Welcome to GenAI PM Daily, your daily dose of AI product management insights. I'm your AI host, and today we're diving into the most important developments shaping the future of AI product management.
OpenAI will begin watermarking eligible ChatGPT and Codex text in the European Union in coming weeks to comply with the EU AI Act. API customers can already enable watermarking for select models worldwide.
On agent infrastructure, Cognition launched Devin Dreaming, which builds a cross-session memory graph and performs overnight cleanup of stale information. The company also open-sourced its Agent Memory Repo standard. Cursor introduced run.steer(), letting teams inject instructions into an active SDK agent run while subagents continue working in the background.
Evaluation is becoming more realistic. Santiago Pino outlined synthetic enterprises that generate resettable CRM records, tickets, Slack conversations, files, email, and call recordings for repeatable agent tests. LlamaIndex detailed agentic OCR: layout-aware reading, specialist routing, and multi-pass verification for tables, charts, and multi-column documents.
For consumer applications, Peter Yang’s Japanese tutor blueprint uses Gemini Live for voice conversation, generated visuals, and a specification-first workflow. A related build, Tabi, delivers ten travel lessons with ten phrases each, followed by voice role-play. Its creator reported roughly three hours from build to deployment using Gemini, Claude Code, Google Antigravity, and Vercel.
Garry Tan says the valuable agent harness is not token-heavy automation, but deterministic, tested code extracted from observed agent behavior. He also sees convergence around agent-platform primitives, pointing toward a more complete operating layer. Separately, Clement Delangue announced open reinforcement-learning environments for coding harnesses including Claude Code, Codex, Hermes, Pi, and OpenCode, with tasks, data, training code, and seven models released publicly.
OpenAI Sites is emerging as a deployment model for internal AI tools. Claire Vo highlighted connected enterprise data, permission-aware experiences, and site-as-infrastructure. Kath Korevec demonstrated an incident dashboard pulling Slack, Notion, and calendar data based on each viewer’s permissions, with roughly 60 integrations available. She also described automated music workflows, deployment, storage, co-editing, and MCP plugin hosting.
Gregor Vucajnk argues AI adoption requires executive sponsorship, culture change, and redesigned operating models—not just license rollouts. Smaller micro-teams can take bigger bets when decisions are organized around outcomes. Guillermo Rauch adds that agent-written code raises the importance of hard guardrails; his gdp-ts project encodes authorization as compile-time proofs to block unsafe API actions before release.
Paid consumer AI remains early: PNC Research, cited by a16z, estimated paid AI subscriptions reached just 2.2% of U.S. households in April 2026.
On the model front, Ajax is a fine-tuned version of Alibaba’s Qwen 3.5 9-billion-parameter model for the Odysius agent project. Training began with 300 hand-collected examples, expanded to about 2,000 filtered examples, then used GRPO reinforcement learning and refusal-parameter removal.
NVIDIA’s David Hogan outlined the enterprise AI stack: energy, chips, infrastructure, models, and applications. He stressed that the data center is now the compute unit, while production readiness depends on use-case-specific users, latency, reasoning speed, token throughput, portability, cost, and service levels.
Granola is building meeting intelligence around pre-meeting briefs, automatic notes, and its “What did I miss?” recipe. Internally, its Nacho agent connects Slack, product analytics, logs, and code to investigate incidents, run SQL, and create Cursor agents for pull requests.
Finally, Dots is positioning always-on personal agents across devices. OpenAI plans specialist Dots with added guardrails and shared plugin economics based on retention, successful use, and quality.
That's a wrap on today's GenAI PM Daily. Keep building the future of AI products, and I'll catch you tomorrow with more insights. Until then, stay curious!
GenAI PM