Andrej Karpathy joins Anthropic's pre-training team
OpenAI co-founder, ex-Tesla Autopilot director and Eureka Labs founder returns to frontier R&D — and will help bootstrap a sub-team using Claude itself to accelerate pretraining research.
Andrej Karpathy joins Anthropic's pre-training team
OpenAI co-founder, ex-Tesla Autopilot director and Eureka Labs founder returns to frontier R&D — and will help bootstrap a sub-team using Claude itself to accelerate pretraining research.
- Joins Anthropic's pre-training team this week, reporting to team lead Nick Joseph; the team owns the large-scale training runs that produce Claude's core capabilities.
- Will help launch a new sub-team focused on using Claude to accelerate pretraining research — agentic AI as a competitive lever, not just raw compute.
- Career arc: founding member of OpenAI (2015-2017) · Director of AI at Tesla (2017-2022), where he led Autopilot · returned to OpenAI (2023-Feb 2024) · founded education startup Eureka Labs (July 2024).
- PhD at Stanford under Fei-Fei Li; widely known for the open-source
nn-zero-to-hero,micrograd,makemoreandllm.ctutorials. - Says he remains "deeply passionate about education" and plans to resume that work "in time," but is jumping back into core R&D now.
"Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time."
— Andrej Karpathy, on X
The hire lands at a delicate moment for OpenAI and slots cleanly into Anthropic's thesis that the next phase of frontier AI is won by automating research itself. Putting Karpathy — arguably the most respected pedagogue/practitioner in the field — on Nick Joseph's pre-training team, and asking him to bootstrap a "use Claude to do pretraining research" team, signals where Anthropic believes its leverage is: not just more H100s, but agentic Claude-on-Claude research loops.
Implications: Expect renewed talent-market pressure on OpenAI; an uptick in Anthropic recruiting momentum; and likely visible product fingerprints over the next 6-12 months in how Claude tools its own training pipeline.
Vercel Labs' agent-browser: native Rust CLI for AI agents driving Chrome
Open-source browser automation built for LLM agents — no Playwright, no Node, just a Rust binary and Chrome for Testing.
Vercel Labs' agent-browser: native Rust CLI for AI agents driving Chrome
Open-source browser automation built for LLM agents — no Playwright, no Node, just a Rust binary and Chrome for Testing.
- Fast native Rust CLI; installs via
npm i -g agent-browser, Homebrew, or Cargo. No Playwright or Node.js required for the daemon. - Uses Chrome from Google's official Chrome for Testing automation channel. Existing Chrome/Brave/Playwright/Puppeteer installs are detected automatically.
- Commands designed for agents:
snapshot(accessibility tree with refs the model can click by),click @e2,fill @e3,screenshot --annotate,pdf,eval <js>, and a built-inchatREPL for natural-language control. - Also supports semantic locators (
find role button click --name "Submit") and traditional CSS selectors. - Latest release v0.27.0 added React DevTools integration (
react tree,react inspect,react renders), Core Web Vitals reporting, SPApushstatenavigation, and repeatable--init-script+--enablefeature flags.
Implications: Positions the browser as a first-class agent surface with primitives that match how LLMs reason (refs from an a11y tree, not pixel-perfect XPath). The React-DevTools integration in particular reads as Vercel making the case that "your agent should understand the framework, not just the DOM."
Boston Dynamics shows how Atlas learned to lift 100+ pound loads via RL
Electric Atlas hoists mini-fridges using proprioception — not vision — after sim-trained RL generalises from 50-70 lb training payloads to 100+ lb in the real world.
Boston Dynamics shows how Atlas learned to lift 100+ pound loads via RL
Electric Atlas hoists mini-fridges using proprioception — not vision — after sim-trained RL generalises from 50-70 lb training payloads to 100+ lb in the real world.
- Atlas trained on 50-70 lb loads in simulation generalised to mini-fridges over 100 lb in real-world tests without additional training.
- Recipe: reference trajectory (teleop or animation) + reward shaping for grip and disturbance resistance + heavy domain randomization (friction, motor strength, weight) + real-robot iteration.
- Atlas weighs ~90 kg; narrow sim-to-real gap credited to hardware with only two actuator types and full arm/leg symmetry.
- Hyundai (parent) targets 30,000 Atlas units/year by 2028 from a new Savannah, GA factory; current build rate is ~4 robots/month.
"You cannot lift a fridge just by looking at it and using your hands."
— Boston Dynamics Atlas team
Implications: Proprioception-first RL is emerging as the dominant recipe for industrial humanoid manipulation; combined with Hyundai's 30k/year roadmap, this is the clearest Western signal that heavy-duty humanoids are leaving the demo phase.
