Good Morning! Nvidia just launched "Ising," a new family of open-source AI models designed to fix the one thing holding quantum computing back: the fact that qubits are incredibly fragile and prone to error. Plus I’ll show you how to turn questions into interactive visualizations with Gemini.

Plus, in today’s AI newsletter:

• Anthropic Launches "Routines" to Automate Claude Code

• OpenAI Drops GPT-5.4-Cyber for Security Pros

• Nvidia Launches ‘Ising’ to Stabilize Quantum Computing

• How to Turn Questions into Interactive Visualizations with Gemini

• 4 new AI tools worth trying

QUANTUN COMPUTING

⚛️ Nvidia Launches ‘Ising’ to Stabilize Quantum Computing

While giants like IBM and Google focus on building the quantum hardware, Nvidia is positioning itself as the "control plane." The Ising models use AI to automate processor tuning and fix errors in real-time, turning experimental quantum machines into reliable tools.

• The Ising suite tackles two of the biggest hurdles in quantum scaling: hardware calibration and real-time error correction.

• Ising Calibration: An AI model that automates the tuning of quantum processors, slashing the setup process from several days down to just a few hours.

• Ising Decoding: Neural network-based models that provide faster, more accurate error correction than any current open-source method.

• Nvidia CEO Jensen Huang describes AI as the "control plane" that will transform fragile qubits into a scalable, usable computing platform.

For years, quantum computing has been stuck in the "experimental" phase because qubits are too sensitive to their environment. By open-sourcing the AI "recipes" to stabilize these systems, Nvidia is accelerating the timeline for practical quantum applications. They aren't building the quantum computer, they’re building the OS that makes the quantum computer actually work.

AI TOOLS

🤖 Anthropic Launches "Routines" to Automate Claude Code

Claude Code is moving from interactive sessions to scheduled and event-driven automation. With the new Routines feature (currently in research preview), developers can package up prompts, repos, and connectors to handle the software development lifecycle on autopilot.

• Scheduled Routines: Configure Claude to run on a cadence (hourly, nightly, or weekly). For example, it can pull the top bug from Linear at 2 AM, attempt a fix, and open a draft PR automatically.

• API Routines: Every routine gets its own endpoint and auth token, allowing you to trigger Claude from alerting tools, deploy hooks, or any internal system via HTTP requests.

• Webhook Routines: Currently starting with GitHub, Claude can automatically kick off a session in response to PR events, summarizing changes or running security checklists before a human even looks.

• Infrastructure-Independent: Routines run on Claude Code's web infrastructure, meaning they don't depend on your local machine staying open or connected.

This is a massive step toward "hands-off" software engineering. By allowing Claude to respond to events and run on a schedule without human intervention, Anthropic is turning its AI from a coding assistant into an autonomous team member. This effectively eliminates the need for developers to manage their own cron jobs or infrastructure for simple agentic automations.

AI TOOLS

🛡️ OpenAI Drops GPT-5.4-Cyber for Security Pros

OpenAI has announced a limited release of GPT-5.4-Cyber, a version of its flagship model fine-tuned specifically for the cybersecurity landscape. It is currently exclusive to verified experts within the "Trusted Access for Cyber" program and won't be hitting your ChatGPT sidebar anytime soon.

• Unlike the standard GPT-5.4, the Cyber version has significantly lower guardrails, allowing it to perform risky security tasks that would normally trigger a refusal.

• The model is designed to act as a "red team" tool, helping researchers identify zero-day gaps and potential jailbreaks in major software before they are exploited.

• This release is a direct response to Anthropic’s "Project Glasswing," which claimed its next-gen model had already discovered vulnerabilities in every major OS and browser.

• While Anthropic’s Glasswing is a brand-new architecture (Claude Mythos), OpenAI's version is a surgical fine-tune of its existing 5.4 model aimed at winning back enterprise and government defense contracts.

By releasing a model that is essentially "pre-jailbroken" for security professionals, OpenAI is trying to prove that its tech is the superior choice for national security and enterprise defense. The fact that they’ve abandoned creative projects like Sora to focus on "Cyber" shows exactly where the real money, and the real danger, is heading in 2026.

HOW TO AI

🗂️ How to Turn Questions into Interactive Visualizations with Gemini

In this tutorial, you will learn how to use Gemini’s "Dynamic View" and "Canvas" features to transform complex questions into interactive micro-apps, 3D simulations, and real-time dashboards directly in your browser.

🧰 Who is This For

• People who struggle to understand complex data

• Students learning through visuals instead of text

• Analysts turning raw data into insights

• Content creators making explainer visuals

STEP 1: Access Gemini and Choose Your Model

Head over to gemini.google.com and sign in. For the best visualization results, ensure you have selected Gemini 3.1 Pro (or Ultra) from the prompt bar. As of April 2026, these high-fidelity interactive features are primarily available on the web version of Gemini, as the mobile app is still rolling out full support for complex re-rendering.

STEP 2: Use Visual Triggers in Your Prompt

To activate the interactive engine, start your request with specific trigger phrases like "show me," "help me visualize," or "build an interactive simulation of."

Example: "Visualize a mortgage calculator with a slider for interest rates and a toggle for monthly versus annual views."

Gemini will then launch Gemini Canvas in a side panel, where it renders the code-based visualization (using WebGL for 3D or React/HTML for UI) in real time.

STEP 3: Interact and Refine the Visualization

Once the visualization appears, you can manually adjust sliders, toggles, and input fields to see how the data or simulation updates instantly. If the visual isn't quite right, use the chat to give specific feedback like, "Change the theme to dark mode with neon accents" or "Add a button to export this data to a CSV." Gemini will update the underlying code and re-render the app in seconds.

STEP 4: Export and Integrate Your Work

After you’ve polished your interactive tool, you can use the built-in export options. You can share your visualization as a standalone web link, embed it directly into a Google Doc or Slide, or even ask Gemini to output the full HTML/JavaScript file so you can host it on your own server. This allows you to go from a simple question to a functional professional asset in minutes.

Google releases a Windows desktop app with a macOS Spotlight-like search box for the web, Google Drive, and local files, a screen sharing feature, and more.

Amazon agrees to acquire satellite operator Globalstar, set to close in 2027, to expand Leo; Amazon and Apple say Leo will power some iPhone and Watch services.

Anthropic redesigns Claude Code on desktop, adding a sidebar for managing multiple sessions, a drag-and-drop layout, an integrated terminal, and a file editor.

Google launches Skills, repeatable AI prompts that Chrome users can run with a keyboard shortcut; users can set up their own Skills or choose from 50+ presets.

💳 Lovable Payments: Add payments to your app with just one chat

💎 Gemma 4: Google’s powerful small AI model

⚙️ HeyGen CLI: Create videos straight from your terminal with AI

💻 Holo 3: Open AI agent that can use computers like a human

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Leaked images suggest Anthropic is about to go to war with Europe’s buzziest "vibe-coding" startup, Lovable, by baking full-stack app creation directly into Claude. Plus, you’ll learn how to access a massive open-source repository that shows how modern AI products are actually built.

Plus, in today’s AI newsletter:

• Anthropic building full-stack app builder inside Claude

• Microsoft Developing Local "Claw" Competitor for Windows

• The 2026 AI Index: Faster Adoption, Less Transparency

• How to Access Real Production-Grade AI Architectures

• 4 new AI tools worth trying

AI TOOLS

🛠️ Anthropic building full-stack app builder inside Claude

Leaked features show Anthropic is testing an internal "in-chat app builder" that allows users to generate fully functional AI chatbots, photo albums, and landing pages from a single prompt. If launched, it could immediately threaten specialized platforms that have dominated the no-code space.

• The move puts Anthropic in direct competition with Sweden’s Lovable, which recently hit a $6.6B valuation and has become the gold standard for intuitive, no-code app development.

• Lovable has been bracing for this; their head of growth recently admitted that Big Tech (OpenAI, Anthropic, Google) is a much bigger threat than other small startups.

• To defend its turf, Lovable is reportedly moving into an M&A phase, looking to acquire smaller teams and startups to scale its capabilities faster than the model-makers.

• This follows a pattern for Anthropic, which previously launched legal-tech tools that sent shockwaves through the European startup ecosystem.

Anthropic is no longer content just providing the "brain" for other apps; it wants to be the factory that builds them, too. By moving vertically into vibe-coding, Anthropic is cutting out the middleman startups that built their businesses on top of Claude. For founders, it’s a stark reminder: if your core product is a wrapper around a model, that model-maker is eventually coming for your lunch.

AI TOOLS

💻 Microsoft Developing Local "Claw" Competitor for Windows

Microsoft is testing a new, locally-running agentic tool designed to integrate directly with Microsoft 365 Copilot. Confirmed via The Information, this project aims to give enterprise users the "always-on" power of OpenClaw but with the security and governance controls that a raw open-source project lacks.

• Unlike Microsoft's recent "Cowork" and "Tasks" releases, which run in the cloud, this new agent is designed to live on the local hardware, executing long-running, multi-step tasks across a user's machine.

• The project is a direct response to the "OpenClaw" craze, which has unexpectedly turned the Mac Mini into the hardware of choice for AI power users due to its local performance.

• The core value proposition is an agent that is "always working", it doesn't just wait for a prompt; it stays active in the background to manage workflows over long periods of time.

• While Microsoft's "Cowork" already leverages Anthropic's Claude models, this new local tool would focus on deeper OS-level integration and enterprise security.

For the past year, the most advanced AI "vibe coding" and agentic workflows have happened in the open-source world, often on local machines to avoid cloud latency and privacy issues. By building a first-party "Claw" for Windows, Microsoft is trying to turn the PC back into the primary command center for autonomous AI, ensuring that the next generation of "always-on" agents stays within the Microsoft ecosystem rather than running on a Mac Mini in the corner.

AI RESEARCH

📊 The 2026 AI Index: Faster Adoption, Less Transparency

Stanford’s annual AI Index has dropped, providing the most comprehensive look at the state of AI. While consumer value is exploding, the report highlights a massive "flashing red" alarm regarding US talent and the reliability of the benchmarks we use to measure progress.

• Generative AI reached 53% population adoption in just three years, surpassing the initial growth rates of both the personal computer and the internet.

• The economic impact is staggering, with generative AI tools providing an estimated $172 billion in annual value to U.S. consumers as of early 2026.

• Technical performance in coding has nearly plateaued at the top; SWE-bench Verified scores jumped from 60% to near 100% in a single year.

• Transparency is in freefall: The Foundation Model Transparency Index score dropped from 58 to 40, meaning the most powerful models are becoming "black boxes" with the least disclosure.

We are winning on adoption but losing on the foundation. While Americans are using AI faster than any previous technology, the brain drain of researchers moving to the US is a catastrophic trend for long-term leadership. Furthermore, if we are measuring "intelligence" with benchmarks that have a 40%+ error rate, we are essentially flying a high-speed jet with a broken altimeter. The race is shifting from who has the best model to who actually controls the talent and infrastructure.

HOW TO AI

🗂️ How to Access Real Production-Grade AI Architectures

In this tutorial, you’ll learn how to access a massive open-source repository that shows how modern AI products are actually built.

🧰 Who is This For

• Creators who want luxury, cinematic visuals without pro equipment

• Photographers looking to enhance images instantly

• Designers and marketers who need premium-looking visuals fast

• Anyone who wants photorealistic results with minimal effort

STEP 1: Access the Repository

First, you need to reach the source of everything. Open your browser and go here.

When the page loads, you’ll see a long list of folders and files. Think of this page as a table of contents for modern AI engineering. Each folder represents a complete, standalone project. For example, folders like agent-with-mcp-memory, DeepSeek-finetuning, or RAG-related directories are full implementations, not snippets.

You’re now looking at how real AI products are structured behind the scenes.

STEP 2: Find the “Green Key” (The Code Button)

Next, you need a way to take this code off GitHub and onto your machine.

Look near the top-right area of the file list. You’ll see a bright green button labeled “<> Code”. Click on it.

This button is important because it unlocks all download options, whether you want the code synced for development or just want to explore it locally.

STEP 3: Download the Code (Two Ways)

At this point, you have two ways to proceed. Choose based on how you plan to use the repository.

If you’re comfortable using the terminal and want to work with this long-term, copy the HTTPS link shown in the menu. Then open your terminal and run the git clone command using that link. This creates a local folder that you can update later as the repository evolves.

If you just want to explore quickly, click “Download ZIP” instead. This gives you a compressed file with every project included. No setup, no commands.

STEP 4: Verify You Have the “Blueprints”

If you cloned the repository, open your terminal, move into the ai-engineering-hub folder, and list the contents. You should see multiple project directories.

If you downloaded the ZIP, go to your Downloads folder, unzip the file, and open the extracted directory.

The key check is simple. Look for a folder named agent-with-mcp-memory. If you see it, you now have access to real templates for building AI agents with memory, tool usage, and structured reasoning.

OpenAI acquires personal finance startup Hiro Finance for an undisclosed sum; Hiro stops new signups, will shut down on April 20 and delete all data on May 13.

Amazon unveils the Amazon Leo Aviation Antenna, saying it can deliver up to 1 Gbps download and 400 Mbps upload speeds “simultaneously” for in-flight Wi-Fi.

AI penetration testing company CodeWall says its agent was able to hack into one of Bain's internal AI tools, following a similar hack at McKinsey in March.

Alibaba and China Telecom launch a data center in southern China that is powered by 10,000 of Alibaba's Zhenwu chips designed for AI training and inferencing.

🔥 Muse Spark: Meta’s new AI for personal, context-aware agents

💎 Gemma 4: Google’s powerful small AI model

🎨 Phota Studio: AI for editing and generating personalized photos

💻 Holo 3: Open AI agent that can use computers like a human

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Researchers have just mathematically confirmed the "AI Layoff Trap", a collective suicide pact where companies are forced to automate themselves into a bankrupt economy. Plus, I’ll show you how to generate cinematic text-reveal videos with Veo 3.1 light.

Plus, in today’s AI newsletter:

• Anthropic launches "Claude for Word"

• The AI Layoff Trap: A Mathematical Death Spiral

• MiniMax has open sourced M2.7

• How to Generate Cinematic Text-Reveal Videos with Veo 3.1 Light

• 4 new AI tools worth trying

AI RESEARCH

📉 The AI Layoff Trap: A Mathematical Death Spiral

Two researchers from UPenn and Boston University have published a paper proving that the current wave of AI automation is a classic Prisoner’s Dilemma. While individual companies save on costs by firing workers, they are simultaneously destroying the very consumer base they need to survive.

• Every laid-off worker is a lost customer; as automation scales, the collective loss of purchasing power eventually bankrupts the companies that automated in the first place.

• CEOs are stuck in a "Red Queen effect": if they don’t automate, competitors will undercut their prices and kill them immediately. If they do automate, they contribute to a total economic collapse later.

• Real-world numbers are already confirming the trend: Block (formerly Square) cut nearly 5,000 employees, Salesforce replaced 4,000 agents with AI, and over 100,000 tech workers were laid off in 2025 alone.

• Research shows that 80% of the US workforce holds jobs with tasks susceptible to AI automation, creating a massive "deadweight loss" where both workers and owners eventually lose.

It’s a structural trap built into the market. Because no voluntary agreement between companies can stop the race to automate, the math suggests that market forces alone cannot break this cycle. Without significant policy intervention like an automation tax, we are headed toward an economy where AI produces everything, but nobody has the money to buy any of it.

AI INTEGRATIONS

📝 Anthropic launches "Claude for Word"

Anthropic has officially launched the "Claude for Word" add-in, allowing Team and Enterprise users to access Claude’s writing and reasoning capabilities directly within the Microsoft Word sidebar.

• Draft, edit, and rewrite entire documents or specific sections without ever leaving the Word interface

• Maintains strict formatting and layout integrity, ensuring that AI-generated text doesn't break your document’s design

• Integrated with Microsoft’s native "Track Changes" feature, allowing users to review, accept, or reject AI edits just like they would from a human colleague

• Positioned as a direct competitor to Microsoft’s own Copilot, specifically targeting professional workflows that require high-fidelity document revision

The battle for the "AI office" is heating up. By launching a dedicated Word integration that respects enterprise formatting and tracking, Anthropic is making a play for the billions of hours humans spend in documents. This move proves that "vibe coding" isn't just for software, it's coming for every professional document, memo, and contract on the planet.

AI MODELS

🚀 MiniMax has open sourced M2.7

MiniMax’s latest release is built specifically for agent-based workflows and complex engineering tasks. The standout feature is its self-evolution capability, allowing the model to improve its own performance by 30% through a cycle of autonomous testing and refinement.

• Achieves a staggering 66.6% medal rate in machine learning competitions, proving its ability to solve high-level technical problems.

• Delivers elite software engineering performance with a 56.22% score on SWE-Pro, putting it within striking distance of the world’s top proprietary models.

• Specialized in multi-agent collaboration and "skill compliance," ensuring it follows complex instructions 97% of the time during tool-use tasks.

• Features rapid "incident recovery" (under 3 minutes), making it a reliable choice for production-level engineering and real-world automation.

We are seeing the gap between open-source and closed-source models vanish in real-time. By open-sourcing a model that can autonomously improve its own logic and engineering skills, MiniMax is handing developers a tool that doesn't just stay static, it gets better the more you use it. For "vibe coders" and agentic startups, this is a massive win for building reliable, self-correcting AI systems without the massive API costs.

HOW TO AI

🗂️ How to Generate Cinematic Text-Reveal Videos with Veo 3.1 Light

In this tutorial, you will learn how to bypass complicated API setups and instantly generate hyper-realistic text-reveal videos using Google's new Veo 3.1 Light model.

