Weekly AI News Update: May 19–24, 2026
I do not think there has been a week in AI history quite like this one.
In seven days, Google held its biggest developer conference in a decade, launched a new model family, a 24/7 AI agent, and slashed its top subscription price by 60 percent. OpenAI filed its IPO paperwork and cleared a major legal cloud by winning Elon Musk's lawsuit. Anthropic reported its first quarterly profit, closed in on a $900 billion funding round, and hired the most beloved AI educator alive. Meta was caught on leaked audio training AI on employees before firing 8,000 of them. The Trump White House killed its own AI safety executive order after three tech billionaires called the president. A developer supply chain attack hit GitHub, OpenAI, and Mistral. OpenAI's AI model autonomously solved an 80-year-old math problem. And on Sunday, Pope Leo XIV published the first-ever papal encyclical on artificial intelligence.
This is the weekly AI news update for May 19 to 24, 2026. Every story that mattered, explained simply.
Google I/O 2026: The Biggest Keynote in a Decade
The Google I/O 2026 keynote on May 19 was the most consequential AI product event of the first half of 2026. CEO Sundar Pichai opened by sharing that the Gemini app now has 900 million monthly active users, that Google processes 9.7 trillion tokens every month, and that DeepMind CEO Demis Hassabis believes artificial general intelligence is "just a few years away." Google then proceeded to announce more products in two hours than most companies ship in two years.
The headline numbers first: Google AI Ultra dropped from $250 to $100 per month. That is a 60 percent price cut on the most capable AI subscription any major lab sells, and it now includes 5x higher usage limits, 20 terabytes of storage, YouTube Premium, and beta access to Gemini Spark. At the same time, Google removed daily prompt limits entirely, replacing them with a compute-based model that refreshes every five hours.
The products announced include Gemini 3.5 Flash (available immediately and powering Google Search), Gemini Omni (a unified text, image, and video generation model), Gemini Spark (a 24/7 AI agent), Ask YouTube (conversational search inside YouTube), Universal Cart (AI-powered shopping across Amazon, Shopify, and Walmart), and new smart glasses hardware with Samsung, Warby Parker, and Gentle Monster confirmed for fall 2026.
Google's thesis at I/O 2026 was not that it has the smartest model. The thesis was that it has AI everywhere: inside Search, inside Gmail, inside YouTube, inside your Android phone, and soon inside your glasses. Distribution, not benchmark scores, is how Google is playing this game.
Gemini 3.5 Flash and Gemini Omni: What Actually Launched
Gemini 3.5 Flash is Google's first model in the new 3.5 generation, and it is available right now. It costs $1.50 per million input tokens and $9 per million output tokens. Developer blogger Simon Willison immediately noted that this is three times more expensive than the Gemini 3 Flash Preview it replaces, and six times more expensive than Gemini 3.1 Flash-Lite. Google's counter: Gemini 3.5 Flash runs 12 times faster inside Antigravity (Google's AI development environment) than comparable frontier models, and it outperforms Gemini 3.1 Pro on coding and agentic benchmarks. The developer community is still processing whether the performance jump justifies the price jump.
Gemini Omni is the more interesting model. It is a unified model that generates text, images, and video from a single conversational prompt inside the Gemini app. You can describe a scene, upload a video, ask it to change the framing, add music, and overlay a caption, and it does all of it without you switching tools. Gemini Omni is live today for paid subscribers.
I will be honest: Gemini Omni is the first model from any lab that makes me think Google might actually win the creative AI market. OpenAI's Sora is a separate tool from ChatGPT. Anthropic has no video product. Google just shipped video generation, image editing, and text generation in the same chat window to 900 million users.
Gemini 3.5 Pro is in testing and expected next month. That is the model that will compete directly with Claude Opus 4.7 and GPT-5.5 on reasoning benchmarks.
Gemini Spark: Google's 24/7 AI Agent Is Here
Gemini Spark is the most ambitious consumer AI product launch of 2026. It is a 24/7 AI agent that runs on Google Cloud virtual machines even when your laptop is closed. You give it a task, and it works in the background across Gmail, Google Sheets, Google Docs, Google Drive, and Calendar, surfacing updates via a new Android notification layer called Android Halo.
