VibeModel: Learn AI

VibeModel is a hands-on way to learn AI: you build real models on your own things and data, and the app tells you honestly whether they work. PROJECT 1 — TEACH AN IMAGE MODEL, ON YOUR PHONE Show it two to five things with the camera. A pre-trained image model (SigLIP 2) already knows how to see; you teach the small part at the end what your things are, right on the phone — the app shows how long it took. Then quiz it on looks it never learned from, in places it has never seen, and get a verdict: did training beat simply matching your photos? Your photos never leave your phone. PROJECT 2 — PREDICT FROM A SPREADSHEET Bring a CSV (or use the sample). Boosted trees train on our servers and are judged on rows held back before training, against two simple baselines. If a column gives the answer away, VibeModel catches the leak and shows what the model scores without it. PROJECT 3 — SORT TEXT BY MEANING Bring tickets, notes or reviews with a label. Word counts and a frozen sentence encoder, each with a trained layer, compete with two no-training baselines — the most common label and word overlap — on held-out rows. SEE WHAT YOU BUILT Every model you build is drawn in 3D from its real structure and your real numbers: SigLIP 2's layers from its own configuration, your examples in feature space, the weights or trees you trained. Nothing is decorative; an Evidence list gives every number and where it came from. HONEST VERDICTS "Use base" is a real result, not a failure: sometimes the pre-trained model already does the job. Every verdict names its baseline, its unit and how many examples it was measured on. Also: a zoo of famous model architectures, drawn from their published configs; model cards you keep; layouts that work at the largest text sizes, and a text Evidence list beside the 3D for VoiceOver.