Live Demos: Open in Any Browser
Live
Token Visualizer
Watch how a language model picks the next word. Real-time probability bars show every candidate token. Adjust temperature and watch the distribution flatten or sharpen. Pick top, sample, or undo, one token at a time.
Live
Token Roulette
One prompt. Five candidate tokens. You decide: pick the top guess, roll the dice with weighted sampling, or undo. Watch how small choices cascade into completely different outputs by the end of the sentence.
Live
Furbish Detector
Prompts engineered to trigger confident hallucination. The detector flags invented methods, fake standards, and sycophantic reversals in real time as the model streams its response. Watch it catch itself in the act.
Workshop Tools: Shown Live at CAST
Repro
Same prompt. Two runs. Temperature 0.8 with no seed versus temperature 0 with a fixed seed. One output drifts, one locks. The reproducibility demo, because "it worked last time" is not a test strategy.
Oracle Problem
When the AI is the system under test, how do you verify its output? Prompts designed to surface the classical oracle problem in a new context: testing a system where the expected result is the thing you cannot know for certain.
Lost in the Middle
Bury a fact at 10%, 50%, or 90% depth in a long document. Ask the model to find it. The retrieval heat map shows where models succeed, and where the middle of a context window becomes a blind spot.
Fleet Race
Three fleet machines. One prompt. Live token-per-second benchmarking of real hardware running real models. A snapshot of what the lab actually runs, not cloud inference, not a benchmark paper.
Neo vs Switch
One question. Two models. Side-by-side streaming responses with no commentary. You watch both arrive in real time and decide which one you would trust in a production context.
Bug Hunt
A live AI agent reads an API spec, hunts defects, and files reproducible bug reports while you watch. Two builds: one buggy, one fixed with a regression snuck in. The 3x rule means flaky doesn't count.
From Zero
Eleven stages. One browser tab. Build a language model from scratch: workbench setup, raw training data, tokenizer construction, the training loop, and every architectural decision in between. The full arc of 3thfloor-2b, told live.
LM Studio Install Guide
Step-by-step setup for Mac (Apple Silicon), Windows, and Mac (Intel). Everything you need is on the USB drive, no internet required. Open this on your own device to follow along at your own pace.