AI News Weekly Summary — May 18–24, 2026 | Google I/O 2026, the Erdős Conjecture, Meta and 8,000 Layoffs

This AI news weekly May 18–24, 2026 covers one of the densest weeks of the year. The loudest story was Google I/O 2026 — the launch of Gemini 3.5 Flash, the multimodal Gemini Omni and a 24/7 agent, Gemini Spark. Right next to it: an internal OpenAI model that, the company says, autonomously disproved an 80-year-old Erdős conjecture.

There was plenty around money and politics too. Meta defended training AI on its employees’ data while announcing 8,000 layoffs, Trump pulled an executive order on pre-release model testing at the last minute, and Anthropic reported its first quarterly operating profit and IPO preparations. (Context: previous week, May 11–17).

1. Google I/O 2026 — Gemini 3.5 Flash, Gemini Omni, the Gemini Spark Agent, a New $100 Ultra Tier

On May 19, Google delivered a very long AI keynote — nearly two hours across Gemini, Search, YouTube, Gmail and hardware. Sundar Pichai said the Gemini app now has over 900 million monthly users (2x year-over-year) and that Google processes 9.7 trillion tokens a month. Demis Hassabis, on stage: “AGI is just a few years away.”

The headline models were Gemini 3.5 Flash (available now, powering Search AI Mode and Antigravity 2.0) and Gemini Omni — a unified text/image/video model fusing Nano Banana and Veo. Among the products: Gemini Spark — a 24/7 agent running on cloud VMs even when your laptop is closed (launching next week for Ultra) — plus Ask YouTube, Universal Cart and “Live” versions for Gmail and Docs. On pricing, Google adds a new $100/month AI Ultra tier, while the previous top plan drops from $250 to $200; daily prompt limits give way to a compute-based model refreshing every five hours.

Google keeps shifting from “we help you search” toward “agents do it for you.” A new, cheaper $100 Ultra tier reads as a strong pricing move — pressure on $200 ChatGPT Pro and on Claude.

2. Anthropic — Round in Progress: $30B at a $900B+ Valuation

Per Bloomberg, Anthropic’s round — at least $30B at a $900B+ valuation — is set to close by the end of May, co-led by Sequoia, Dragoneer, Greenoaks and Altimeter; no final term sheet was signed as of May 16.

The scale is striking: in February 2026 the company was valued at $380B. CEO Dario Amodei says the capital is for compute — the Amazon Web Services and Google Cloud commitments coming online through 2027. This is infrastructure-scale capital, not growth-stage fundraising.

3. An OpenAI Model Disproves an Erdős Conjecture — 80-Year-Old Problem Solved

On May 20, OpenAI announced that one of its internal reasoning models autonomously disproved the Erdős unit distance conjecture (posed in 1946). Per OpenAI, a 125-page proof establishes an infinite family of planar point configurations that beat the grid arrangements long believed optimal; the model reached for algebraic number theory (the Golod–Shafarevich criterion).

Mathematicians reviewed the work: Fields medalist Tim Gowers called it “a milestone in AI mathematics,” and Princeton’s Will Sawin quantified the best configurations as scaling like n^(1+δ), δ ≥ 0.014.

By OpenAI’s account, this is one of the first times AI has autonomously solved a prominent open problem at the heart of a field of mathematics — the model wasn’t trained for it or guided step-by-step. It’s exactly the kind of progress Jack Clark forecast at Oxford this week.

4. Elon Musk Loses His OpenAI Lawsuit — Unanimous Verdict in Under Two Hours

On May 19, a federal jury in Oakland unanimously rejected all of Elon Musk’s claims against OpenAI and Sam Altman. The jury deliberated for less than two hours after eleven days of trial. The verdict: claims barred by the statute of limitations — Musk filed too late.

Musk co-founded OpenAI in 2015, left the board in 2018, and argued the for-profit conversion abandoned the nonprofit mission. OpenAI countered there was never a binding nonprofit commitment and that the suit was tactical. In the background: xAI has dissolved as an independent company, folding into SpaceX as the SpaceXAI division. Altman’s reply on X: “Thank you.”

