FOLLOW YOUR CURIOSITY

What happens if…?

Choose an experiment by topic, refresh an idea, or take an optional side path. Registered experiments link to their teaching lessons; prototypes are marked as work in development.

Build a saved operating plan

Build something you can evaluate

Understand Neural Networks PROJECT

Classifier workbench

Train, compare, test, and explain a small network.

Build AI Systems PROJECT

Puzzle-agent workbench

Compare a learned policy with exact search and evaluate new rules.

Experiments by topic

134 experiments

How does a network learn?

Understand Neural Networks

Your first neuron

Turn inputs, a weight, and a bias into a prediction.

Try experiment Read the lesson Standalone playground

Follow the slope of a mistake

Turn errors into a loss curve, then find out why the size of a learning step matters.

Try experiment Read the lesson

Make a problem easier to see

Build XOR with two hidden neurons and discover what nonlinear activations add.

Try experiment Read the lesson

Send the error backward

Trace a mistake through two weights, calculate both gradients, and then update them together.

Try experiment Read the lesson

Learn from small batches

Inspect individual gradient arrows, average a mini-batch and compare learning under explicit example and update budgets.

Try experiment Read the lesson

Use momentum and adaptive steps

Compare gradient descent, accumulated momentum and Adam while inspecting their state and actual parameter changes.

Try experiment Read the lesson

Schedule the learning rate

Compare warmup, step decay and cosine decay, then choose a checkpoint using declared development checks and stopping rules.

Try experiment Read the lesson

Initialize and normalize

Inspect activation and gradient scales, break hidden-unit symmetry, and compare batch and layer normalization in training and evaluation modes.

Try experiment Read the lesson

Trace disappearing and exploding gradients

Follow exact derivatives through depth and time, compare identity residual paths and inspect globally clipped parameter updates.

Try experiment Read the lesson

Make numerical computation trustworthy

Break naive softmax, repair its log-domain loss and check gradients with explicit perturbation sizes and tolerances.

Try experiment Read the lesson

Choose complexity using development evidence

Compare model capacity, weight penalties, dropout and training duration using saved classifier candidates before opening the test set.

Try experiment Read the lesson

Compare manual and automatic gradients

Build a small computation graph, inspect each backward contribution and compare automatic gradients with manual formulas and finite differences.

Try experiment Read the lesson

What has the network learned?

Understand Neural Networks

Learn the pattern, not the examples

A model can fit every training point and still fail on the next one. Find out how to tell.

Try experiment Read the lesson

What does this neuron notice?

Draw an image with 25 pixels, inspect each weight, and see what makes a detector respond.

Try experiment Read the lesson

Catch a model taking a shortcut

Change the background, keep the shape, and investigate what a perfect score was hiding.

Try experiment Read the lesson

Slide a detector across an image

Reuse one filter at many positions and trace how stride, borders and pooling change the result.

Try experiment Read the lesson

Teach useful invariances

Train on transformed images, check whether labels survive, and compare failure modes under a matched budget.

Try experiment Read the lesson

Compare convolution with image attention

Compare local filters with patch attention, then test which changes in image arrangement remain visible.

Try experiment Read the lesson

Turn sound into features

Connect samples, local frequency measurements and a spectrogram, then expose aliasing and window tradeoffs.

Try experiment Read the lesson

Order changes meaning

Train a small recurrent predictor, follow its hidden state and expose what a bag of events forgets.

Try experiment Read the lesson

Learn from relationships between examples

Train an encoder from paired views, diagnose collapsed features and test what a simple classifier can still recover.

Try experiment Read the lesson

Reuse a representation

Compare a new network with frozen and partly adapted features under matched target-label budgets.

Try experiment Read the lesson

Learn with objects and relationships

Train shared graph messages, change directed relationships and test whether reordering objects changes their predictions.

Try experiment Read the lesson

Locate and segment objects

Draw boxes and masks, expose merged and missed objects, and evaluate a tiny pixel classifier on real images.

Try experiment Read the lesson

Estimate depth and pose

Move paired cameras, reconstruct points and challenge pose estimates with wrong correspondences.

Try experiment Read the lesson

How does AI process language?

Understand Neural Networks

Choose the pieces a model predicts

Fit character, word and subword tokenizers; inspect unknown text and a count-based next-token baseline.

Try experiment Read the lesson

Give words a place in space

Explore vectors, similarity, and why a word’s representation can change with context.

Try experiment Read the lesson

Compare, normalize, combine

Edit queries, keys, and values, and trace the weighted sum that lets tokens share context.

Try experiment Read the lesson

Follow a token through a Transformer

Inspect a tiny numerical decoder from token IDs to the next-token distribution.

