Learning Brief — 2026-05-18

What we covered

• AI news: Google's I/O Positioning, NVIDIA's Video Generation Tools, and the Agent Evaluation Shift
• PM news: Anthropic's Claude Code Shift: HTML Over Markdown Signals How AI Agents Are Changing Product Building
• PM learning: HTML is the new Markdown: How AI tooling changes what you build with

Mental model

The interface you build in determines what problems become solvable; optimize for interface constraints that unlock new user behaviors, not just better outputs.

Summary

Google is entering its annual I/O developer conference positioned as a clear third player in the foundation model race, signaling a strategic recalibration in how it's competing against OpenAI and Anthropic in the AI landscape. NVIDIA released Cosmos Predict 2.5 with LoRA and DoRA fine-tuning capabilities, enabling developers to customize video generation models for specific domains like robotics without full retraining—lowering the barrier to building specialized video AI products. IBM Research and collaborators launched the Open Agent Leaderboard, establishing standardized benchmarks for evaluating agentic AI systems, addressing a critical gap in how teams can objectively compare agent architectures and performance.

Anthropic engineers just shared something subtle but telling about how they're building with Claude Code: they've moved from markdown to HTML as the primary interface language for AI-generated specs and micro-apps. On the surface, that's a technical choice. But there's a real PM lesson buried in there.

Here's what's happening. When you're building tools for AI agents to use—not humans—the interface language matters differently than it does for documentation. Markdown was designed for human readability and portability. HTML gives you semantic structure, styling, and interactivity that AI models can actually parse and act on more reliably. It's not about what's easier for us to read. It's about what creates the most actionable output for the next agent in the chain.

This is a canary in the coal mine for product teams. As AI becomes embedded deeper into your workflows, you're going to face similar decisions: do you optimize your data structures, APIs, and outputs for human consumption, or for machine consumption? The answer increasingly is: both, but in different layers.

What Anthropic is signaling is that they're building a product ecosystem where AI agents are first-class citizens in the workflow. They're not bolting AI onto existing human-centric tools. They're rethinking the primitives from the ground up—choosing formats that let Claude Code generate, iterate, and hand off work to the next step with minimal friction.

If you're a senior PM thinking about AI-native products or how to integrate AI deeper into existing workflows, this is the mindset shift to watch. It's not "how do we make AI explain things better to humans?" It's "how do we make our product structure native to how AI agents actually work?" That's where the competitive advantage sits right now.

Here's the thing that matters for how you think about AI products going forward: the medium you build in shapes what becomes possible, and that medium changes faster than most PMs realize.

Thariq Shihipar from Anthropic makes a deceptively simple observation — HTML replaced Markdown as the native format for Claude Code interactions. On the surface, that's a technical detail. But what that means in practice is that engineers stopped thinking about text outputs and started thinking about interactive interfaces. The constraint changed. And when constraints change, what people build changes with it.

This is the mental model worth sitting with: your product's affordances don't just enable features, they reshape what problems your users think they can solve. When the tool shifted from "here's text you read" to "here's an interactive spec you can edit live," the entire category of problems that felt solvable expanded. Suddenly you're not managing outputs — you're managing micro-applications.

Why does this matter to you as a PM? Because you're probably still thinking about AI features as "better search" or "smarter classification" or "faster writing." But the engineers who are winning aren't optimizing within the old medium. They're asking what becomes possible when you change the interface itself.

Think about it like this: Markdown forced a certain kind of thinking — linear, read-only, output-focused. HTML opens bidirectional interaction. That's not a small difference. That's the difference between a tool that generates and a tool that lets you iterate. The move here is to stop asking "how do I make AI better at X" and start asking "what interface would let my users do something they couldn't do before?"

This connects directly to how you prioritize. If you're measuring success by "did the AI output improve," you're optimizing for the wrong thing. You should be measuring whether the interface unlocked new behaviors. Did users start using this in ways we didn't predict? Are they solving problems they previously thought were impossible?

For this week: pick one AI feature you're currently shipping or planning. Don't ask how to make the model smarter. Instead, ask what interface change would make the output actually useful enough that people change their workflow. What would have to be true about the interaction layer — not the model — for this to matter?