Andrej Karpathy joins Anthropic's pre-training team to use Claude to accelerate Claude
The OpenAI co-founder and former Tesla AI lead lands at Anthropic's core training group — explicitly to build a team that uses Claude itself to compress pretraining research cycles.
Andrej Karpathy joins Anthropic's pre-training team to use Claude to accelerate Claude
The OpenAI co-founder and former Tesla AI lead lands at Anthropic's core training group — explicitly to build a team that uses Claude itself to compress pretraining research cycles.
- Karpathy starts this week on the pre-training team, the group running the massive training runs that produce Claude's core capabilities.
- His remit: launch a new sub-team focused on Claude-using-Claude to accelerate research — automating parts of the AI R&D loop.
- Announcement on X: "I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative."
- Latest in a string of senior hires; pairs with Anthropic's growing reputation as a magnet for technical talent post-OpenAI.
"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."
— Andrej Karpathy, on X
Implications: The "AI to build AI" loop is now an explicit org structure inside Anthropic. If Karpathy's group ships a working Claude-accelerated pretraining pipeline, the iteration cadence gap between Anthropic and competitors widens — and OpenAI's brain drain narrative gets another data point.
Cursor Composer 2.5 matches Opus 4.7 / GPT-5.5 at one-tenth the cost — built on Kimi K2.5
79.8% SWE-Bench Multilingual, $0.50/M input. The wedge: 25× more synthetic training tasks plus targeted RL on top of Moonshot's open-source checkpoint.
Cursor Composer 2.5 matches Opus 4.7 / GPT-5.5 at one-tenth the cost — built on Kimi K2.5
79.8% SWE-Bench Multilingual, $0.50/M input. The wedge: 25× more synthetic training tasks plus targeted RL on top of Moonshot's open-source checkpoint.
- Launched May 18; matches Claude Opus 4.7 and GPT-5.5 on SWE-Bench Multilingual (79.8%) and CursorBench v3.1 (63.2%).
- Built on Moonshot AI's open-source Kimi K2.5 checkpoint, post-trained with 25× more synthetic tasks and a new targeted reinforcement-learning technique.
- Pricing $0.50/M input, $2.50/M output — Opus 4.7 sits at $5/$25.
- Cursor 3 "Glass" already integrates Composer 2.5 into the Agents Window for parallel multi-agent flows.
Implications: "Open-weight checkpoint plus heavy domain post-training" is now a credible recipe for matching closed frontier on coding. Every IDE that doesn't ship its own model in 2026 is going to look expensive.
Anthropic expands Project Glasswing — Claude Security in public beta for vuln triage
Cyber-verified teams get Claude scanning codebases, triaging vulnerabilities and generating fixes; an explicit lane for open-source defenders.
Anthropic expands Project Glasswing — Claude Security in public beta for vuln triage
Cyber-verified teams get Claude scanning codebases, triaging vulnerabilities and generating fixes; an explicit lane for open-source defenders.
- Claude Security goes from limited preview to public beta for verified security teams.
- Use cases: codebase scanning, vulnerability triage, automatic fix generation, sharing fixes back to open-source maintainers.
- Pairs with the self-hosted sandboxes + MCP tunnels ship last week — keeps source and findings inside customer infra.
Implications: Anthropic is doubling down on regulated verticals where its safety story is a sales advantage rather than a friction point. Watch which security vendors choose to embed Claude vs. building in-house.
Gemini 3.5 Flash ships everywhere — Google bets the next AI wave on agents, not chatbots
Frontier-level intelligence with Flash speed, available the same day in Gemini app, AI Mode in Search, the API, Antigravity 2.0 and Android Studio.
Gemini 3.5 Flash ships everywhere — Google bets the next AI wave on agents, not chatbots
Frontier-level intelligence with Flash speed, available the same day in Gemini app, AI Mode in Search, the API, Antigravity 2.0 and Android Studio.
- Outperforms Gemini 3.1 Pro on coding and agentic benchmarks; leads multimodal understanding (84.2% CharXiv Reasoning).
- Ships day-one to Gemini API, AI Studio, Android Studio, Gemini Enterprise + the consumer Gemini app and AI Mode in Search.
- Paired with Antigravity 2.0 — Google's agent-first dev platform, now a standalone desktop app for orchestrating parallel cohorts of agents.
- Google claims aggregate enterprise savings could top $1B/year vs. prior frontier models.