5. Meta Trained AI on Employees’ Data — Then Fired 8,000 of Them the Same Day

On May 19, leaked audio from an April 30 Meta all-hands surfaced, in which Mark Zuckerberg defends the Model Capability Initiative (MCI) — tracking employee activity across Gmail, Google Chat, the Metamate assistant and VS Code to train Meta’s models on “how really smart people work.” The same day, roughly 8,000 employees got layoff notices.

Zuckerberg insisted the data is anonymous and not used for performance review or surveillance. Employees organized an internal protest and petition; “train your replacement culture” and “Black Mirror episode” spread online. Meta is spending $115–135B in 2026 capex — nearly double the prior year.

The anonymity claim is probably technically true. The ethics question is more about whether it was appropriate to gather people’s work patterns without explicit informed consent, for a purpose tied to replacing their roles. It’s one of the main reasons the political backlash against AI is growing.

6. Trump Pulls His AI Safety Executive Order — After Talks with Musk, Zuckerberg and Sacks

On May 21, hours before the signing ceremony, the White House scrapped an executive order establishing a voluntary 90-day pre-launch model review (with NSA involvement in classified testing). Invitations had gone out. Per Axios sources, the main reason was simple: Trump “just hates regulation,” and former AI czar David Sacks deemed it “unnecessary.”

By the same accounts, between Wednesday night and Thursday morning, Sacks, Elon Musk and Mark Zuckerberg spoke with the president directly, warning the review would slow AI development. Trump, publicly: “I think it gets in the way — 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 order was drafted partly in response to Claude Mythos discovering zero-days at scale — a concrete security risk. The administration chose the “don’t regulate” frame anyway. You can read it as a sign that informal CEO access to the president weighs more than weeks of interagency work.

7. Anthropic Posts Its First Operating Profit, OpenAI Files an S-1, a $45B SpaceX Deal

Around May 21–22, three IPO-adjacent threads converged. Anthropic revealed it is tracking to its first-ever quarterly operating profit, projecting $10.9B in Q2 2026 revenue (up 130% from $4.8B in Q1). OpenAI is preparing a confidential S-1 with the SEC. And SpaceX’s prospectus disclosed that Anthropic pays SpaceX $1.25B per month through May 2029 for GPU compute — a $45B total contract.

Frontier AI firms are assumed to be permanently loss-making, so Anthropic turning operationally profitable while still training frontier models reshapes the IPO narrative for both labs.

8. Cursor Composer 2.5 — A Coding Model Built on Kimi K2.5, Trained on Colossus 2

Cursor shipped Composer 2.5 — built on Kimi K2.5 and trained on 25x more synthetic coding tasks than its predecessor. Per The Decoder, it matches Claude Opus 4.7 and GPT-5.5 on coding benchmarks at a far lower price, and is tuned for long agentic tasks (instruction adherence across many files, reduced context drift).

Elon Musk noted under the launch that it was “partially trained on Colossus 2” — SpaceXAI’s supercomputer. Anthropic had earlier secured Colossus 1 for Claude Code; now Kimi K2.5 and Composer 2.5 use Colossus 2.

9. GitHub Hacked — 3,800 Internal Repos via a Poisoned Nx Console Extension

On May 20, GitHub confirmed ~3,800 internal repositories were exfiltrated. The vector: a trojanized Nx Console VS Code extension (2.2 million installs), live on the Visual Studio Marketplace for 18 minutes on May 18. That was enough to silently pull a credential stealer — 1Password vaults, Claude Code configs, npm and GitHub tokens, and AWS keys.

The group is tracked as TeamPCP (UNC6780). GitHub’s CISO said there’s no evidence of impact to customer data stored outside internal repos.