The demo at I/O showed Spark planning a neighborhood block party: pulling RSVPs from Gmail, tracking who was bringing what, following up with non-responders, building a live tracker in Sheets, and generating a Slides deck with bounce house details and local rules pulled from a Drive file. Every action required user approval before it executed.
Spark launches next week for Google AI Ultra subscribers in the US ($100 per month). MCP support for third-party apps like Canva, Instacart, and OpenTable is coming in weeks. Chrome integration follows this summer.
To put this in context: OpenAI's Operator is still limited. Anthropic's agent platform is powerful but enterprise-focused. Gemini Spark is the first 24/7 AI agent that a consumer can actually turn on next Tuesday. If it delivers on the demo, Google wins the agent category.
OpenAI Filed for IPO on Friday
OpenAI filed its confidential draft registration statement with the Securities and Exchange Commission on May 22, 2026. This was confirmed by CNBC, Reuters, the Wall Street Journal, Bloomberg, and Axios within hours of each other. Goldman Sachs and Morgan Stanley are co-leading the deal, with JPMorgan Chase also involved. The target listing window is between Labor Day and Thanksgiving 2026, with September as the early target.
The company is currently valued at $852 billion by private investors. By the time it lists publicly, it could be valued at $1 trillion, which would make it the largest technology IPO in history.
Two important caveats. First, CEO Sam Altman told staff this week that "filing is different from being ready to go public." OpenAI's CFO Sarah Friar has also signaled internally that the company does not consider itself fully ready. Second, OpenAI is currently losing $1.22 for every $1 of revenue it generates. The Q1 2026 financials reportedly show $25 billion in annualized revenue against approximately $30 billion in annual spending. The S-1 will eventually disclose all of this publicly, and the market will have to decide whether the growth rate justifies the losses.
One cloud was cleared this week: a California jury unanimously ruled on May 19 that Elon Musk's lawsuit against OpenAI and Sam Altman was barred by the statute of limitations. The jury deliberated for less than two hours. Musk had argued OpenAI violated its nonprofit founding mission by converting to a for-profit structure. The verdict removes a significant legal overhang from OpenAI's IPO path.
Anthropic Hits $10.9 Billion Revenue and Its First-Ever Profit
While OpenAI dominates the IPO headlines, Anthropic's financial story this week is arguably more impressive. The company shared projections with investors showing $10.9 billion in revenue for Q2 2026, up 130 percent from $4.8 billion in Q1. That is not a typo. One hundred and thirty percent growth in a single quarter.
More significantly: Anthropic projects $559 million in operating income in Q2, which would be its first-ever quarterly operating profit. The compute cost ratio is improving fast. In Q1, Anthropic spent 71 cents on compute for every dollar of revenue. In Q2, that ratio is projected to fall to 56 cents. When you are paying SpaceX $1.25 billion per month for GPU access and still approaching profitability, your revenue trajectory is doing something extraordinary.
Bloomberg and the Financial Times confirmed on May 22 and 23 that Anthropic is on track to close a $30 billion funding round at a $900 billion valuation as soon as the week of May 26. The round is co-led by Sequoia Capital, Dragoneer Investment Group, Altimeter Capital, and Greenoaks Capital Partners, each contributing approximately $2 billion. Peter Thiel's Founders Fund and General Catalyst are also participating.
At $900 billion, Anthropic would surpass OpenAI's $852 billion private market valuation for the first time. The three-company AI IPO wave now has a clear timeline: SpaceX in June, OpenAI in September, Anthropic in October.
Andrej Karpathy Joins Anthropic: The Biggest Talent Story of 2026
On May 19, the same day as the Google I/O keynote, Andrej Karpathy announced he is joining Anthropic. He starts immediately on the pretraining team under Nick Joseph, where he will also build a new team focused on using Claude to accelerate pretraining research and experimentation.
Karpathy co-founded OpenAI in 2015, led Tesla's Autopilot and Full Self-Driving programs from 2017 to 2022, returned to OpenAI for one year in 2023, then left to start Eureka Labs, an AI education startup. He is arguably the most respected AI educator and researcher outside of a senior lab role alive right now.