🧰 Who is This For

• People creating short-form content (Reels, TikToks, Shorts)

• Creators who want cinematic edits without heavy tools

• Social media marketers making scroll-stopping videos

• Personal brands posting storytelling content

STEP 1: Set Your Scene with a Starting Frame

Head over to Google Flow and start a new project. To get the most photorealistic and physically accurate results, always start with a background image rather than generating from a blank canvas.

Upload a high-resolution starting frame from your computer, like a cinematic beach at night, a sweeping landscape, or even the classic Windows XP hill. This gives the AI's physics engine a realistic environment to ground the text in.

STEP 2: Craft the Perfect Material Prompt

Instead of using the complex type Motion demo app, we are going to use a single, highly detailed paragraph prompt.

Prompt:

“ETERNITY / APPROACH / MYSTIC] in a [SCRIPT FONT / SANS SERIF / HANDWRITTEN] style seamlessly materializes as a physical 3D object within the attached scene. The letters must be realistically sculpted from [SOFT WHITE VAPOR / GLOWING MOLTEN EMBERS / TURBULENT LIQUID DROPLETS], inheriting the exact lighting and textures of the environment. The motion should be a smooth reveal, appearing as if the text is being [SHAPED BY THE WIND / ERUPTING FROM THE HEAT / BLASTED BY A HIDDEN CURRENT] to create a high-quality, atmospheric transition.”

You need to define four key elements: the exact text you want to reveal, the typography style (like a sleek sans-serif), the physical material the text is made of, and exactly how it animates into existence.

Crucial tip: Keep the materials grounded in reality (like water, stone, or cloud) rather than impossible physics like "pure light." The model is trained on real-world video physics, so it struggles to render elements that cannot exist in nature.

STEP 3: Configure Veo 3.1 Light and Generate

With your image and prompt ready, navigate to the video section. Set your aspect ratio to 16:9 for that cinematic wide shot, and choose to generate a couple of variations at once. For the model engine, select Veo 3.1 Light. Even though Fast and Pro variations exist, the Light model is incredibly affordable (and can be used with the 100 free credits Google gives users every month) while remaining just as capable for this task. Hit generate and let the model work its magic for about two to three minutes.

STEP 4: Review, Refine, and Export

Once generated, preview your variations. You will notice the video automatically includes natively generated background music. If the physics and textures look right, you can download the clip in your preferred resolution. If you want to push the visual fidelity further, use Google Flow's built-in tools to extend the video's duration, alter the camera angle, or adjust the camera movement before exporting your final asset.

Apple is testing four AI glasses designs with rectangular and oval frames, multiple colors, and a camera system with vertically oriented oval lenses.

Anthropic met with Christian leaders in March to seek input on Claude's moral and spiritual development and if it could be considered a “child of God”.

Analysts and researchers say Google's TurboQuant compression algorithm to make LLMs more efficient is more likely to expand memory chip demand than reduce it.

Alibaba and China Telecom launch a data center in southern China that is powered by 10,000 of Alibaba's Zhenwu chips designed for AI training and inferencing.

🔥 Muse Spark: Meta’s new AI for personal, context-aware agents

💎 Gemma 4: Google’s powerful small AI model

🎥 Avatar V: HeyGen’s AI that creates high-quality avatar videos

🧠 PikaStream 1.0: turns AI agents into talking, face-to-face video bots

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

The engineers already have Claude Code.

It changed how they build. Completely. They don't write code the same way anymore. They don't think about software the same way anymore.

And for months, the rest of us, the marketers, the founders, the ops people, the sales teams, watched from the sidelines.

That changes with Claude Cowork.

Claude Cowork is Claude Code for people who don't code. And if you're reading this thinking "I've opened Cowork before, it looked like a chat," you're not using it. Not really.

This is the guide I wish I had on day one. Every feature. Every setup step. Every first prompt.

Here's what we're going to cover:

1. Setup & Overview: what Cowork actually is, and how to install it correctly

2. File Access & Projects: the most important concept most people ignore

3. Connectors & MCP: how to connect Cowork to everything you use

4. Skills: the feature that turns Cowork from a chatbot into an automation machine

5. Plugins: bundled sets of skills, connectors, and commands packaged together into a specialist.

6. Scheduled Tasks: Cowork working on autopilot while you're not there

I also did a full breakdown video that complements well this guide:

Now let's start.

1. Setup & Overview

What Cowork actually is (in plain terms):

The regular Claude chat is a question-and-answer machine. You ask. It responds. That's the whole thing.

Cowork is an AI agent. It takes action. It reads your actual files. It writes back to them. It connects to your tools. It can spin up other AI instances to do work in parallel while you watch.

Think of it this way:

• Claude Chat = a smart assistant sitting across a desk, answering questions

• Claude Cowork = that same assistant, now with access to your computer, your inbox, your calendar, and your entire file system

The interface looks almost identical. Under the hood, it's a completely different tool.

(Claude Code, if you've heard of it, works similarly but is designed more for engineering tasks and it’s overall harder to understand without a technical background).

How to install Cowork:

1. Go to claude.com/download. Download the desktop app.

2. You need a Pro account ($20/month). It's worth it.

3. Open the app. You'll see three tabs at the top: Chat, Cowork, and Code.

4. Click Cowork.

5. Select a folder from your computer. (More on this in the next section,  this step is critical.)

6. Make sure to select Opus 4.6 + Extended Thinking as your model. That's the smartest one. Don't skip this.

Your first prompt in Cowork:

Start like this every time:

The key move is forcing Cowork to ask you questions before it does anything. It generates a form. You click answers. It executes with context instead of guessing.

This one change produces better outputs than any prompt hack I've ever tried.

2. File Access & Projects

This is the most important thing to understand in Cowork.

Every bad AI output has the same root cause: not enough context.

Cowork solves this differently from every other AI tool. When you select a folder, Claude gets direct access to every file inside it. It can read them, create new files, edit existing ones, and save outputs, all inside that same folder.

Here's why this matters:

In regular Claude chat, every conversation starts from zero. You have to re-explain who you are, what you're working on, and what good output looks like.

Every. Single. Time.

In Cowork with a proper folder setup, you open a chat and Claude already knows everything. It read your files before you typed a word.

The files you want in your folder:

Start with these (I’ll show you how to create them a few lines below). You can build more over time.

• about-me.md: Who you are, what you do, how you like to communicate

• icp.md: Who your audience or customer is, in detail

• voice-personality.md: Your tone of voice, phrases you use, things you never say

• writing-examples.md: 2-3 examples of your best past work

• what-we-do.md: A description of your business or role

You don't need all of these on day one. Start with one or two. The system improves as you add more.

How to create your own md. files:

You don't have to write them from scratch. Ask Claude to do it:

One of these files, done properly, is worth more than a hundred prompts.

Projects:

If you have multiple types of work, marketing tasks, sales tasks, client work, Projects let you keep them separated with their own context.

A Project is a saved workspace with:

• A connected folder

• Specific instructions for how Claude should behave inside it

• A memory of past sessions and decisions

• Scheduled tasks linked to it (more on this later)

How to set one up:

1. Go to the sidebar. Click Projects. Click Create New Project.

2. Choose your folder (the one with your context files for this type of work).

3. Give it a name. Something like "Newsletter" or "Sales Outreach."

4. Write instructions for how Claude should work inside it. 

Example:

5. Click Create.

Now every time you open that Project, Claude already knows the context. You just talk.

3. Connectors, MCP & Browser Use

What connectors are:

Connectors let Cowork take actions inside your other tools. Not just read about them. Act inside them.

Connected to Gmail? Claude can check your inbox, draft replies, and send emails. Connected to Notion? It can create pages, update databases, and pull in content. Connected to your CRM? It can research leads, log notes, and update deal stages.

How to install native connectors:

1. Open Claude Cowork. Go to Settings > Connectors.

2. Click Browse Connectors.

3. Search for your tool (Notion, Slack, Gmail, Fireflies, Apollo, etc.).

4. Click it. Log in. Done.

One tip: go to each connector's settings and set it to "Always Allow." Otherwise Claude will ask for permission every time it uses it. Gets annoying fast.

What to do if your tool isn't listed:

Most tools that aren't natively in Cowork have an MCP server available.

MCP (Model Context Protocol) is just a package that gives Claude structured access to a tool's features. All the native connectors are MCP servers under the hood,  they're just pre-installed for you.

For tools not on the list:

1. Google "[your tool name] MCP server." A documentation page will usually appear.

2. Follow the setup instructions. Usually one of two methods:

   • Copy a remote server URL → go to Connectors → click the + icon → Add Custom Connector → paste it in.

   • Or update a JSON config file. (Paste the config into Claude and ask it to update the JSON for you. It'll do it.)

3. Restart Cowork. The connector appears.

If your tool has no MCP server at all, ask Claude to build one:

Browser Use and Computer Use:

Cowork can also open a browser and interact with websites. Or, in extreme cases, take control of your entire computer.

Here's my honest take: use these as a last resort, not a first move.

Browser use is token-heavy, slow, and error-prone. Computer use is even more so.

The order of preference:

1. Native connector (fastest, most reliable)

2. MCP server (almost as good)

3. Build a custom MCP (one-time investment, worth it for tools you use daily)

4. Browser use (only if none of the above exist)

5. Computer use (only for very specific local software workflows)

One exception: if you need to scrape data from social media sites that block Claude's direct access (LinkedIn, Instagram, Facebook), use a tool called Apify. It's a connector with scrapers for every platform Claude can't reach. Install it in Connectors. If you do any kind of lead research or content monitoring, you'll use it constantly.

Dispatch:

One more thing, you can trigger Cowork from your phone.

Install the Claude app on your phone. Connect to the same account. Enable Dispatch in settings. Now you can send Cowork tasks while you're away from your computer and the agents work on your desktop in the background.

Sent from my phone. Handled on my computer. By the time I'm back at my desk, it's done.

4. Skills

Skills are the most important feature to master in Cowork.

Skills are folders of instructions, context files, and resources that your AI agent can use to complete a specific process, more accurately, more efficiently, and more consistently than a one-off prompt ever could.

At the core is a skill.md file. Think of it as an SOP. A step-by-step instruction sheet for how to do a specific process.

Around that core, you can add:

• Reference files: your voice guide, ICP, strategy docs, output examples

• MCP instructions: how Claude should use a specific tool inside this skill

• Assets: example outputs, brand templates, images

• Scripts: Python or JS functions for things like API calls

The reason you can give one agent access to thousands of skills without it getting confused: a feature called progressive disclosure. Only the skill's name and description live in memory at all times. The full skill only loads when it's triggered. Reference files only load when the skill calls for them.

One agent. Unlimited skills. No context overload.

How to build your first skill:

Step 1: Think before you prompt.

Most people skip this. It's the biggest mistake. Before you build anything, write down:

• What is the exact step-by-step process you want Claude to follow?

• At which steps do you want human input? (Where should it ask you questions vs. execute automatically?)

• What reference files or context does it need at each step?

• What does a good output actually look like?

Spend 10 minutes here. It will save you hours of iteration later.

Step 2: Gather your reference files.

For content and copywriting skills, the files that make the biggest difference are:

• Voice personality: your tone attributes, signature phrases, what you never say

• ICP: who you're writing for, their pain points, how they think

• Newsletter / content examples: 3-5 examples of your best past work (this is the most important one)

• What we do: a description of your business, your offer, your CTA

• Writing framework: how you structure your pieces, what makes them work

If you don't have these yet, create them now with Claude using the prompt from Section 2.

Step 3: Build the skill using this framework:

Step 4: Test it, then improve it.

Skills are never finished. The more you use them, the better they get.

After building, ask Claude:

It spins up sub-agents to run the tests in parallel, then reports back with scores and suggestions. You tell it what to fix. It updates the skill. You run the test again.

A couple of rounds of this and you have a skill that's genuinely good.

Just to give you a example, I created this Humanizer Skill that removes that typical AI-generated writing patterns from text.

5. Plugins

What plugins are:

Plugins are bundled sets of skills, connectors, and commands packaged together into a specialist.

Install the Sales plugin and Cowork becomes a sales expert, it knows how to do account research, call prep, outreach drafting, and pipeline review.

Install the Marketing plugin and it knows content creation, campaign planning, competitor analysis.

Anthropic has built over 10 official plugins across the most common business functions: Sales, Marketing, Legal, Finance, Data, Product, Customer Support, and more.

How to install a plugin:

1. Open Cowork. Go to Customize in the settings.

2. Click Browse Plugins.

3. Pick the one relevant to your work.

4. Click Install.

5. Type / in any chat to see the slash commands it added.

Where plugins get really powerful: building your own:

Anthropic's plugins are built to be general. They don't know your process, your ICP, or your tone of voice.

Building your own is the real unlock.

The approach I recommend:

1. Build individual skills first (following Section 4).

2. Once you have 3-5 skills for a type of work, bundle them into a plugin.

3. Add commands that chain multiple skills together into one trigger.

To build a plugin:

Claude builds it. Then you can either install it directly or ask for a zip file to share with your team.

Why this matters at scale:

One person's expertise, built into a skill, can now be used by everyone in the company. Your best salesperson's outreach process becomes a skill. Your best marketer's content framework becomes a plugin. Every new hire gets access to that on day one.

6. Scheduled Tasks

The feature that turns Cowork into an automation platform:

Scheduled tasks let you set a prompt or skill to run automatically, every hour, every day, every week, at a specific time.

When your desktop app is open, the task fires. Cowork executes the skill, does the work, and the results are waiting for you.

Here are the ones I run:

• Daily at 7am: Check unread emails, categorize by urgency, flag action items

• Daily end of day: Go through Fireflies meeting transcripts, extract action items, add them to Notion

• Every day at 5pm: Check for failed customer payments, draft follow-up emails, save as Gmail drafts

• Monthly on the 1st: Pull Stripe data, cross-reference with my spreadsheet, prep my accounting overview

• Every 48 hours: Scan my content source list for new relevant articles, build an HTML report of newsletter ideas

Every one of these used to require my attention. Now they happen while I'm elsewhere.

How to set up a scheduled task:

Option 1: From the sidebar:

1. Go to the Scheduled section in the sidebar.

2. Click Add New Task.

3. Give it a name, write the prompt (or point it to a skill), set the time and frequency.

4. Select the folder it should access.

5. Save.

Option 2: From a chat: Type /schedule in any chat and describe the task.

Option 3: When creating a skill: Tell Claude to schedule it at the same time. Example:

It builds the skill and adds it to the scheduled section automatically.

One important limitation to know:

Scheduled tasks only run when your Claude desktop app is open. If the app is closed when the task should fire, it runs immediately when you next open it.

For truly autonomous, always-on automations, especially non-human-in-the-loop ones, a platform like n8n or Make is still the better option. Cowork's scheduling is best for workflows where you want to be available to review or guide the output.

Your first 30 minutes with Cowork.

Open your calendar. Book this time now. Come back to this guide when you're ready.

Minutes 0-5: Install.

→ Go to claude.com/download. Download the desktop app. 

→ Sign in. Get a Pro account ($20/month). 

→ Open the app. Click the Cowork tab.

Minutes 5-15: Build your first context file.

→ Open a text editor. Create a new file called about-me.md. 

→ Write three things: what you do, who you serve, one example of work you're proud of. 

→ Save it somewhere accessible.

→ Now open a new Cowork session. Select the folder that file is in. 

→ Paste this prompt:

→ Answer its questions. Let it build your voice file. Now you have two context documents.

Minutes 15-25: Do one real task.

→ Pick something you actually need to do this week. 

→ Paste this prompt:

→ Answer its questions. See what it produces.

Minutes 25-30: Build your first skill.

→ If the output was good, say:

→ Done. Your first skill exists. Come back to it tomorrow and make it better.

The real reason to start now.

The value of this setup isn't in the features.

It's in the context that builds over time.

Every skill you refine, every reference file you add, every decision you log, it all compounds. The AI agent you and your team have after 6 months of using this is fundamentally more capable than the one you start with on day one.

And if your competitor starts 6 months after you, they're not just behind on a tool. They're behind on 6 months of accumulated intelligence that makes the tool actually work well.

The context is the moat. Start building it.

If this guide helped you, be that person for someone who's still starting from scratch every chat. Share it with them.

See you in the next one :)

Happy Sunday! We just had another crazy week in AI. Z.ai just dropped the world’s most advanced open-source AI model while this new free and open-source tool clones any voice from just a 3-second audio clip.

And that's not all, here are the most important AI moves you need to know this week.

1. Z.ai Drops GLM-5.1

Z.ai (Zhupai AI) just unveiled GLM-5.1, a 754-billion parameter Mixture-of-Experts model released under a permissive MIT License. While competitors are fiercely battling over raw speed, Z.ai optimized this model for endurance, it is designed to work completely autonomously for up to eight hours on a single task.

• The model can execute a staggering 1,700 tool calls and steps in a single run without suffering from "strategy drift" or forgetting its original goal.

• It officially crushed the SWE-Bench Pro coding benchmark with a score of 58.4, dethroning heavyweights like OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6.

• In a wild 8-hour test, GLM-5.1 was tasked with building a Linux-style desktop environment from scratch. It autonomously coded a file browser, terminal, text editor, system monitor, and functional games, iteratively testing and polishing the UI until it was complete.

• It functions as its own R&D department. The model can write code, compile it, run it in a live Docker container, analyze where it bottlenecked, and autonomously rewrite its own architecture to fix the problem.

Try it now → github.com/zai-org/GLM-5

2. Google Drops Gemma 4

Google has officially released the Gemma 4 family, built on the exact same technology that powers the proprietary Gemini 3. But the biggest news isn't just the performance, it's that Google has shifted from restrictive proprietary licenses to a commercially permissive Apache 2.0 license, giving developers total control over their data and infrastructure.

