Today we look at how fast local AI video tools are improving and what that means for anyone who wants to create without waiting on big cloud servers. We also break down how a full GPT model can fit inside a spreadsheet so you can see every part working. Plus we share two free tools you can try right now that turn maps and neural networks into something you can actually explore.

The Big Story

Imagine wanting to make a short video for a school project or a funny TikTok. A few years ago you would send your idea to a big company’s servers and wait. Now an open-weights model called MiniMax H3 can create the whole thing on your own laptop, including sound, in about three minutes.

Open-weights means the model’s numbers and rules are public, so anyone can run it locally instead of needing an internet connection to a company’s computers. Ethan Mollick tested it by asking for “an otter using a laptop on an airplane” and got a short clip that already looks surprisingly natural.

The test used a consumer-grade computer with no special hardware beyond what many people already own. The output included both moving images and matching audio generated together. Less than two years earlier, reaching this level of quality required sending the request to remote data centers and paying for each second of processing time.

This matters because local tools keep your prompts and creations private—no one else sees what you’re making. It also removes the cost and wait times that come with cloud services. For students or creators who want to experiment without limits, that changes what’s possible on an ordinary laptop.

The speed of progress is the real headline. Less than two years ago this quality would have required expensive cloud time. Now the same result runs at home. That shift puts video creation in the same category as text and image tools that already live on personal devices.

If you have a reasonably powerful computer, you can look for MiniMax H3 downloads and try generating your own short scene. Start with something simple like “a cat riding a skateboard in a park” and see how the model handles motion and sound. The experiment itself shows how quickly these tools are moving from “research demo” to “something you can run yourself.” Source: x.com

Explain Like I'm 14

You know how a recipe card tells you exactly what to do with each ingredient and in what order? A spreadsheet version of a tiny GPT does the same thing, except every cell is one tiny step in the recipe.

The creator built a working nanoGPT with only about 85,000 parameters and put every number, every multiplication, and every addition into visible cells. You can watch the input text turn into numbers, watch those numbers get multiplied by weights, and watch the final numbers turn back into predicted words—all without any hidden code.

Because everything sits in plain sight, you can click on a cell and see exactly which earlier cells fed into it. It is like opening the hood of a car and finding every wire, every bolt, and every connection labeled and reachable.

The spreadsheet recreates the full GPT architecture, so every layer of attention and every feed-forward step appears as ordinary spreadsheet formulas. When you change one input number, you can trace how that single change ripples through dozens of later cells until it affects the final word prediction.

The point is not to run this tiny model for real work. The point is to remove the mystery. When you see the same operations happening at a huge scale inside ChatGPT or Claude, you already know the basic pattern: turn words into numbers, do lots of math, turn numbers back into words.

That single spreadsheet makes the “black box” feel a lot less black. Once you understand the small version, the big versions stop feeling like magic and start feeling like very large, very fast versions of the same recipe.

Cool Stuff & Try This

Turn any city into a clean poster with one click An open-source tool lets you pick any place in the world and instantly generate a minimalist map poster from its street layout. No design skills needed—just choose a city and the tool pulls public map data to create something you could print or use as a phone background.

Go to the tool’s website, type in your hometown or a city you want to visit, and hit generate. You can adjust line thickness or colors if you want, then download the result. It is a fun way to see how streets form patterns you never notice when you are walking around.

See the whole family tree of neural networks at once A single graphic organizes dozens of different neural network designs into one clear map. Instead of just hearing the words “neural network,” you can see how the common ones connect and branch out.

Open the image and zoom in on areas that interest you. You will notice that some designs are built for images, others for text, and some try to combine both. It is like looking at a subway map of AI ideas instead of trying to remember every stop by name.

Quick Bits

Different ways to give an AI its own computer One experiment compares three setups: letting the AI use your actual laptop, giving it a temporary online machine that resets after each use, or giving it a persistent online machine that keeps its files between sessions. Each choice changes what the AI can remember and do safely.

AI video on your own machine is getting fast The otter-on-a-plane clip took roughly three minutes to generate locally with sound. That speed on consumer hardware shows how quickly open video models are catching up to cloud-only tools.

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.