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.
9 min read52 articles
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.
9 min readPick a model and a GPU to see how weights, KV cache and batch size set memory, speed and cost.
8 min readAsk a question and watch a language model search, read and calculate its way to an answer, step by step.
7 min readRun a real detector on a photo or your webcam, then tune the thresholds that turn 1,917 guesses into a few boxes.
7 min readSmall visual labs for consensus, indexes, embeddings, classifiers, and language-model sampling.
2 min readReading OpenAI’s Navier–Stokes construction through its core geometry, energy scaling and smooth forcing.
6 min readType any labels and let CLIP pick one for a photo, then search images with plain text.
5 min readRound a real detector’s weights to 8, 4 or 2 bits and see which layers, and which boxes, break first.
6 min readEmbed every passage on this site in your browser, then compare what PCA, t-SNE and UMAP keep and lose.
5 min readUse matched seeds and original diffusion trajectories to separate conditioning, scheduler choices and evaluation claims.
5 min readReproduce write skew, compare serializable validation with a shared conflict row, and connect long snapshots to version retention.
5 min readSearch real 200D embeddings and separate IVF routing, scalar quantization, refinement, filtering and projection losses.
6 min readUse Docker Compose to see load balancing, network isolation, statelessness, and the difference between a layer and a tier.
3 min readDefine rollout gates, preserve schema compatibility, and make rollback a routine operation.
4 min readUse full-vocabulary GPT-2 snapshots to test decoding order, display truncation errors and the difference between local and sequence scores.
5 min readTransfer learning with a frozen MobileNet embedding and a k-nearest-neighbor head, and where the trick breaks down.
5 min readRetrain a linear classifier under class weighting, oversampling, undersampling, and SMOTE, and watch accuracy diverge from minority recall.
6 min readStep through boosting rounds on noisy 1D data and see shrinkage and tree depth trade convergence speed for stability.
5 min readDrag a starting point across a bowl, a ravine, and a Rosenbrock valley and watch gradient descent, momentum, RMSProp, and Adam disagree.
9 min readMeasure retrieval recall per question and compare pipeline changes against a versioned evaluation set.
2 min read · NotebookSeparate representation, optimization and generalization with controlled classification experiments.
4 min readQuorums, log freshness, commit rules and why a leader cannot take its own word for a read.
5 min readTrace splits, redistribution and range scans, then account for the storage costs hidden by big-O.
6 min readUse a decoder roofline to reason about weight reuse, cache traffic, quantization, and serving latency.
4 min readFollow the residual stream, test causal routing, and budget the KV memory behind long-context inference.
4 min readConnect local image features to receptive fields, sampling, aliasing, and per-layer compute costs.
4 min readInspect signed responses, boundary conditions, dilation, and the spectral behavior of a small filter.
4 min readSeparate advection, numerical diffusion, and pressure projection, then measure Jacobi convergence.
4 min readA Python notebook showing how a small slow cohort changes tail latency while the average stays low.
2 min read · NotebookFrom curved light paths to sparse radio measurements: the geometry and reconstruction behind the EHT image.
6 min readCorrelate alerts without losing source events, clock uncertainty, or the ability to reject a bad grouping.
2 min readMeasure Python object overhead, choose the right storage representation, and account for conversion costs.
2 min readUse Go’s syntax tree to turn dependency direction, naming rules, and yesterday’s production bug into CI checks.
3 min readDesign platform defaults around recurring developer tasks, explicit ownership, and measurable recovery.
2 min readHow a small SRE team used CDK constructs, operating guides, and guardrails instead of building an internal PaaS.
4 min readA scheduled plan, an exit code with three meanings, and enough deduplication to keep drift alerts useful.
3 min readCentralize small configuration backups with Ansible, save snapshots only when files change, and let systemd report the result.
3 min readSpatial grounding depends on preserving coordinates, local visual tokens, and training the decoder to speak in regions.
5 min readA LangGraph loop that separates retrieval failures from writing failures and retries each one differently.
3 min readTerraform creates the platform, Argo CD reconciles workloads, and Karpenter supplies the nodes they actually need.
3 min readSeparate the three-year decisions from next month’s tuning knobs, and the document becomes much easier to trust.
2 min readChoose test boundaries by risk, feedback speed, and maintenance cost—not by chasing a flattering coverage number.
3 min readTool protocols and agent protocols solve different boundaries; the alphabet soup gets easier once the arrows are drawn.
3 min readA practical plan for learning Rust and Kubernetes without pretending the first purchase needs to be a tiny data center.
3 min readA practical map of retrieval quality, answer quality, and the tests needed before a grounded answer earns trust.
3 min readBuild Workload Identity Federation with Terraform and let GitHub Actions impersonate a service account safely.
2 min readKeep UI-shaped reads fast and domain-shaped writes honest, without deploying another service for sport.
3 min readHow an image becomes a sequence of patches, and why attention changed the computer-vision toolbox.
2 min readUse Terraform and OIDC to exchange short-lived workflow tokens for narrowly scoped AWS credentials.
2 min readPhi-3.5 Vision and Florence-2 look similar at the input, but they are built for rather different jobs.
3 min readA small state machine that writes a workflow, checks its own work, and knows when to ask a human.
3 min readA Raspberry Pi, WireGuard, and one mildly hostile captive portal walk into an apartment network.
5 min read