10. TeamPCP — A Supply-Chain Worm: OpenAI, Mistral, the European Commission, 170+ npm Packages

The GitHub breach is one stop on a longer campaign. The “Mini Shai-Hulud” worm began May 11 by compromising TanStack’s router ecosystem and spread to 170+ npm packages and two PyPI packages (CVE-2026-45321, CVSS 9.6).

Confirmed victims beyond GitHub: OpenAI (two employee devices; a macOS app-signing certificate revocation on June 12), Mistral AI (one device; a $25,000 Monero extortion demand) and the European Commission’s public site. Developers who had the Nx Console extension should rotate all credentials immediately.

11. Jack Clark at Oxford — a Nobel in 12 Months, 60% Odds of Self-Improving AI by 2028

On May 20, Anthropic co-founder Jack Clark delivered the Cosmos Lecture at Oxford. He predicted AI would help produce a Nobel-worthy discovery within 12 months, bipedal robots would start assisting tradespeople within two years, and fully AI-run companies would generate millions in revenue within 18 months.

The loudest claim: a 60%+ chance that by the end of 2028 an AI model will fully train its successor (recursive self-improvement). Clark said “intelligence explosion” is now in official Anthropic research documents, and acknowledged a “non-zero chance” AI could lead to catastrophe.

12. Andrej Karpathy Joins Anthropic

On May 21, Andrej Karpathy — OpenAI co-founder and former Tesla Autopilot lead — joined Anthropic to rebuild its pretraining research team, citing “the next few years at the LLM frontier as especially formative.”

It’s one of the highest-profile AI talent moves of 2026 — Karpathy chose Anthropic over OpenAI, Google and xAI. Context: Anthropic openly wants to use AI to accelerate AI research.

13. Gemini Connects to Adobe, Canva and CapCut — a Creative Studio in a Chat Window

Around I/O, Google announced three creative platforms integrating into the Gemini app: Adobe (the Firefly agent, reaching Photoshop, Premiere, Illustrator), Canva (“Magic Layers” — open a Gemini-generated image in Canva with editable layers) and CapCut (video and image editing from inside Gemini). The idea: generate in Gemini and hand assets straight to professional tools without switching apps.

The Canva integration is furthest along and live in beta for Gemini AI Ultra subscribers; Adobe and CapCut arrive in the coming weeks.

14. NextEra Buys Dominion for $67B — The Largest US Utility Merger Ever

NextEra Energy announced a $67B acquisition of Dominion Energy — the largest US utility merger ever — citing AI-driven power demand as the primary rationale. AI data centers are projected to consume 15–25% of US electricity by 2030, and the current grid can’t support that.

NextEra, which runs North America’s largest renewables portfolio, is buying Dominion to expand generation and transmission for hyperscale workloads. It’s another sign that power availability, not model capability, can be the main constraint on scale today.

15. The Vatican Previews Its First AI Encyclical — With an Anthropic Co-Founder

On May 18, the Vatican announced that Pope Leo XIV will present his first encyclical on May 25 — Magnifica Humanitas (“Magnificent Humanity”) — alongside Christopher Olah, co-founder of Anthropic. It centers on “the protection of the human person in the age of artificial intelligence” and was signed on May 15, exactly 135 years after Leo XIII’s Rerum Novarum on workers’ rights.

The parallel is deliberate: Leo XIV frames AI as the Industrial Revolution of our era. Olah’s presence — an interpretability researcher — signals the Church’s interest in the gap between AI capability and AI transparency.

16. Anthropic Acquires Stainless — Better SDKs for the Claude API

On May 18, Anthropic confirmed it acquired Stainless, a startup building high-quality software development kits (SDKs) for API products. Stainless built the SDKs for the OpenAI API, Cloudflare and Merge.

It signals investment in the developer-experience layer around the Claude API — SDK quality meaningfully drives adoption. Claude’s Python, TypeScript, Java, Go and Ruby libraries should improve noticeably over the next 6–12 months.

17. CAISI — All Five Frontier Labs Now Under Pre-Release Review

The US Commerce Department’s CAISI finalized pre-release evaluation agreements with all five frontier labs: OpenAI, Anthropic, Google DeepMind, Microsoft and xAI. Every major US model release now passes a government evaluation before public launch.

