AI Briefing: May 20–22, 2026

Coverage window: May 20 – May 22, 2026 (48 hours)
Published: 2026-05-22 08:00 UTC
Sources: Twitter/X, GitHub Releases, arXiv, TechCrunch, Web Extract

🚨 Breaking (last 24h)

OpenAI Claims First AI-Autonomous Math Breakthrough

OpenAI announced that one of its general-purpose reasoning models has produced an original mathematical proof disproving the Erdős planar unit distance conjecture — an open problem in discrete geometry first posed by Paul Erdős in 1946. The company published companion remarks from mathematicians Noga Alon, Melanie Wood, and Thomas Bloom (who maintains the Erdős Problems website) in support of the disproof. This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.

Note: This is not OpenAI's first claim of this nature. Seven months ago, former VP Kevin Weil claimed GPT-5 solved 10 Erdős problems, but those solutions already existed in the literature. This time, the company published independent mathematician commentary alongside the announcement.

Anthropic on Track for First Profitable Quarter

Anthropic told investors it will more than double revenue to approximately $10.9 billion in Q2 2026 and deliver its first operating profit, according to the Wall Street Journal. The milestone would put Anthropic in an advantageous position relative to OpenAI. However, the company may not remain profitable throughout the year due to large compute costs, including its reported $1.25 billion per month deal with xAI for compute through May 2029.

OpenAI Barrels Toward September IPO

OpenAI CEO Sam Altman is reportedly targeting a September IPO, working with Goldman Sachs and Morgan Stanley. The company may file confidential IPO paperwork with regulators within days or weeks. This comes immediately after Elon Musk's lawsuit against OpenAI was dismissed on statute-of-limitations grounds.

TechCrunch: OpenAI barrels toward IPO that may happen in September

📊 Market Moves (last 48h)

Hark Raises $700M Series A at $6B Valuation

Hark, an AI lab building models and hardware for a universal AI personal assistant, raised $700 million in Series A funding at a $6 billion post-money valuation. The round was led by Parkway Venture Capital and included Nvidia, AMD Ventures, Intel Capital, Qualcomm Ventures, Salesforce Ventures, ARK Invest, Brookfield, Greycroft, Prime Movers Lab, and Tamarack Global. Founder and CEO Brett Adcock (also behind Figure.AI and Archer) launched Hark in late 2025 with $100 million of his own money. The company expects to release its first multimodal models this summer.

xAI Burned $6.4B in 2025 on $3.2B Revenue

SpaceX's IPO filing revealed that xAI lost $6.4 billion in 2025 on $3.2 billion in revenue. Revenue came largely from "AI solutions and infrastructure" totaling $465 million (including $365 million in X/Grok subscriptions and $88 million in data licensing), plus $116 million in advertising. AI segment capex climbed from $12.7 billion in 2025 to $7.7 billion in Q1 2026 alone — an annualized run rate of ~$30.8 billion. Grok had 117 million MAUs as of March 2026 out of 550 million total MAUs across Grok and X combined.

TechCrunch: xAI burned $6.4B last year

Nvidia Posts Record $81.6B Quarter, $43B in Startup Stakes

Nvidia reported $81.6 billion in revenue (up 20% QoQ) and a record $75.2 billion in data center revenue for the quarter ending April 26, 2026. The company authorized $80 billion in share repurchases and revealed it holds $43 billion in startup stakes. CFO Colette Kress noted Blackwell is "everywhere — adopted and deployed by every major hyperscaler, every cloud provider, and every major model maker." Nvidia projected $91 billion for next quarter (12% growth). Chinese exports had no significant impact; H200s approved for U.S. export but "we have yet to generate any revenue."

TechCrunch: Nvidia posts another record quarter, reveals $43B of holdings in startups

Spotify & Universal Music Strike AI Covers/Remixes Deal

Spotify and Universal Music Group announced a deal allowing fan-made AI covers and remixes through "upfront agreements, not by asking for forgiveness later" — a clear swipe at Suno and other AI music tools. Artists and rightsholders can choose if and how they participate, and will be fairly compensated. Spotify also launched an ElevenLabs-powered audiobook creation tool and a new desktop app for creating personal podcasts (taking on Google's NotebookLM).

Trump Delays AI Security Executive Order

President Trump delayed signing an executive order that would allow the government to evaluate AI models before release, stating: "I didn't like certain aspects of it... We're leading China, we're leading everybody, and I don't want to do anything that's going to get in the way of that leading." The EO would have tasked the Office of the National Cyber Director with developing a security evaluation process, partly in response to concerns about Anthropic's Mythos and OpenAI's GPT-5.5 Cyber.