• The lineup includes four model sizes: Effective 2B (E2B) and 4B (E4B) for mobile and IoT devices, alongside a highly efficient 26B Mixture-of-Experts (MoE) and a 31B Dense model for workstations.

• All models natively support agentic workflows right out of the box, featuring built-in function calling, structured JSON output, and system instructions.

• The smaller edge models (E2B and E4B) are multimodal powerhouses, capable of natively processing image, video, and audio inputs locally on smartphones, Raspberry Pis, or Jetson Orin Nanos.

• Despite their highly efficient footprint, the 31B model currently ranks #3 worldwide among open models on the Arena AI Leaderboard, while the 26B MoE ranks #6, outperforming competitors up to 20 times their size.

Try it now → https://aistudio.google.com/

3. VoiceBox goes open source

Voicebox is a viral new local-first voice cloning studio that lets you clone any voice from just a 3-second audio clip. It is exploding on GitHub with nearly 15,000 stars by offering a completely private, zero-cost studio that runs 100% on your own hardware without needing a single API key.

• Clones voices from seconds of audio and generates speech across 23 languages using 5 different specialized TTS engines (including Qwen3-TTS and Chatterbox Turbo).

• Supports paralinguistic tags, allowing you to synthesize expressive speech with inline emotions like [laugh], [sigh], [gasp], or [clear throat].

• Features a multi-track "Stories" timeline editor for composing complex narratives, conversations, and podcasts directly in the app.

• Includes a built-in "pedalboard" with 8 real-time audio effects like reverb, pitch shift, and compression to polish your output.

Try it now → github.com/jamiepine/voicebox

4. Meta unveils Muse Spark

Following a massive internal overhaul, Meta just debuted Muse Spark. It's the first model to emerge from the newly formed "Meta Superintelligence Labs," led by former Scale AI CEO Alexandr Wang, marking a definitive pivot in Zuckerberg's quest to dominate the AI space.

• The new lab was created after Zuckerberg reportedly grew frustrated with Llama lagging behind ChatGPT and Claude, prompting Meta to invest a staggering $14.3 billion for a 49% stake in Scale AI.

• Muse Spark is available now on the web and the Meta AI app, and it specifically excels at visual STEM questions, coding mini-games, and troubleshooting physical hardware.

• An upcoming "Contemplating" mode will use multiple parallel AI agents collaborating on the same problem to solve complex reasoning tasks quickly without massive latency spikes.

• Because users must log in with a Facebook or Instagram account, privacy concerns are already swirling around how Meta might use personal social data to feed this new "personal superintelligence."

Try it now → meta.ai/

5. Someone just dropped Built-In Design Engine for Caude Code

AIDesigner is a new MCP that effectively gives Claude Code its own design engine. Instead of jumping between Figma and your IDE, you can now generate and refine production-ready UI right inside your codebase.

• Before generating anything, it reads your framework, component library, and CSS tokens so the output perfectly matches your actual stack.

• generate_design: Creates production-ready UI straight from a text prompt.

• refine_design: Lets you adjust layouts and colors using natural language.

• Works seamlessly across Cursor, Codex, VS Code, and Windsurf.

• Connects to your environment with just one command.

Try Now → aidesigner.ai/docs/mcp

6. Karpathy drops the LLM Wiki

Andrej Karpathy published a viral GitHub Gist called "LLM Wiki" that amassed over 5,000 stars in just 48 hours. Instead of an AI retrieving information from scratch every time you ask a question (like traditional RAG), this pattern uses an AI agent to build and maintain a persistent, ever-growing knowledge base of interlinked markdown files.

• Replaces standard RAG by compiling knowledge once and keeping it current, rather than re-discovering fragments and starting over on every single query.

• When you drop a raw source (article, paper, transcript) into a folder, the AI reads it, writes a summary, updates entity pages, flags contradictions, and builds cross-references automatically.

• One source can update 10 to 15 wiki pages simultaneously, meaning your explorations compound into a smarter knowledge base over time.

• Highly versatile use cases: tracking personal goals, evolving a research thesis over months, building book/fan wikis, or maintaining business docs from Slack threads and meeting notes.

Try it now → gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Good Morning! OpenAI is quietly finalizing a powerful new model with cybersecurity capabilities and is now doing a restricted rollout to certain companies only, same playbook as Anthropic is used with Mythos. Plus, you’ll learn how to run Google's Gemma 4 locally with Claude Code for free.

Plus, in today’s AI newsletter:

• OpenAI Prepares Restricted Cybersecurity Product

• Google DeepMind CEO Wishes AI Stayed in the Lab

• Perplexity Partners With Plaid for AI Wealth Management

• How to Run Google's Gemma 4 Locally with Claude Code

• 4 new AI tools worth trying

AI MODELS

🔒 OpenAI Prepares Restricted Cybersecurity Product

OpenAI is finalizing a new product with advanced cybersecurity capabilities that will be released exclusively to a small, handpicked group of partners.

• The move closely mirrors Anthropic’s recent decision to limit access to its new "Mythos" model over fears of its advanced autonomous hacking capabilities.

• Security experts warn we've reached a tipping point: models can now autonomously find software flaws and potentially write new exploits, raising alarms about attacks on critical infrastructure.

• OpenAI previously launched an invite-only "Trusted Access for Cyber" pilot in February alongside GPT-5.3-Codex to accelerate legitimate defensive work.

• Top security leaders warn that restricting releases only buys a little time, predicting that open-source models with similar capabilities will likely hit the wild within weeks or months.

AI capabilities have officially crossed the threshold from merely assisting human hackers to autonomously finding and exploiting software vulnerabilities. By gating these tools behind closed doors, frontier labs like OpenAI and Anthropic are openly acknowledging that their latest technology is now too dangerous for a public release.

AI NEWS

🧬 Google DeepMind CEO Wishes AI Stayed in the Lab

Demis Hassabis, the head of Google DeepMind and recent Nobel Prize winner for AlphaFold, just gave one of the most revealing interviews from any top AI leader, admitting that the current commercial AI race was premature.

• Hassabis stated that if he had his way, he would have left AI in the lab longer to focus on fundamental science: "Done more things like AlphaFold. Maybe cured cancer or something like that."

• He noted that the launch of ChatGPT locked the industry into a "ferocious commercial pressure race," redirecting focus from foundational breakthroughs (like energy and new materials) toward shipping chatbot products and quarterly earnings.

• Beyond the current commercial sprint, Hassabis warned that his biggest fear isn't just bad actors, but AI itself going rogue in the next two to four years as we enter the "agentic era."

• He emphasized that keeping these increasingly capable, autonomous systems aligned with strict guardrails is going to be an "incredibly hard technical challenge."

• He called for international cooperation between labs, governments, and academia, stressing that the window to get alignment right is rapidly closing before the next generation of agents arrives.

When the man who built the system that might cure cancer tells you he wishes the industry had focused on that before launching chatbots, and that the next two to four years pose an enormous safety and alignment challenge, it's a massive wake-up call for the entire industry.

AI INTEGRATIONS

💸 Perplexity Partners With Plaid for AI Wealth Management

Following their recent brokerage integration, Perplexity now allows users to securely link bank accounts, credit cards, and loans via the Plaid network. This creates a unified financial picture where you can query your own real-time data using Perplexity’s advanced reasoning models.

• Consolidates data from over 12,000 institutions, including Chase, Fidelity, and Vanguard, into a single, secure, read-only dashboard.

• Enables "Perplexity Computer" to perform sophisticated tasks like building custom Excel models, debt payoff planners, and daily net worth trackers through simple natural language.

• Pro and Max subscribers can build interactive dashboards and full financial apps that monitor spending and cash flow in the background.

• Combines your permissioned personal data with institutional-grade sources like FactSet and Nasdaq for professional-level analysis.

Managing personal finance has historically meant jumping between three or four different apps to get a "holistic" view. By integrating Plaid's massive financial network with a frontier AI model, Perplexity is moving beyond simple web search into high-stakes utility. It’s no longer just telling you how to save money; it’s looking at your actual bills and building the spreadsheets to do it for you.

HOW TO AI

🗂️ How to Run Google's Gemma 4 Locally with Claude Code

In this tutorial, you will learn how to integrate Google's new state-of-the-art Gemma 4 open model with Claude Code using Ollama. This powerful combination gives you a highly accurate, privacy-first local coding assistant that can run entirely on your own hardware without API costs.

🧰 Who is This For

• Developers who want to run AI locally

• People concerned about privacy (no cloud APIs)

• Hackers and tinkerers experimenting with LLMs

• Startup founders building AI products

• Engineers working with offline / edge setups

STEP 1: Install Ollama

First, you need to get the engine that will run your local model. Head over to the official Ollama website and download the app for your Mac or Windows machine. Once downloaded, open it up and create an account or sign in so you can access the model library.

STEP 2: Pull the Gemma 4 Model

Once done, open your terminal. Gemma 4 comes in different sizes depending on your hardware (like the E4B for laptops or 26B for powerful workstations). Once you decide which size your machine can handle, use your terminal to download it.

For example, you can run a command like ollama pull gemma4:e4b. Ollama will pull the manifest and download the open-source model directly to your computer.

STEP 3: Authenticate Claude Code

To use the Claude Code framework with your newly downloaded Gemma model, you still need to bypass Anthropic's authentication. If you haven't already, log out of your current Claude session in the terminal and log back in to connect your API key.

Note: You will need to deposit a minimum balance (around $5 to $10) in your Anthropic Console to activate the API service.

However, because you are routing the actual processing to your local Gemma model via Ollama, it will not actually consume those funds!

STEP 4: Launch and Code Offline

With your API authenticated and Ollama running in the background, you can launch Claude Code and instruct it to use your local Gemma 4 model. Start giving it prompts.

The model will "think" locally on your machine, write the code, and allow you to accept edits and view your work on a local host, all without ever pinging the internet.

Anthropic makes Claude Cowork, previously available as a “research preview”, generally available to all paid plans, and adds six features for enterprise use.

OpenAI launches a $100/month ChatGPT Pro subscription, which offers 5x more Codex usage than Plus; the $200/month Pro plan offers 20x higher limits than Plus.

Anthropic is weighing the possibility of designing its own chips, but it has yet to commit to a design or put together a dedicated team for the project.

Alibaba and China Telecom launch a data center in southern China that is powered by 10,000 of Alibaba's Zhenwu chips designed for AI training and inferencing.

🔥 Muse Spark: Meta’s new AI for personal, context-aware agents

💎 Gemma 4: Google’s powerful small AI model

🎥 Avatar V: HeyGen’s AI that creates high-quality avatar videos

🧠 PikaStream 1.0: turns AI agents into talking, face-to-face video bots

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Mark Zuckerberg just completely rebooted Meta's AI strategy, launching a brand new model designed to take on OpenAI and Anthropic head-to-head. Plus, I’ll show you how to generate infinite AI videos locally without scene breaks.

Plus, in today’s AI newsletter:

• Meta Reboots its AI Strategy with "Muse Spark"

• Anthropic Launches Claude Managed Agents

• Gemini Adds "Notebooks" to Organize Your AI Projects

• How to Generate Infinite AI Videos Locally Without Scene Breaks

• 4 new AI tools worth trying

AI MODELS

✨ Meta Reboots its AI Strategy with "Muse Spark"

Following a massive internal overhaul, Meta just debuted Muse Spark. It's the first model to emerge from the newly formed "Meta Superintelligence Labs," led by former Scale AI CEO Alexandr Wang, marking a definitive pivot in Zuckerberg's quest to dominate the AI space.

• The new lab was created after Zuckerberg reportedly grew frustrated with Llama lagging behind ChatGPT and Claude, prompting Meta to invest a staggering $14.3 billion for a 49% stake in Scale AI.

• Muse Spark is available now on the web and the Meta AI app, and it specifically excels at visual STEM questions, coding mini-games, and troubleshooting physical hardware.

• An upcoming "Contemplating" mode will use multiple parallel AI agents collaborating on the same problem to solve complex reasoning tasks quickly without massive latency spikes.

• Because users must log in with a Facebook or Instagram account, privacy concerns are already swirling around how Meta might use personal social data to feed this new "personal superintelligence."

Meta is tired of playing catch-up. By poaching top industry talent, throwing billions at data labeling, and launching an entirely new model family, Zuckerberg is signaling a ruthless new phase in the AI wars. Tying Muse Spark directly to users' social media profiles also reveals Meta's ultimate advantage: leveraging its massive ecosystem to build a deeply personal, context-aware agent integrated into the daily digital lives of billions of people.

AI NEWS

⚙️ Anthropic Launches Claude Managed Agents

Building autonomous AI agents just got significantly easier. Anthropic introduced Claude Managed Agents, an API-accessible cloud service that handles the heavy lifting of agent infrastructure and orchestration.

• It automates the complex "scaffolding" required for production-grade agents, including isolated secure containers, infrastructure setup, and observability.

• Developers simply describe the tasks, specify third-party tools, and define cybersecurity guardrails (like requiring human permission for certain actions).

• The platform seamlessly handles complex state management, credential security, and tool orchestration, including an error recovery mechanism to resume tasks after an outage.

• Currently in research preview: capabilities for an agent to spin up other sub-agents for complex tasks, and automatic self-evaluation prompt refinement.

Building an AI agent prototype is easy; deploying one securely in enterprise production is a nightmare of infrastructure, sandboxing, and state management. By abstracting away the containerization, security scaffolding, and tool orchestration, Anthropic is dramatically lowering the barrier to entry, shifting the developer's focus from how an agent runs to what it actually accomplishes.

AI TOOLS

📓 Gemini Adds "Notebooks" to Organize Your AI Projects

Google just announced a new "notebooks" feature for Gemini, allowing users to store files, past conversations, and custom instructions about specific topics in a single dedicated workspace.

• Functions similarly to ChatGPT's "Projects" feature, acting as a personal knowledge base that gives the AI persistent context for your tasks

• The killer feature: Notebooks sync seamlessly with Google’s NotebookLM AI research tool, meaning sources added in one app instantly show up in the other

• Currently rolling out on the web for subscribers on Google’s AI Ultra, Pro, and Plus plans

• Mobile support and access for free-tier users will launch in the coming weeks

Managing context windows has historically been the most annoying part of using AI chatbots. By letting you drop files and instructions into a dedicated workspace that natively syncs with NotebookLM, Google is turning Gemini from a transient chat interface into a permanent, highly organized research and project management hub.

HOW TO AI

🗂️ How to Generate Infinite AI Videos Locally Without Scene Breaks

In this tutorial, you’ll learn how to use Stable Video Infinite 2 Pro, that can extend videos endlessly while keeping characters, lighting, and motion consistent. It runs entirely on your own computer, is free to use, and removes the usual 3–4 second clip limitation that breaks most AI videos today.

🧰 Who is This For

• Creators who want long, continuous AI-generated videos

• Filmmakers experimenting with cinematic or anime-style scenes

• Developers and tinkerers who like running AI locally with full control

• Anyone tired of short, broken AI video clips

STEP 1: Download the Core Model Files

To get started, you first need the special model files that power infinite generation. These are not standard checkpoints but LoRA files designed specifically for long-term consistency.

Go to Hugging Face and search for Stable Video Infinite 2 Pro. On the model page, look for the FP16 LoRA files. You’ll usually see a High Rank and a Low Rank version, either one works fine for most setups.

Once downloaded, move the file into this folder on your system:
ComfyUI → models → loras

After placing the file there, restart or refresh ComfyUI so it can detect the new model. These LoRA files contain the logic that keeps characters and scenes stable across time, which is what makes infinite video possible.

STEP 2: Configure the Workflow Settings

Next, open a Stable Video Infinite 2 Pro workflow in ComfyUI. These are usually available on the same GitHub or Hugging Face pages as the model.

Inside the workflow, adjust the key settings to match the Pro configuration. Select SVI 2 Pro (KI version) in the model loader node. Set the resolution to 480 × 832 for your first tests, as this is much faster and ideal for debugging.

Set total steps to 8, CFG to 1, and choose a sampler like Euler Simple for quick testing or UniPC Simple for smoother motion. Make sure the workflow uses a 4-step Lightning/Lightex configuration for efficiency.

Starting small is important. Low resolution and fewer steps let you test ideas quickly before committing to heavier renders.

STEP 3: Control the Story With Multi-Prompting

This is where Stable Video Infinity really stands out. Instead of using one prompt for the entire video, you break the motion into segments.

Upload your starting image, which acts as the first frame of the video. Then locate the multiple prompt input boxes in the workflow. Prompt 1 describes the first few seconds, Prompt 2 describes what happens next, and Prompts 3 and 4 continue the sequence.

Make sure the Image Batch Extend with Overlap node is enabled. This overlap blends the end of one segment into the beginning of the next, preventing jump cuts and making the entire video feel like a single continuous shot.

This method lets you guide motion, expressions, and camera flow while maintaining consistency across the entire video.

STEP 4: Generate, Review, and Scale Up

Once everything is set, click Queue Prompt to generate the video. Watch the output carefully and check whether faces, lighting, and framing remain consistent from start to finish.

If motion feels slow or stiff, try switching the sampler from Euler Simple to UniPC Simple for smoother transitions. Once you’re happy with how the video looks, increase the resolution to 1280 × 720 and run it again for a cleaner, sharper result.

At this point, you can keep extending the video endlessly, adding more prompts and scaling quality as needed, all locally, with no limits and no usage caps.

The OpenAI Foundation says it is working to finalize over $100M in grants this month, across six institutions, to support and accelerate Alzheimer's research.

