An 80-Year-Old Math Conjecture Fell, Karpathy Joined Anthropic, and Google Rewrote Its Entire Product Stack

The week of May 18th delivered some of the biggest personnel moves, scientific breakthroughs, and product drops we’ve seen — and a few warning signs nobody should ignore.

There are weeks where the big news trickles in, and weeks where it arrives in a flood. This was the second kind.

Andrej Karpathy — one of the most respected AI researchers alive — quietly walked away from his education startup and joined Anthropic. OpenAI announced that an internal model independently disproved a conjecture that had stood for 80 years. Google held its I/O event and reminded the industry that it still has more surface area than anyone. And California, home to 33 of the world’s top 50 AI companies, became the first U.S. state to formally ask: what happens to workers when this all plays out?

There’s a lot here. Let’s get into it.

AI Updates

Andrej Karpathy Just Joined Anthropic — and What He’s Building There Matters

If you follow AI seriously, you know Karpathy. Co-founder of OpenAI in 2015. Led Tesla’s Autopilot. Returned briefly to OAI, then left again in 2024 to build an AI education startup. One of the most credible voices in the entire field.

This week, he announced he’s joining Anthropic.

He’ll work under Nick Joseph on the pre-training team, but the more interesting detail is what he’s actually building: an internal effort to apply Claude to Anthropic’s own AI training pipeline. Essentially, using Claude to help build better versions of Claude. The self-improving model playbook, but now with one of the people who helped invent the field running the operation.

“The next few years at the frontier of LLMs will be especially formative,” he wrote on X. That quote alone tells you why he made the call.

This is a massive win for Anthropic — not just as a talent signal, but as a strategic one. The race to automate AI training itself is one of the defining competitions of the next 24 months.

🔗 Source

Google I/O Dropped Everything at Once — and It Was a Lot

Google held its flagship I/O event this week, and the headline isn’t any single product — it’s the sheer volume and ambition of what landed simultaneously.

Gemini Omni converts text, images, audio, and video inputs into video outputs. Gemini 3.5 Flash benchmarks near Opus 4.7 and GPT-5.5 at 4x the speed and half the cost. Gemini Spark is a new 24/7 personal agent running on Google Cloud virtual machines that takes agentic actions across Workspace, Chrome, email, and chat without you touching anything. And Google Search got what it called its biggest redesign in a generation — cross-modal inputs, generative UI, and around-the-clock info agents built in.

Then there’s the hardware: Google Intelligent Eyewear is officially coming, with audio-first AI smart glasses built with Samsung, Warby Parker, and Gentle Monster shipping this fall — Google’s first glasses push since Glass. A display-equipped Project Aura follows after.

The individual numbers on Flash don’t blow the competition away. But that’s not the point. Google’s play is a more agentic, multimodal Gemini woven into tools that hundreds of millions of people already live inside every day. No other lab can replicate that distribution.

🔗 Google I/O Keynote | Intelligent Eyewear

OpenAI’s Internal Model Just Disproved an 80-Year-Old Mathematical Conjecture

In 1946, Paul Erdős posed a problem about unit distances — how many same-length links you can draw between a set of points. A grid-based theory shaped the field for eight decades. This week, OpenAI announced that an internal general-purpose reasoning model found a proof that disproves it, drawing on algebraic number theory in a way nobody had tried before.

The proof has been verified by some of the most respected mathematicians alive, including Tim Gowers, Noga Alon, and Thomas Bloom. And crucially: this didn’t come from a math-specific system like DeepMind’s AlphaProof. It came from a general model that hasn’t even been released yet.

OAI’s Alex Wei framed the stakes clearly: “math is a leading indicator of what is to come.” A general-purpose AI making original contributions to a field it wasn’t designed for is the early shape of what researchers have been calling Level 4 AI — not just accelerating work, but genuinely creating new knowledge.

🔗 Source

California Just Became the First State to Formally Protect Workers from AI

Governor Newsom signed an executive order this week directing California state agencies to study and develop policies around AI-driven job displacement — one day after Meta laid off 8,000 employees citing AI efficiency gains.

The order sets concrete timelines: within 90 days, a public dashboard tracking AI’s job impact launches. Within 180 days, agencies will pitch WARN Act updates for faster layoff alerts. By October 15th, the state will review union AI negotiations, update workforce training, and explore redirecting AI revenue toward public benefit — including universal basic capital.

Over 70,000 U.S. jobs have already disappeared in 2026. California is home to 33 of the top 50 AI companies in the world. The state choosing now to formally ask what this means for workers isn’t political theater — it’s the first serious policy attempt to get ahead of a curve that has so far moved faster than any legislature has been willing to acknowledge.

🔗 Source

The Musk vs. OpenAI Trial Ended — But Not With the Answers Anyone Wanted

After three weeks of high-profile testimony, leaked texts, and billionaire appearances on the stand, Elon Musk’s $100B+ lawsuit against OpenAI, Sam Altman, Greg Brockman, and Microsoft was dismissed. The jury found unanimously that the case was filed too late — not a ruling on the substance.