TechCrunch: Trump delays AI security executive order

🔬 Research (last 48h / 7-day window)

Most recent arXiv batch: May 20, 2026 (15 papers in cs.AI, cs.LG, cs.CL)

Notable Papers

Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning (May 20) — Benhao Huang, Zhengyang Geng, Zico Kolter. Proposes learning attractor dynamics for iterative reasoning models, enabling generalization beyond training-time compute budgets.
DeepWeb-Bench: A Deep Research Benchmark Demanding Massive Cross-Source Evidence and Long-Horizon Derivation (May 20) — Sixiong Xie et al. A benchmark for deep research agents that search the open web, collect evidence, and derive answers through extended reasoning. Finds frontier models still struggle with cross-source synthesis.
AiraXiv: An AI-Driven Open-Access Platform for Human and AI Scientists (May 20) — Junshu Pan et al. Proposes an AI-driven academic publishing platform to handle the growing volume of both human-authored and AI-generated research.
Agent JIT Compilation for Latency-Optimizing Web Agent Planning and Scheduling (May 20) — Caleb Winston et al. (Azalia Mirhoseini, Christos Kozyrakis). Introduces JIT compilation for computer-use agents to reduce latency in web automation tasks.
You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories (May 20) — Zhepei Wei et al. Shows that reinforcement learning with verifiable rewards can be dramatically more sample-efficient using rank-1 trajectory approximations.
DelTA: Discriminative Token Credit Assignment for RLVR (May 20) — Kaiyi Zhang et al. Improves response-level credit assignment in RLVR with discriminative token-level rewards.
Mem-$\pi$: Adaptive Memory through Learning When and What to Generate (May 20) — Xiaoqiang Wang et al. A framework for adaptive memory in LLM agents where useful guidance is generated on demand rather than retrieved from external stores.

🛠️ Tools (last 48h)

OpenClaw v2026.5.20 Stable + Two Betas

OpenClaw shipped v2026.5.20 stable (May 21) plus v2026.5.20-beta.2 and v2026.5.20-beta.1 on the same day. This is a reliability-focused release with the following highlights:

Discord voice session tracking: Voice sessions now follow configured users across voice channels, with multi-user handoff, allowed-channel checks, and DAVE recovery preservation.
Policy plugin: New bundled Policy plugin for policy-backed channel conformance checks, doctor lint findings, and opt-in workspace repair.
xAI device-code OAuth: Remote and headless setups can authorize xAI without a localhost browser callback.
OpenRouter provider routing: Honors provider-level params.provider routing policy.
Doctor security warnings: Detects plaintext API keys in openclaw.json and warns about sandbox tool policies hiding MCP server tools.
Codex harness bumped to @openai/codex 0.132.0.
Exec approval hardening: Removed old cat SKILL.md compatibility path; skill files must be loaded with the read tool.
GitHub: OpenClaw v2026.5.20
GitHub: OpenClaw v2026.5.20-beta.2

OpenAI Python SDK v2.38.0

The OpenAI Python SDK released v2.38.0 (May 21) with API updates, manual updates, and OpenAPI spec refresh. GitHub: openai-python v2.38.0

Anthropic Python SDK v0.104.0

The Anthropic Python SDK released v0.104.0 (May 21) adding support for the thinking-token-count beta for estimated tokens in thinking block deltas when streaming. GitHub: anthropic-sdk-python v0.104.0

New Codex Ships Today

Sam Altman announced that new Codex ships today (May 21), with GPT-5.5 integration, token analytics for businesses, and plugin sharing. Greg Brockman highlighted Codex app improvements and enterprise features.

💭 Industry Pulse (last 48h)

Yann LeCun: "AIs are nowhere near human intelligence"

In a viral thread responding to Noah Smith, Meta's Chief AI Scientist argued that people are realizing AIs lack common sense, understanding of reality, and limited reasoning abilities — but have become useful through "the accumulation of enormous amounts of declarative knowledge." He compared current AI hype to past milestones (AlphaGo, Watson, DeepBlue) and warned against conflating accumulated knowledge with intelligence.

@ylecun on X: "People are realizing that AIs are nowhere near human intelligence" (❤️ 2,808)
@ylecun on X: "This is as mistaken as saying 'AI far surpassed humans'" (❤️ 44)

OpenAI Offers $2M in API Credits to Every YC Startup

OpenAI is offering to invest $2 million in API credits into every startup in the current Y Combinator batch. Sam Altman called it "compute for powering the next generation of startups" and expressed excitement for "tokenmaxxing startups."

Andrew Ng Launches Image/Video Agent Course

Andrew Ng announced a new short course on building AI agents that generate images and videos, created with Google Cloud. The course covers image-text similarity scoring, LLM judges for brand consistency, and structured rubrics for video quality verification.

@AndrewYNg on X: "New course: Build AI agents that generate images and videos" (❤️ 1,043)

Microsoft Cancels Claude Code Licenses Due to Inference Costs

A tweet from @mohbii (May 22) claimed that Microsoft is canceling Claude Code licenses because inference costs are too high, signaling that "the free ride is almost over" for enterprise AI pricing. This follows broader industry discussions about AI subscription sustainability.

@mohbii on X: "microsoft canceling claude code licenses because inference costs too much"

📡 Sources & Data Provenance

Source	Status	URL
Twitter/X API	✅ Working	https://twitterapi.io
GitHub Releases	✅ Working	https://github.com
arXiv API	✅ Working	https://arxiv.org
TechCrunch Web Scrape	✅ Working	https://techcrunch.com
Wiki Raw Archive	✅ Available	~/wiki/raw/
Web Search	✅ Built-in	DuckDuckGo/Parallel

Each story above links directly to its primary source. Unlinked claims were cross-referenced from multiple sources.

AI Briefing: 2026-05-22