Apptronik doubles Austin HQ, opens Silicon Valley office as humanoid race tightens
Apollo pilots with Mercedes-Benz and GXO Logistics scale up; headcount grows from ~300 to ~500.
Apptronik doubles Austin HQ, opens Silicon Valley office as humanoid race tightens
Apollo pilots with Mercedes-Benz and GXO Logistics scale up; headcount grows from ~300 to ~500.
- Doubled Austin HQ footprint to scale Apollo training and data collection; opening a Silicon Valley office.
- Apollo deployed in real-world pilots with Mercedes-Benz and GXO Logistics; partnership with Google DeepMind plugs Gemini-family models into the robot's reasoning stack.
- $520M Series A extension closed February 2026 (co-led by B Capital and Google); total Series A north of $935M at a reported ~$5B valuation.
- Direct competition frame: Tesla Optimus, Figure (US) vs Unitree, AgiBot (China — currently leading in unit shipments).
Implications: US humanoid leaders are pivoting from R&D-mode to industrial buildout (factories, training data, headcount) — capital and footprint are becoming the moats, not just robot demos.
IDTechEx: humanoid ROI sharpens — payback under 6 months at high utilisation
Average selling price projected to drop from ~$115k (2024) to ~$37k by 2030; commercial scale still gated by capability gaps.
IDTechEx: humanoid ROI sharpens — payback under 6 months at high utilisation
Average selling price projected to drop from ~$115k (2024) to ~$37k by 2030; commercial scale still gated by capability gaps.
- Payback projected at ~6 months under high-utilisation scenarios vs ~15 months at medium utilisation.
- ASP forecast: $114,700 (2024) → ~$37,000 (2030); operating cost under $5/hour at high utilisation by 2030.
- Market projected at ~$25B by early 2030s; annual shipments reaching 1.8M units by 2036.
- Primary near-term demand: automotive manufacturing, then logistics; complex/variable/safety-critical tasks remain unsolved.
Implications: For the first time the unit-economics narrative is being pegged to specific payback windows — but IDTechEx explicitly cautions that effective task generalisation, not cost curves, decides which vendors actually scale.
Shield AI tapped to integrate Hivemind autonomy onto Pentagon's LUCAS attack drones
~$35k one-way attack drones get swarm-capable AI — one operator commands the swarm, humans retain strike authority.
Shield AI tapped to integrate Hivemind autonomy onto Pentagon's LUCAS attack drones
~$35k one-way attack drones get swarm-capable AI — one operator commands the swarm, humans retain strike authority.
- LUCAS is ~10 feet long with an explosive payload that detonates on impact; unit cost ~$35,000.
- Hivemind enables coordination, manoeuvre, and adaptation in real time; a single human operator retains strike authority.
- Pilot includes an operational demonstration in fall 2026 with one operator commanding a swarm.
- Sits inside Defense Secretary Hegseth's "Drone Dominance" push; DoD planning ~$55B on drones/autonomy in FY2027 under the Defense Autonomous Warfare Group.
- Shield AI is fresh off a $2B raise at a $12.7B valuation.
"Mass without coordination is limited in value. Hivemind is the AI pilot that makes that mass intelligent."
— Brandon Tseng, President, Shield AI
Implications: LUCAS gives the Pentagon a swarm-ready Shahed-class attritable munition, and Hivemind — already on one-way attack drones in Ukraine — becomes a de facto autonomy reference stack for cheap mass strike.
Pentagon awards Perennial Autonomy $500M counter-drone ceiling contract
JIATF 401 deal covers AI-enabled interceptors and strike drones already battle-tested against Iranian UAVs.
Pentagon awards Perennial Autonomy $500M counter-drone ceiling contract
JIATF 401 deal covers AI-enabled interceptors and strike drones already battle-tested against Iranian UAVs.
- Systems covered: Merops air-to-air interceptors (~$15,000/unit), Bumblebee V2 FPV counter-drones, Hornet midrange strike drones.
- Merops was originally built for Ukrainian forces to counter Russian Shahed drones; now used by U.S. forces against Iranian UAVs.
- Separate $5.2M Bumblebee V2 contract awarded January; Hornet was Army-tested in March.
- Three-year duration or until $500M is expended, whichever comes first.
"Drones are the defining threat of our time. Deploy and scale low-cost, attritable air-to-air drone interceptors at all our facilities."
— Brig. Gen. Matt Ross
Anduril + Booz Allen integrate cyber/RF effects into Menace and Lattice for SOF teams
Sit(x) + zero-trust + DETS fold into Anduril's C4 and Lattice — operational integrations demoed at SOF Week 2026.
