Monday, May 18, 2026

What happened: Pope Leo XIV will present his first papal encyclical focused on artificial intelligence on May 25, with Anthropic co-founder Christopher Olah invited as a guest speaker at the presentation.

Key details:

• Event date: May 25, 2026
• Pope Leo XIV delivering first encyclical on artificial intelligence
• Christopher Olah, Anthropic co-founder, invited as guest speaker
• Represents Vatican's formal engagement with frontier AI policy

Why it matters: This marks the first formal papal teaching document on AI, signaling that religious institutions are treating AI governance as a moral and ethical priority. Anthropic's invitation of Olah as speaker suggests the company is positioning itself as a trusted voice in AI ethics discussions with global institutions beyond government and industry.

Practical takeaway: Watch for the full text of the encyclical on May 25 to understand how religious authority frames AI's role in society, as it may influence broader ethical frameworks in AI policy discussions.

What happened: Anthropic will brief leading global finance ministries and central banks on critical vulnerabilities in the global financial system's cyber defenses that its Claude Mythos Preview AI model has uncovered.

Key details:

• Briefings target leading finance ministries and central banks worldwide
• Focus is on cyber vulnerabilities discovered by Claude Mythos Preview
• Represents proactive regulatory engagement by Anthropic on AI-discovered security gaps
• Demonstrates Claude Mythos's capability to identify complex financial infrastructure weaknesses

Why it matters: This signals AI models are now discovering security vulnerabilities faster than human-led security reviews, and regulators need visibility into AI-discovered threats to critical infrastructure. It also positions Anthropic as a trusted partner in financial system security and represents a shift toward AI vendors directly advising government on national security implications of their own models.

Practical takeaway: Financial institutions should monitor announcements from this regulatory briefing, as disclosed vulnerabilities may require immediate patching of treasury and banking infrastructure.

What happened: A new class of AI models called World Action Models addresses a fundamental weakness in robotics AI by enabling robots to understand and simulate how the physical world changes in response to their actions, not just learn which movements match which camera images.

Key details:

• World Action Models tackle core limitation: robotics models learned movement-to-vision correspondence but lacked understanding of how world state changes from actions
• A new survey organizes approximately 100 papers into two distinct architectural approaches
• Models can learn from ordinary video that contains no robot action labels—data previously useless for traditional robotics AI
• Key advantage: ability to learn from unlabeled video makes training more scalable

Why it matters: This represents a major shift in robotics AI capability—models that understand world dynamics can plan more effectively, reason about unintended consequences, and generalize to new situations. Using unlabeled video dramatically reduces the expensive data collection burden that has limited robotics progress, potentially accelerating embodied AI deployment.

Practical takeaway: Robotics teams should evaluate World Action Model approaches for new projects, especially those where you have access to large unlabeled video datasets but limited labeled robot interaction data.

What happened: According to an analysis by The Information, Anthropic and OpenAI now capture 89 percent of revenue among the top AI startups, while the broader AI startup sector generates $80 billion in total revenue.

Key details:

• Anthropic and OpenAI capture 89% of top AI startup revenue combined
• Total AI startup revenue across the sector: $80 billion
• Analysis conducted by The Information
• Reflects extreme market concentration at the frontier

Why it matters: The dominance of two companies in capturing AI revenue demonstrates massive competitive moats built on compute scale and model performance. This concentration threatens ecosystem diversity and raises questions about whether smaller AI companies can compete or achieve profitability in a market dominated by these two players. It also limits venture capital allocation opportunities in the AI space.

Practical takeaway: AI startups should focus on vertical specialization or unique data moats rather than competing on general-purpose model capabilities, as horizontal AI competition is effectively decided by Anthropic and OpenAI's scale advantage.

What happened: Cursor released Composer 2.5, an AI coding model that achieves performance equivalent to Anthropic's Opus 4.7 and OpenAI's GPT-5.5 on coding benchmarks while operating at a significantly lower cost.

Key details:

• Composer 2.5 is built on Kimi K2.5 as its base foundation
• The model was trained on 25x more synthetic tasks compared to its predecessor
• Achieves parity with Opus 4.7 and GPT-5.5 on benchmarks
• Offered at a fraction of the price of competing models

Why it matters: Cost-competitive frontier-level AI coding models reduce developer reliance on expensive proprietary solutions from OpenAI and Anthropic, democratizing access to high-performance coding assistance. This reflects the growing efficiency of AI coding tools and indicates competition is driving down pricing across the market.

Practical takeaway: Evaluate Cursor's Composer 2.5 if cost per token is a key factor in your coding assistant selection, as it now offers frontier-level performance without frontier-level pricing.

What happened: A coalition of conservative organizations led by Humans First has called on President Donald Trump to issue an executive order requiring mandatory safety testing for frontier AI models before they are publicly deployed or shipped.

Key details:

• Coalition led by Humans First organization
• Open letter addressed to President Donald Trump
• Requests executive order mandating safety testing requirement
• Focuses on frontier AI models specifically
• Represents cross-partisan convergence on need for AI safety guardrails

Why it matters: This demonstrates that AI safety regulation is no longer a left-leaning concern—conservative political groups are now advocating for mandatory testing, suggesting bipartisan consensus is forming around pre-deployment safety requirements. This could accelerate movement toward government-mandated safety standards rather than industry self-regulation.

Practical takeaway: AI developers should anticipate potential future regulatory requirements for pre-deployment safety testing and evaluate whether current testing practices would satisfy government-mandated standards.