The first thing that stands out about Discover AI is not the promise of teaching AI, but the fact that it does so without requiring an account and without sending telemetry. The app, available on GitHub, collects 51 hands-on labs in six thematic groups: it starts from the basics—what a neuron is, how gradient descent works—and moves to topics such as attention, RAG, agents, fine-tuning, quantization, and serving. Some labs go beyond animating concepts: the attention lab runs a small trained transformer with 1.37 million parameters inside the app, while the training labs train a character-level model live as you move sliders.

This choice to run real models locally is more than a technical detail. It signals that a significant part of AI learning can now happen on ordinary hardware, without depending on cloud credits or dedicated GPUs. For teams evaluating on-premise or air-gapped deployments, the ability to experiment with fine-tuning and quantization in an isolated environment is a concrete lever: it builds operational understanding of memory, latency, and cost constraints that text-only tutorials cannot convey. It is no accident that the app works entirely offline: the constraint of not transmitting data becomes part of the experience, not a limitation to overcome.

In a context where data sovereignty and regulatory compliance push many organizations to consider self-hosted solutions, educational tools that run without a network are not a quirk: they train people to think of AI as a workload that can live inside a controlled perimeter, not just as an API to call.

The project was built with the help of Claude and GPT Astra for writing lesson text, and Grok for a second pass on references. The author admits there will be mistakes and invites users to report them or open a pull request. It's a detail that shows how generative AI tools are being used to produce educational material about AI itself, but the core value remains in the interactive labs, not the texts.

Structurally, the existence of such a tool strengthens a trend: AI technical knowledge is shifting from abstract model understanding to practical competence in running, small-scale training, and measuring behavior. This shift rewards those who can read the implications of VRAM, quantization, and throughput, and pushes practitioners toward the model lifecycle in deployment contexts. It may not be enough to train a professional, but it reduces the distance between paper theory and the reality of a process running locally.

The GitHub repository is open to contributions for new interactive labs, better visualizations, or tweaks that make cause and effect clearer in existing labs.