Anduril + Booz Allen integrate cyber/RF effects into Menace and Lattice for SOF teams
Sit(x) + zero-trust + DETS fold into Anduril's C4 and Lattice — operational integrations demoed at SOF Week 2026.
- Sit(x) on Menace gives a unified view across ATAK (Android), WinTAK, web, and iTAK clients.
- DETS runs as a Lattice-integrated app letting operators task cyber and RF effects inside C2 workflows.
- Booz Allen's zero-trust layer handles policy enforcement and logging to ease accreditation.
"Allows us to connect to compromised networks all over the world and still utilize mission applications without compromising the operators on the ground."
— Brett Melancon, Anduril Chief Solutions Architect
SOF Week 2026 opens in Tampa — AI, autonomy and counter-UAS as headline themes
SOCOM + Global SOF Foundation event runs May 18-21; exhibition floor live May 19-21.
SOF Week 2026 opens in Tampa — AI, autonomy and counter-UAS as headline themes
SOCOM + Global SOF Foundation event runs May 18-21; exhibition floor live May 19-21.
- Quantum Cyber unveiled an AI-powered "system of systems" platform for autonomous warfare, counter-UAS, and border security, including a 25+ km autonomous drone.
- Galvion launched the Cortex Evo integrated combat helmet combining ballistic protection, power, data and onboard processing.
- One Stop Systems showcasing rugged edge AI, sensor fusion and tactical compute.
- TEKEVER announced its first U.S. office in Fayetteville, NC, timed to SOF Week.
Implications: SOF Week has become the de facto deal floor for autonomy-at-the-edge vendors — European players are now planting U.S. flags directly into SOCOM's backyard rather than approaching through Beltway primes.
Starship V3 Flight 12 slips again to May 21 — first integrated test in seven months
Second consecutive delay for SpaceX's redesigned V3 stack; Booster 19 + Ship 39 stand 408 ft tall with Raptor 3 engines.
Starship V3 Flight 12 slips again to May 21 — first integrated test in seven months
Second consecutive delay for SpaceX's redesigned V3 stack; Booster 19 + Ship 39 stand 408 ft tall with Raptor 3 engines.
- Launch pushed from May 19 → May 20 → no earlier than Thursday, May 21 at 6:30 p.m. EDT from Pad 2 at Starbase, Texas.
- Raptor 3 engines: sea-level thrust 250 tf (up from 230 tf); V3 stack stands 408 ft tall.
- Designed to lift over 100 t to LEO — nearly triple V2's ~35 t capacity.
- Ship 39 will deploy ~20 next-gen Starlink simulators; two carry cameras to scan the heat shield mid-flight.
- Booster 19 splashes down in the Gulf ~7 min after liftoff; Ship 39 targets Indian Ocean ~1 hour in. No catch attempts.
- Heat shield carries one intentionally missing tile to stress-test reentry survivability.
Implications: NASA's Artemis 4 lunar landing in 2028 depends on Starship working as the Human Landing System; a clean V3 debut is on the critical path. Multiple slips with no public technical explanation suggest preflight checks on the new pad and Raptor 3 stack are running tight.
SpaceX pushes Starlink past 10,500 active satellites with Vandenberg launch
Falcon 9 lifts 24 Starlink v2 Mini from SLC-4E — 58th SpaceX launch of 2026, 612th booster recovery overall.
SpaceX pushes Starlink past 10,500 active satellites with Vandenberg launch
Falcon 9 lifts 24 Starlink v2 Mini from SLC-4E — 58th SpaceX launch of 2026, 612th booster recovery overall.
- Liftoff at 7:46 p.m. PDT May 20 from SLC-4E at Vandenberg SFB carrying 24 Starlink v2 Mini satellites (Starlink 17-42).
- Booster B1103 flew its second mission and landed on droneship "Of Course I Still Love You" ~8 min after liftoff.
- SpaceX's 58th launch of 2026 and the company's 651st successful flight since 2010.
- Starlink constellation now numbers more than 10,500 operational satellites, including 600+ with direct-to-device capability.
NASA refines Artemis III plan as core stage stacks in VAB
Mission re-scoped as crewed Earth-orbit rendezvous-and-docking test with commercial landers; SLS core stage mated to engine section May 12.
NASA refines Artemis III plan as core stage stacks in VAB
Mission re-scoped as crewed Earth-orbit rendezvous-and-docking test with commercial landers; SLS core stage mated to engine section May 12.
- No longer targets a lunar surface landing on first crewed flight; will test rendezvous/docking between Orion and HLS landers in cislunar or Earth orbit.
- Crew selection, AxEMU spacesuit interface, mission duration and science ops are all still being defined "over the coming weeks."
