Good morning! Here's what's happening in AI today:
Moonshot AI launches Kimi K3, a 1 million token open frontier model
Grok adds Automations for scheduled AI tasks that run on their own
tldraw launches an offline whiteboard app with no account needed
How to build an AI agent and make it better the more you use it
And 4 new AI tools worth trying today
TOGETHER WITH SUPERBLOCKS
Most AI app builders charge you top-dollar rates for every little task and bill in credits you can't predict. Superblocks routes the easy work to free open-source models and saves the expensive models for jobs that actually need them. Same quality, smaller bill. And what you build is the real deal: working apps on your company's real data, safe enough that IT actually approves.
With Superblocks:
Cut your AI bill without cutting quality
Build real apps from plain-English prompts
Keep everything inside your company's own cloud
AI MODEL
Moonshot AI released Kimi K3, a 2.8 trillion parameter open weight model built for long running agentic coding and self evolving workflows, positioning it as one of the most ambitious open releases from a lab outside the largest US players.
Kimi K3 handles a full 1 million token context window, meaning it can hold an entire codebase or a long, dense document in memory at once without losing track of earlier details as the task continues.
A new attention system called Kimi Delta Attention lets the model decode up to 6.3 times faster in long context situations, addressing one of the biggest practical slowdowns that comes with working across very large inputs.
The model is already live today on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with full open weights set to release by July 27 so developers can eventually run and modify it themselves.
Most frontier models keep their weights closed, so Kimi K3 stands out by promising full open access alongside genuinely competitive performance. For anyone building AI agents or coding tools, a faster, longer context model that you can inspect and modify changes what's possible outside the big labs. This is one of the largest open model releases in recent memory.
AI AGENT
🤖 Grok Launches Automations Feature
Grok introduced Automations, a feature that lets you describe a task once in plain language and have Grok run it automatically on a schedule or trigger, rather than waiting for you to prompt it each time.
Describe any job in plain language, whether it's a daily check-in, a recurring report, or an ongoing research task, and Grok sets up the schedule or trigger for you without any manual configuration.
Once running, Grok completes the task entirely on its own in the background and reports back to you only when the work is actually done.
This is built for ongoing, repetitive work like monitoring a topic, tracking changes, or compiling recurring summaries, removing the need to open the app and re-trigger the same request every single day.
This turns Grok from a chat tool you actively prompt into something that works in the background for you. For anyone juggling repetitive research or monitoring tasks, this removes the need to manually trigger the same request every day.
AI TOOL
tldraw launched tldraw offline, a free file based desktop app that brings their canvas tool directly to your computer with no account, login, or server connection required to use it.
Work fully offline with no account needed at all, and every file stays local on your machine instead of syncing to a server somewhere else.
AI agents like Claude Code or Codex can read, edit, and add scripts directly to your tldraw files, letting your coding tools actually participate in the same canvas you're sketching or diagramming on.
Agents can write persistent scripts into a file, so a single tldraw file can behave like its own small app every time you open it, carrying its logic along with it.
This is one of the first canvas tools built specifically to work alongside AI coding agents rather than around them. If you sketch, diagram, or whiteboard as part of your workflow, this gives your AI tools a way to actually participate in that process.
HOW TO AI
I've been teaching programming for over ten years, and the most consistent thing I see isn't a lack of intelligence. It's a lack of feedback. Students who improve are the ones who test themselves, see exactly where they failed, and fix that specific gap.
Most people set up a system prompt, run it for a week, and slowly drift away because the outputs aren't quite right and they don't know why. That's a system design problem. Here are five patterns that fix it.
1) Define what "done well" actually means
Before building anything, write down what a great output looks like, specific enough that someone else could judge pass or fail. Without a clear target, you can't tell if your agent is actually improving or just feels like it is. Put a short quality standard in your project instructions: what does a 10/10 output look like, what does a fail look like.
2) Keep the context clean
The most common cause of agent degradation isn't a bad model, it's a bloated system prompt trying to hold every rule for every possible task at once. The fix is progressive disclosure: keep the system prompt to what the agent always needs, and package everything else, specific policies, workflows, procedures, as a skill it loads only when the task calls for it.
3) Build the testing loop
Write down five tasks your agent needs to handle. Run them, score each pass or fail, and for every failure write one sentence on why it failed. Fix that specific thing, then run the five tasks again. Anthropic's own team used this loop to take an inventory agent from 62% to 92% accuracy, same model, better system.
4) Let memory accumulate
An agent that starts from zero every session doesn't improve, it repeats the same mistakes. In Settings → Memory, turn on both "Search and reference past chats" and "Generate memory from chat history." Review what gets stored once a week, delete what's wrong, add what's missing. After a few months, you stop re-explaining yourself.
5) Build a habit of reviewing sessions, not just running them
Anthropic's own agents use a feature called Dreaming, a scheduled process that reviews past sessions and memory between runs and rewrites what the agent knows, so the next session starts smarter than the last one. It's currently limited to developers building on Claude Managed Agents. The manual version is open to everyone: regularly review what your agent got right and wrong, and update its memory and instructions accordingly.
None of these require a new model. The agent that gets better over time is the one where every session leaves it slightly more capable than the last.
P.S. You can access all my prompts, workflows and AI trainings if you upgrade here
Gemini Notebook rebranded from NotebookLM, keeping the same app and mission with a new name tied to Google's broader AI lineup.
VS Code released version 1.129, adding terminal commands directly in chat and letting agents delegate work across sessions.
Gemma 4 launched on LiveKit Inference, hitting 354ms time to first audio for real time voice agents.
🔧 Firecrawl: Free web scraping and search for AI agents, no API key needed.
📊 Bolt Slides: Open source tool letting any coding agent generate slides.
💻 Codex: Review pull requests and send edits without leaving the platform.
🎬 Google Vids: Prompt to swap backgrounds, adjust lighting, and clean footage.
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