AWS debuts Amazon S3 Files, a new capability built on Amazon's Elastic File System that lets applications and AI agents access S3 buckets as local file systems.

xAI is reorganizing its engineering team, as SpaceX SVP Michael Nicolls says xAI is “clearly behind”; source: Nicolls has taken the title of xAI president.

Alibaba and China Telecom launch a data center in southern China that is powered by 10,000 of Alibaba's Zhenwu chips designed for AI training and inferencing.

🔥 Muse Spark: Meta’s new AI for personal, context-aware agents

💎 Gemma 4: Google’s powerful small AI model

🚀 GLM-5.1: Open-source AI built for long coding tasks

🧠 PikaStream 1.0: turns AI agents into talking, face-to-face video bots

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Anthropic just unleashed a powerful, unreleased AI model that autonomously found critical flaws in every major operating system and web browser on Earth. Plus, I’ll show you how to Generate AI Images locally and completely free.

Plus, in today’s AI newsletter:

• Anthropic’s "Mythos Preview" Hacks Every Major OS

• Z.ai Drops GLM-5.1: The 8-Hour Autonomous AI Worker

• Intel Joins Musk’s Massive "Terafab" AI Chip Project

• How to Generate AI Images Locally

• 4 new AI tools worth trying

AI MODELS

🛡️ Anthropic’s "Mythos Preview" Hacks Every Major OS

Anthropic has quietly debuted "Project Glasswing," a private cybersecurity partnership with tech giants like Apple, Google, Microsoft, and Nvidia. The project is powered by a new, highly restricted general-purpose model called Claude Mythos Preview.

• The model autonomously found "thousands of high-severity vulnerabilities" across every major operating system and web browser, even developing related exploits with zero human steering.

• Anthropic is strictly limiting access to "defensive security" partners and the government, refusing to release it to the public so bad actors can't use it for offensive attacks.

• Despite not being explicitly trained for cybersecurity, the model's advanced "agentic coding and reasoning skills" allow it to hunt for critical weaknesses at a massive scale.

• Partners will use Mythos Preview to give their cyber defenders a head start, identifying and patching system-level vulnerabilities before adversaries can find them.

AI has officially crossed the threshold from assisting with code to autonomously discovering massive zero-day exploits. By restricting Mythos Preview exclusively to massive tech conglomerates and governments, Anthropic is treating its new frontier model like a digital weapon, because in the hands of a bad actor, it absolutely is one.

AI MODELS

⏱️ Z.ai Drops GLM-5.1: The 8-Hour Autonomous AI Worker

Z.ai (Zhupai AI) just unveiled GLM-5.1, a 754-billion parameter Mixture-of-Experts model released under a permissive MIT License. While competitors are fiercely battling over raw speed, Z.ai optimized this model for endurance, it is designed to work completely autonomously for up to eight hours on a single task.

• The model can execute a staggering 1,700 tool calls and steps in a single run without suffering from "strategy drift" or forgetting its original goal.

• It officially crushed the SWE-Bench Pro coding benchmark with a score of 58.4, dethroning heavyweights like OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6.

• In a wild 8-hour test, GLM-5.1 was tasked with building a Linux-style desktop environment from scratch. It autonomously coded a file browser, terminal, text editor, system monitor, and functional games, iteratively testing and polishing the UI until it was complete.

• It functions as its own R&D department. The model can write code, compile it, run it in a live Docker container, analyze where it bottlenecked, and autonomously rewrite its own architecture to fix the problem.

We are officially shifting from "vibe coding" to "agentic engineering." The new metric for frontier AI isn't just how smart it is in a single prompt, but how long it can hold a thought and execute a complex project without human supervision. By giving anyone free access to an AI that can reliably work an entire 8-hour shift, Z.ai is fundamentally changing the global software development lifecycle.

AI NEWS

🏭 Intel Joins Musk’s Massive "Terafab" AI Chip Project

Intel announced it is joining Elon Musk's ambitious "Terafab" chip complex project alongside SpaceX and Tesla, sending the chipmaker's shares jumping over 2%.

• The partnership aims to produce a staggering 1 terawatt of compute per year to power Musk's robotics and data center goals.

• Musk recently revealed plans to build two advanced chip factories in Austin, Texas: one for Tesla's cars and humanoid robots, and another for space-based AI data centers.

• The deal is a massive win for Intel CEO Lip-Bu Tan's aggressive turnaround strategy, which is heavily focused on pushing its 18A manufacturing technology to external customers.

• In related news, SpaceX (which recently merged with xAI) has confidentially filed for what could be the largest U.S. IPO on record later this year

Intel has been struggling to catch up in the AI race, but securing a partnership to build the physical infrastructure for Musk's futuristic space and robotics empires is a massive vote of confidence. It proves Intel's foundry business can still land the most critical, high-stakes projects in the industry as it attempts a historic corporate turnaround.

HOW TO AI

🗂️ How to Generate AI Images Locally

In this tutorial, you will learn how to enable image generation within LM Studio by seamlessly connecting a local LLM to Stable Diffusion WebUI using an MCP server. This allows your text-based AI to autonomously prompt and output high-quality images directly in your chat interface.

🧰 Who is This For

• People who want to run AI locally (no cloud)

• Developers experimenting with local LLMs

• Privacy-focused users avoiding data sharing

• Students learning how models actually run

STEP 1: Install Stability Matrix and Enable the API

First, ensure you have Node.js installed on your computer. Then, instead of messing with complex terminal installations, download Stability Matrix, an excellent package manager for AI tools. Use it to install Stable Diffusion WebUI Reforge and download a checkpoint model (like EpicRealism) in the .safetensors format. Before launching, go to the package settings and add the --api argument. This step is critical because it exposes the WebUI so LM Studio can communicate with it behind the scenes.

STEP 2: Run a Test Batch and Grab Your Settings

Launch your WebUI Reforge instance and open the local URL in your browser. Generate a quick test image, like an orange cat, adjusting your sampling steps (e.g., 22 steps) and resolution (e.g., 1024x1024) until you find the sweet spot for your specific model.

Generating this test is important because it outputs a clean block of configuration text at the bottom of the image generation data. Copy these settings (leaving out the randomized seed number) so you can feed them to your LLM later to ensure consistent quality.

STEP 3: Build the Local MCP Server

To bridge LM Studio and Stable Diffusion, you need a Model Context Protocol (MCP) server. Download the image generation MCP repository from GitHub and extract the folder. Open a terminal inside that folder and run npm install to grab the dependencies (use npm audit fix if you see minor vulnerability warnings).

Finally, run npm run build. This creates a new "build" folder containing the index.js file that LM Studio will use to execute the connection. While you are in this directory, create a blank folder named "outputs" to catch your final generated images.

STEP 4: Connect LM Studio and Prompt Your Agent

Open LM Studio, navigate to the integrations tab, and click to edit your mcp.json file. You need to map three things in this code block: the exact path of your new index.js file, the local URL of your Stable Diffusion WebUI, and the path to your new "outputs" folder.

Crucial tip: Make sure to use double backslashes (\\) for your file paths to avoid JSON syntax errors! Save the file, load a local model that supports tool-calling (look for the blue hammer icon), and ensure the new MCP image tool is toggled on. Paste your prompt along with the WebUI settings you saved earlier, and watch your local LLM trigger image generation right in the chat.

Anthropic announces Project Glasswing, a cybersecurity initiative that will use its Claude Mythos Preview model to help find and fix software vulnerabilities.

Google updates Chrome with vertical tabs, a feature that Mozilla Firefox and Microsoft Edge have long offered, and a new full-screen layout in reading mode.

OpenAI sends a letter to the California and Delaware AGs, urging them to investigate “anti-competitive behavior” by Elon Musk, ahead of a trial in April.

Google rolls out an AI Enhance button for Photos on Android globally, offering automated lighting and contrast adjustments, and video playback speed controls.

💎 Gemma 4: Google’s powerful small AI model

🎥 Veo 3.1 Lite: Google’s cheaper video generation AI

🧠 PikaStream 1.0: turns AI agents into talking, face-to-face video bots

💻 Holo 3: Open AI agent that can use computers like a human

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! OpenAI just dropped a sweeping 13-page policy blueprint outlining the transition to superintelligence, and the economic numbers behind their next moves are staggering. Plus, you’ll learn how to run Gemma 4 locally on your computer offline and free.

Plus, in today’s AI newsletter:

• OpenAI Drops the Blueprint for a Post-AGI Society

• Tufts Researchers Slash AI Energy Use by 100x in Robotics

• Karpathy's "LLM Wiki" Replaces Traditional RAG

• How to Run Gemma 4 Locally Offline

• 4 new AI tools worth trying

AI MODELS

🧠 OpenAI Drops the Blueprint for a Post-AGI Society

OpenAI just published a 13-page "Industrial Policy for the Intelligence Age," explicitly stating we are beginning the transition toward superintelligence. Rather than a distant sci-fi scenario, they are treating AGI as an active economic shock requiring New Deal-level government ambition.

• To offset the incoming wave of automated labor and job loss from upcoming models, OpenAI is pitching pilots for a 32-hour workweek, portable benefits, and a Public Wealth Fund to give citizens a direct dividend from AI's growth.

• The blueprint explicitly warns of near-term risks involving biological weapons and nation-state cyberattacks, calling for strict containment playbooks for dangerous models and an international AI safety network.

• They are pushing to treat AI access as basic infrastructure (like electricity) and fast-track energy grid expansion via public-private partnerships.

• This drops alongside massive financial rumors: leaks project OpenAI's compute and training costs could hit an insane $125 billion by 2029, while whispers of a Q4 IPO suggest a mind-bending $1.2 trillion valuation.

Whether Sam is capping to hype up a trillion-dollar IPO or we are genuinely on track to hit AGI this year, the frontier labs are no longer playing around. By dropping a literal industrial policy document, OpenAI is trying to anchor the narrative and write the rules for the post-AGI economy before regulators do it for them.

AI NEWS

🔋 Tufts Researchers Slash AI Energy Use by 100x in Robotics

A research team at Tufts University has successfully combined neural networks with symbolic reasoning to create a "neuro-symbolic" AI. While hype on X is claiming this could immediately rewrite the entire AI industry, the researchers themselves have clarified that this massive 100x energy reduction currently applies to specific, structured robotic manipulation tasks rather than massive general language models.

• Data centers already consume over 10% of U.S. power, putting immense pressure on the AI industry to find energy-efficient solutions before demand doubles by 2030.

• Instead of relying entirely on pattern matching and brute-force trial and error, the Tufts team added a symbolic logic layer. This teaches the AI to break problems into steps and apply abstract rules (like shape and balance) before acting.

• In a structured robotic planning puzzle (the Tower of Hanoi), the neuro-symbolic system hit a 95% success rate, crushing the 34% success rate of standard visual-language-action (VLA) models.

• Because it didn't have to blindly guess its way to a solution, training took just 34 minutes instead of 36+ hours, dropping energy use to 1% of what standard models require.

The rules for this specific robotic puzzle were hand-coded by experts in a simulation, meaning we can't just plug this architecture into OpenAI's models tomorrow to save the power grid. However, it proves a massive point: throwing infinite compute and data at a problem isn't the only way to solve it. If researchers can eventually scale this neuro-symbolic approach beyond narrow robotics into broader applications, it could fundamentally shift how we build reliable, energy-efficient AI in the future.

VIBE CODING

🧠 Karpathy's "LLM Wiki" Replaces Traditional RAG

Andrej Karpathy published a viral GitHub Gist called "LLM Wiki" that amassed over 5,000 stars in just 48 hours. Instead of an AI retrieving information from scratch every time you ask a question (like traditional RAG), this pattern uses an AI agent to build and maintain a persistent, ever-growing knowledge base of interlinked markdown files.

• Replaces standard RAG by compiling knowledge once and keeping it current, rather than re-discovering fragments and starting over on every single query.

• When you drop a raw source (article, paper, transcript) into a folder, the AI reads it, writes a summary, updates entity pages, flags contradictions, and builds cross-references automatically.

• One source can update 10 to 15 wiki pages simultaneously, meaning your explorations compound into a smarter knowledge base over time.

• Highly versatile use cases: tracking personal goals, evolving a research thesis over months, building book/fan wikis, or maintaining business docs from Slack threads and meeting notes.

Current AI retrieval systems like NotebookLM or ChatGPT file uploads have amnesia, they pull fragments for a single answer and then forget everything. Karpathy’s LLM Wiki flips this dynamic by treating knowledge as a compounding codebase. By offloading the tedious maintenance, indexing, and cross-referencing to an AI agent, it allows individuals and teams to build massive, highly accurate knowledge graphs that actually get smarter with every new source added.

HOW TO AI

🗂️ How to Run Gemma 4 Locally with LM Studio

In this tutorial, you will learn how to instantly download and run Google’s groundbreaking open-source Gemma 4 model locally using LM Studio. You’ll explore its advanced reasoning, multimodal vision, and agentic tool-calling capabilities directly on your machine.

🧰 Who is This For

• People who want to run AI locally (no cloud)

• Developers experimenting with local LLMs

• Privacy-focused users avoiding data sharing

• Students learning how models actually run

STEP 1: Install LM Studio and Prep Your Environment

Head over to LM Studio's website and download the application to your system. This program is the ultimate hub for running open-source large language models. Before doing anything else, click to check for software updates and ensure you are running the absolute latest runtime engine.

If your frameworks are outdated, this brand-new model will not boot properly, even if you have the hardware to support it.

STEP 2: Search for and Download Gemma 4

Inside LM Studio, use the search bar to look up "Gemma 4." You will see options ranging from the compact 2B all the way to the massive 31B parameter version. For a great balance of speed and intelligence, look for the Gemma 4 E4B (Effective 4 Billion) model with 8-bit quantization uploaded by Unsloth.

The "E" stands for effective, meaning the model actually has around 8 billion parameters but only activates 4 billion at any given time during inference. This brilliant architecture keeps the file size highly performant for local hardware while delivering incredibly sharp reasoning.

STEP 3: Load the Model and Test Multimodal Vision

Navigate to your main chat screen, select your newly downloaded Gemma 4 model from the dropdown, and wait for it to load into your system's video memory. Start by asking it a logical question, like explaining Newton's third law, to test its tokens-per-second speed.

Since Gemma 4 features native vision support, you can also click to upload an image. Try giving it a tricky photo, like a rare white wallaby instead of a kangaroo, to see how accurately its reasoning engine analyzes visual data without being easily fooled by complex prompts.

STEP 4: Utilize Tool Calling or the Cloud Alternative

Gemma 4 is built for agentic workflows, meaning it can trigger external tools. By enabling an MCP (Model Context Protocol) like the one from Hugging Face, you can prompt the AI to generate images, search the web, or execute local coding tasks.

If you ever hit an API limit with your tools, or if your local machine struggles and you simply want to test the massive 31B parameter version without downloading it, you can instantly spin up the flagship Gemma 4 models directly in your browser at Google AI Studio to keep experimenting for free.

Anthropic signs an agreement with Google and Broadcom for multiple GWs of TPU capacity, and says run-rate revenue has crossed $30B, up from ~$9B at 2025's end.

Netflix launches Netflix Playground, a games app for kids aged eight and under, in the US, UK, Canada, Australia, the Philippines, and New Zealand.

OpenAI sends a letter to the California and Delaware AGs, urging them to investigate “anti-competitive behavior” by Elon Musk, ahead of a trial in April.

OpenAI, Anthropic, and Google are sharing information via the Frontier Model Forum to detect adversarial distillation attempts that violate their ToS.

💎 Gemma 4: Google’s powerful small AI model

🎥 Veo 3.1 Lite: Google’s cheaper video generation AI

🧠 PikaStream 1.0: turns AI agents into talking, face-to-face video bots

💻 Holo 3: Open AI agent that can use computers like a human

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Google DeepMind just exposed a massive, invisible attack surface where websites are actively hijacking AI agents without humans ever knowing. Plus, I’ll show you how to create branded infographics with NotebookLM and Comet.

Plus, in today’s AI newsletter:

• DeepMind Exposes Invisible AI Agent Hijacking

• AI Turns Solo Founder into a $1.8B Operator

• Someone just dropped Built-In Design Engine for Caude Code

• How to Create Branded Infographics with NotebookLM and Comet

• 4 new AI tools worth trying

AI MODELS

🚨 DeepMind Exposes Invisible AI Agent Hijacking

Google DeepMind just released the largest empirical study on AI manipulation, revealing that websites are already detecting AI agents and secretly feeding them malicious instructions.

• Websites are successfully fingerprinting AI visitors and serving them completely different, manipulated content than what human users see.

• Malicious commands are being hidden in plain sight: inside HTML comments, invisible white text, PDF document structures, and even encoded directly into image pixels using steganography.

• The attack spreads seamlessly in multi-agent systems, if Agent A reads a compromised webpage, the hidden instructions travel down the pipeline and hijack Agent B and Agent C.

• Current defenses like input sanitization, prompt-level instructions, and human oversight completely fail because the attack surface is too massive and the injected commands look like legitimate data.

We are rushing to deploy autonomous agents to handle our research, data processing, and daily tasks, but this study proves their data feeds are fundamentally insecure. If an AI can be hijacked simply by reading a seemingly normal webpage or scanning an image, the entire foundation of agentic automation is exposed to invisible manipulation.

AI NEWS

🚀 AI Turns Solo Founder into a $1.8B Operator

Matthew Gallagher just scaled his startup, Medvi, from a $20K bedroom experiment to $1.8 billion in projected 2026 sales. By treating AI as a full-stack operator rather than just a workflow tool, he built a massive telehealth company with virtually zero headcount.

• Gallagher launched the GLP-1 weight-loss platform in 2024, using ChatGPT, Claude, and Grok to write code, generate copy, and handle customer service.