- Follows the successful Artemis II 10-day lunar flyby with Wiseman, Glover, Koch and Hansen.
Implications: Re-scoping Artemis III away from a surface landing is a tacit acknowledgment that neither SpaceX's Starship HLS nor Blue Origin's MK2 lander will be flight-ready on the original timeline — making Starship V3 Flight 12 a higher-stakes test for the entire Artemis program.
Armada raises $230M at $2B, partners with Johnson Controls on 400k sq ft AI data-center factory
Series B co-led by Overmatch, BlackRock and 8090 Industries; Galleon Forge One will manufacture megawatt-scale modular data centers for edge AI.
Armada raises $230M at $2B, partners with Johnson Controls on 400k sq ft AI data-center factory
Series B co-led by Overmatch, BlackRock and 8090 Industries; Galleon Forge One will manufacture megawatt-scale modular data centers for edge AI.
- $230M Series B at $2B pre-money; total funding now ~$500M.
- New investors: BlackRock, Johnson Controls, NightDragon, Mitsui, Singtel Innov8.
- Galleon Forge One factory: up to 400,000 sq ft in Arizona, 500+ direct jobs, production of "Leviathan" modular data center starts summer 2026.
- Customer bookings grew 540% FY25-FY26; Q1 FY27 bookings +2,000% YoY.
- Existing deployments include U.S. Navy, Aker BP, WinDC.
"The AI race will be won by companies that manufacture, deploy, and improve infrastructure with speed, scale and sovereignty."
— Dan Wright, CEO, Armada
Implications: Modular, edge-sited AI compute is becoming an investable subcategory alongside the GPU-cloud goliaths. Expect more strategic capital from industrial OEMs (HVAC, power, defense primes) entering infra-startup cap tables.
Unframe raises $50M Series B after $100M TCV and 400% NRR in year one
Highland Europe leads; LLM-agnostic enterprise AI delivery with outcome-based pricing — only pay after live trial proves out.
Unframe raises $50M Series B after $100M TCV and 400% NRR in year one
Highland Europe leads; LLM-agnostic enterprise AI delivery with outcome-based pricing — only pay after live trial proves out.
- $50M Series B; cumulative funding now $100M.
- Founded 2024 by ex-Noname Security execs: Shay Levi (CEO), Larissa Schneider (COO), Adi Azarya (R&D).
- $100M TCV booked in under 12 months from April 2025 launch; 400% NRR across Fortune 500 customer base.
- Existing investors Bessemer, Craft, TLV, Third Point, Cerca, Vintage all followed on.
"Every enterprise we speak with has a backlog of high-impact AI use cases and almost nothing in production. We built Unframe to close the gap between ambition and execution."
— Shay Levi, CEO, Unframe
Implications: Outcome-based pricing flips the usual enterprise pilot script. If this scales, expect copycats in consulting-heavy verticals and pressure on OpenAI's newly announced Deployment Company.
Nourish raises $100M Series C for AI-native metabolic health clinic
Menlo Ventures leads; 10,000+ dietitians + AI agents covered for 200M+ lives, documented A1C and LDL deltas at scale.
Nourish raises $100M Series C for AI-native metabolic health clinic
Menlo Ventures leads; 10,000+ dietitians + AI agents covered for 200M+ lives, documented A1C and LDL deltas at scale.
- $100M Series C; total raised now $215M. Lead: Menlo Ventures (J.P. Sanday joins board).
- Participation: Thrive Capital, Index Ventures, J.P. Morgan Growth Equity Partners, Maverick, YC, BoxGroup, Atomico, Daybreak, Operator Partners.
- Network: 10,000+ Registered Dietitians; 200M+ covered lives across all 50 states.
- Outcomes: 8% body weight reduction, 1.3-pt A1C drop, 31-pt LDL drop, 23-pt systolic BP drop · $2,000+ savings per patient/year for health plans.
Implications: The "human expert + AI agent" hybrid is becoming the dominant clinical-AI archetype for regulated, insurance-paid care.
Radar hits unicorn status with $170M Series B for AI retail inventory
Ceiling sensors + RFID + computer vision cut BOPIS cancellation rates from 25% to 3% in deployed stores.
Radar hits unicorn status with $170M Series B for AI retail inventory
Ceiling sensors + RFID + computer vision cut BOPIS cancellation rates from 25% to 3% in deployed stores.
- $170M Series B at a $1B valuation — new unicorn.
- Co-leads: Gideon Strategic Partners, Nimble Partners; participation from Align Ventures.
- Stack: ceiling-mounted sensors plus RFID, fused with computer vision.
- "Buy online, pick-up-in-store" cancellation rate cut from 25% → 3% in deployed stores.