Five things to do with the language model already in your Chrome
Chrome can run Gemini Nano on your own machine. Here is what it is good at, where it quietly goes wrong, and live examples you can run on this page.
Essays, notebooks, and field notes connected by machine learning.
Chrome can run Gemini Nano on your own machine. Here is what it is good at, where it quietly goes wrong, and live examples you can run on this page.
Pick a model and a GPU to see how weights, KV cache and batch size set memory, speed and cost.
Ask a question and watch a language model search, read and calculate its way to an answer, step by step.
Run a real detector on a photo or your webcam, then tune the thresholds that turn 1,917 guesses into a few boxes.
Small visual labs for consensus, indexes, embeddings, classifiers, and language-model sampling.
Type any labels and let CLIP pick one for a photo, then search images with plain text.
Round a real detector’s weights to 8, 4 or 2 bits and see which layers, and which boxes, break first.
Embed every passage on this site in your browser, then compare what PCA, t-SNE and UMAP keep and lose.
Search real 200D embeddings and separate IVF routing, scalar quantization, refinement, filtering and projection losses.
Use full-vocabulary GPT-2 snapshots to test decoding order, display truncation errors and the difference between local and sequence scores.
Transfer learning with a frozen MobileNet embedding and a k-nearest-neighbor head, and where the trick breaks down.
Retrain a linear classifier under class weighting, oversampling, undersampling, and SMOTE, and watch accuracy diverge from minority recall.
Step through boosting rounds on noisy 1D data and see shrinkage and tree depth trade convergence speed for stability.
Drag a starting point across a bowl, a ravine, and a Rosenbrock valley and watch gradient descent, momentum, RMSProp, and Adam disagree.
Separate representation, optimization and generalization with controlled classification experiments.
Use a decoder roofline to reason about weight reuse, cache traffic, quantization, and serving latency.
Follow the residual stream, test causal routing, and budget the KV memory behind long-context inference.
Connect local image features to receptive fields, sampling, aliasing, and per-layer compute costs.
Inspect signed responses, boundary conditions, dilation, and the spectral behavior of a small filter.
How an image becomes a sequence of patches, and why attention changed the computer-vision toolbox.
Phi-3.5 Vision and Florence-2 look similar at the input, but they are built for rather different jobs.