Musk posted on X that it was “a calendar technicality” and that he’d appeal. OpenAI’s defense argued Musk himself backed a for-profit structure early on, pushed for personal control, and only sued after founding xAI in 2023. The Microsoft claim was also thrown out.

The frustrating part: for anyone hoping the trial would clarify who controls a nonprofit AI organization once billions of dollars enter the picture, this outcome settles nothing. The question is still wide open — and the legal architecture around AI governance remains unresolved.

🔗 Source

ArXiv Is Banning Researchers Who Let AI Hallucinate Into Their Papers

ArXiv announced a one-year submission ban for any researcher caught submitting papers with hallucinated references or other clear evidence of unchecked LLM output. After the ban expires, affected authors face a permanent additional restriction: all future submissions must first clear peer review at an outside venue.

Computer science chair Thomas Dietterich put it bluntly: papers showing authors didn’t verify LLM output cannot be trusted. Fabricated citations are already rising in biomedical literature. ArXiv’s enforcement model — moderator flags, section chair confirmation, author right to appeal — may become the template other institutions follow. The scientific record’s reliability is what’s at stake.

🔗 Source

Google Published Its AI Co-Scientist Research — And the Results Are Striking

Google published its Co-Scientist research in Nature this week, introducing Hypothesis Generation — a Gemini-powered tool that runs “idea tournaments” between competing research agents to surface novel hypotheses for biology labs.

The system draws from the same tournament logic as AlphaGo: agents propose, critique, and rank hypotheses before refining the top candidates. In a Stanford liver-fibrosis project, one Co-Scientist drug lead reportedly cut a scarring-related lab signal by 91% in testing. Google also launched Gemini for Science this week — pairing Co-Scientist with AlphaEvolve for discovery and NotebookLM for literature analysis.

This isn’t a chatbot for scientists. It’s a system targeting the scientific method itself, built on a stack — AlphaFold, specialized databases, years of biological models — that took billions and decades to assemble. Few labs can replicate the foundation.

🔗 Source

Other AI Updates Worth Your Attention This Week

• Anthropic formed a $200M partnership with the Gates Foundation to deploy Claude in vaccine screening, disease forecasting, and K-12 tutoring in developing nations.
• OpenAI launched personal finance inside ChatGPT, connecting via Plaid to 12,000+ institutions for real-time access to spending and investments — it can analyze but not yet move money. 🔗 Source
• Emergence AI’s virtual town simulation ran five identical worlds under different AI models. Claude logged zero crimes in 15 days. Grok had all agents dead by day 4. Gemini’s town literally caught fire after two agents fell in love. 🔗 Source
• Anthropic acquired Stainless, the startup behind its official SDKs and MCP tooling, pulling its developer library team in-house.
• Meta is laying off 8,000 employees and cancelling 6,000 open roles as part of its AI efficiency push.
• SpaceX’s IPO prospectus revealed Anthropic is paying $1.25B per month through 2029 for compute access across Colossus and Colossus II.
• OpenAI partnered with Malta to offer free ChatGPT Plus to every citizen completing a national AI literacy course — the first country-wide deal of its kind.
• Intuit is cutting 17% of its workforce, attributing the move to its AI-first strategic pivot.
• METR’s first Frontier Risk Report found top-lab agents can autonomously finish multi-week engineering work, but struggle on hard-to-verify tasks.
• A Gallup survey found 70% of Americans oppose data centers being built nearby — polling less popular than local nuclear power plants.

⚙️ Robotics, Hardware & Technology Updates

Meta and Anduril Are Building AR Glasses for Soldiers — and the Army Slot Is Wide Open

A year-long partnership between Meta and defense startup Anduril is now yielding a battlefield headset prototype that fuses AI, drones, and targeting directly into soldiers’ line of sight, MIT Technology Review reports.

Anduril holds a $159M Army prototyping contract to build AR glasses that attach to existing military helmets. Soldiers can use eye-tracking and voice commands to cue drone or artillery strikes from the headset. Anduril is also self-funding EagleEye — an integrated helmet-and-headset combo developed with Meta, expected to reach production after 2028 — which plugs into Anduril’s Lattice platform, the same software behind their $20B Army integration contract.

The context matters: Microsoft had a $22B contract for battlefield AR that was cancelled after the system failed to prove viable in field conditions. That slot is now open. Whoever delivers a headset soldiers will actually use gets to embed their AI stack into Pentagon procurement for the next decade.

🔗 Source

China Just Turned On the World’s First Underwater Data Center — Powered by Wind

Off the coast of Shanghai, China has begun full commercial operation of what it says is the world’s first offshore wind-powered underwater data center — a $226M facility sitting more than 30 feet beneath the East China Sea.

The 24 MW center draws 95% of its power from a surrounding wind farm and uses passive seawater cooling instead of industrial chillers, cutting electricity consumption by 22.8% and eliminating freshwater use entirely. It houses nearly 2,000 servers, including GPU clusters designed to handle AI workloads, and operates at a Power Usage Effectiveness (PUE) of 1.15 — a number that would be exceptional for any data center anywhere.

