I know I’m late with today’s edition, and I apologise but I was following Google I/O…

Six stories this week, one direction. Frontier labs stopped racing on the model and started buying the floor underneath. Anthropic bolted agents inside your perimeter and took the SDK generator that ships every Claude library.

Google's I/O answer was a model that creates anything from anything. Mistral pulled Vienna into Paris for the factory floor. Amazon got into AI radio. And the OSS layer answered with Hermes, the agent that gets better the more you use it.

No Short Signals this edition. The six full stories pull their weight.

Çelik

The same I/O keynote that introduced Gemini Omni also retired the search results page most of us grew up on. Google reframed Search around an AI-powered intelligent search box: longer conversational queries supported, AI Overviews answering follow-ups inside AI Mode starting Tuesday, information agents that track changes in the background 24/7 launching this summer, and generative UI that builds interactive widgets and mini-apps on the fly using Gemini Flash 3.5 plus the Antigravity agentic platform. Head of Search Liz Reid called it the biggest change to the entry point of the web in 25 years.

The numbers behind the move are the punch line. AI Overviews already touches 2.5 billion monthly users; the conversational AI Mode tops a billion. ChatGPT, the comparison everyone reaches for, sits at 900 million weekly. Google is not chasing OpenAI; it is rewiring the surface that two and a half billion people land on first.

For any CEE operator whose pipeline runs on inbound traffic and SEO, the read is brutal. Referrals to publishers were already down on AI Overviews; this collapses them further. The fix this quarter is to write for the answer surface, not the link surface. Get your content into the AI Overview citation slot and into whatever information-agent feeds your customers will be running, or watch the traffic line keep bending the wrong way.

Anthropic shipped self-hosted sandboxes and MCP tunnels live from its Code with Claude developer event in London on May 19. Self-hosted sandboxes (public beta) let agent tool calls run on your own infrastructure or with managed providers like Cloudflare, Daytona, Modal, or Vercel, while Anthropic keeps the orchestration loop on its side. MCP tunnels (research preview) open a single outbound connection from inside your network so private databases, internal APIs, and ticketing systems become tools the agent can call without anything being exposed to the public internet.

The New Stack framed the move as Anthropic locking down agent infrastructure for the enterprise crowd that has been quietly running pilots and refusing to ship anything that needed data to leave their perimeter. Coffee-meet conversations with CISOs and CIOs across CEE have been pointing at this exact wall for nine months. The wall just came down.

If you run a team in a regulated industry or sell into one, opt into the public beta this sprint, not next quarter. The first ops team in your category that ships a real internal agent against a private MCP server will own the reference customer slot.

You don't have to read everything — just the right thing. 1440's daily newsletter distills the day's biggest stories from 100+ sources into one quick, 5-minute read. It's the fastest way to stay sharp, sound informed, and actually understand what's happening in the world. Join 4.5 million readers who start their day the smart way.

Join for free today!

Demis Hassabis took the I/O stage in Mountain View on May 19 and introduced Gemini Omni, a generative model that converts text, photos, audio, and existing video into new video. Conversational editing through voice command, characters and backgrounds swappable mid-shot, and a physics layer that holds up under prompts like "a claymation explainer of protein folding" running as a stop-motion clip with voice-over.

The positional shift is the real story. Veo and Sora taught the market to expect text-to-video. Omni is text-or-image-or-audio-or-video-to-video, with Gemini's reasoning sitting behind the generation. Hassabis told CNBC the long arc is "any output from any input," and the demo lined up with the claim.

Gemini Omni Flash launches this summer for Google AI Plus, Pro, and Ultra subscribers in the Gemini app and Flow. Free inside YouTube Shorts and YouTube Create. If you run brand or content for a CEE company, the cost floor for a finished short video collapses again this quarter. Re-budget what your team produces, not what it spends.

A day before the London event, Anthropic acquired Stainless, the company that has generated every official Anthropic SDK since the API launched in 2022. Stainless turns an OpenAPI spec into native libraries across TypeScript, Python, Go, Java, Kotlin, plus CLIs and MCP servers, all designed to feel idiomatic in each language.

The reasoning is on the tin. Anthropic invented MCP to make agent connectivity possible. Stainless is the factory that produces the wrappers agents call to do anything. Folding the two together lets Anthropic compress the path from "we shipped a new endpoint" to "every Claude agent on the planet can call it tomorrow."

For founders shipping on top of Claude, the read-through is direct. The integration layer between your product and the model is now an Anthropic product. Plan for richer first-party SDKs, more MCP servers landing pre-built, and less room for third-party glue. If your moat was "we wrote the cleanest Claude wrapper," your moat is on borrowed time.

Mistral AI announced on May 19 that it had acquired Vienna's Emmi AI, with sources placing the price near €300M in cash and stock. Emmi builds physics-AI models that simulate airflow, heat transfer, and material stress, and ran Austria's largest 2025 funding round at €15M. It is Mistral's second 2026 acquisition, after Koyeb.

Arthur Mensch told reporters the buy is about positioning Mistral as the preferred partner for aerospace, automotive, and semiconductor manufacturers. Where OpenAI and Anthropic chase enterprise software workflows, Mistral is leaning into the factory floor, where the substrate is CAD files, finite-element models, and tolerances rather than tickets and pull requests.

For CEE operators, this is the second acquisition in eighteen months where a Paris frontier lab pulled a Central European deep-tech team onto its roadmap. The pattern is clear and the read is uncomfortable: if your team is doing world-class physics, materials, or simulation work from Vienna, Prague, or Ljubljana, the buyer is now Paris before it is San Francisco. Price your option to stay independent against that.

While the four labs above bought and bolted their adjacent layers, the OSS counterpoint is Hermes Agent, an open-source agent framework that wraps the usual reason-and-act loop with a learning layer. Successful task runs get extracted as named skills, those skills get refined as new outcomes arrive, and a persistent user model carries preferences and decisions across sessions. The agent compounds in capability the longer it runs against a given task set.

The positioning is sharp. OpenClaw, the OSS leader most operators reach for first, optimizes for broad reactive tool use. Hermes is built for the opposite bet: repetitive structured task types where an agent that learns is worth the architectural overhead. Closed labs are racing to own the perimeter, the SDK, the physics, and the audio. OSS is racing to own the loop that gets smarter every time you use it.

For builders this is the option labs cannot easily sell you. If your team runs the same kinds of tasks at the same kinds of customers every week, the value of an agent that remembers what worked last time is real. Test Hermes on one workflow this month. If it accelerates on its own, you have a defensible moat the foundation models will not ship for you.