Implications: The price/perf bar for "good enough for production agents" just dropped again. Anthropic and OpenAI now have to either match Flash on speed-per-dollar or differentiate on something the agent ecosystem actually pays for — sandbox security, integrations, vertical wins.
Pope Leo XIV launches the first papal AI encyclical alongside Anthropic's Chris Olah on May 25
"Magnifica Humanitas" lands 135 years to the day after Rerum Novarum — the Vatican explicitly framing AI as the new "social question."
Pope Leo XIV launches the first papal AI encyclical alongside Anthropic's Chris Olah on May 25
"Magnifica Humanitas" lands 135 years to the day after Rerum Novarum — the Vatican explicitly framing AI as the new "social question."
- Encyclical "Magnifica Humanitas" published May 25, 2026 — Pope Leo XIV's first major teaching document.
- Signed May 15 — 135 years to the day after Leo XIII's "Rerum Novarum" on industrial-era workers' rights.
- Presented at the Vatican's Synod Hall alongside Christopher Olah (head of interpretability research, Anthropic).
- Context: Pentagon recently labeled Anthropic a "supply chain risk" — the first such designation against a US AI company — over the lab's refusal to loosen autonomous-weapons safeguards.
Implications: Anthropic gets a global moral platform that compounds with its safety-first commercial narrative — a counterweight to losing the Pentagon contract. Watch which other AI labs publicly engage with the encyclical's framing and which stay silent.
California jury tosses Musk's lawsuit against OpenAI — case beyond the statute of limitations
Nine-member advisory jury needed 90 minutes; Judge Gonzalez Rogers dismissed; Musk's team announced they will appeal.
California jury tosses Musk's lawsuit against OpenAI — case beyond the statute of limitations
Nine-member advisory jury needed 90 minutes; Judge Gonzalez Rogers dismissed; Musk's team announced they will appeal.
- Verdict on May 18 — jury found Musk's 2024 filing beyond statute of limitations.
- Musk sued Altman, Brockman and OpenAI in Feb 2024, alleging they "stole a charity" by shifting to a for-profit structure.
- Marc Toberoff (Musk's lead attorney) says they will appeal.
- Removes a key legal overhang from OpenAI's restructuring path.
Implications: OpenAI clears the biggest founder-vs-board legal cloud over its capped-profit corporate structure. The appeal is pending, but for now investors price the for-profit conversion as legally durable.
Schaeffler signs for thousands of wheeled humanoids — first units live in Germany by year-end
London startup Humanoid (founded 2024) lands one of the largest humanoid rollouts on record: phased deployment from Dec 2026, scaling to 1,000+ robots and 1M+ actuators by 2032.
Schaeffler signs for thousands of wheeled humanoids — first units live in Germany by year-end
London startup Humanoid (founded 2024) lands one of the largest humanoid rollouts on record: phased deployment from Dec 2026, scaling to 1,000+ robots and 1M+ actuators by 2032.
- Binding phased deployment agreement between Humanoid (London, founded by Artem Sokolov in 2024) and Schaeffler (Germany).
- Phase 1: Dec 2026 – Jun 2027 at Herzogenaurach (box-handling) and Schweinfurt (integration + real-world validation).
- Schaeffler becomes Humanoid's preferred actuator supplier for five years; deal includes 1M+ actuators.
- One of the largest disclosed humanoid commitments to date — and the youngest company to land a deal at this scale.
Implications: Wheeled humanoids — cheaper, safer, faster to certify than bipedal — just got a marquee industrial reference. Expect Schaeffler-style framework deals to multiply in European auto / industrial supply chains through 2027.
JAL starts a two-year humanoid trial at Haneda — Unitree G1s handle baggage on the tarmac
Japan's labour-shortage answer: $15,400-per-unit Unitree G1s loading cargo and cabin-cleaning across a 60M-passenger hub.
JAL starts a two-year humanoid trial at Haneda — Unitree G1s handle baggage on the tarmac
Japan's labour-shortage answer: $15,400-per-unit Unitree G1s loading cargo and cabin-cleaning across a 60M-passenger hub.
- JAL × GMO AI & Robotics deploying two Unitree G1 units at Haneda Airport, two-year trial.
- G1 specs: 130 cm, 35 kg, 2 m/s, ~2 hr battery — small, cheap, low-stakes.
- Tasks: baggage / container handling on the tarmac, cabin cleaning.