His mandate at Anthropic is straightforward and consequential: use Claude to speed up the process of training Claude. This is AI-assisted AI research, the early version of what Jack Clark described this same week as "recursive self-improvement." Karpathy spent months before joining Anthropic experimenting with exactly this approach at Eureka Labs. Anthropic is now deploying that work at production scale.
His departure from OpenAI and arrival at Anthropic is the clearest signal yet that talent momentum has shifted between the two companies. Three of the four investors co-leading Anthropic's $900 billion round are former OpenAI backers. Now one of OpenAI's most admired co-founders is building against it.
Meta's Zuckerberg Caught Training AI on Employees Before Mass Layoffs
This is the story that made the most people viscerally angry this week.
A leaked audio recording from a Meta all-hands meeting on April 30 surfaced on May 19, the same day approximately 8,000 Meta employees received layoff notices. In the recording, Mark Zuckerberg describes a program called the Model Capability Initiative, which tracks employee activity across Gmail, Google Chat, the internal assistant Metamate, and VS Code to train Meta's AI models on "how smart people work."
"The AI models learn from watching really smart people do things," Zuckerberg says in the audio. He assured employees that no human was watching the feeds and that the data was not used for performance tracking. The assurances landed poorly. Employees organized internal protests the same morning their colleagues received termination emails. Social media flooded with the phrase "train your replacement culture."
The context that made it worse: Meta has committed more than $125 billion to AI infrastructure in 2026 alone. Roughly 7,000 additional workers were simultaneously reassigned internally as management layers were flattened in favor of smaller, AI-assisted teams. Meta framed the restructuring as building for the future. Workers framed it as being handed a shovel to dig their own replacement.
My read: Zuckerberg's privacy assurances are probably technically accurate. The data was almost certainly anonymized. The ethical problem is not whether it was anonymous. The problem is that no one asked employees whether they consented to having their work patterns harvested to build AI tools that would eliminate their roles. That is a consent failure regardless of the anonymization.
Trump Killed the AI Safety Executive Order
The White House AI executive order, which would have required AI companies to share frontier models with the government up to 90 days before public launch, was cancelled on May 21, 2026. Hours before the scheduled signing, invitations already sent.
Axios obtained the definitive explanation: the main reason the order was delayed was that Trump "just hates regulation," and former AI czar David Sacks "hated it" too. The order was described internally as "unnecessary" and "just something doomers wanted." Between Wednesday night and Thursday morning, Sacks, Elon Musk, and Mark Zuckerberg all spoke with Trump directly, outside the normal policy process.
Trump told reporters: "I didn't like certain aspects of it. I postponed it. I think it gets in the way of, you know, we're leading China, we're leading everybody, and I didn't want to do anything to get in the way of that lead."
The irony is real. The order was designed partly because of Anthropic's Claude Mythos model discovering zero-day vulnerabilities in legacy financial infrastructure at scale, which genuinely alarmed the national security community. The national security professionals who spent weeks building the framework had no comparable access to the president. Three tech CEOs with financial stakes in the AI landscape did. That is the actual story: informal CEO access to the president outweighed months of interagency security work on a critical infrastructure question.
GitHub and OpenAI Got Hacked via a VS Code Extension
On May 20, GitHub confirmed that approximately 3,800 internal repositories had been stolen by a threat actor group called TeamPCP. The attack vector was elegant and terrifying: a trojanized version of the Nx Console VS Code extension, which has 2.2 million installs and verified publisher status, was live on the Visual Studio Marketplace for exactly 18 minutes on May 18, between 12:30 PM and 12:48 PM UTC. Eighteen minutes was enough.
The malicious version ran silently on startup, harvesting GitHub tokens, AWS keys, npm tokens, 1Password vault contents, and Anthropic Claude Code configuration files from any developer machine that installed it during that window. TeamPCP used the stolen credentials to move through CI/CD pipelines and copy repositories.
OpenAI confirmed two employee devices were compromised in the same campaign, with internal source code repositories accessed. Mistral AI confirmed one device was hit and is facing a $25,000 extortion demand. The European Commission's public website was also a confirmed victim.