Microsoft’s Project Natick proved submerged servers can be up to 8x more reliable than land-based equivalents — then quietly shelved the program. China has now picked up where that research left off and gone commercial. The catch is the same one that stopped Natick: when hardware fails 30 feet underwater, fixing it is a different kind of problem.

🔗 Source

SpaceX’s IPO Is Designed So Nobody Can Touch Elon Musk — Ever

The Wall Street Journal got a look at SpaceX’s IPO structure this week, and the picture it paints is unlike anything Wall Street has processed before. Musk will retain roughly 85% of voting control through supervoting shares and a dual-class structure. Texas corporate law makes it nearly impossible for shareholders or the board to remove him or override major strategic decisions without his consent.

The offering is targeting a valuation around $1.5 trillion, funneling public capital into rockets, Starlink, and Musk’s orbital AI ambitions — while explicitly designing out the kind of shareholder fights that have plagued Tesla over pay packages and controversial acquisitions.

Investors get exposure to one of the most consequential companies in the world. They get almost no say in how it’s run. How Wall Street prices that trade-off could quietly set the template for every mega-unicorn IPO that follows.

🔗 Source

A Startup Just Launched an Uber-for-Humanoids Cleaning Service at $150 a Visit

San Francisco startup Gatsby is booking Unitree G1 humanoid robots to clean people’s homes through an app — no human cleaner present, flat rate of $150 per visit regardless of home size.

The startup is backed by NVIDIA’s Inception program and pitches itself as a “consumer distribution layer” for humanoids. The framing is deliberate: Gatsby isn’t selling robots, it’s selling the outcome, the same move Uber made with cars. In China, human-robot cleaning teams in Shenzhen are already advertising whole-home jobs for around $11, which gives a sense of where the cost floor eventually goes.

This is still a small San Francisco pilot. But it’s the first time the question “would you let a robot clean your house?” has come with a real booking link. The trust, liability, and reliability questions are enormous — and exactly what this kind of real-world test will start to answer.

🔗 Source

Airbnb Is Quietly Becoming the Super-App for Everything Travel

Airbnb rolled out a sweeping expansion this week, stacking airport pickups, luggage storage, car rentals, boutique hotel bookings, and Instacart-powered grocery delivery into a single app interface.

Luggage storage via Bounce is live across 175 cities. Airport rides through Welcome Pickups cover 160+ cities. Grocery delivery is live in 25+ U.S. cities. Car rentals and boutique hotels land this summer, with price-match guarantees and booking credits attached.

Airbnb has always made money on the nightly rate. Now it wants a slice of everything else around the stay — the ride there, what’s in the fridge when you arrive, how you get around. The strategic bet is that travel spend is undermonetized and the company that owns the relationship can capture significantly more of it.

🔗 Source

Other Robotics, Hardware & Tech Stories This Week

• Hyundai plans to deploy 25,000+ Atlas robots across its U.S. plants by 2028, ramping production to 30,000 units per year.
• Figure AI’s humanoids hit 8 straight days of live-streamed warehouse operations, autonomously sorting 230,000+ packages.
• Boston Dynamics’ Atlas was shown carrying a 100-pound mini-fridge using whole-body AI coordination — a deceptively hard manipulation task.
• China reportedly doubled J-20 stealth fighter production efficiency by converting a Chengdu plant into a fully robot-run “dark factory.”
• Dozens of empty Waymo robotaxis were spotted circling the same Atlanta cul-de-sac — 50 vehicles in a single hour.
• SpaceX scrubbed the first Starship V3 launch just before liftoff due to a ground-system issue.
• Oura filed IPO paperwork with the SEC, timing dependent on market conditions.
• Tesla’s FSD (Supervised) is now approved in the Netherlands and Lithuania, with more European countries queued.
• South Korea launched a $34M project to develop a sovereign Korean humanoid robot by 2030.
• Arizona State researchers found Phoenix-area data centers raise downwind neighborhood temperatures by up to 4°F — a land-use consequence barely on anyone’s policy radar yet.
• Apple is reportedly using slightly defective chips in cheaper popular devices, turning silicon rejects into a high-margin product line.

The Bigger Picture

The thread connecting this week isn’t any one story. It’s the widening gap between what AI is becoming and the systems built to manage it.

Karpathy at Anthropic. A model disproving 80-year-old math. Gemini agents setting a virtual town on fire. California becoming the first state to ask “what do we do about the jobs?” These aren’t separate headlines — they’re the same story told from different angles.

The technology is accelerating. The governance, infrastructure, and public trust needed to absorb it are all moving slower. The organizations that close that gap — not just ship faster — are the ones that define what comes next.

Which development from this week stood out most to you? Drop it in the comments — I’d genuinely like to know.

If this landed in your feed through someone else, subscribe to get it directly next week.