- Per-unit cost ~$15,400 — cheap enough that the airport can experiment without a multi-year ROI gate.
Implications: Service-sector humanoid rollouts in labour-short markets (Japan, Korea, EU) likely scale faster than manufacturing pilots — lower task complexity, public-facing PR value, and cheap hardware.
Pentagon hands Eric Schmidt's Perennial Autonomy a $500M counter-drone IDIQ
JIATF-401 picks the company behind Merops (4,000+ Russian-drone intercepts in Ukraine) for enterprise-wide counter-UAS ops — Merops, Bumblebee, Hornet.
Pentagon hands Eric Schmidt's Perennial Autonomy a $500M counter-drone IDIQ
JIATF-401 picks the company behind Merops (4,000+ Russian-drone intercepts in Ukraine) for enterprise-wide counter-UAS ops — Merops, Bumblebee, Hornet.
- Joint Interagency Task Force 401 (JIATF-401) awards a 3-year IDIQ to Perennial Autonomy for AI-enabled counter-UAS systems.
- Product line: Merops interceptors, Bumblebee quadcopters, Hornet mid-range strike drones — already in use across US Central Command.
- Company is the rebrand of Schmidt's earlier "White Stork" / "Project Eagle" defense lab; Merops has 4,000+ combat intercepts in Ukraine since mid-2024.
- The Army bought 13,000 Merops units in the first days of the conflict at ~$15k each — Driscoll told lawmakers in April.
Implications: The Pentagon is wiring counter-drone as enterprise procurement, not bespoke. Combat-proven kit + IDIQ structure = the fastest possible path from startup demo to fielded systems. Expect more rebadged Ukraine-tested platforms through FY27.
DAWG funding plan still framing the FY27 conversation — $54.6B vs $226M today
A 100× expansion of autonomous-warfare spend, with $53.6B routed through the defense reconciliation package instead of base appropriations.
DAWG funding plan still framing the FY27 conversation — $54.6B vs $226M today
A 100× expansion of autonomous-warfare spend, with $53.6B routed through the defense reconciliation package instead of base appropriations.
- Defense Autonomous Working Group asks $54.6B for FY27 vs. $226M current allocation.
- Only $1B sits in the base budget; $53.6B comes via the proposed $350B reconciliation package.
- Joint Chiefs Chairman Caine: autonomous weapons are "a key and essential part of everything we do."
- Petraeus & Flanagan call it "the largest single commitment to autonomous warfare in history."
Implications: The Perennial Autonomy IDIQ is the first big disbursement under the DAWG narrative. Watch which winners follow — Shield AI, Anduril, Palantir are best positioned in the slipstream.
Starship V3 takes flight — first launch nails re-entry and Indian Ocean splashdown
After Thursday's hydraulic-pin scrub, Ship 39 cleared the tower on Friday at 6:30 PM EDT, deployed 22 mock satellites and survived re-entry with no heat-shield burnthrough. Booster failed the boost-back and crashed into the Gulf as expected.
Starship V3 takes flight — first launch nails re-entry and Indian Ocean splashdown
After Thursday's hydraulic-pin scrub, Ship 39 cleared the tower on Friday at 6:30 PM EDT, deployed 22 mock satellites and survived re-entry with no heat-shield burnthrough. Booster failed the boost-back and crashed into the Gulf as expected.
- Liftoff at 6:30 PM EDT (22:30 UTC) on May 22 from Starbase, Texas — Flight 12, debut of the scaled-up V3.
- Upper stage deployed 22 mock Starlinks; two carried external cameras streaming heat-shield footage on re-entry.
- Re-entry was clean — no heat-shield burnthrough, in stark contrast to earlier V2 test flights.
- Super Heavy booster failed its boost-back burn and crashed into the Gulf of Mexico (not planned for recovery on this flight).
- One upper-stage engine didn't ignite, putting the trajectory "within bounds" but off the originally targeted line.
- Soft Indian Ocean splashdown, expected tip-over and explosion — same end state as planned.
"The upper stage achieving its primary objectives despite the booster's difficulties — a dramatic balancing act of wins and losses."
— CNN coverage, Flight 12
Implications: SpaceX gets the milestone it needed ahead of an expected IPO: V3 cleared the pad, survived re-entry, and proved the new heat-shield architecture works. Booster boost-back still needs work, but that's the next flight, not a blocker. Artemis and commercial Starship manifests re-anchor.
