Seroter's Daily Reading — #788 (May 20, 2026) — Seroter's Daily Reading

Listen: https://blossom.nostr.xyz/5a6f7fd87737991e09c1d2ee93ee13357d41d16dcf235c25d8a37cb7a3446458.mpga

Seroter's Daily Reading, episode 788, May 20, 2026. The last day of Google I/O is over, and Seroter got to host a wonderful panel discussion, attend a couple sessions, and meet some great people. So naturally, today's reading list is dominated by Google I/O announcements.

Let's start with what Google is calling the biggest changes to Search in decades. Google is entering the era of A new era for AI Search, where you can create, where you can create, customize, and manage multiple AI agents right in the Search experience. They are starting with information agents that run in the background 24/7, intelligently reasoning across information to find exactly what you need at exactly the right moment. The example they give is apartment hunting — you can brain dump all your requirements and your agent continuously scans for listings that meet your needs, notifying you when something pops up. Or tracking when your favorite pro athletes announce a sneaker collab. These information agents launch first for Google AI Pro and Ultra subscribers this summer.

Google also announced Gemini for Science, a collection of tools designed to expand the scale and precision of scientific exploration. The core idea is that scientific knowledge is growing so fast that individual researchers struggle to see the full picture, and AI can help bridge that gap. The three main experimental tools are Hypothesis Generation built with Co-Scientist, which simulates the scientific method through a multi-agent idea tournament to generate and evaluate hypotheses. Computational Discovery built with AlphaEvolve and ERA, an agentic research engine that generates and scores thousands of code variations in parallel, allowing scientists to test novel modeling approaches in fields like solar forecasting or epidemiology that would take months to navigate manually. And Literature Insights built with Google NotebookLM, which searches scientific literature and structures results into tables with custom searchable attributes for side-by-side analysis. Enterprise partners like BASF are already using AlphaEvolve to optimize supply chains, and organizations like Daiichi Sankyo and the U.S. National Labs are using Co-Scientist to accelerate research. Validation papers for ERA and Co-Scientist are published today in Nature. They are also launching Science Skills, a specialized bundle integrating over 30 major life science databases including UniProt, AlphaFold, and AlphaGenome. Using these skills on agentic platforms lets researchers perform complex workflows like structural bioinformatics and genomic analyses in minutes rather than hours.

Google also unveiled Managed Agents in the Gemini API, which lets you spin up an agent with a single API call — one that can reason, use tools, and execute code in an isolated, ephemeral Linux environment. This is powered by the new Antigravity agent built on Gemini 3.5 Flash. The pitch is that building a production-grade agent used to mean managing complex infrastructure, scaffolding, and isolated sandboxes. With Managed Agents, Google is abstracting that away so you can focus on product experience and agent behavior. The Antigravity agent provisions a remote Linux sandbox where you can reason, plan and call tools, execute code and manage files, and browse the web. Each interaction creates or resumes an environment with all files and state intact.

Google AI Studio at I/O 2026 is becoming a real powerhouse. It now has direct access to Google Workspace, so you can build dashboards on top of your Sheets data, create tools that organize Drive, or work with documents your team already lives in — all without leaving AI Studio. You can export directly to Google Antigravity, bringing your conversation history, project files, and secrets with you. There's also new custom asset generation using a model called Nano Banana for generating images on the fly, and a new edit tool that lets you annotate and iterate on your app right in the preview window. And perhaps most interestingly, Google is bringing the full build-mode experience to your phone with a new mobile app available for pre-registration. You can start on mobile on the go, then go deep in the flow when you're back at your desk. Create and share your next idea in minutes, all from your phone.

And speaking of design, Google also announced new ways to vibe design with Stitch, transforming your design experience into a live, collaborative partnership with the Stitch Agent. Whether you start with a text prompt, use your voice, or bring existing codebase and design files, you can now design, stream, and steer iterations in real time. When you're ready, you can generate a shareable link via Google AI Studio, or export your screens into Google Antigravity to plug in backend logic, or publish directly to the web with Netlify. These updates are available globally starting today.

Then there's Agent Executor, Google's open source distributed agent runtime. The pitch is own your agents, models, and compute. With Agent Executor, enterprises have maximum flexibility to maintain sovereignty over workloads and keep proprietary workflows within their self-managed compute and custom sandboxes. You can prevent vendor lock-in by deploying on your own infrastructure, bring your own harness and agents since it's harness-agnostic, and fully control execution including MCPs, skills, and other agents running on your own data plane. They also announced Agent Substrate, a new open source project built on Kubernetes that introduces a new level of abstraction for handling agents at scale. Standard Kubernetes is optimized for thousands of long-running services, but agents are nonlinear programs that wait for external inputs, creating different demands. Agent Substrate is designed for the chatter of millions of sub-second tool calls that would otherwise overwhelm a standard control plane.

