Build / Code pathway · train, tinker, run
The camp's own tools are simulations on purpose — frozen examples and hand-authored distributions, so the mechanism stays visible and nothing depends on an account or a server. When you're ready to train or run a real model, these are the places to go. Bring the same loop: predict, change one thing, compare, name what it did.
For everyone who asked about local models and building from scratch
In sessions we won't set up accounts, install software, or host models — the core teaching moments never depend on live AI, and that's a feature (reliability, cost, consent). What we will do: Count the Next Token already builds a tiny language model from scratch, visibly, with arithmetic — that's the "how does training work" answer in miniature. Everything below is for going further on your own machine, on your own terms, between sessions or after the camp. Most of it is free; the listings say what each one needs.
The single best follow-up to this camp: train an image, sound, or pose classifier on your own examples in two minutes, then break it. Class A is you with glasses, Class B without — now test a friend. It's the Dataset Balance Simulator with real stakes: skewed examples in, skewed behavior out, and you chose the examples. Training runs in your browser; your webcam images don't need to leave the machine.
Train text, image, and number classifiers and then use them inside Scratch projects — a label-it-yourself game becomes a working program. Built for classrooms, with teacher-administered access and ready-made lesson worksheets. The natural next step from the younger-learners adaptations.
Label fish vs. trash to train a classifier, then watch it inherit your labeling choices — bias as a first-hand experience for elementary students, with the ethics discussion built in.
The machine guesses your doodle as you draw — pure what-does-the-machine-see energy. Then open the dataset and inspect millions of other people's drawings: why does it expect that shape for "shoe"? Because that's what it was shown.
Friendly machine learning for the p5.js world this community already lives in: image classification, body and hand pose, sound, and Teachable Machine model imports — a sketch that responds to your gestures is an afternoon, not a semester. Start in the p5.js Web Editor with the ml5 examples; this is the Build/Code pathway's natural home.
A tiny neural network you can watch train: add neurons, change the data, and see the decision boundary bend in real time. The one link on this page about learning itself — what "training" looks like while it happens. Overfitting becomes something you can see: give it too much capacity and watch it memorize the noise.
The classic creative-ML instrument: demonstrate a few gesture-to-output examples and play the mapping live — webcam to synth, motion to visuals. For the artists in the cohort who want training to feel like rehearsing rather than programming.
Run open-weight language models (Llama, Mistral, Gemma families and more) entirely on your own computer — one install, one command, no account, nothing sent anywhere. The honest answer to "do I have to pay someone to use this?": small models run fine on a recent laptop; what you trade is capability for ownership. Re-run the camp's Session 1 experiments against a local model and compare.
The same local-model idea with a friendly desktop interface instead of a terminal — browse models, download, chat, and inspect token-by-token output. The gentler on-ramp if the command line is a barrier.
Where open models, datasets, and demos actually live. Two camp-relevant uses: try models in the browser via Spaces without installing anything, and — in the spirit of "Excavating AI" — read datasets directly: browse what's actually inside the training data everyone argues about, and check a model's card against the camp's Model Card Builder.
One rule carries over
Whatever you open from this page, run it like the camp runs everything: predict before you press the button, change one variable, and name what the machine did — a default, a failure, or a pattern. The tools change; the method is the take-home.
For everyone who heard the API is out of reach
An API just means calling the model from your own code or a notebook instead of a chat box. The common belief that this needs an expensive subscription is out of date: every major provider now has a free tier or starter credits, and pay-as-you-go pricing is fractions of a cent for the small jobs a classroom runs. Here's where to start without a credit card — and what an "agent" actually is once you can make one call.
The lowest-friction door: sign in, generate an API key in the browser, and call a current model on a free daily quota — no billing setup. Best first stop for "I want to build with it, not just chat." Paste the key into a Colab notebook and you've moved from consumer to maker in an afternoon.
All three give starter credits or a free tier, then charge per token rather than a flat subscription — a whole class's worth of short calls is cents, not a monthly fee. The camp's Tokenizer shows why: you pay by the token, so short prompts stay cheap. The honest answer to "do APIs have a free way in?" is yes — and even paid use is far closer to a bus fare than a subscription.
An agent is just a model in a loop that can call tools — search the web, run code, hit another API — and decide what to do next, until a stopping rule says it's done. Strip the hype and it's the camp's own method automated: predict, act, check, repeat. You don't need a framework to see it — give a model one tool and let it choose when to reach for it. Once one API call works, an agent is that call, in a loop, with a decision.