Bret Taylor's Sierra closes $950M Series E at $15.8B — triples its valuation in 18 months
Tiger Global and GV lead; Sierra now has 40% of the Fortune 50 as customers and $150M+ ARR. Total capital past $1B.
Bret Taylor's Sierra closes $950M Series E at $15.8B — triples its valuation in 18 months
Tiger Global and GV lead; Sierra now has 40% of the Fortune 50 as customers and $150M+ ARR. Total capital past $1B.
- Round led by Tiger Global and Google's GV, with Benchmark, Sequoia, Greenoaks participating.
- Valuation tripled from $4.5B 18 months ago; total capital across rounds now over $1B.
- Sierra claims 40% of the Fortune 50 as customers; agents handle billions of interactions across mortgage, insurance, customer service.
- Bret Taylor also chairs OpenAI's board; co-founded Sierra with ex-Google Labs lead Clay Bavor.
Implications: Enterprise AI agents is becoming a "few clear winners" race rather than a long-tail platform — Sierra is increasingly the default for Fortune 500 buyers. Decagon, Cresta, Parloa now have a much steeper hill on outbound sales.
Exa closes $250M Series C at $2.2B — AI-native web search infra for agents gets fully funded
The "Google for agents" thesis now backed by serious capital — Exa joins Parallel Web Systems as the two well-funded contenders for agent-search infra.
Exa closes $250M Series C at $2.2B — AI-native web search infra for agents gets fully funded
The "Google for agents" thesis now backed by serious capital — Exa joins Parallel Web Systems as the two well-funded contenders for agent-search infra.
- Series C announced May 20: $250M at $2.2B post-money valuation.
- Direct competitor to Parallel Web Systems ($230M total, $2B post-money).
- Sells web search + retrieval APIs purpose-built for AI agents, not human users.
- Trend: foundation-model labs no longer treat web search as something to build in-house — they buy the layer.
Implications: Agent-grade web search is now a real category with two well-funded leaders. Expect M&A pressure on smaller agent-tooling shops (Browserbase, Apify, Brightdata) before Q4.
Nvidia Q1 FY27 prints $81.6B but the stock dips — $80B buyback can't move the needle
EPS $1.87 vs $1.78 expected, Data Center +92% YoY, dividend bump and $80B buyback — and shares still close down 0.9% Friday because the bar was pre-priced.
Nvidia Q1 FY27 prints $81.6B but the stock dips — $80B buyback can't move the needle
EPS $1.87 vs $1.78 expected, Data Center +92% YoY, dividend bump and $80B buyback — and shares still close down 0.9% Friday because the bar was pre-priced.
- Revenue $81.6B vs. $79.2B consensus; net income $58.3B vs. $42.9B expected; EPS $1.87 diluted.
- $80B stock buyback authorized; dividend payout edged up 1¢ — Nvidia plans to return ~half of free cash flow to shareholders in 2026.
- Stock dipped 0.9% on the print; market wanted bigger forward guidance.
- PHLX Semi Index +70% YTD in 2026; global semi revenue forecast to hit $1.32T (+64% YoY).
Implications: Even Nvidia's best print can't beat the expectations curve anymore. AMD's MI450 deals with OpenAI and Meta plus AMD's +114% YTD show that the multipolar AI-silicon market is now real — software vendors planning for "Nvidia only" need to revise.
Wall Street's AI-chip love rotates from Nvidia to Intel, AMD, Micron
PHLX Semi +70% YTD; AMD +114% on MI450 wins; CNBC tracks the rotation from sole-leader Nvidia to a multi-vendor AI silicon basket.
Wall Street's AI-chip love rotates from Nvidia to Intel, AMD, Micron
PHLX Semi +70% YTD; AMD +114% on MI450 wins; CNBC tracks the rotation from sole-leader Nvidia to a multi-vendor AI silicon basket.
- AMD shares up 114% YTD 2026; OpenAI and Meta both deploying MI450 GPUs in H2 2026.
- Intel and Micron riding the same rotation — Wall Street narrative shifts to "broader silicon ecosystem."
- Global semi revenue forecast: +64% to $1.32T in 2026.
- Big-tech AI capex on track to clear $700B in 2026 vs. ~$400B in 2025.
Implications: "Multi-vendor by default" replaces "Nvidia by default" in 2026 procurement. For software vendors building on accelerators, cross-platform support is now table stakes, not a nice-to-have.