The same hands-on labs our cohort trains on — tokenizers, chunking, prompts — free in your browser, no account, no tracking scripts. This is how we teach: from the inside.
Each one answers a question people usually hand-wave: what does the model actually see, how does a document become retrievable, what makes a prompt work — and how right are you when you're sure?
Each one owns a single reveal — a bill, a ratio, a radar, a grid — that makes an invisible mechanic visible. Pure client-side, nothing to install, nothing to sign. All ninety now ship with the ◆ X-ray reveal and an export bar.
Sampling, embeddings and retrieval need real model calls, so they live behind a free account — along with the full training program.
Temperature, top-p and top-k against a live model — watch the probability distribution reshape as you drag the sliders.
Turn words into vectors and see meaning as geometry: scatter plots, similarity heatmaps, and king − man + woman ≈ queen.
Run keyword vs semantic vs hybrid retrieval side-by-side on a real index, with citations and abstention behavior.
This is 1% of the training. The full program — daily reps, incident drills, an AI shadow engineer, a verifiable record — is free for founding cohorts.