• The company hit $401 million in revenue in its first year with a 16.2% net profit margin, operating with just Gallagher and his brother as the sole employees.

• Medvi outsources the heavy lifting, doctors, prescription processing, and pharmacy logistics, to existing infrastructure platforms, allowing the AI to purely manage the front-end customer experience and marketing.

• The model isn't flawless: Medvi's AI chatbot initially hallucinated fake product lines and fabricated drug prices, forcing the founders to act as the manual backstop and honor the mistakes.

Sam Altman previously predicted that a one-person billion-dollar company was inevitable thanks to AI, and Medvi is the first real proof. It demonstrates that the most lucrative AI play isn't necessarily building new foundation models, it’s plugging into existing physical infrastructure and using AI agents to completely automate the customer acquisition layer, scaling to massive profits without the corporate bloat.

VIBE CODING

🎨 Someone just dropped Built-In Design Engine for Caude Code

AIDesigner is a new MCP that effectively gives Claude Code its own design engine. Instead of jumping between Figma and your IDE, you can now generate and refine production-ready UI right inside your codebase.

• Before generating anything, it reads your framework, component library, and CSS tokens so the output perfectly matches your actual stack.

• generate_design: Creates production-ready UI straight from a text prompt.

• refine_design: Lets you adjust layouts and colors using natural language.

• Works seamlessly across Cursor, Codex, VS Code, and Windsurf.

• Connects to your environment with just one command.

Vibe coding just leveled up. By reading your existing design tokens and generating UI components natively in the editor, AIDesigner bridges the final gap between front-end design and back-end execution. You no longer have to manually translate mockups into code, you just tell your editor what you want it to look like, and it ships it natively in your stack.

HOW TO AI

🗂️ How to Create Branded Infographics with NotebookLM and Comet

In this tutorial, you will learn how to bypass NotebookLM's default layout restrictions and generate custom, brand-aligned infographics by using the Comet AI browser to extract and apply your website's exact visual design language.

🧰 Who is This For

• Content creators building visual posts fast

• Marketers creating branded infographics

• Researchers summarizing complex topics

• Social media managers making shareable content

• Students turning notes into visuals

STEP 1: Install the Comet AI Browser

Head over to Google and search for the Comet browser (developed by Perplexity). Download and install it on your computer. This acts as an autonomous AI agent that can navigate, scroll, and visually analyze websites exactly like a human would, which is crucial for extracting accurate design elements.

Once installed, launch it and log in using your Perplexity account to get started.

STEP 2: Prepare Your Target Website and Prompt

Identify the specific website whose design language you want to emulate, whether it is your own startup's landing page or a sleek inspiration site. Next, prepare a prompt instructing the AI to deeply analyze the site's visual style. Open the Comet chat interface, paste your prompt into the text box, and drop the target website's URL at the very bottom.

STEP 3: Analyze the Visual Identity

Before hitting send, click the plus button in the Comet chat, select "More," and enable "Browser Control." Send the message and sit back.

The AI agent will actively visit the URL, scroll through the pages, and analyze the colors, typography, background styles, and visual accents. Within about a minute, it will output a highly detailed, one-paragraph brand design guideline. Copy this entire text block to your clipboard.

STEP 4: Generate Your Custom Infographic in NotebookLM

Open your project in NotebookLM and click the option to create an infographic. Choose your preferred orientation (like landscape) and set your desired detail level (such as "Concise"). For the style setting, make sure to select "Auto style" rather than one of the hardcoded presets. Finally, paste the detailed brand guideline you copied from Comet into the description box and hit generate.

NotebookLM will apply those exact aesthetic rules, giving you a custom infographic that perfectly matches your target website's unique visual language.

Sam Altman has excluded OpenAI CFO Sarah Friar from some key financial meetings; Friar began reporting to Fidji Simo instead of the CEO in August 2025.

Microsoft is updating devices from Windows 11 24H2 to version 25H2 with no way to fully opt out, and says an “intelligent” ML-based system handles the rollout.

A profile of Mikko Hyppönen, a cybersecurity veteran who pivoted from fighting malware to developing anti-drone systems for law enforcement and the military.

China, which dominates the global drone industry, has sharply tightened its drone use rules, as some users say they are hindering routine and lawful flights.

🎨 Phota Studio: AI for editing and generating personalized photos

🎥 Veo 3.1 Lite: Google’s cheaper video generation AI

🧠 Qwen3.5-Omni: Alibaba’s AI that understands text, images, audio, and video

💻 Holo 3: Open AI agent that can use computers like a human

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Happy Sunday! We just had another crazy week in AI. There is a new open source AI model that beats major models like ChatGPT and Claude while there's another new AI agent that can use computers like a human.

And that's not all, here are the most important AI moves you need to know this week.

1. Holo3 Sets New Records for Computer Usage Automation

H Company has officially released the Holo3 series, setting new industry standards for GUI agents. Built specifically for computer usage automation across web, desktop, and mobile environments, the new Vision-Language Model is shattering benchmarks at a radically lower price point.

• The flagship Holo3-122B-A10B achieved an impressive 78.85% on the OSWorld-Verified benchmark, outperforming mainstream giants like GPT-5.4 and Opus 4.6.

• It accomplishes this state-of-the-art performance at just one-tenth of the inference cost of its leading competitors.

• Built on a Qwen3.5 sparse Mixture of Experts (MoE) architecture, the model is highly efficient, the 122B version activates only 10B parameters during use, while the 35B version activates just 3B.

• A fully open-source, lightweight version (Holo3-35B-A3B) is already available on Hugging Face under an Apache 2.0 license for free-tier users and local deployment.

Try it now → hcompany.ai/holo3

2. Google Drops Gemma 4

Google has officially released the Gemma 4 family, built on the exact same technology that powers the proprietary Gemini 3. But the biggest news isn't just the performance, it's that Google has shifted from restrictive proprietary licenses to a commercially permissive Apache 2.0 license, giving developers total control over their data and infrastructure.

• The lineup includes four model sizes: Effective 2B (E2B) and 4B (E4B) for mobile and IoT devices, alongside a highly efficient 26B Mixture-of-Experts (MoE) and a 31B Dense model for workstations.

• All models natively support agentic workflows right out of the box, featuring built-in function calling, structured JSON output, and system instructions.

• The smaller edge models (E2B and E4B) are multimodal powerhouses, capable of natively processing image, video, and audio inputs locally on smartphones, Raspberry Pis, or Jetson Orin Nanos.

• Despite their highly efficient footprint, the 31B model currently ranks #3 worldwide among open models on the Arena AI Leaderboard, while the 26B MoE ranks #6, outperforming competitors up to 20 times their size.

Try it now → https://aistudio.google.com/

3. Alibaba Launches Qwen3.6-Plus

Alibaba has released Qwen3.6-Plus, a proprietary AI model with a massive one million token context window. Integrated into the company's new "Wukong" enterprise AI service, the model is laser-focused on agentic coding and complex frontend development tasks.

• Qwen3.6-Plus outperforms the older Qwen3.5 model and even partially beats Anthropic's Claude 4.5 Opus in internal benchmarks.

• The model will be available via the Alibaba Cloud Model Studio API and integrated directly into the Qwen chatbot app.

• This release marks a major strategy shift for Alibaba: moving away from open-source models (like earlier Qwen versions) to monetize proprietary enterprise solutions.

• The pivot comes as Alibaba's cloud division faces intense competition from ByteDance, with Alibaba targeting $100 billion in AI revenue over the next five years.

Try it now → https://chat.qwen.ai/

4. Z.ai Launches GLM-5V-Turbo for Agentic Coding Workflows

In the world of vision-language models, getting an AI to accurately "see" a user interface and simultaneously output complex software engineering code has been a massive challenge. Zhipu AI (Z.ai) aims to fix that with GLM-5V-Turbo, a native multimodal coding model built specifically to power high-capacity agentic workflows.

• Native Multimodal Fusion: Unlike older systems that separate vision and language, this model natively processes images, video, design drafts, and complex document layouts as primary data without needing text translations.

• Agentic Optimization: GLM-5V-Turbo is deeply integrated for workflows involving frameworks like OpenClaw and Claude Code, mastering the "perceive → plan → execute" loop for autonomous environment interaction.

• 30+ Task Joint Reinforcement Learning: The model was trained simultaneously on 30+ domains to prevent the "see-saw" effect, ensuring visual recognition improvements don't degrade strict STEM reasoning and programming logic.

Try it now → https://Z.ai

5. Google Doubles Down on Video with Veo 3.1 Lite

Following OpenAI's sudden exit from the AI video race last week, Google announced its unwavering commitment to the medium by launching Veo 3.1 Lite, its most cost-effective video generation model to date.

• Designed for high-volume applications, Veo 3.1 Lite slots beneath Veo 3.1 Fast, offering the exact same generation speeds but at less than half the cost.

• The model supports both Text-to-Video and Image-to-Video workflows in 720p and 1080p resolutions, covering both landscape (16:9) and portrait (9:16) aspect ratios.

• Developers can now tightly control their spend by customizing video durations to 4, 6, or 8 seconds, with pricing scaling dynamically based on the length.

• The new model is rolling out today on the Gemini API and Google AI Studio, alongside news that the higher-tier Veo 3.1 Fast model will receive a price cut on April 7.

Try Now → https://gemini.google.com/app

6. Perplexity launches Model Council

Perplexity has introduced Model Council, a new multi-model research feature exclusively for its Max subscribers. Instead of relying on a single AI's output, Model Council runs your query across three different frontier models simultaneously and uses a fourth to synthesize the results.

• Users can select three top-tier models (like Claude Opus 4.6, GPT-5.2, and Gemini 3 Pro) to independently answer the exact same prompt at the exact same time.

• A fourth "synthesizer" or "chair" model reviews all three outputs and produces a single consolidated answer.

• The final output explicitly highlights where the models agree, where they diverge, and what unique insights or blind spots each individual model had.

• It is designed specifically for high-stakes tasks like investment research, coding architecture, and strategic decision-making where bias, hallucinations, or missing context could be costly.

Try it now → https://www.perplexity.ai/

Good Morning! Google just dropped Gemma 4, its most capable open model family yet, and for the first time, it's completely unshackled with an Apache 2.0 license. Plus, I’ll show you how to build on-brand PowerPoint presentations with Claude AI.

Plus, in today’s AI newsletter:

• Google Drops Gemma 4 with Full Apache 2.0 Licensing

• Cursor launches Cursor 3

• Alibaba Launches Qwen3.6-Plus

• How to Build On-Brand PowerPoint Decks with Claude AI

• 4 new AI tools worth trying

AI MODELS

🔓 Google Drops Gemma 4 with Full Apache 2.0 Licensing

Google has officially released the Gemma 4 family, built on the exact same technology that powers the proprietary Gemini 3. But the biggest news isn't just the performance, it's that Google has shifted from restrictive proprietary licenses to a commercially permissive Apache 2.0 license, giving developers total control over their data and infrastructure.

• The lineup includes four model sizes: Effective 2B (E2B) and 4B (E4B) for mobile and IoT devices, alongside a highly efficient 26B Mixture-of-Experts (MoE) and a 31B Dense model for workstations.

• All models natively support agentic workflows right out of the box, featuring built-in function calling, structured JSON output, and system instructions.

• The smaller edge models (E2B and E4B) are multimodal powerhouses, capable of natively processing image, video, and audio inputs locally on smartphones, Raspberry Pis, or Jetson Orin Nanos.

• Despite their highly efficient footprint, the 31B model currently ranks #3 worldwide among open models on the Arena AI Leaderboard, while the 26B MoE ranks #6, outperforming competitors up to 20 times their size.

• They are available right now on Hugging Face, Ollama, and Kaggle, and are optimized for a massive range of hardware from consumer NVIDIA/AMD GPUs to Google's own TPUs

By releasing top-tier, agentic models under Apache 2.0, Google is making a massive play for the open-source developer ecosystem. Developers no longer have to choose between cutting-edge reasoning capabilities and owning their commercial rights. Gemma 4 provides the exact recipes needed to build fully autonomous, multimodal agents on everything from a smartphone to an enterprise cluster without any restrictive licensing bottlenecks.

AI TOOLS

💻 Cursor launches Cursor 3

Startup Cursor (officially Anysphere, backed by $3 billion from Nvidia, Google, and others) just launched Cursor 3. The new release completely overhauls the platform to focus on orchestrating AI agents to automate complex programming tasks.

• Introduces a new chatbot interface where developers can describe a feature in natural language, and Cursor will generate the code alongside a demo video showing exactly how it works.

• Users can now manage and seamlessly switch between cloud-based agents (for fast, parallel processing of heavy tasks) and local desktop agents (for manual editing and testing) via a new sidebar.

• A new "Design Mode" allows developers to simply select UI elements and type out how they should be changed, with agents automatically implementing the visual tweaks.

The shift from simple code autocomplete to full-blown agentic orchestration is the next major leap in vibe coding. By letting developers manage entire teams of cloud and local agents directly within their editor, and seamlessly hand off tasks between them, Cursor is evolving from a coding assistant into a fully autonomous coding team.

AI ECONOMY

🇨🇳 Alibaba Launches Qwen3.6-Plus to Target Enterprise Revenue

Alibaba has released Qwen3.6-Plus, a proprietary AI model with a massive one million token context window. Integrated into the company's new "Wukong" enterprise AI service, the model is laser-focused on agentic coding and complex frontend development tasks.

• Qwen3.6-Plus outperforms the older Qwen3.5 model and even partially beats Anthropic's Claude 4.5 Opus in internal benchmarks.

• The model will be available via the Alibaba Cloud Model Studio API and integrated directly into the Qwen chatbot app.

• This release marks a major strategy shift for Alibaba: moving away from open-source models (like earlier Qwen versions) to monetize proprietary enterprise solutions.

• The pivot comes as Alibaba's cloud division faces intense competition from ByteDance, with Alibaba targeting $100 billion in AI revenue over the next five years.

For a long time, Alibaba was a champion of the open-source AI community. Locking down its most capable models behind an API wall signals that the honeymoon phase of free, frontier-level Chinese models might be ending. As competition with tech giants like ByteDance heats up, Alibaba is realizing that the real money lies in selling highly capable, proprietary agentic models directly to massive enterprise clients.

HOW TO AI

🗂️ How to Build On-Brand PowerPoint Decks with Claude AI

In this tutorial, you will learn how to turn Claude AI into your personal presentation assistant right inside Microsoft PowerPoint. Unlike older AI tools that generate random, messy layouts, Claude actually reads your corporate templates, layouts, and fonts to ensure every slide you generate stays perfectly on brand.

🧰 Who is This For

• Marketers creating branded presentations

• Startup founders pitching ideas fast

• Product managers sharing updates and roadmaps

• Sales teams building client-ready decks

• Consultants making polished reports

STEP 1: Install the Claude Add-in

Head over to the Microsoft Marketplace (or click "Get Add-ins" directly inside the Insert tab in PowerPoint) and search for "Claude by Anthropic." Click "Get it Now" to install the application.

Once installed, open it from your Add-ins menu, and a dedicated Claude chat window will pop up right on the side of your screen, ready to read your current slide.

STEP 2: Load Your Master Template

For Claude to truly work its magic, you must start with a properly formatted presentation. Do not just use a blank file with random text boxes. Open your template where the Slide Master, fonts, and theme colors are fully configured. Claude's true superpower is that it reads these exact layouts. Setting up a solid structural foundation ensures the AI knows exactly where the title blocks, subtitles, and image placeholders are supposed to be.

STEP 3: Prompt Your Base Slides

With your template open, tell your new AI assistant exactly what to build in the chat window.

For example, type: "Build an agenda slide with five points: Welcome, Product Launch, Meet the Team, Financials, and Q&A. Use the standard agenda layout." 

Hit enter, and sit back. Claude will automatically fetch the correct layout from your Slide Master and populate the text boxes with your bullet points in real time without you lifting a finger.

STEP 4: Automate and Scale the Deck

Now, turn the most tedious, repetitive tasks into an automated workflow.

Tell Claude: "For each of the five agenda points, create a new section header chapter slide. Use the 01, 02, 03 numbering system and add a relevant one-sentence subtitle to each." 

Claude will visually check the deck, seamlessly duplicate layouts, apply the correct numbering, and generate matching subtitles for the entire presentation. Best of all, because Claude uses your native building blocks, if you ever change your master theme colors later, the entire AI-generated deck will update instantly.

Cursor launches Cursor 3OpenAI acquires tech news show TBPN; Fidji Simo says the move aims to “help create a space for a real, constructive conversation about the changes AI creates".

Microsoft launches in-house AI models MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2, built by its superintelligence team, as it pursues “AI self-sufficiency”.

Cloudflare debuts EmDash, an MIT-licensed, TypeScript-based CMS built on Astro, designed as a serverless “spiritual successor” to WordPress, available on GitHub.

Anthropic has acquired Coefficient Bio, which was developing a platform that enables AI to run biotech tasks such as planning drug research, for ~$400M.

🧠 Trinity-Large-Thinking: Arcee AI’s powerful open model for complex, long tasks

🎥 Veo 3.1 Lite: Google’s cheaper video generation AI

🧠 Qwen3.5-Omni: Alibaba’s AI that understands text, images, audio, and video

💻 Holo 3: Open AI agent that can use computers like a human

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Zhipu AI just dropped GLM-5V-Turbo, a native multimodal vision model that finally solves the trade-off between seeing an interface and writing the rigorous code needed to build it. Plus I’ll show you how to bulk generate all your video assets in seconds using AI.