The EU is in separate talks with Anthropic over access to the Mythos model, and the UK’s AI Safety Institute published updated red-teaming guidance ahead of I/O. For enterprises it’s a trust signal; for developers, a slightly longer window between announcement and API availability.

18. Other Important Signals of the Week

• Amazon Alexa+ — on-demand AI-generated podcasts with two synthetic co-hosts; a rival to NotebookLM.
• Igor Babuschkin (xAI co-founder) — raising up to $1B for a new AI research startup at up to a $5B valuation.
• Intuit — cutting ~3,000 jobs (~8%) while pivoting to AI agents; estimated savings over $500M annually.
• Chinese models (Kimi K2.6, DeepSeek V4, GLM-5.1, Qwen 3) — approaching 60% of OpenRouter traffic; China leads the open-weights tier today.
• Apple–Google — Thomas Kurian confirmed a more personalized, Gemini-powered Siri arriving later in 2026; full launch expected with iPhone 18.

Weekly Takeaway

The week of May 18–24, 2026 lined up around one axis: distribution and compute versus raw test scores. At I/O, Google didn’t unveil a Mythos-killer — it shipped fast, cheap Gemini 3.5 Flash, the Gemini Spark agent and a new $100 Ultra tier, betting on billions of users. Anthropic and OpenAI are playing a different game: profitability, IPOs and compute contracts ($45B for SpaceX, the $67B NextEra–Dominion merger), because energy and GPU availability can be the bottleneck now.

The second, less comfortable side of the week is the tension around adoption. An OpenAI model disproves an 80-year-old Erdős conjecture and Jack Clark puts 60% odds on self-improving AI by 2028 — while Meta trains AI on the data of staff it’s firing, the TeamPCP worm hits the software supply chain (GitHub, OpenAI, Mistral), and Trump, under pressure from CEOs, pulls his safety-testing order. Capability is moving fast; the institutions meant to govern it, clearly slower.

Źródła / Sources (wspólne dla obu wersji):

• Google I/O 2026: Everything Google announced at I/O 2026 — 9to5Google, Google I/O 2026 Roundup — MacRumors, With Gemini 3.5 Flash, Google bets on agents — TechCrunch, Google I/O 2026 collection — Google Blog, How Google plans to win the AI war — Axios
• Anthropic — runda, zysk, IPO: AI News May 18, 2026 — Build Fast with AI, AI News May 22, 2026 (Anthropic profitable, OpenAI files IPO) — Build Fast with AI
• OpenAI — hipoteza Erdősa: Model disproves discrete geometry conjecture — OpenAI, 80-year-old geometry mystery cracked — Interesting Engineering
• Musk vs OpenAI: Federal jury delivers verdict in Musk’s lawsuit — Fox Business
• Jack Clark / Oxford: Anthropic sells Claude’s promise while warning of dangers — Time, Jack Clark intelligence explosion — Axios
• Meta — MCI i zwolnienia: Leaked audio: Zuckerberg defends employee tracking — TechStory
• Regulacje / rozporządzenie: Why Trump’s AI executive order was pulled — Axios, Government to test AI models before launch — CNN
• Cyberbezpieczeństwo / TeamPCP: GitHub internal repositories breached via Nx Console — The Hacker News, GitHub confirms 3,800 repos stolen — VentureBeat
• Cursor / kreatywne integracje / Stainless: Cursor releases Composer 2.5 — Techmeme, Gemini to add Adobe, Canva, CapCut — eWeek, Anthropic acquires Stainless — Anthropic
• Watykan / encyklika: Pope Leo to present AI encyclical with Anthropic co-founder — National Catholic Reporter
• Erdős / matematyka, energetyka, pozostałe: AI News May 20, 2026 (I/O, Musk, NextEra) — Build Fast with AI, AI News May 23, 2026 (Erdős, Meta, GitHub) — Build Fast with AI