Practical AI, explained and tested.
We explore useful models, try small experiments, and share enough detail for you to follow along—including the cases that did not work.
Watch on YouTube · Browse tutorials · Read roundups & notes
Choose a question, watch the experiment, then inspect the companion. Each test records its own setup and limits; results from different tasks are not a shared model leaderboard.
| Question | Watch | Follow along |
|---|---|---|
| How do I run Laya locally? | Installation and three demos | Setup and examples |
| Can Ornith recover when a coding tool fails? | Ornith versus Qwen | Saved tests and reproduction guide |
| What can a small local model handle? | MiniCPM5 1B and 2B | Configurations and recorded outputs |
| What are we teaching a language model to do? | The first theory lesson | Python examples and course roadmap |
For more comparisons, browse the channel's videos. For explanations you can build on, start with the language-model course.
| You want to… | Start here |
|---|---|
| Try your first local AI example | Beginner setup guide |
| Install Laya and classify requests | Laya walkthrough |
| See Needle turn text into proposed commands | Needle demo |
| Understand the Dream-RSI experiment | Episode notes and results |
| Run the advanced Dream-RSI experiment | Standalone source, setup and tests |
| Catch up without installing anything | Reading archive |
Technical index links directly to entry points, pinned dependencies, saved outputs and test limits. Each guide identifies the platform actually tested. No private account, paid subscription or RUNTIME service is needed for the included local examples.
Code and guides live in the tutorial library. Articles and newsletters live in the reading archive. Each runnable guide includes synthetic examples, setup steps, sources and limitations. A recorded result is evidence for that run, not a promise of the same accuracy or speed on every machine.
Have a correction or a question? Use the relevant repository's Issues page. Please share a small synthetic example rather than private documents or logs. Include the model version, hardware, prompt, expected behavior and actual output when reporting a different result. Small reproducible cases help us improve the tests and decide what to investigate next.