Shifting gears, there's a really thoughtful piece from Itamar Gilad on what he calls 3 Hyped AI PM Archetypes + 1 We're Missing. He walks through the AI PM — someone with deep knowledge of LLM training, inference, fine-tuning, and RAG. He thinks this is mostly hype. Yes, one in seven open PM positions now ask for AI skills, the fastest growing category at 465% since 2023, but that makes sense because the category barely existed before. More likely AI becomes a specialty like security or fintech rather than a requirement for all PMs. Then there's the Developer PM archetype — a PM who works like a developer, using coding agents like Claude Code or Cursor to do product work. He thinks this is less hype, and actually happening now. Coding agents are good at remembering context, searching across many files, and have unique skills like spawning sub-agents. He acknowledges that PMs coding prototypes for validation is a helpful skill, but cautions against prototypes as requirements — a working prototype already provides too many answers and can derail discussion into implementation details. Then there's the No PM archetype, the idea that PMs become obsolete. He thinks this is almost entirely hype, mostly from people who misunderstand what product management is about. Though he acknowledges the delivery-focused product owner role is vulnerable — one who mainly translates roadmaps into backlogs and stories, because AI is already producing such artifacts at a fraction of the time and cost. The draftsmanship analogy is apt here — when CAD appeared, most draftsmen lost their jobs not because the work became impossible, but because architects and designers could use the tools themselves. His fourth archetype is the AI-empowered PM — one who helps put AI to use across all company functions, not just coding or spec generation. Starting from what the organization actually needs rather than what the technology can do. He prefers this model because it begins with the needs of the org and the product manager and sees how AI can help, rather than projecting backward from technology capabilities.

GitHub quietly reminded users that security is their responsibility too. They are scaling back their bug bounty program. Their senior security researcher notes that not every valid submission represents a meaningful security risk — some reports identify hardening opportunities or documentation gaps. Many reports also describe scenarios where someone experiences an undesirable outcome after interacting with malicious content, like cloning a malicious repo or asking an AI tool to analyze untrusted code. These are well-written and technically accurate, but they misunderstand where the security boundary lies. When an attack requires the victim to actively seek out and engage with attacker-controlled content, the security boundary is the user's decision to trust that content. Fair point, though one hopes this doesn't discourage meaningful security research.

Flutter and Dart developers got some good news. The team announced experimental support for Dart in Cloud Functions for Firebase in a piece titled The Flutter missing link: Why full-stack Dart changes everything. For years, Flutter developers have been forced into a language mismatch when extending their apps to the cloud — context switching into TypeScript, Go, or Python with different concurrency models. The new Shared Package pattern moves business logic and data models into a standalone Dart package, eliminating the manual duplication that plagues traditional stacks. When you update a field in your shared package, that change propagates across the entire stack. Dart binaries execute immediately without a warm-up period, as fast as 10 milliseconds, thanks to ahead-of-time compilation. A native Dart binary can be as small as 10 megabytes versus a traditional SDK footprint of 211 megabytes. The Firebase CLI abstracts away infrastructure complexity entirely — a single command handles compilation and deployment. This is experimental stage, requiring Dart SDK 3.9 or higher and Firebase CLI v15.15.0, with support currently focused on HTTPS and callable functions. But the direction is clear: full-stack Dart, one language to rule them all.

Georg Schwarz has a detailed post titled From Kubernetes Dev Setup to Production: What Actually Changes on what it actually means to take a Kubernetes deployment from development setup to production platform. He worked on diff/docs, a hosted deployment of La Suite Numerique Docs, and walked through the transformation. The starting point already worked for development — local minikube, local certificates, generated credentials, bundled development dependencies, manual Helm sequencing. The production target was repeatable delivery, guarded change management, recovery paths, monitoring, and policy controls. He organized the work into four stages: make the building blocks work, make the product work end-to-end, move change control into Git and restructure environments around reconciliation, then make operations observable, recoverable, and sustainable. Key moves included GitOps through Flux where Git becomes the deployment API, secrets encrypted with SOPS, database backups with restore-check automation, and observability that answers operational questions rather than just creating dashboards. The interesting insight is that most problems appeared at the seams between components — PostgreSQL worked, object storage worked, the gateway worked, but the product still needed integration fixes before it behaved correctly. Media handling, for example, is rarely just upload file to S3 — it also involves URL generation, gateway paths, authorization behavior, content-type handling, and how the frontend renders stored assets. His conclusion is that production is not a state you reach once; it is a habit.

And finally, Blackstone and Google launched a new joint venture in a piece titled Blackstone, Google launch new compute-as-a-service venture to provide to provide data center capacity and Google Cloud TPUs under a compute-as-a-service offering. Blackstone is committing five billion dollars in initial equity capital, with the first 500 megawatts of capacity expected online in 2027. Google supplies the hardware, including TPUs, software, services, and technical expertise. This comes as enterprise spending on cloud infrastructure reaches historic levels — 129 billion dollars in Q1 2026 alone, a 35 billion dollar year-over-year increase. The TPU cloud company is taking a page from neocloud providers, offering compute-as-a-service as companies bid for enterprise spend on compute power. Five neocloud providers are now among the top 30 cloud providers, and neocloud revenues are anticipated to reach 400 billion dollars by 2031. Prior to this joint venture, Google and Broadcom had signed an agreement with Anthropic to add multiple gigawatts of TPU capacity starting in 2027, and Google partnered with CoreWeave to launch CoreWeave Interconnect linking the AI cloud provider to Google Cloud for cross-cloud training and inference.

That's episode 788. Lots of Google I/O, naturally — search agents, scientific research tools, managed agents, AI Studio updates, Stitch design, Agent Executor, Science Skills, and a new TPU cloud venture with Blackstone. Outside of Google, the Itamar Gilad piece on AI PM archetypes stands out as the most thought-provoking read, especially his argument for the AI-empowered PM rather than the AI PM or No PM paths. The Kubernetes production-readiness piece and the Dart full-stack news are more niche but deeply practical. Hope you found something useful.