👋 Hello hello,
Two of the biggest AI companies on earth just made completely opposite moves this week, and both were about the same exact thing: watermarks. Anthropic started tagging everything Claude writes so it can be traced back later. Google did the reverse, giving people a button to hide the AI stamp on their images and videos. Same technology, two opposing philosophies. Alibaba dropped an open model small enough to run on your laptop that's beating models many times its size.
🚨🚨 This week we’re hosting live sessions and it’s free for all. We’ll be building AI workflows from scratch, in real-time. If you haven’t attended any of these sessions, this is your chance! 🚨🚨
Let’s get into it.
🔥🔥🔥 Three Exclusive AI Updates
Every response Claude generates will soon carry a hidden watermark, baked in to comply with the EU's new AI Act. It works by nudging low-stakes word choices (like picking "grey" over "overcast") according to a hidden key, instead of picking randomly. You'll never notice it. There's nothing added to the text, no hidden characters, and it doesn't cost extra tokens or slow anything down.
It also can't be traced back to you, your org, or your specific chat. It only answers one question after the fact: was Claude likely involved in writing this? Anthropic isn't doing this alone either. Most major model makers signed the same EU code of practice, so expect this to become standard across the board, not a Claude-only thing.
Google just added a toggle that lets you hide the visible watermark on images, video, and music made with Gemini. Flip "Show watermark" off in settings and it applies going forward to everything from Nano Banana images to Omni videos and Lyria tracks.
Here's the catch: the invisible SynthID watermark and C2PA metadata stay embedded no matter what you choose. You can still ask Gemini or Google Search "is this AI generated?" and get a straight answer. The visible stamp is optional now. The paper trail underneath it isn't. Rolling out over the next few days, with some regions requiring a paid Ultra plan to access the toggle.
Alibaba released Qwen3.8-27B as open weights, and it's a serious one. It's a native multimodal model, meaning it handles text and images out of the box, with a 262K token context window that stretches to 1M. At just 27 billion parameters, it's already outperforming Qwen's own larger 3.7-Plus model on real coding and office tasks, and it's licensed under Apache 2.0, so you can actually build on it commercially without friction.
The bigger picture: this is the kind of capability that used to require a massive, closed, expensive model. Now it fits on a single high-end GPU..
🔥🔥 Two Tools To Try
1. 📱 Port22
If you run coding agents like Claude Code, Codex, or OpenCode on your Mac and hate babysitting your terminal, this is worth a look. Port22 mirrors your active sessions to your phone in real time. When an agent needs your approval to edit a file, your phone buzzes and you tap the actual option it offered instead of guessing at a keystroke later. Free for one Mac and two sessions, full features included.
2. 📄 Blume
An open-source documentation framework for anyone who's ever hated setting up a docs site. Drop a folder of markdown files in, run one command, and you get a full site with search, theming, and SEO already handled. It's also built to be read by AI tools out of the box, with an MCP server that lets Claude Code, Cursor, or VS Code query your docs directly. MIT licensed.
🔥 One Pro Tip For Claude Usage
Claude Code bills you per token, and a messy session quietly costs way more than a clean one. A few small habits fix that:
Run /clear between unrelated tasks so old context doesn't get resent (and rebilled) on every single turn.
Pick your model and effort level before you start. Switching mid-session resets your cache and gets billed at full price.
@-mention files by name instead of describing them. Claude attaches the file directly instead of spending a step searching for it.
Add quiet flags to noisy commands, like test runners that print hundreds of lines, so that output doesn't sit in your context for the rest of the session.
Run /compact before you step away from your keyboard. The cache expires after an hour, and it's much cheaper to summarize while it's still active.
🎤 Join Us Live This Friday
This week we’re opening shop for everyone! We’re hosting a free AI workflow building session this Friday! Join us live, build in real time, and get your questions answered on the spot.
Can’t make it live? Don’t worry, we’ll send out the recording after the session to everyone who RSVP’ed.
Event details:
📅When: Friday at 10AM PT/1PM ET
📍Where: Virtual, Google Meet
After you RSVP, you’ll be sent a link with the event details. Hope to see you there!
Did you find today's post useful?
Until next time,
Team @PracticalyAI
