Tools

Across the semester you will meet many tools. The list below is not exhaustive, and any of it may change as the field moves. The point is to learn the categories, so that you can evaluate the next tool that appears.

Everything listed here is a tool rather than a model. Every tool wraps one or more models and adds an interface, a system prompt, safety filters, a data policy, and a price. The model behind a tool can change without the name changing. The entries below therefore name only the tool, while your practice log names both, with the model’s version or the date. Chapter 2 sets out the difference.

Apps made for this book¶

The book also ships with its own small apps: single-page, self-contained tools built to make one idea in a chapter concrete. They run in your browser, need no account, and send no data anywhere. The sixteen below are ready to use.

• Training loop playground: watch a tiny model’s loss fall step by step as it trains.
• Next-token sampler: compare greedy, temperature, and top-p sampling on the same next-token distribution.
• Tokeniser explorer: see how a sentence splits into tokens, and why Norwegian text costs more tokens than English.
• Forward diffusion explorer: watch an image dissolve into noise, then step back to see how diffusion models learn to reverse it.
• Inference energy estimator: estimate the energy and water cost of running a prompt through a model of a given size.
• Provenance inspector: inspect the metadata a generated file carries, and what it does and does not prove.
• Word-vector explorer: explore how word embeddings place related words near each other in vector space.
• Markov melody generator: generate a short melody from a Markov chain trained on a handful of tunes.
• Agent loop simulator: step through an agent’s plan, act, and observe loop one tool call at a time.
• Rhythm-bot: map a tapped rhythm onto a simple generative pattern, live in the browser.
• Evolve a picture: breed a drawing by choosing favourites, biomorph style.
• Seed and guidance grid viewer: step through a grid of generations that vary one knob at a time, with the settings shown beside each cell.
• Frame-consistency inspector: step through a generated clip frame by frame with a difference view beside it, so you can see exactly when and where the drift begins.
• Splat viewer: load a Gaussian splat capture and orbit it, with a control that thins the splats so you can see how little geometry is actually there.
• Translation sketchpad: draw a line and hear it as pitch and loudness, then sing or play into the microphone and see the contour drawn back, with a rule-based text description as a third data type.
• Sensor-to-sound mapper: connect a phone or laptop sensor to a sound parameter, and compare a mapping you can learn with one you cannot.

Tool categories¶

Hardware to borrow¶

The following equipment is available through the fourMs Lab. Ask the course coordinator:

• a VR headset
• a 360 camera
• a phone gimbal
• an Arduino kit with sensors
• a small robot platform