Good Morning! Here's what I have for you in today's newsletter:
OpenAI starts watermarking ChatGPT and Codex text in the EU
Krea brings its agent into After Effects for motion graphics and resizing
Cursor lets developers steer SDK agents while they run
Run a free private AI on your Mac and ask it about your files
4 new AI tools worth trying today
AI TOOLS
OpenAI expanded its watermarking work to text. An invisible watermark called textGrain will be added to eligible ChatGPT and Codex output in the EU as part of its efforts to meet the EU AI Act's transparency rules.
textGrain subtly adjusts word choices without adding any visible symbols. A detector with a secret key can identify the pattern, according to a report co-authored with Penn and Yale researchers.
API customers worldwide can enable watermarking on select models, though it's off by default. In ChatGPT and Codex, the rollout will reach all EU plans over the coming weeks.
Detector access starts with approved researchers and expert organizations. OpenAI says editing, translation, and short passages can make the watermark harder to detect, so its absence doesn't prove a person wrote the text.
If you use ChatGPT or Codex in the EU, eligible text will soon carry a watermark you won't see when reading or copying it. Elsewhere, API developers choose whether to enable it. A detector can still miss the mark, so its result shouldn't settle the question of who wrote something.
AI TOOLS
Krea launched Krea Agent inside Adobe After Effects. You can ask it to create motion graphics, resize videos, and work with sound and images directly in your project. It's free to try.
Describe the motion graphic you want, and the agent builds it in an After Effects composition without making you open another app.
The same agent can resize videos for different formats, add sound, and place generated images into your project.
The integration adds to Krea Agent's existing tools, which include a video editor, custom apps, and the ability to use Pinterest boards as creative references.
Need a lower third, a vertical version of your video, or background audio? You can request it inside the After Effects project you're already editing. That could save a few trips between apps, along with the exports and reimports needed to bring everything back together.
AI AGENTS
Cursor updated its SDK so developers can give agents new instructions while they're working. The run.steer() method adds a message to the agent's next turn without stopping the task already in progress.
If a subagent is mid-task when you steer the run, it continues in the background. Once finished, it sends its result back to the parent agent in a follow-up turn.
stream() and wait() now continue until all background subagents finish. Custom tools also support MCP annotations such as readOnlyHint and destructiveHint, which identify read-only and potentially destructive actions.
Developers can supply their own system prompt while still loading Cursor's rules, skills, and tool schemas. Access to the new features is rolling out account by account.
A change in instructions no longer means stopping an agent and starting over. You can redirect the run while its subagents finish their work. Tool annotations also give the model more context about what a call will do, including whether it could delete or change something. Check whether your account has access yet.
HOW TO AI
Install Ollama, run a Gemma model, and ask about your files without a subscription.
You can run an AI model on your Mac for free. Once you've downloaded it, it works offline and processes your questions and files on your machine. You'll need at least 8 GB of memory and around 10 GB of free disk space.
Step 1: Check your memory and pick a model
Open the Apple menu, then About This Mac, and look for Memory. Run gemma3:1b on 8 GB, gemma4:e2b on 16 GB, and gemma4:e4b on 24 GB or more. Skip any model with "cloud" in its name, since those require a sign-in.
Step 2: Install Ollama and get your first answer
Download Ollama for macOS from ollama.com/download and open it. Choose a local model and skip signing in. In Terminal, run:
ollama run gemma4:e2bOnce you see >>>, ask a question. Then turn off Wi-Fi and ask again.
Step 3: Fix the short memory
Below 24 GB, Ollama starts with a 4k-token context window, so a long document may exceed it. Move the context slider in Ollama's Settings up one notch. Send a message, then run ollama ps. Aim for 100% GPU, and lower the slider if the model splits onto the CPU.
Step 4: Ask about your own files
Drag a PDF or text file into the Ollama app. Tell the model to answer only from the file, support each answer with an exact quote, and reply "Not in this document." when the answer isn't there. Test it with questions you already know.
Step 5: Lock it to local and keep it tidy
Disable cloud features, then restart Ollama:
[ -f ~/.ollama/server.json ] || echo '{"disable_ollama_cloud": true}' > ~/.ollama/server.jsonManage models with ollama list, ollama rm, and ollama stop.
P.S. You can access all my AI courses, trainings, prompts library, and $500+ worth of AI tool discounts here.
Cognition introduced Dreaming for Devin, which remembers how you work across sessions and reviews those memories overnight to remove outdated details and spot new patterns. It also introduced Agent Memory Repo, a proposed open standard for agent memory.
Obsidian released version 1.14 to everyone. It adds Quick Capture for instant notes on iOS, highlight colors, editing for Markdown files outside your vault, macOS Quick Look previews, and Kanban grouping in Bases.
Cohere launched North 2, its enterprise agent platform, with more than 15 new features. These include on-premises and air-gapped deployment, Slack, SharePoint, Jira, and GitHub connectors, reusable agents, and memory across sessions.
🎬 HyperFrames Studio: a video editor where an agent creates the first draft, then makes changes based on your drawings or comments on individual frames.
📣 Lightreel: suggests Meta ad hooks, creative ideas, and account improvements when you ask what ads to run.
☀️ Upstage: Solar Mini 4 is free on Nous Portal for two weeks, with 3B active parameters, 35B total, and a 512K context window.
🔗 Together AI: launched Together Link to run open models in coding tools, track spending, and automatically choose a cheaper or stronger model for each task.
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
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial
✓ $500+ worth of AI tools