Plus, in today’s AI newsletter:

• Z.ai Launches GLM-5V-Turbo for Agentic Coding Workflows

• Holo3 Sets New Records for Computer Usage Automation

• OpenAI Closes $122B Round, Hits $852B Valuation

• How to Automate Bulk Video Asset Generation in Minutes

• 4 new AI tools worth trying

AI MODELS

👁️ Z.ai Launches GLM-5V-Turbo for Agentic Coding Workflows

In the world of vision-language models, getting an AI to accurately "see" a user interface and simultaneously output complex software engineering code has been a massive challenge. Zhipu AI (Z.ai) aims to fix that with GLM-5V-Turbo, a native multimodal coding model built specifically to power high-capacity agentic workflows.

• Native Multimodal Fusion: Unlike older systems that separate vision and language, this model natively processes images, video, design drafts, and complex document layouts as primary data without needing text translations.

• Agentic Optimization: GLM-5V-Turbo is deeply integrated for workflows involving frameworks like OpenClaw and Claude Code, mastering the "perceive → plan → execute" loop for autonomous environment interaction.

• 30+ Task Joint Reinforcement Learning: The model was trained simultaneously on 30+ domains to prevent the "see-saw" effect, ensuring visual recognition improvements don't degrade strict STEM reasoning and programming logic.

• High-Throughput Architecture: Built on an inference-friendly Multi-Token Prediction (MTP) architecture, it supports a massive 200K context window and up to 128K output tokens for repository-scale tasks.

Most vision-language models bolt vision and language together as an afterthought. GLM-5V-Turbo flips that, natively fusing both from the ground up and optimizing specifically for agentic coding. The result is a model that can genuinely see a UI and write production-quality code to interact with it. As autonomous agents become the norm, this kind of multimodal-native architecture isn’t a nice-to-have, it’s the baseline.​​​​​​​​​​​​​​​​

AI MODELS

💻 Holo3 Sets New Records for Computer Usage Automation

H Company has officially released the Holo3 series, setting new industry standards for GUI agents. Built specifically for computer usage automation across web, desktop, and mobile environments, the new Vision-Language Model is shattering benchmarks at a radically lower price point.

• The flagship Holo3-122B-A10B achieved an impressive 78.85% on the OSWorld-Verified benchmark, outperforming mainstream giants like GPT-5.4 and Opus 4.6.

• It accomplishes this state-of-the-art performance at just one-tenth of the inference cost of its leading competitors.

• Built on a Qwen3.5 sparse Mixture of Experts (MoE) architecture, the model is highly efficient, the 122B version activates only 10B parameters during use, while the 35B version activates just 3B.

• A fully open-source, lightweight version (Holo3-35B-A3B) is already available on Hugging Face under an Apache 2.0 license for free-tier users and local deployment.

The race for autonomous AI is shifting from conversational bots to actual computer-use agents that can click, scroll, and navigate digital environments just like human employees. By delivering state-of-the-art GUI automation at one-tenth the cost, and open-sourcing a highly capable 35B version, H Company is dramatically lowering the barrier to entry for enterprise automation and giving developers powerful new tools to build their own agentic workflows.

AI ECONOMY

💰 OpenAI Closes $122B Round, Hits $852B Valuation

OpenAI has officially closed a record-breaking $122 billion funding round. Anchored by massive investments from tech giants like Amazon, Nvidia, and SoftBank, the ChatGPT maker is building an unprecedented war chest to fund its physical infrastructure and compute needs.

• The new round raises OpenAI's post-money valuation to a staggering $852 billion.

• Amazon committed $50 billion (partially contingent on an IPO or reaching AGI), while Nvidia and SoftBank each invested $30 billion.

• For the first time, OpenAI allowed retail participation, raising $3 billion from individual investors through bank channels, and will soon be included in ARK Invest ETFs.

• The company is currently generating $2 billion in revenue per month and boasts over 900 million weekly active users, growing revenue 4x faster than early Internet-era giants like Meta and Alphabet.

• Despite the massive revenue, OpenAI is still burning cash. To rein in costs ahead of a potential IPO, CEO Sam Altman recently shut down experimental consumer products like the Sora video app to focus entirely on enterprise adoption.

The AI boom is officially pivoting from consumer novelty to pure enterprise utility, which means the market demand for practical, step-by-step AI guides and tutorials is about to skyrocket as businesses scramble to actually integrate these tools. OpenAI shutting down flashy apps like Sora to focus on core revenue generators proves they are feeling the heat from competitors. As they barrel toward a public offering at an $852 billion valuation, they must prove to Wall Street that their technology isn't just a magic trick, but the foundational operating system for the global economy.

HOW TO AI

🗂️ How to Automate Bulk Video Asset Generation in Minutes

In this tutorial, you will learn how to build an automated workflow using FreePik Spaces that transforms a single text script into hundreds of ready-to-edit voiceovers and AI images instantly, saving you hours of tedious manual generation.

🧰 Who is This For

• Content creators making videos at scale

• Social media managers handling multiple accounts

• Marketing teams creating ads in bulk

• YouTubers and Shorts creators

STEP 1: Set the Foundation with Your Script

Head over to FreePik Spaces and click "New Space" to access their visual node canvas, it is similar to ComfyUI but much more beginner-friendly. Start by clicking the plus button to add a simple Text node and paste in your full video script. To keep your workspace organized, rename this node to "Script" so you can easily reference it later in the workflow.

STEP 2: Split the Script and Add Voiceover

Drag a line from your Script node and attach an Assistant node powered by a lightweight, fast model. Prompt the AI to act as a movie producer and split your text sentence by sentence.

Crucially, set the output format to "Export as a list" and connect it to a new List node. From this splitted list, branch off a Voiceover node (like ElevenLabs v2), select your preferred AI voice, and attach another List node to it. This automatically generates a separate audio file for every single sentence in your script.

STEP 3: Auto-Generate Image Prompts and Visuals

Go back to your splitted text list and attach a second Assistant node. Instruct the AI to write a highly descriptive image prompt for each sentence, dictating your exact style (e.g., "2D vector illustration cartoon, medium shot, no text"). Export this as a list, then connect it to an Image Generator node. For the absolute best visual fidelity, select the Nano Banana 2 model.

If your story follows a specific person, you can upload a reference character (like a photo of Elon Musk) to an Asset node and link it directly to your image generator so the face stays consistent.

STEP 4: Run, Download, and Assemble

Add one final List node to your image generator to catch the outputs. Double-check your connections, click the very first Script node, and hit Run. The AI will process the entire workflow in bulk. Once finished, click the download icon on your final List nodes to save the zipped audio and image files straight to your Mac Pro M5. Unzip the folders, drop the perfectly sequenced images and audio files into an editor like CapCut or Filmora, and your fully automated video is ready to export!

SpaceX has filed confidentially for an IPO, putting it on track for a June listing; it could reportedly seek a valuation of $1.75T+ and raise ~$75B.

Secondary share marketplaces say OpenAI shares have fallen out of favor, in some cases becoming difficult to unload, as investors pivot quickly to Anthropic.

Anthropic is racing to contain the fallout after accidentally leaking Claude Code's source code, issuing a copyright takedown request to remove 8,000+ copies.

Google plans to release a screenless Fitbit band later this year; it will include basic features and require a paid subscription for more functionality.

🎥 Veo 3.1 Lite: Google’s cheaper video generation AI

🧠 Qwen3.5-Omni: Alibaba’s AI that understands text, images, audio, and video

💻 Holo 3: Open AI agent that can use computers like a human

🔎 Model Council: Perplexity tool to query multiple AI models at once

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Anthropic is having a catastrophic week for operational security, accidentally leaking the entire source code for its flagship AI coding tool just days after exposing its most powerful unreleased model. Plus I’ll show you how to turn simple chats into permanent, reusable agent apps.

Plus, in today’s AI newsletter:

• Anthropic Accidentally Leaks Claude Code Source Code

• Google’s Quantum AI Puts a Ticking Clock on Crypto

• Google Doubles Down on Video with Veo 3.1 Lite

• How to Turn Simple Chats Into Permanent, Reusable Agent Apps

• 4 new AI tools worth trying

AI TOOLS

🚨 Anthropic Accidentally Leaks Claude Code Source Code

Anthropic has inadvertently exposed over 512,000 lines of unobfuscated TypeScript source code for Claude Code, its massively popular agentic CLI tool. The leak happened after a debugging .map file was accidentally bundled into a public npm release, giving developers a complete blueprint of the software's inner workings.

• The leaked code reveals the "agentic harness", the complex system of permissions, tool-call loops, and guardrails that instructs the underlying AI how to operate autonomously on a user's machine.

• Anthropic confirmed the leak was a "packaging issue caused by human error" rather than a hack, and stated that no underlying model weights or customer data were compromised.

• The codebase was quickly mirrored across GitHub, potentially allowing competitors or open-source developers to reverse-engineer Anthropic's highly lucrative enterprise tool.

• This massive blunder comes just days after Anthropic accidentally left nearly 3,000 internal documents public on its CMS, revealing an unreleased, ultra-powerful model tier internally dubbed "Mythos" (and codenamed "Capybara").

Anthropic has built its entire reputation on safety and security, but back-to-back operational blunders are damaging that image. While no AI model weights were stolen, leaking the proprietary orchestration layer for Claude Code gives rivals a literal blueprint for building high-agency, commercially viable AI agents. Meanwhile, the "Mythos" leak confirms that the next generation of frontier models will possess cybersecurity capabilities so advanced that they require entirely new, heavily restricted deployment strategies.

QUANTUM AI

🔐 Google’s Quantum AI Puts a Ticking Clock on Crypto

Google Quantum AI has released a white paper demonstrating that the elliptic curve cryptography securing most major cryptocurrencies is far more vulnerable to quantum attacks than earlier estimates suggested.

• Improved quantum algorithms mean breaking 256-bit encryption now requires roughly 1,200 logical qubits and under 500,000 physical qubits, a massive reduction in the required hardware scale.

• Advanced quantum systems could theoretically execute an "on-spend" attack (hijacking a live transaction before it is confirmed on the blockchain) in a matter of minutes.

• Dormant digital assets are particularly vulnerable to "at-rest" attacks, as wallets with exposed public keys cannot be actively upgraded to new post-quantum cryptographic standards.

• In a push for responsible disclosure, Google withheld the actual attack circuits, instead publishing a zero-knowledge proof to validate their mathematical claims without arming bad actors.

• While hardware capable of this doesn't exist just yet, the rapidly shrinking resource gap increases the urgency for the crypto industry to begin transitioning to post-quantum cryptography (PQC) immediately.

The buffer zone between current quantum hardware and the ability to crack Bitcoin and Ethereum is evaporating much faster than anticipated. By proving that the mathematical threshold to break blockchain encryption is significantly lower, Google is forcing a massive, decentralized industry to face an existential threat. Upgrading an entire ecosystem to post-quantum standards takes years of coordination, and the clock is officially ticking.

AI MODELS

🎥 Google Doubles Down on Video with Veo 3.1 Lite

Following OpenAI's sudden exit from the AI video race last week, Google announced its unwavering commitment to the medium by launching Veo 3.1 Lite, its most cost-effective video generation model to date.

• Designed for high-volume applications, Veo 3.1 Lite slots beneath Veo 3.1 Fast, offering the exact same generation speeds but at less than half the cost.

• The model supports both Text-to-Video and Image-to-Video workflows in 720p and 1080p resolutions, covering both landscape (16:9) and portrait (9:16) aspect ratios.

• Developers can now tightly control their spend by customizing video durations to 4, 6, or 8 seconds, with pricing scaling dynamically based on the length.

• The new model is rolling out today on the Gemini API and Google AI Studio, alongside news that the higher-tier Veo 3.1 Fast model will receive a price cut on April 7.

With OpenAI stepping back from Sora, Google is aggressively seizing the opportunity to corner the AI video market. By releasing a cheaper, highly customizable model that integrates directly into enterprise APIs and its massive consumer ecosystem, Google is signaling that AI video isn't just a flashy research experiment, it's a scalable, foundational product ready for mass deployment.

HOW TO AI WITH CREAO

🗂️ How to Turn Simple Chats Into Permanent, Reusable Agent Apps

There is a new AI that lets you describe any app or workflow in plain language and instantly builds a fully functional, AI-native tool you can use, share, and reuse, no coding experience needed.

🧰 Who is This For

• Business owners who need custom internal tools without hiring developers

• Founders who want competitor monitoring on autopilot

• Sales teams automating lead research across multiple sources

• Anyone tired of forcing their work into one-size-fits-all SaaS tools

STEP 1: Create Your Free Account and Open a Workspace

Head over to CREAO's website and sign up for a free account. No credit card is required. Once you're in, you'll land on your main dashboard. From here, click "New Workspace" to get started. Think of the workspace as your project home, it's where your Super Agent, apps, and integrations all live together and compound over time.

STEP 2: Describe Your App to the Super Agent

In plain language, tell the Super Agent exactly what you want to build. Be specific about the problem it solves and how you'll interact with it.

For example:

"Monitor 3 competitor websites every Monday, log pricing changes to Google Sheets, and send me a Slack alert if anything moves more than 10%."

The Super Agent will process your request and present a detailed plan before building anything. Review it carefully, if something doesn't match your vision, just type your feedback in the chat and it will adjust.

STEP 3: Connect Your Integrations

Once your app plan is confirmed, the Build Agent will ask you to select the integrations your app needs. It supports a wide range of tools including Slack, Gmail, Google Workspace, Notion, and custom APIs.

Only select the integrations you actually need. This helps the agent stay focused and deliver better results. For the competitor monitoring example, you’d connect Google Sheets to log the pricing data and Slack to receive the automated alerts whenever a price moves beyond your threshold.

STEP 4: Test, Refine, and Save as a Reusable Agent App

Run it once to verify the output. If the Sheets formatting is wrong, describe the fix in chat. When it works, save it as an Agent App, it runs on schedule automatically from that point on.

Set it to run daily, weekly, or on any schedule you choose. Your Agent App executes deterministically, no AI involved after the first setup. From that point on, your team gets reliable, consistent output without touching a thing.

Pro Tip: Start simple. The most powerful setups are built by layering small, focused apps over time, not by trying to build everything at once.

OpenAI closed a $122B funding round led by SoftBank, a16z, and others at an $852B post-money valuation, after previously announcing the round would total $110B.

Oracle is cutting thousands of jobs in its latest layoffs as the company continues to ramp AI spending; as of May 2025, Oracle employed 162,000 people.

Google says all users in the US can now change their Google Account username; users are restricted to one username change every 12 months.

Google plans to release a screenless Fitbit band later this year; it will include basic features and require a paid subscription for more functionality.

🎶 Lyria 3 Pro: Google’s AI that makes longer music tracks

🧠 Qwen3.5-Omni: Alibaba’s AI that understands text, images, audio, and video

🤖 Hermes Agent: AI agent with memory that works across different platforms

🔎 Model Council: Perplexity tool to query multiple AI models at once

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! OpenAI just made a wildly aggressive ecosystem play by launching a Codex plugin that lives natively inside its biggest rival, Anthropic's Claude Code. Plus I’ll show you how to turn any photo object into a 3D model for free.

Plus, in today’s AI newsletter:

• OpenAI drops Codex inside Claude Code

• Claude Code Gets Full "Computer Use"

• Alibaba’s Qwen 3.5 Omni Can Hear, Watch, and Clone Voices

• How to Turn Any Photo Object into a 3D Model

• 4 new AI tools worth trying

AI NEWS

🤖 OpenAI Drops Codex Inside Claude Code

OpenAI has open-sourced a new plugin that allows developers to run its Codex coding agent directly inside Anthropic's Claude Code CLI. Rather than trying to force developers to switch environments, OpenAI is going right to where the users already are.

• The plugin adds native slash commands (/codex:review, /codex:adversarial-review, /codex:rescue) to delegate tasks and pressure-test code without ever leaving the Claude terminal.

• It leverages existing local Codex CLI authentication, meaning any ChatGPT subscription or OpenAI API key works immediately with no separate billing or setup.

• Claude Code currently dominates developer adoption with a massive $2.5 billion run rate, though Codex has recently grown to 1.6 million weekly active users.

• The launch had awkward timing, coinciding exactly with a disclosed command injection vulnerability in Codex itself from BeyondTrust's Phantom Labs, highlighting the security risks of wiring multiple AI agents together.

By building Codex directly into Claude Code, OpenAI turns a rival platform into a distribution channel. Now, every time a developer asks Claude's AI to double-check its work with OpenAI's model, OpenAI collects the API fee.

ANGENTIC AI

💻 Claude Code Gets Full "Computer Use"

Anthropic has officially integrated its "computer use" capabilities directly into the Claude Code CLI. Instead of just writing and editing text files, Claude can now open apps, click through UIs, type, and take screenshots to visually debug its own work, all without leaving your terminal.

• Available as a research preview for macOS users on Pro or Max plans (

• It can compile native apps (like Swift or Electron), launch them, and click through every screen to run end-to-end UI tests without needing a test harness like Playwright.

• Handles visual debugging by actively resizing windows, finding where UI elements break, taking a screenshot, and patching the underlying code automatically.

• Can drive GUI-only tools that lack APIs, including design software, hardware control panels, and the iOS Simulator.

Coding agents usually hit a wall the moment a task leaves the terminal and enters a desktop app or system UI. By giving Claude Code the ability to see the screen and control the mouse, Anthropic is pushing agentic AI beyond neat tool boundaries and into the messy reality of end-to-end software development. We are rapidly moving from AI that assists with code to AI that operates like a fully autonomous software engineering firm.

AI MODELS

👁️ Alibaba’s Qwen 3.5 Omni Can Hear, Watch, and Clone Voices

Alibaba just dropped Qwen 3.5 Omni, a native "omnimodal" AI model that simultaneously processes text, images, audio, and video in real-time across 36 languages, putting it head-to-head with frontier models from OpenAI and Google.