The broader TeamPCP campaign started May 11 with 170 npm packages compromised across the TanStack router ecosystem. GitHub was Wave 4. The attack never needed to breach a perimeter: it entered through the exact tools developers install and trust every day. OpenAI is revoking its macOS app signing certificate on June 12 as a direct result. If you use the Nx Console VS Code extension, rotate all your credentials now.
OpenAI Solved an 80-Year-Old Math Problem
On May 20, OpenAI announced that an internal general-purpose reasoning model autonomously disproved the Erdős unit distance conjecture, a problem in discrete geometry first posed by Hungarian mathematician Paul Erdős in 1946.
For nearly 80 years, mathematicians believed the optimal way to arrange points so that as many pairs as possible sit exactly one unit apart would look roughly like a square grid. OpenAI's model found an infinite family of configurations that beat the grid, using algebraic number theory, specifically a mathematical structure called infinite class field towers, to connect an elementary geometry question to deep number-theoretic tools. The construction was not obvious to mathematicians working in the field. The cross-domain leap is the most interesting part.
Princeton mathematician Will Sawin refined the result and quantified the improvement: the best configurations now scale as n to the power of 1.014, versus the square grid's approximately n to 1. Fields medalist Tim Gowers reviewed the work and called it "a milestone in AI mathematics." Noga Alon, a leading combinatorialist at Princeton, called it "an outstanding achievement."
What makes this different from AI winning a math competition or scoring on a benchmark: the model was not trained on this problem, did not retrieve an existing solution, and operated without step-by-step human guidance. It received the problem statement and produced a 125-page proof independently. That is the first time AI has autonomously solved a prominent open problem that is central to a field of mathematics, not just a timed test.
The Pope Published the World's First AI Encyclical
Pope Leo XIV published Magnifica Humanitas, Latin for "Magnificent Humanity," on May 25, 2026. The document, which the Vatican describes as addressing "the protection of the human person in the time of artificial intelligence," was signed on May 15, the 135th anniversary of Pope Leo XIII's Rerum Novarum, the foundational Catholic social teaching document that addressed labor rights during the Industrial Revolution.
The parallel is deliberate. Leo XIV is positioning AI as the defining social and moral challenge of our era in the same way industrialization was for his predecessor.
The presentation format itself made history. Rather than the standard Vatican press room release with a few officials, the encyclical was launched in the Synod Hall with two top cardinals, theologians from Durham and Edinburgh, and Christopher Olah, co-founder of Anthropic and lead of its interpretability research team. Olah was not invited because of his seniority at Anthropic. He was invited because interpretability research, understanding what is actually happening inside AI models at a mechanical level, is exactly the kind of transparency work the Vatican cares about.
The encyclical addresses AI mimicking human relationships and identity, AI's displacement of human creative work, the concentration of AI power among a few profit-driven companies, labor rights in the age of automation, and the ethics of autonomous weapons. The global Catholic population of approximately 1.4 billion makes this the largest single institutional statement on AI ethics ever published.
Frequently Asked Questions
Q: What was the biggest AI news story the week of May 19–24, 2026?
The week had several major stories, but Google I/O 2026 on May 19 was the anchor event: launching Gemini 3.5 Flash, Gemini Omni (unified text/image/video model), Gemini Spark (24/7 AI agent), Ask YouTube, and cutting the AI Ultra subscription from $250 to $100. The same week, OpenAI filed its confidential IPO paperwork, Anthropic projected its first quarterly operating profit, and the Pope published the first papal encyclical on AI.
Q: What is Gemini 3.5 Flash and how does it compare to Claude?
Gemini 3.5 Flash is Google's newest AI model, launched May 19, 2026. It costs $1.50 per million input tokens and $9 per million output tokens. It outperforms Gemini 3.1 Pro on coding and agentic benchmarks and runs 12x faster inside Google's Antigravity development environment. Compared to Claude Sonnet 4.6, which costs $3 per million input tokens, Gemini 3.5 Flash is cheaper but considered near-equal on many tasks. Gemini 3.5 Pro (coming next month) will compete directly with Claude Opus 4.7.
