Google's MASSIVE "AI Ultra" | Mariner, Astra, Veo 3, jules, Diffusion etc (FIRST LOOK)

Okay, here's a rundown of the new AI stuff Google just announced.

Google unveiled a bunch of new AI products and updates at their Google I/O event, covering everything from video generation and coding assistants to advanced research tools and a new high-tier subscription service called "Google AI Ultra" to access the cutting edge features. The announcements include significant improvements and entirely new capabilities, hinting at a push towards more agentic and multi-modal AI experiences.

Here are the key things discussed:

• Veo 3: This is Google's new AI video generation model. The big news is that it can now generate videos with sound, including dialogue, background noises, and sound effects. This adds a whole new layer of narrative possibility to AI-created videos.
• Google AI Subscription Tiers: Google has updated its AI subscription plans. The lower tier, previously called Google AI Advanced, is now Google AI Pro. The major new addition is Google AI Ultra, which is positioned as the VIP plan for users who want access to the latest and greatest AI features as soon as they're released.
• Google AI Ultra: This new top-tier plan is pretty pricey, expected to cost $250 a month, though there's currently an introductory offer of $125 a month for the first three months. Subscribing to Ultra gets you early and exclusive access to a ton of premium AI tools.
• What you get with Ultra: The Ultra plan includes the Gemini app, exclusive access to 2.5 Pro Deep Think, Veo 3 (the new video model), Flow (the AI filmmaking tool with Veo 3 access), WISC (image-to-video creation), the highest usage limits, new NotebookLM features, Gemini integration in Google apps (Gmail, Docs, Vids), early access to Project Mariner, a YouTube Premium individual plan, and a massive 30 terabytes of storage across Google Photos, Drive, and Gmail. I am particularly excited about the AI-specific benefits included in this tier.
• Gemini Diffusion: Google announced a text diffusion model specifically for code. This is surprising because diffusion models are usually associated with generating images or other continuous data, not discrete code. The fact that it can apparently generate working code in just three seconds is described as "insane" and "weird" due to the nature of diffusion models. This is definitely something I want to try out.
• Jules: This is an AI coding agent powered by Gemini 2.5 Pro. Jules is designed to work asynchronously across your code repository, meaning you can give it tasks like fixing bugs or refactoring code, and it will work on them in the background while you do other things. It also offers "CodeCasts," a daily podcast summarizing recent code commits, which sounds pretty neat. It currently seems to be under heavy load due to the recent announcement, and at first glance, it looks a bit similar to OpenAI's Codex. I plan to test this out.
• NotebookLM Updates: NotebookLM, which was already useful for summarizing and analyzing documents, is getting some cool new features. It will soon be able to do video overviews, which is a step up from its existing audio overview capability. I've seen some leaked previews of this feature and am hoping it rolls out soon.
• NotebookLM & Deep Research File/Image Uploads: A significant update is the ability to upload your own files and images directly into Deep Research within NotebookLM. This includes connecting to Google Drive and Gmail, making it much easier to analyze personal documents and data. I've been using Google Deep Research more recently and am finding it very impressive, especially now with this new file upload feature. Both Google's and OpenAI's deep research tools are good in different ways, and I'm still figuring out which is better for which tasks.
• Deep Research & Canvas Integration: Deep Research reports can now be easily transformed into custom web pages using a "Create" button, making it simple to share your research findings.
• Project Astra / Gemini Live: This is Google's vision for a universal AI assistant. Gemini Live on Android devices will soon have camera and screen-sharing capabilities, allowing the AI to see what you're seeing (via the camera) or what's on your screen to provide assistance. The demo for this looked impressive, and I'm curious to see how well it works in the real world.
• Gemini Live & App Integration: In the coming weeks, Gemini Live will connect with other Google apps like Calendar, Keep, Tasks, and Maps, making it a more integrated personal assistant on Android.
• Imagen 4 (Imagine 4): This is Google's text-to-image application. It's mentioned that everyone can now create images for free within the Gemini app using this technology.
• Flow: This is the AI video editing tool that provides access to Veo. It features something called "Flow TV" which cycles through generated AI videos, essentially like random AI video channels. These examples, often generated with Veo 2, look really good and show off what the tool can do, especially with the "ingredients to video" feature mentioned as a premium capability. The Ultra subscription grants access to Veo 3 within Flow.
• Testing Flow with Veo 3: I tried creating a video in Flow using Veo 3 (after realizing it had to be manually enabled), prompting for "a live tiger made out of snow sneaking through snow." The results were pretty good, although the tiger's movements were a bit robotic in one version. Upscaling to 1080p is also an option.
• Project Mariner: This is an agentic research tool included in the Ultra membership. The idea is that you give it a complex task (like finding all Google I/O announcements and adding them to NotebookLM), and it goes out and performs the steps needed, acting like an autonomous agent using a built-in browser.
• Testing Project Mariner: I tried using Project Mariner to find Google I/O announcements. It navigated websites, handled a cookie banner (though it annoyingly asked for permission first), and performed Google searches. It later successfully copied top Reddit AI news into an online notepad.
• Project Mariner Status & Challenges: Project Mariner is described as a research preview and is expected to make mistakes. It's a beta tool currently being worked on. A key challenge highlighted is giving the agent permission to access logged-in services (like NotebookLM or personal accounts), which raises security and trust issues. Building AI agents that can effectively use a computer like a human is notoriously difficult, and while Mariner looks promising based on a brief test, more extensive testing is needed.
• Overall Impression: Many of the newly announced tools are still experiencing issues, likely due to high demand after the announcement. However, the features I was able to test, particularly the visual and agentic capabilities, look very interesting and promising. I plan to delve deeper into each of these tools in separate videos soon.