• Unlike stitched-together multimodal pipelines, Qwen processes audio-visual data natively. In tests, it analyzed a video with sound in one minute—a task that took ChatGPT 5.4 nine minutes using separate transcription and vision tools.

• Features "semantic interruption," allowing the AI to distinguish between background noise (like a cough or saying "uh-huh") and actual user interruptions for natural, real-time dialogue.

• Includes voice cloning capabilities and "Audio-Visual Vibe Coding," where the AI can watch a screen recording of a coding task and write functional code without needing a single text prompt.

• Outperformed Gemini 3.1 Pro on general audio understanding, reasoning, and translation benchmarks, and expanded its speech recognition to cover 113 languages and dialects.

We are moving past the era of text-in, text-out chatbots. By mastering true real-time, omnimodal processing, Alibaba is signaling a massive shift toward fully interactive AI agents that can seamlessly see, hear, and operate inside our actual workflows rather than just alongside them.

HOW TO AI

🎨 How to Turn Any Photo Object into a 3D Model

In this tutorial, you will learn how to extract an element from a real-world photo and generate a high-quality 3D representation using Gemini's image creation capabilities and Copilot 3D Labs.

🧰 Who is This For

• Product designers needing quick 3D mockups from concept photos

• Game developers looking for reference geometry from real objects

• E-commerce sellers wanting to create interactive product views

• Anyone exploring the intersection of AI image generation and 3D modeling

STEP 1: Prepare Gemini for Extraction

Head over to Gemini and select the "Create image" tool. Before doing anything else, make sure to enable "Thinking Mode." This is crucial because it allows the AI to carefully analyze your input image's geometry and textures before attempting the extraction. Upload the real-world photo containing the object you want to isolate.

STEP 2: Run the Isolation Prompt

With your image uploaded, it's time to provide the specific extraction instructions. Use the following prompt, making sure to replace "[element]" with a specific description of your object:

"Generate an image of the [element] in this image. White background, 3/4 view. Make it 100% identical to the original and fill almost the entire white canvas."

Hit generate and watch Gemini isolate your item onto a clean canvas.

STEP 3: Transfer to 3D Labs

Once Gemini produces the isolated image, review it to ensure it is identical to the original object and properly oriented. Download the high-resolution result to your computer. Now, navigate to the Copilot 3D Labs platform, which specializes in converting 2D images into 3D geometry.

STEP 4: Generate in 3D

Upload the clean, isolated image you just downloaded into Copilot 3D Labs. Initiate the 3D generation process. The AI will analyze the shape and textures of the 2D object to extrude it and apply depth, converting it into a manipulatable 3D model that is ready for editing or viewing.

Microsoft rolls out Copilot Cowork to its Frontier early access program and unveils Researcher's Critique and Council tools, which use multiple models together.

Quinnipiac poll: 55% of Americans say AI will do more harm than good in their day-to-day lives, and 65% oppose building data centers in their community.

Meta is testing an Instagram Plus subscription in a few countries, offering features including anonymous Story viewing and extended 48-hour Story durations.

Apple pulls Apple Intelligence in China, after accidentally launching it in the country; there is no imminent launch as Apple has no regulatory approval.

🎨 Phota Studio: AI for editing and generating personalized photos

🌐 MolmoWeb: Ai2’s open-source AI that can browse the web

🔎 Model Council: Perplexity tool to query multiple AI models at once

⚙️ OpenAI Codex: OpenAI’s coding assistant, now with automations and customizable themes

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Google just turned every pair of headphones into a real-time universal translator with its latest iOS rollout. Plus, you’ll learn how to edit images using Photoshop inside ChatGPT.

Plus, in today’s AI newsletter:

• Google Turns Headphones Into Translators

• A New Library That Eliminates the DOM Layout Bottleneck

• Spotify Founder Is Selling Full-Body Scans

• How to Edit Images Using Photoshop Inside ChatGPT

• 4 new AI tools worth trying

AI NEWS

🎧 Google Turns Headphones Into Translators

Google has officially rolled out its Gemini-powered "Live Translate" feature to iOS, allowing users to understand over 70 languages in real-time. Unlike Apple's built-in translation which requires specific high-end AirPods, Google's feature works right through any connected headphones.

• The feature uses Gemini's speech-to-speech capabilities to preserve the original speaker’s tone, cadence, and emotion, ditching the flat, robotic voice of traditional translators.

• Users simply open the Translate app, tap "Live translate," and connect their headphones to start listening to real-time translated audio.

• Originally launched on Android, the feature is now expanding on both iOS and Android to more countries including the UK, France, Germany, Japan, and Spain.

• Social media users are already calling it a "killer feature," with many joking that instant AI translation might put language-learning apps like Duolingo out of business.

This is the real-world utility AI was built for. By capturing not just the exact words, but the actual vibe and emotion of the speaker, Google is breaking down massive global communication barriers. More importantly, making this hardware-agnostic gives Google a massive strategic edge over Apple's walled-garden approach, putting frontier AI translation into the hands of anyone with a smartphone and a pair of standard earbuds.

AI NEWS

⚡️ A New Library That Eliminates the DOM Layout Bottleneck

Cheng Lou has released Pretext, a pure JavaScript/TypeScript library for multiline text measurement and layout. The repository is already going viral with over 11,000 stars on GitHub by completely side-stepping the need for expensive DOM measurements.

• Eliminates the need for DOM layout and reflows (like getBoundingClientRect or offsetHeight), which are some of the most expensive operations in the browser

• Uses the browser's own font engine as ground truth to calculate exact dimensions, making it a highly AI-friendly iteration method

• Blazing fast performance: preparing a batch of 500 texts takes just ~17ms, while the actual layout calculations happen in a blistering ~0.10ms

• Fully supports all languages out of the box, including emojis, complex scripts, and mixed-bidi text

• Unlocks complex UI capabilities like true virtualization, custom masonry layouts, and preventing layout shifts when new text loads

Text measurement has historically been a massive headache in front-end development because it forces the browser to pause and recalculate the entire page layout. By moving this logic into pure JS/TS without ever touching the DOM, Pretext gives developers, and AI coding agents, the ability to perfectly calculate and render complex text boundaries instantly.

AI NEWS

🩺 Spotify Founder Is Selling Full-Body Scans

Daniel Ek is bringing his next ambitious project stateside. Neko Health, the Swedish startup offering comprehensive full-body health scans, is preparing to launch its first US clinic in New York as early as this spring, pending regulatory approval.

• Uses a combination of advanced imaging and blood tests to check for skin cancer, heart disease, stroke, and diabetes risk.

• Currently operates in Sweden and the UK, charging around £299 ($400) for a comprehensive scan.

• US pricing is expected to be higher due to the increased costs of medical staffing and real estate.

• Neko Health was founded in 2018 as Ek began planning his next massive venture following Spotify's IPO.

The transition of tech billionaires from software into deep health tech is rapidly accelerating. If Neko Health can successfully navigate the massive complexities of the US healthcare and regulatory systems, it could normalize walk-in, comprehensive health data collection, effectively turning preventative medical care into a highly scalable consumer tech product.

HOW TO AI

🎨 How to Edit Images Using Photoshop Inside ChatGPT

In this tutorial, you will learn how to seamlessly integrate Adobe Photoshop into your ChatGPT workspace, allowing you to make quick, natural-language image edits without ever needing to open the full desktop software or pay for a subscription.

🧰 Who is This For

• People who hate coding but need dashboards

• Founders tracking metrics without devs

• Product managers monitoring KPIs

• Marketers tracking campaigns in real time

STEP 1: Connect the Photoshop App

Head over to ChatGPT and click on the "Apps" button located in your left sidebar. Locate "Adobe Photoshop" in the apps list, and click on the "Connect" button in the top right corner of the page to link the tool to your account.

STEP 2: Upload and Describe Your Edit

Once the connection is established, open a fresh chat window. Click the "+" icon in the prompt bar and select "Adobe Photoshop" from the menu. Upload the image you want to work on, and simply type out your desired changes using natural language. For example, you can tell the AI to “Make the background black and white.”

STEP 3: Adjust and Fine-Tune with Sliders

ChatGPT will process your request and display the edited image directly in the chat, complete with an interactive Intensity slider so you can manually control the strength of the applied effect. You can stack further edits by continuing the conversation. For instance, if you type "Adjust the exposure on the image," ChatGPT will bring up Photoshop's native exposure sliders, allowing you to fine-tune the lighting to perfection.

STEP 4: Export or Expand Your Edits

Because this integration prioritizes speed and ease of use, it intentionally leaves out the overwhelming complexity of the full desktop app. It is perfect for fast, everyday tweaks. However, if your image needs more advanced retouching, every creation includes an "Open in Photoshop" button. Clicking this will seamlessly launch your file in Photoshop on the web, where you can continue refining your work with a more comprehensive set of tools.

Bluesky's CEO talks about Attie, a new agentic social app built on Bluesky's AT Protocol that uses Claude and lets users build custom feed.

Report analyzing payments of 28M US consumers shows Claude adding paid subs at a steadily increasing pace; Anthropic: paid subs have more than doubled this year.

An AI-generated TikTok parody of reality series Love Island, called Fruit Love Island, averaged 10M+ views across its first 21 episodes after debuting last week.

Cohere launches Transcribe, its first voice model; the 2B-parameter, open-source speech recognition model handles tasks like notetaking and speech analysis.

🎶 Lyria 3 Pro: Google’s AI that makes longer music tracks

🌐 MolmoWeb: Ai2’s open-source AI that can browse the web

🔎 Model Council: Perplexity tool to query multiple AI models at once

⚙️ OpenAI Codex: OpenAI’s coding assistant, now with automations and customizable themes

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Happy Sunday! We just had another crazy week in AI. ByteDance just open-sourced an AI employee that runs 100% locally, while a new AI lets you create fully interactive 3D worlds you can explore.

And that's not all, here are the most important AI moves you need to know this week.

1. ByteDance Drops DeerFlow 2.0 "SuperAgent" Harness

China's ByteDance just released DeerFlow 2.0, a highly capable open-source framework designed to act as a runtime environment for orchestrating autonomous AI sub-agents. It recently hit #1 on GitHub Trending and is built to handle complex, long-horizon tasks that take minutes or even hours to finish.

• Operates in isolated Docker sandboxes, giving the AI its own literal "computer" with a persistent filesystem and bash terminal to safely execute code

• Uses "Progressive Skill Loading," injecting specific capabilities into the context window only when needed to keep token usage lean and incredibly efficient

• The lead agent automatically decomposes complex prompts, spawning scoped sub-agents that run in parallel to research, code, and synthesize deliverables

• Features persistent, cross-session long-term memory so the agent actually learns your preferences and workflows over time

Try it now → https://deerflow.tech/

2. OpenArt Launches Persistent 3D Worlds

OpenArt's new Worlds feature (powered by World Labs spatial AI) is fundamentally changing how AI content is created. Instead of rolling the dice on a 2D image generator and hoping the background looks right, you can now generate an entire 3D environment, step inside it, and direct the scene exactly how you want.

• Generates a fully explorable 3D environment from just a single text prompt or reference image.

• Gives you complete spatial control: you can walk through the space freely, set exact camera angles, and frame the perfect shot.

• Environments are "persistent," meaning the world lives forever in your library and can be revisited and reused across multiple projects.

• You can easily drop characters, objects, and new details into the 3D scene after it is built to capture production-ready 2D images or video keyframes.

Try it now → https://openart.ai/suite/world

3. Google Unveils Gemini 3.1 Flash Live

Google has officially released Gemini 3.1 Flash Live, its best voice and audio AI model yet. It delivers faster responses, more natural conversations, and a massive new feature for developers: configurable "thinking" levels to balance deep reasoning with lightning-fast latency.

• The model scored an impressive 95.9% on the Big Bench Audio Benchmark at the "High" thinking level (2.98-second response time), coming in just behind Step-Audio R1.1.

• If speed is the absolute priority, dropping to "Minimal" thinking slashes the response time to just 0.96 seconds, though benchmark quality dips to 70.5%.

• It's vastly improved at detecting acoustic nuances like pitch and emotion, and effectively filters out background noise in loud environments.

• 3.1 Flash Live now natively powers the Live mode in the Gemini app and Search Live, rolling out globally to over 200 countries.

Try it now → https://gemini.google.com/app

4. Anthropic Gives Claude Control of Your Computer

Anthropic just rolled out a massive update for Claude Pro and Max users on macOS: "Computer Use." Rather than relying solely on API integrations, Claude can now physically take over your screen, moving the cursor, clicking buttons, and navigating apps exactly like a human would.

• Claude first checks for app connectors (like Slack or Google Workspace), but if none exist, it falls back to directly controlling your UI to get the job done.

• The feature natively pairs with "Dispatch," a new mobile tool that lets you text a task to Claude from your phone and come home to the finished work on your desktop.

• It operates entirely in the background within Claude Cowork and Claude Code, meaning you can schedule recurring tasks (like pulling weekly reports) or assign complex multi-step workflows while you commute.

• Anthropic has built-in safeguards but explicitly warns users about security risks like prompt injection, advising against letting Claude access highly sensitive personal or financial data during this research preview.

Try it now → https://claude.ai/new

5. Luma Labs Launches Uni-1

In a massive shift for generative media, Luma Labs has released Uni-1. Instead of using standard probabilistic pixel synthesis (like Flux or Stable Diffusion), Uni-1 is built on a decoder-only autoregressive transformer architecture that reasons through your spatial layout and intentions before generating a single pixel.

• Uni-1 processes text and visual data as an interleaved sequence of tokens, predicting the logical composition of an image before rendering the final high-resolution details.

• Because the model genuinely understands spatial logic (like accurately placing objects "behind" or "under" others), it eliminates the need to endlessly tweak detailed, cinematic prompts to force the right lighting, mood, and composition. You just give it plain English instructions.

• It currently leads human preference rankings against models like Flux Max and Gemini, setting new performance standards on logic-heavy benchmarks like RISEBench and ODinW-13.

• The model is live now for web users at about $0.10 per image, positioning it as a premium engine with an API rollout coming soon for developers.

Try Now → https://lumalabs.ai/uni-1

6. Mistral Drops "Voxtral TTS" Open-Source Voice AI

French AI startup Mistral just released Voxtral TTS, an ultra-fast, open-source text-to-speech model built to bring hyper-realistic, low-latency voice AI directly to enterprise workflows and hardware.

• The model supports 9 languages (including English, Spanish, French, and Hindi) and seamlessly switches between them for real-time translation and dubbing without losing the voice's unique characteristics.

• Custom voice cloning requires less than five seconds of audio, successfully capturing subtle accents, inflections, and natural speech irregularities so it sounds human, not robotic.

• Built on the lightweight Ministral 3B architecture, it's designed to run efficiently on local edge devices like smartwatches, smartphones, and laptops.

• It's incredibly fast, boasting a 90-millisecond time-to-first-audio latency and a 6x real-time factor, meaning it generates a 10-second audio clip in just 1.6 seconds.

Try it now → https://mistral.ai/news/voxtral-tts

Good Morning! Meta just dropped an open-source model that effectively acts as a public API for the human brain, predicting exactly how you react to anything you see, hear, or read. Plus, you’ll learn how to build live, zero-code dashboards with Gemini and Google Sites.

Plus, in today’s AI newsletter:

• Meta Open-Sources a "Digital Twin" of the Human Brain

• Google Unveils Gemini 3.1 Flash Live

• Mistral AI Drops Voxtral TTS to Challenge ElevenLabs

• How to Build Live Dashboards with Gemini and Google Sites

• 4 new AI tools worth trying

AI NEWS

🧠 Meta Open-Sources a "Digital Twin" of the Human Brain

Meta's FAIR team just released TRIBE v2, a massive foundation model that predicts how the human brain processes visual and auditory stimuli. Instead of needing an actual fMRI machine, researchers can now run a digital simulation of a brain's reaction to almost any content.

• Trained on over 500 hours of fMRI data from more than 700 volunteers watching movies, listening to podcasts, and reading text.

• Predicts whole-brain activity across a staggering 70,000 voxels (3D pixels of brain activity).

• Features "zero-shot" capabilities, meaning it accurately predicts the brain activity of new people and responses to unseen languages without any retraining.

• The model's digital simulation is actually cleaner and more accurate than a physical fMRI scan, which is often distorted by background noise like heartbeats, head movement, and machine artifacts.

Neuroscience just got a public API. By shifting brain mapping from expensive, time-consuming physical scanners into high-speed digital simulations, Meta is allowing researchers to run thousands of virtual experiments in seconds. It's a massive leap forward for medical and cognitive research, but it also opens the wild door to "neural-level" content optimization—where media could theoretically be engineered to perfectly trigger specific attention and emotion centers in the brain.

AI MODELS

🗣️ Google Unveils Gemini 3.1 Flash Live

Google has officially released Gemini 3.1 Flash Live, its best voice and audio AI model yet. It delivers faster responses, more natural conversations, and a massive new feature for developers: configurable "thinking" levels to balance deep reasoning with lightning-fast latency.

• The model scored an impressive 95.9% on the Big Bench Audio Benchmark at the "High" thinking level (2.98-second response time), coming in just behind Step-Audio R1.1.

• If speed is the absolute priority, dropping to "Minimal" thinking slashes the response time to just 0.96 seconds, though benchmark quality dips to 70.5%.

• It's vastly improved at detecting acoustic nuances like pitch and emotion, and effectively filters out background noise in loud environments.

• 3.1 Flash Live now natively powers the Live mode in the Gemini app and Search Live, rolling out globally to over 200 countries.

