In partnership with
In Today’s Issue:
🎬 Gemini Omni 1.1 Flash makes AI video a priced commodity
🛡️ 100+ companies sign an urgent AI cyber defense letter
🐉 Tencent's Hy4 preview storms the open-source frontier
🦆 Microduck: the $399 robot you retrain yourself
🏛️ Washington keeps almost regulating AI
✨ And more AI goodness…
⚡ The Signal
The frontier stopped being a chat window this week: it prices your video by the second, patrols your networks, and walks across your desk.
Google made 4K AI video a line item at three cents a second for drafts and thirty for the final cut. More than 100 companies, from OpenAI and Anthropic to Microsoft and Oracle, signed one letter warning that AI-powered cyber attacks will become far more widespread within months. Tencent pushed a 770-billion-parameter open model within three points of the closed frontier. And a Hugging Face company started selling a $399 robot you retrain at home, on the day Nvidia reportedly agreed to buy its parent for $12.9 billion. The guardrails for all that autonomy are being built in public, and in a hurry.
All the best,
Kim Isenberg
(Google)
🎬 Google makes AI video a commodity
Google released Gemini Omni 1.1 Flash, video generation it calls "production-ready for professional use via the Gemini API". The model extends scenes up to 40 seconds, interpolates between a chosen first and last frame, and upscales output to 1080p or 4K; a new 360p draft mode renders previews 60 percent faster at a third of 720p's cost. It is available in the Gemini API and AI Studio, with consumer access via Flow and the Gemini app for subscribers.
👉 tl;dr: AI video is now priced like a utility, by the second and by the resolution.
(OpenAI)
🛡️ Over 100 companies sign a cyber defense letter
More than 100 organizations, including OpenAI, Anthropic, AWS, Google, Microsoft and Oracle, signed an open letter calling for a "global surge in cyber defense". The letter warns that "in the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated as models around the world become increasingly capable", putting hospitals, utilities and core internet infrastructure at risk. It asks companies to fix their highest-risk vulnerabilities now and governments to fund defense for essential services.
👉 tl;dr: The AI industry is publicly bracing for machine-speed attacks it expects within months.
(Tencent)
🐉 Tencent open-sources Hy4 preview, its biggest model yet
Tencent released Hy4 preview, a 770-billion-parameter Mixture-of-Experts model (49B active per token) with a 1-million-token context window, open-weight under Apache 2.0. Tencent calls the jump from Hy3 "the largest generation-over-generation gain we've measured": on the agentic coding benchmark DeepSWE it scores 64.3 against Hy3's 28.0, and its 85.4 on Terminal Bench 2.1 lands three points behind GPT-5.6 Sol's 88.3. Weights are on Hugging Face; APIs run on Tencent Cloud and OpenRouter at about $0.83 per million input tokens.
👉 tl;dr: China's biggest platform company just pushed open-source within sight of the closed frontier.
🎬 Watch This
SemiAnalysis founder Dylan Patel joined Dwarkesh Patel this week to argue that two labs will soon control most of the world's workforce. Patel tracks GPU supply chains and lab compute budgets for a living, which makes his case for a sharply concentrating frontier worth 77 minutes of your time.
Posted Thursday alongside OpenAI's open letter on collective cyber defense (see In The News).
OpenAI appears to be building a Codex that never clocks out. Code in the public repo describes a "Persistent mode" where the agent will "continue working until put to sleep" and queue up its own follow-up tasks; OpenAI confirmed to Wired it is testing the feature, with no launch planned.
Hugging Face Puts a Robot Lab on Your Desk
The Takeaway
👉 Microduck is a 25 cm, $399 open-source biped from Pollen Robotics, Hugging Face's robotics arm; pre-orders opened Thursday, shipping before Christmas.
👉 It ships with seven trained behaviors, from walking and kicking to roller skating, and stands back up on its own when it falls.
👉 The whole stack is Apache 2.0: SDK, MuJoCo simulation and RL training code on GitHub; retraining runs on your laptop or on Hugging Face Jobs.
👉 The timing writes its own story: Nvidia reportedly agreed to buy Hugging Face for $12.9 billion the same day.
The cheapest ticket into real robot training now costs $399 and looks like a duck. Pollen Robotics, the robotics arm of Hugging Face, opened pre-orders on Thursday for Microduck, a 25-centimeter, 800-gram biped with 15 motors, a camera, LiDAR and two inertial sensors, running its control policy 50 times a second on the robot itself. It walks, kicks, grabs, roller-skates and picks itself up when it falls, and it ships before Christmas.