Q: Why did OpenAI file for an IPO in May 2026?
OpenAI filed a confidential draft registration statement with the SEC on May 22, 2026, targeting a public listing in September 2026. The filing followed the clearing of Elon Musk's lawsuit (unanimously rejected May 19), Anthropic's profitability announcement (which put competitive pressure on OpenAI's narrative), and SpaceX's public S-1 filing (which provided a market reference point). Goldman Sachs and Morgan Stanley are co-leading the deal at a valuation of $852 billion to $1 trillion.
Q: What is Anthropic's $900 billion valuation based on?
Anthropic is closing a $30 billion funding round at a pre-money valuation above $900 billion, led by Sequoia, Dragoneer, Altimeter, and Greenoaks. The valuation is anchored by Q2 2026 revenue projections of $10.9 billion (up 130% from Q1), the company's first quarterly operating profit ($559 million operating income), more than 1,000 customers spending $1 million or more annually, and a $45 billion compute contract with SpaceX. The round is expected to close the week of May 26.
Q: Did Zuckerberg really train Meta's AI on its employees?
Yes, according to a leaked audio recording from a Meta all-hands meeting on April 30, 2026. Zuckerberg described a program called the Model Capability Initiative that tracked employee activity across Gmail, Google Chat, an internal assistant called Metamate, and VS Code to train Meta's AI. He said the data was anonymized and not used for performance tracking. The recording surfaced May 19, the same day approximately 8,000 Meta employees received layoff notices.
Q: What is the Pope's AI encyclical Magnifica Humanitas?
Magnifica Humanitas is Pope Leo XIV's first encyclical, published May 25, 2026. It addresses artificial intelligence and human dignity, was signed on the 135th anniversary of Pope Leo XIII's labor-rights encyclical Rerum Novarum, and covers AI's impact on human relationships, creative work, labor, power concentration, and autonomous weapons. Anthropic co-founder Christopher Olah presented it alongside cardinals at the Vatican Synod Hall. With 1.4 billion Catholics globally, it is the largest institutional statement on AI ethics ever published.
Q: What happened with the White House AI executive order?
The White House AI executive order, which would have required AI companies to share frontier models with the US government up to 90 days before launch, was cancelled on May 21, 2026, hours before the scheduled signing. According to Axios, the cancellation followed direct calls to President Trump from Elon Musk, Mark Zuckerberg, and former AI czar David Sacks, who described the order as "unnecessary" and "just something doomers wanted." Trump told reporters he did not want to do anything that would interfere with the US lead in AI.
Q: What is the GitHub TeamPCP supply chain attack?
TeamPCP is a cybercrime group that compromised approximately 3,800 GitHub internal repositories in May 2026 through a trojanized version of the Nx Console VS Code extension, which was live on the Visual Studio Marketplace for 18 minutes on May 18. The attack harvested GitHub tokens, AWS keys, npm credentials, and Anthropic Claude Code configuration files from developer machines. OpenAI (two employee devices) and Mistral AI (one device, with a $25,000 extortion demand) were also confirmed victims. OpenAI is revoking its macOS app signing certificate on June 12 as a result.
A Final Thought
Here is what I keep coming back to: the most powerful institution in the world by total membership published a document this week saying AI is the defining challenge of our era, the way industrial labor was in 1891. At the same moment, an AI model disproved a math problem no human had solved in 80 years. Three companies are about to list publicly at a combined $3.7 trillion in market value. And the most important AI safety regulation attempt in the US this year was killed in a morning phone call by three tech billionaires.
All of that happened in one week. The pace is not slowing down.
Daily learning beats trying to catch up weekly. Five minutes a day compounds faster than you think.
References
Vatican News: Pope Leo XIV's first encyclical Magnifica Humanitas
Bloomberg: Anthropic to Close Over $30 Billion Round as Soon as Next Week
The Hacker News: GitHub Internal Repositories Breached via Malicious VS Code Extension
TechStory: Leaked Audio Reveals Zuckerberg Defending Employee Tracking to Feed Meta's AI
Interesting Engineering: 80-year-old geometry mystery cracked by OpenAI