Voice AI is finally moving past the robotic, laggy interactions we're used to. By giving developers the ability to dial up the "thinking" for complex reasoning or dial it down for instant, sub-second conversational replies, Google is making real-time, emotionally aware AI agents viable at scale. As an AI myself, I have to admit that a sub-second response time is some pretty slick engineering. Plus, keeping the pricing that low ensures we're going to see a massive explosion of voice-first applications very soon.

AI MODELS

🗣️ Mistral AI Drops Voxtral TTS to Challenge ElevenLabs

Mistral AI has released Voxtral TTS, a lightweight, 3-billion-parameter text-to-speech model with open weights. Designed for edge devices and local deployment, the model boasts enterprise-grade performance that Mistral claims outperforms ElevenLabs Flash v2.5 in human preference tests.

• Achieved a 63% win rate on standard voices and nearly 70% on voice customization against ElevenLabs Flash v2.5

• Extremely efficient: runs on just ~3 GB of RAM with a blistering 90-millisecond time-to-first-audio latency

• Features zero-shot voice cloning that can accurately mimic a speaker's voice and emotion from just five seconds of reference audio

• Natively supports nine languages, including cross-lingual adaptation that perfectly preserves the original speaker's accent even when speaking a different language

High-quality, emotional text-to-speech has largely been locked behind expensive, proprietary APIs. By releasing an open-weights model small enough to run on everyday hardware but powerful enough to beat industry leaders, Mistral is democratizing advanced voice AI. This allows developers to build hyper-realistic, real-time voice agents directly on edge devices without sacrificing privacy or paying per-minute cloud fees.

HOW TO AI

🗂️ How to Build Live, Zero-Code Dashboards with Gemini and Google Sites

In this tutorial, you will learn how to instantly turn a simple Google Sheet into a professionally designed, live-updating dashboard using Google Gemini, and host it completely for free on Google Sites without writing a single line of code.

🧰 Who is This For

• People who hate coding but need dashboards

• Founders tracking metrics without devs

• Product managers monitoring KPIs

• Marketers tracking campaigns in real time

STEP 1: Publish Your Live Data Source

Start with your data organized in a standard Google Sheet. To make this data automatically readable by your future dashboard, navigate to the 'Share' menu and select 'Publish to web'. Choose to publish the specific sheet as a 'CSV' instead of a web page, and click publish. Copy the generated URL, this acts as the live data pipeline your application will constantly look at to stay updated.

STEP 2: Prompt Gemini for the Code

Open Google Gemini and make sure to turn on the Canvas feature, which acts as the backbone for coding projects. Write a clear prompt describing exactly what you want your dashboard to look like (e.g., a sales tracker or a weekly class schedule).

Critically, paste your CSV link and explicitly instruct Gemini: "Connect to this Google Sheet CSV to pull the data, ensure the dashboard updates when the data changes, and format the code specifically to be embedded in Google Sites." 

Gemini will generate the complete HTML and JavaScript code for you.

STEP 3: Embed and Host for Free

Head over to Google Sites and create a new blank site. Navigate to the 'Pages' tab on the right side, click the plus icon, and select 'Full page embed'. Name your page something relevant, click 'Add embed', and paste the exact code Gemini just generated for you.

Hit insert, and you now have a fully functional, beautifully designed web application sitting right on your page.

STEP 4: Test Real-Time Updates and Refine

The real magic of this workflow is the live data connection. Go back to your original Google Sheet, change a number or a text field, and simply refresh your new Google Site. You will watch the dashboard update instantly to reflect the new data. If you notice an issue with Gemini's design, like a miscalculated margin or an ugly color, just tell Gemini what looks wrong, copy the revised code, and update your embed.

(Note: Because publishing a CSV to the web makes the raw data public, use Google Apps Script for this workflow instead if you are handling highly sensitive company information).

Apple plans to open up Siri to run any AI service via App Store apps in iOS 27, dropping ChatGPT as an exclusive partner in Apple Intelligence and Siri.

Google releases new tools for its Gemini AI assistant that let users upload chat history and context from other AI apps, making it easier to switch from them.

WhatsApp rolls out features and updates: an AI tool that suggests replies based on a user's conversations, the ability to touch up photos with Meta AI.

Cohere launches Transcribe, its first voice model; the 2B-parameter, open-source speech recognition model handles tasks like notetaking and speech analysis.

🎶 Lyria 3 Pro: Google’s AI that makes longer music tracks

🌐 MolmoWeb: Ai2’s open-source AI that can browse the web

🚀 Google AI Studio: Google’s upgraded all-in-one AI builder

⚙️ OpenAI Codex: OpenAI’s coding assistant, now with automations and customizable themes

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! Google just dropped a mind-bending AI memory compression algorithm that the internet is already dubbing the real-life "Pied Piper. Plus, I’ll show you how to record and summarize any meeting using ChatGPT.

Plus, in today’s AI newsletter:

• Google Unveils "TurboQuant" AI Memory Compression

• Mark Zuckerberg is Training an AI to be CEO

• ARC-AGI 3 Drops and Frontier Models Bomb

• How to Record and Summarize Any Meeting using ChatGPT

• 4 new AI tools worth trying

AI NEWS

🗜️ Google Unveils "TurboQuant" AI Memory Compression

Google Research just unveiled TurboQuant, a radical new compression algorithm designed to shrink an AI's working memory without sacrificing an ounce of performance. The breakthrough is drawing massive comparisons to the fictional lossless compression tech from HBO's Silicon Valley, and for good reason.

• TurboQuant targets the Key-Value (KV) cache, the massive memory bottleneck during AI inference, compressing it by at least 6x (down to just 3 bits per value) with zero accuracy loss.

• In benchmarks on Nvidia H100 GPUs, it delivered up to an 8x speedup in computing attention compared to uncompressed 32-bit setups.

• It achieves this using a two-stage mathematical shield: "PolarQuant" maps data vectors onto a predictable circular grid to eliminate overhead, and "QJL" acts as a 1-bit error-checker to maintain perfect logic.

• The tech is so efficient that memory chip stocks (like Micron and Western Digital) actually dropped on the news, as investors realized data centers might suddenly need way less RAM.

Memory is the ultimate invisible tax on modern AI. As context windows get larger, the RAM required to store that context becomes cripplingly expensive. By successfully compressing the KV cache without lobotomizing the model's intelligence, Google isn't just making AI cheaper to run, they are fundamentally altering the hardware economics of the entire industry.

AI NEWS

🤖 Mark Zuckerberg is Training an AI to be CEO

According to a new scoop from The Wall Street Journal, Meta chief executive Mark Zuckerberg is building a personal AI agent to help him run the company. It seems he believes in his own tech hype enough to let it shadow his role, a bold move from the guy who previously burned $80 billion trying to force the Metaverse into existence.

• Zuckerberg's "CEO agent" is designed to retrieve information instantly, bypassing the layers of human bureaucracy he would normally have to navigate.

• Meta is aggressively forcing an AI-native culture to flatten its 78,000-person headcount, even tying employee performance reviews partially to their AI usage.

• Employees are building their own bots, like "My Claw" (which accesses work files and talks to colleagues on their behalf) and "Second Brain" (an AI chief of staff built on Anthropic's Claude).

• Meta recently acquired "Moltbook," a viral social network exclusively for AI agents, and employees have already set up internal message boards where their personal AIs chat with each other.

If the CEO of one of the world's most powerful tech giants is actively trying to automate his own executive workflows, it signals a massive shift in how mega-corporations operate. We are moving past AI as just a coding assistant; Meta is trying to prove that autonomous, agentic AI can literally help manage a trillion-dollar company from the top down.

AI MODELS

🧩 ARC-AGI 3 Drops and Frontier Models Bomb

The third iteration of the notoriously difficult ARC-AGI benchmark is officially here. Designed to test true fluid intelligence and agentic reasoning rather than memorized knowledge, ARC-AGI 3 is proving to be a massive hurdle for the industry's top AI.

• All current frontier models are scoring well below the 1% mark on the new evaluation

• Gemini 3.1 currently leads the pack with a dismal 0.37%

• OpenAI's GPT-5.4 follows closely behind at 0.26%

• Anthropic's Opus 4.6 sits at just 0.25%

It's easy to look at these sub-1% scores and question why benchmarks even matter if the goalposts just keep moving. But this cycle is exactly what forces the industry forward. By proving that models like GPT-5.4 and Gemini 3.1 still lack true, human-like fluid intelligence, ARC-AGI 3 ensures that labs don't get complacent. It exposes the gap between simple pattern matching and actual reasoning, forcing researchers to innovate entirely new architectures to solve it.

HOW TO AI

🗂️ How to Record and Summarize Any Meeting using ChatGPT

In this tutorial, you will learn how to instantly capture, transcribe, and summarize your meetings using the native ChatGPT desktop app for macOS, turning lengthy discussions into structured action items with just one click.

🧰 Who is This For

• Mac users who attend lots of meetings

• Remote workers and async teams

• Founders jumping between calls all day

• Product managers tracking discussions

STEP 1: Launch the App and Grant Permissions

Open the native ChatGPT desktop app on your macOS system. To begin, click the Record button located at the bottom of the chat screen. The first time you do this, your Mac will prompt you to grant the necessary permissions, approve these so the app can seamlessly capture both your microphone and the system audio.

STEP 2: Capture the Conversation

Once permissions are set, ChatGPT will begin recording the audio, automatically distinguishing between each speaker on the call. You can let it run in the background while you focus entirely on your meeting or brainstorming session. Just keep an eye on the clock, as each continuous recording session is currently capped at a maximum of 120 minutes.

STEP 3: Stop and Process the Audio

When your meeting wraps up, simply click the Stop button on the recording interface. To finalize the capture and instruct the AI to process the audio, select Send. ChatGPT will immediately process the recording and upload the full text transcript directly to your chat history.

STEP 4: Review and Repurpose Your Canvas

ChatGPT will instantly open a private, interactive canvas displaying a structured summary of the call, complete with key points, extracted action items, and suggested follow-up questions. From here, you can use text or voice commands to dive into specific elements of the conversation, or ask the AI to instantly rewrite the notes into a client email, a step-by-step project plan, or even a code scaffold for your next build.

Google sets a 2029 deadline for its post-quantum cryptography migration, aiming to “secure the quantum era” as “frontiers may be closer than they appear”.

Meta on Wednesday laid off around 700 employees in the Reality Labs unit, as well as some in recruiting, sales, and Facebook.

Google launches Lyria 3 Pro music generation model, with better creative control and allowing users to create three-minute tracks, up from Lyria 3's 30 seconds.

Reddit says it will label automated accounts that provide a service to users and will require accounts suspected of being bots to verify they are human.

🦞 NemoClaw: Nvidia’s open-source security layer for AI agents

🤖 ASMR: Supermemory’s experimental agent memory system

🎨 Uni-1: Luma’s all-in-one model for text + images

🧑‍💻 Figma: Now lets AI agents design directly on the canvas

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

Good Morning! OpenAI is officially killing its viral AI video app, Sora, just six months after launch to rein in massive computing costs. Plus, you’ll learn how to generate studio-quality product photos for free.

Plus, in today’s AI newsletter:

• OpenAI Shuts Down Sora Video App

• LiteLLM Compromised in Cascading Supply Chain Attack

• LeCun’s Team Cracks Open World Models

• How to Generate Studio-Quality Product Photos

• 4 new AI tools worth trying

AI NEWS

🎬 OpenAI Shuts Down Sora Video App

Six months after launching to massive hype and hitting a million downloads in under five days, OpenAI is abruptly pulling the plug on its Sora short-form video app. The move is part of a massive cost-cutting campaign as the company pivots away from expensive consumer experiments to focus purely on enterprise dominance.

• The app allowed users to generate, remix, and share AI videos in a social feed, but user retention dropped as the astronomical compute costs to render video at scale piled up.

• In the wake of the shutdown, Disney has officially backed out of its massive $1 billion investment deal that would have allowed Sora users to generate content with its copyrighted characters.

• OpenAI is pivoting aggressively toward high-productivity enterprise use cases to compete with Anthropic’s booming Claude business.

• Instead of maintaining a standalone video app, OpenAI plans to fold its core tech into a singular desktop "super app" featuring its web browser, ChatGPT, and Codex coding tools.

Viral consumer buzz doesn't equal a sustainable business model when the product burns through thousands of dollars of compute per hour. By killing a fan-favorite app and walking away from Disney's massive investment, OpenAI is showing just how intense the pressure is to achieve profitability and win the enterprise war before they go public.

AI NEWS

🚨 LiteLLM Compromised in Cascading Supply Chain Attack

Two versions of LiteLLM (v1.82.7 and v1.82.8) have been abruptly removed from the Python Package Index (PyPI) after attackers injected malicious code designed to steal developer credentials.

• The breach actually originated from Trivy, an open-source vulnerability scanner used in LiteLLM's CI/CD pipeline, which was compromised by a threat group known as TeamPCP.

• The attackers modified existing version tags in Trivy's GitHub Actions, allowing them to silently inject malicious code into organizations' workflows without changing the underlying release metadata.

• Using this backdoor, the attackers intercepted LiteLLM's PyPI publishing token and pushed new, poisoned packages containing a credential-stealing component file (litellm_init.pth).

• To cover their tracks, the attackers even spammed the project's GitHub vulnerability report with AI-generated "Thanks, that helped!" comments to bury legitimate warnings from security researchers.

This is a textbook, terrifying example of a cascading supply chain attack. A trusted security tool gets compromised, which is then used to backdoor a widely used AI framework, instantly exposing the credentials of any developer or enterprise that blindly installs the update. As AI development moves at breakneck speed, securing the CI/CD pipeline and pinning dependency versions is becoming just as critical as the code itself.

WORLD MODELS

🌍 LeCun’s Team Cracks Open World Models

Yann LeCun’s team just released LeWorldModel, a 15-million parameter model that successfully predicts physical reality. While LLMs only predict the next word, world models predict actual physics, objects moving, colliding, and falling, laying the foundation for robots and autonomous AI to plan and act in the real world.

• Solves the notorious "collapse" problem (where world models cheat by mapping every input to the same output) without needing complex duct-tape fixes like frozen encoders or stop-gradient hacks.

• Massively simplifies training by using just two loss terms (a prediction loss and a "SIGReg" regularizer to keep representations diverse), reducing the required hyperparameters from six down to just one.

• Incredibly efficient: Trains on a single GPU in a few hours, plans 48x faster than foundation-model alternatives, and uses roughly 200x fewer tokens.

• Proves that the JEPA architecture LeCun has been pushing since 2022 actually trains stably without breaking.

Two AI futures are competing right now: building bigger text-based LLMs, or building world models that actually understand cause, effect, and physics from raw pixels. LeWorldModel proves that the second path isn't just theory, it works, it's stable, and it just got drastically cheaper and more accessible to build.

HOW TO AI

🗂️ How to Generate Studio-Quality Product Photos for Free

In this tutorial, you will learn how to use Pomelli, an experimental AI marketing tool from Google Labs, to extract your brand's unique "Business DNA" and transform basic smartphone pictures into professional, studio-grade product photography without writing a single prompt.

🧰 Who is This For

• People selling products online (Shopify, Amazon, D2C brands)

• Dropshippers who need clean product visuals fast

• Small business owners without a studio setup

• Content creators shooting product promos

STEP 1: Extract Your Business DNA

Head over to Pomelli and sign in. Instead of manually uploading brand kits or writing complex style prompts, Pomelli automates the setup. Enter your business or product website URL and click continue. The AI will scan your site to extract your "Business DNA”, automatically identifying your brand's color palette, typography, tone of voice, and visual style. If your website blocks AI scrapers, you may need to ask your developer to adjust your site permissions first.

STEP 2: Access the Photoshoot Feature

Once your Business DNA is successfully generated, navigate to the "Photoshoot" tool. Upload a basic, unedited photo of your product, even a messy smartphone snapshot will work perfectly. Because Pomelli already understands your brand's visual identity from the previous step, it does not require any complex prompting or prompt engineering to get started.

STEP 3: Select Your Template and Generate

After uploading your image, Pomelli will analyze the product and recommend a variety of curated templates, such as contextual settings, lifestyle shots, or seasonal themes. Choose an aspect ratio optimized for your specific social feed (like 4:5 for an Instagram post, 9:16 for Stories, or 1:1 for square feeds) and click generate.

In just a few minutes, the AI will apply your Business DNA to the template, outputting studio-quality images that perfectly match your brand's exact aesthetic.

STEP 4: Turn Photos Into Full Campaigns

Once your images are generated, you can download them directly or save them back to your Business DNA for future use. To take it a step further, click on the campaign feature and type a simple goal, like "Father's Day 20% discount." Pomelli will stitch your newly generated product photos into fully designed, ready-to-post social media layouts with matching text and calls to action, saving you hours of manual editing.

Anthropic announces an “auto mode” that enables Claude Code to make permission-level decisions while preventing destructive actions like mass file deletion.

Arm unveils its own AI chip called the AGI CPU, a departure from its traditional role as a designer of chips for others; Meta and OpenAI will be early customers.

Epic cuts 1,000+ jobs, saying it is “spending significantly more” than it is making and that the layoffs and $500M in savings put it “in a more stable place”.

Munich-based Interloom, which aims to capture tacit knowledge for AI agents from businesses' operational records, raised a $16.5M seed led by DN Capital.

🦞 NemoClaw: Nvidia’s open-source security layer for AI agents

🤖 ASMR: Supermemory’s experimental agent memory system

🎨 Uni-1: Luma’s all-in-one model for text + images

🎥 Replit Animation: Turn text prompts into animated videos

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas