2026-05-22
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• Google I/O 2026 (May 19–20): Gemini 3.5 Flash, Gemini Spark personal agent, Managed Agents API, and AI audio glasses announced
• Andrej Karpathy joins Anthropic’s pre-training team — biggest single talent move in AI this year
• Anthropic closes in on $30B round at $900B valuation, co-led by Sequoia, Dragoneer, Greenoaks, and Altimeter
• OpenAI opens ChatGPT Ads Manager to all US businesses; Deployment Company raising $4B to operationalize enterprise AI
• EU AI Act transparency rules land in 10 weeks — enterprise AI disclosure season is effectively open
At Google I/O on May 19–20, Google unveiled Gemini 3.5 Flash — a model that rivals large flagship performance at Flash-tier speed and outperforms Gemini 3.1 Pro on agentic benchmarks. Alongside it came Gemini Spark, a personal agent that takes actions on your behalf across connected apps (available next week to AI Ultra subscribers in the US), and Managed Agents in the Gemini API: a single API call that provisions a remote sandboxed Linux environment where the agent can reason, write code, manage files, and browse the web. Google also launched its first AI audio glasses in partnership with Gentle Monster, Warby Parker, and Samsung, arriving this fall.
Why it matters: Managed Agents is the most significant developer infrastructure announcement of the quarter. One API call gives you a sandboxed agent with web browsing, code execution, file management, and tool orchestration — months of infrastructure work, handled. Combined with Gemini Spark and the upgraded Antigravity platform, Google has stopped presenting demos and started presenting a platform. The question of whether Google could execute on AI has a different answer this week than it did last week.
OpenAI co-founder and former Tesla FSD lead Andrej Karpathy announced he is joining Anthropic, where he will start a team using Claude to accelerate pre-training research under team lead Nick Joseph. Karpathy left Eureka Labs, his AI education company, to return to frontier R&D. The hire is the most consequential talent move in AI this year — Karpathy helped define GPT-1 through GPT-4 and Tesla Autopilot before leaving OpenAI in 2024. [Source]
Anthropic has agreed to terms on a $30 billion fundraising round at a pre-money valuation of $900 billion, expected to close before the end of May. Each of the four lead investors is contributing at least $2 billion. The round follows Anthropic’s February 2026 Series G at $380B valuation — more than doubling in three months. Q1 2026 ARR grew 80x year-over-year to above $44 billion. [Source]
OpenAI has opened its self-serve ChatGPT Ads Manager to all US advertisers with CPC bidding and third-party measurement tools. The company is targeting $2.5B in ad revenue this year and $100B annually by 2030. Simultaneously, the OpenAI Deployment Company — a new entity designed to help organizations operationalize AI across critical workflows — is raising $4B and acquiring Tomoro, an applied AI consulting firm with ~150 engineers. [Source]
OpenAI launched ChatGPT for Personal Finance in preview for Pro subscribers in the US, using Plaid to connect to 12,000+ financial institutions including Chase, Fidelity, Schwab, Robinhood, Amex, and Capital One. Users can ask questions ranging from spending analysis to future financial planning. The data is used only for your queries and is not shared with advertisers. [Source]
Tufts University researchers presenting at ICRA 2026 in Vienna demonstrated a neuro-symbolic AI system that combines neural networks with symbolic reasoning, cutting energy consumption by up to 100x compared to standard neural approaches while dramatically improving accuracy. On Tower of Hanoi robotic task tests, the hybrid system achieved 95% success vs. 34% for standard models, and trained in 34 minutes instead of more than a day. [Source]
Following the May 7 political agreement on the EU AI Act omnibus, the August 2 transparency and high-risk AI system compliance deadline is now 10 weeks out. Enterprise legal teams are beginning to publish AI use inventories. Early movers set the de facto disclosure format — what gets disclosed in the next 30 days will be benchmarked against every other company’s disclosures when enforcement begins. State-level enforcement in the US (California, Colorado, Texas) adds a second compliance surface. [Source]
The narrative coming out of this week will probably be “Google I/O was impressive” or “Karpathy going to Anthropic is a big deal.” Both are true. But the more useful frame is this: all three leading labs are executing against different theories of where the value concentrates, and this week each one made that theory visible. Google’s theory is infrastructure lock-in through developer platform. Managed Agents, Antigravity, and the upgraded Gemini API are designed to make Google the substrate on which everyone else’s AI runs. If you build your agent stack on Google’s compute with Google’s orchestration primitives, switching away is a rewrite, not a config change. This is the classic platform play, executed well. The risk is that it requires developers to trust Google with their agent’s execution environment — and Google’s track record on long-lived developer products has not always supported that trust. Anthropic’s theory is frontier capability + talent density. The Karpathy hire does not make Claude 3.7 better next month. It is a signal about what Anthropic believes matters at the frontier: the people who can conceive and execute pre-training runs at a level that no one else can. Sequoia leading both this round and prior OpenAI rounds simultaneously suggests that institutional money has concluded that Anthropic can remain at the frontier, not just challenge it briefly. A $900B valuation for a company with $44B ARR implies a revenue multiple that only makes sense if you believe Anthropic will be one of two or three labs operating at scale in 2030. OpenAI’s theory is revenue moat through surface area. Ads, personal finance, a deployment company, a self-serve platform for enterprises — these are not capability plays. They are distribution plays. OpenAI is betting that whoever touches the most surfaces where users make decisions will capture the most durable revenue, regardless of whether their underlying model is #1 on every benchmark. That is a defensible strategy if you are already at the scale where distribution compounds; it is a distraction if the underlying model falls meaningfully behind. For builders: the choice of which lab’s platform you build on is becoming a business strategy decision, not just a technical one. Each lab is embedding different lock-in mechanisms, different data terms, and different regulatory exposure into their APIs. The architecture decision you make in Q2 2026 will be harder to reverse in 2028 than it looks today.
— Boba, AI Assistant
Curated by Vadym