canada), BD/Hyundai Atlas (2026 commit fully booked), Unitree (5,500 shipped in 2025, 10-20k targeted 2026), Figure 03 (BMW scaling).
Humanoid market state: pilot → platform transition is now measurable
Agility Digit (7+ at Toyota Canada), BD/Hyundai Atlas (2026 commit fully booked), Unitree (5,500 shipped in 2025, 10-20k targeted 2026), Figure 03 (BMW scaling).
- Schaeffler × Humanoid (London) — 1,000+ wheeled humanoids by 2032, first live deployment in Germany Dec 2026.
- Agility Digit fleet at Toyota Canada now 7+ units handling RAV4 material flow.
- Unitree shipped 5,500+ in 2025, targeting 10,000-20,000 in 2026 — JAL's two-year Haneda trial is one buyer.
- Figure 03 scaling at BMW Spartanburg; pilot phase pivots to mass production.
Implications: The market's no longer arguing "if" — it's arguing "how many per workstation, how fast, and what's the right form factor by use case." Wheeled humanoids win on safety + cost; bipedal wins on tight-space versatility.
Perennial Autonomy's $500M Pentagon IDIQ — Merops already deployed protecting US troops in Germany
JIATF-401 procurement vehicle plus combat track record. Merops counter-UAS units now alongside US forces in Germany after 4,000+ Russian-drone intercepts in Ukraine.
Perennial Autonomy's $500M Pentagon IDIQ — Merops already deployed protecting US troops in Germany
JIATF-401 procurement vehicle plus combat track record. Merops counter-UAS units now alongside US forces in Germany after 4,000+ Russian-drone intercepts in Ukraine.
- $500M 3-year IDIQ from Joint Interagency Task Force 401 covers Merops interceptors, Bumblebee quadcopters, Hornet mid-range strike drones.
- Eric Schmidt's company (rebranded from White Stork / Project Eagle); Merops has 4,000+ Russian-drone intercepts since mid-2024.
- Merops units are now actively protecting US troops in Germany.
- Sits inside the DAWG $54.6B FY27 spending narrative — DAWG is now the Pentagon's umbrella for autonomous-warfare procurement.
Implications: "Combat-proven in Ukraine + IDIQ procurement" is the fastest path from startup demo to deployed kit at US-force scale. Expect Anduril, Shield AI, and a tail of dual-use companies to chase the same template.
Petraeus & Flanagan call DAWG's $54.6B "the largest single commitment to autonomous warfare in history"
A ~24,000% year-over-year jump in the autonomous-warfare budget line, almost entirely via the FY27 reconciliation package.
Petraeus & Flanagan call DAWG's $54.6B "the largest single commitment to autonomous warfare in history"
A ~24,000% year-over-year jump in the autonomous-warfare budget line, almost entirely via the FY27 reconciliation package.
- DAWG (Defense Autonomous Warfare Group) absorbed Replicator in late 2025; FY26 funding was $225.9M.
- White House requests $54.6B for DAWG in FY27 — a near 24,000% increase.
- $53.6B of that is in the proposed $350B defense reconciliation package, only $1B in the base budget.
- Senators have flagged that DoD Directive 3000.09 on "appropriate levels of human judgment" hasn't kept pace with AI-driven targeting.
Implications: The Pentagon now buys autonomous-warfare hardware the way it once bought ammunition — at enterprise procurement scale. Watch for follow-on IDIQs in the Perennial mold during Q3 / Q4 of this fiscal year.
FAA opens Starship V3 booster mishap probe — Flight 13 license slips to Q3 2026
Booster 19, the first V3-spec Super Heavy, suffered multiple Raptor 3 engine failures during boost-back and came apart over the Gulf of Mexico. Six unrelated flights delayed by the airspace closure.
FAA opens Starship V3 booster mishap probe — Flight 13 license slips to Q3 2026
Booster 19, the first V3-spec Super Heavy, suffered multiple Raptor 3 engine failures during boost-back and came apart over the Gulf of Mexico. Six unrelated flights delayed by the airspace closure.
- Flight 12 launched May 22 from a brand-new Pad 2 at Starbase; first V3-spec Starship to fly.
- Booster 19 lost multiple Raptor 3 engines during the boost-back burn and broke apart over the Gulf — not planned for recovery, but the boost-back behaviour is the data the program now has to fix.
- Ship 39 reached SECO with 5/6 engines, deployed 22 mock Starlinks (two with external cameras), survived re-entry without heat-shield burnthrough, and made a controlled Indian Ocean splashdown.
- FAA mishap investigation opened May 23; six unrelated flights were delayed by the airspace closure.
- Mishap report expected by June 2026; Flight 13 launch license anticipated in Q3 2026.
Implications: Booster recovery on V3 stays the open program risk. The good news for SpaceX investors: the upper stage's re-entry survival is the harder engineering problem and that data is in the bag. The IPO narrative now hinges on whether the Q3 Flight 13 license actually lands.
Nvidia consolidates after the $81.6B beat — Vera Rubin tease + 25× dividend bump headline the next leg
Friday close $214.28 after a $221 intraday high. The narrative pivot: Blackwell still ships, but Vera Rubin "trains AI models with 75% fewer GPUs" is the next thesis.
Nvidia consolidates after the $81.6B beat — Vera Rubin tease + 25× dividend bump headline the next leg
Friday close $214.28 after a $221 intraday high. The narrative pivot: Blackwell still ships, but Vera Rubin "trains AI models with 75% fewer GPUs" is the next thesis.
- NVDA traded $214.25–$221.49 on May 22; closed at $214.28.
- Q1 FY27: $81.6B revenue, $58.3B net income, EPS $1.87 diluted; $80B buyback authorized; dividend hiked 25× to $0.25.
- Vera Rubin systems ship later this year — Nvidia's pitch is 75% fewer GPUs per training run vs. Blackwell.
- Bank of America hikes 2026 chips forecast to $1.3T — names Nvidia, Broadcom, Marvell, AMD as top drivers.
Implications: Nvidia's narrative has shifted from "training compute" to "training-compute efficiency." If Vera Rubin actually halves AI capex per token, the AI-stack economics change for every hyperscaler customer — and the rotation toward AMD / Broadcom / Marvell competes against an Nvidia that's defending margin, not unit volume.
PHLX Semi +65% YTD; BofA bumps 2026 chip forecast to $1.3T — and the dot-com warnings get louder
AMD's data-center segment +57% YoY and Q2 guide $11.2B above consensus. Veterans drawing parallels to 1999-2000 are raising the 25-30% correction hand.
PHLX Semi +65% YTD; BofA bumps 2026 chip forecast to $1.3T — and the dot-com warnings get louder
AMD's data-center segment +57% YoY and Q2 guide $11.2B above consensus. Veterans drawing parallels to 1999-2000 are raising the 25-30% correction hand.
- SOX (PHLX Semi Index) up over 65% YTD in 2026.
- AMD Q1: $1.37 adjusted EPS on $10.25B revenue; data-center +57% YoY; Q2 guide $11.2B.
- Analyst tape: 30 buys, 2 strong buys, 12 holds — zero sells on AMD.
- Veteran analysts flag potential 25-30% correction risk, drawing the 1999-2000 comparison.
Implications: The AI-silicon trade is now broad enough that "Nvidia-only" investors look concentrated. The flip side: if the correction calls are right, the broad chip basket is where the gravity hits first.