(Pollen Robotics)
What Pollen is really selling is the training loop. Its pitch: "Fun out of the box. Yours to retrain." Every behavior is a reinforcement-learning policy, meaning the duck learned its moves by trial and error in a physics simulation before running on the real hardware: "Trained in sim, deployed on the real robot." The full stack, SDK, MuJoCo simulation and training code, is open source under Apache 2.0; teach it a new trick on your laptop, or rent the compute on Hugging Face Jobs and flash the result to the duck.
(Pollen Robotics)
The skeptical read: $399 buys a dev kit with seven canned behaviors and a Christmas ship date. The strategic read is bigger. Microduck turns Hugging Face's open-model culture into hardware in the same week The Information reported that Nvidia agreed to buy Hugging Face for $12.9 billion, with pre-orders and the acquisition report landing on the same Thursday. If the deal closes, the little duck becomes Nvidia's cheapest on-ramp to embodied AI.
Why it matters: Robot learning had a hardware price floor in the thousands; a $399 sim-to-real kit makes training robots a hobbyist skill, and hobbyist skills have a habit of becoming industries.
Sources:
🔗 https://pollen-robotics.com/microduck/
🔗 https://www.cnbc.com/2026/08/27/nvidia-hugging-face-acquisition.html
🔗 https://arstechnica.com/ai/2026/08/report-nvidia-to-acquire-ai-model-repository-hugging-face-for-13-billion/
Your Competitors Just Hired an Entire GTM Team. No Humans.
AI agents are doing the work of full GTM teams. Pipeline, content, customer service. The startups using them aren't waiting for headcount approvals.
Get the free Practical Guide to Agentic GTM for Startups and see exactly where to start.
The chart: Google's pricing table for Gemini Omni 1.1 Flash, in dollars per second of generated video: 360p costs $0.03, 720p $0.10, 1080p $0.15 and 4K $0.30, next to Veo 3.1 Lite at $0.05 to $0.08 and Veo 3.1 Fast at $0.10 to $0.30. The older Gemini Omni Flash offered only 720p, at the same $0.10.
The lesson: Video generation now has a price ladder instead of a price. One model spans a cheap 360p drafting tier and 4K delivery, so studios can iterate at three cents a second and pay ten times that only for the final render. That pricing structure, more than any benchmark, is what makes AI video production-ready.
The caveat: Per-second pricing flatters short clips: a minute of 4K costs $18, and a real workflow with drafts and retries multiplies that, so "cheap" depends on how many takes you throw away.
🏛️ Washington Keeps Almost Regulating AI
⚡ Bottom line: The White House's draft executive order creating an AI oversight body has stalled, The Information reported Thursday.
💡 Why it matters: The US still has no federal AI law, no dedicated AI regulator, and a policy office that keeps losing its leaders.
🔎 What it means: The practical rulebook for frontier AI is being written by state legislatures and the labs themselves.
The story in one sentence: the White House drafted an executive order to create a self-regulatory organization, described in coverage as "a public-private organization to supervise the release of AI models", and by Thursday The Information reported the draft has stalled. That is quietly remarkable: self-regulation was supposed to be the version of oversight the industry could live with.
(Getty Images via TechCrunch)
It is a familiar ending. In May, an earlier AI executive order was shelved within a day after former AI czar David Sacks phoned the president with objections, with Elon Musk and Mark Zuckerberg reportedly making similar calls. Sacks had already given up the czar title in March; by July, TechCrunch was describing the government's AI standards office as a "revolving door".
The vacuum is not empty. The White House's March framework asked Congress for a federal AI law by year's end, built around six priorities including preemption of state AI laws; Congress has not delivered one. The states are moving anyway: in July, Illinois became the first state to require annual independent third-party safety audits of frontier AI developers, starting in 2028.
(Getty Images via TechCrunch)
Whether this reads as dysfunction or as a lucky escape depends on your politics. Either way, the most consequential AI rules in America right now are being written in state capitols and inside the labs' own safety teams, while Washington's drafts keep dying before the signing ceremony.
Never worry about roaming again
Stay connected on every trip with Saily eSIM plans. From beach vacations to business travel, access data in 200+ destinations.
VIP perks available.
Activate instantly upon arrival.
Download SAILY in your app store and use code newsletter15 at checkout to get an exclusive 15% off your first purchase.
Chat support available 24/7. Get a full refund if your device isn’t eSIM compatible.