Try experiment Read the lesson

Predict the next token

Shift inputs and targets, inspect causal visibility, calculate cross-entropy and compare reference versus generated prefixes.

Try experiment Read the lesson

Prepare a text corpus

Inspect text provenance, repair duplicate leakage, compare filters and control training exposure.

Try experiment Read the lesson

Train a tiny language model

Train a causal attention decoder, inspect actual gradients, restore checkpoints and compare position-level losses against a count baseline.

Try experiment Read the lesson

Choose how to sample

Compare greedy, temperature, top-k and top-p decoding on trained checkpoints with fixed random draws and separate quality diagnostics.

Try experiment Read the lesson

Adapt a pretrained model

Compare full fine-tuning with a rank-two output update, inspect its arithmetic and measure new versus retained behavior.

Try experiment Read the lesson

Learn from preferences

Train a response ranker from editable pairwise labels, examine rater disagreement and expose a polished but false proxy exploiter.

Try experiment Read the lesson

Connect instruction tuning, preferences and verified rewards

Compare actual supervised, DPO-style and exact expected-reward updates on a finite response set; expose a weak verifier and inspect a fixed-reference penalty.

Try experiment Read the lesson

Use a pretrained model responsibly

Inspect token IDs, context budgets, genuine recorded completions, output schemas and source support.

Try experiment Read the lesson

Evaluate language tasks with evidence

Design a rubric and reserved task suite, inspect actual model failures, and distinguish token loss, source support and task success.

Try experiment Read the lesson

Why does AI need computation?

Understand Neural Networks

Spend a limited compute budget

Balance model size and training data, distinguish active weights from stored weights, and measure a small browser computation.

Try experiment Read the lesson

Package a model as a component

Export and reload a real local model bundle; check named inputs, fitted preprocessing, dependency hashes and output meanings.

Try experiment Read the lesson

Measure and optimize inference

Profile actual local inference, separate cold and warm traffic, and inspect precision error before choosing an optimization.

Try experiment Read the lesson

Version data and experiments

Reconstruct a result from local data, code, configuration and artifact records; locate changed preprocessing.

Try experiment Read the lesson

Deploy with a reversible plan

Compare local and service boundaries, inspect shadow and staged traffic, and restore stable routing without erasing history.

Try experiment Read the lesson

Test the complete system

Build a test suite and expose what unit, numerical, data, integration and behavior checks catch or miss.

Try experiment Read the lesson

Monitor what changed

Replay changing traffic, request delayed labels and allocate limited alert reviews.

Try experiment Read the lesson

Respond to an incident

Trace a release regression, contain its impact and verify recovery without rewriting the incident history.

Try experiment Read the lesson

Data audits and training diagnosis

Understand Neural Networks

What should the machine predict?

Define inputs, outputs, users and success before choosing a model. Test how a changed target changes the labels.

Try experiment Read the lesson

Examples are measurements

Inspect missing values, copied requests, unit mismatches and conflicting labels before treating a table as evidence.

Try experiment Read the lesson

Decide where AI belongs

Compare a rule, supplied candidate predictions and simulated human review, including costs and unfinished work.

Try experiment Read the lesson

Repair data without leaking answers

Audit real flower measurements, separate related rows, and turn a tempting score into a defensible claim.

Try experiment Read the lesson

Repair a broken training run

Inspect predictions, gradients and actual parameter changes. Compare a failure and repair from the same checkpoint.

Try experiment Read the lesson

Different mistakes have different costs

Move a threshold, count different errors, and defend a decision using precision, recall and explicit costs.

Try experiment Read the lesson

How much should a small result convince you?

Draw repeated samples, inspect uncertainty and discover why more biased data can strengthen the wrong claim.

Try experiment Read the lesson

Who receives the errors?

Trace a limited review budget through two channels and compare who is missed, reviewed unnecessarily or absent from the evidence.

Try experiment Read the lesson

Use data with a purpose

Choose justified fields, examine provenance and permission, and build a practice data card with explicit handling limits.

Try experiment Read the lesson

Explain what your model is for

Turn results into bounded claims, identify missing evidence and draft a model card linked to your actual experiment.

Try experiment Read the lesson

Simple models and useful baselines

Understand Neural Networks

Start with a simple baseline

Fit majority, nearest-neighbor and shallow-tree predictors; compare the same validation cases before adding complexity.

Try experiment Read the lesson

Fit a line and explain its errors

Move a line over licensed flower measurements, inspect residuals and compare a fitted model with a mean baseline and shuffled labels.

Try experiment Read the lesson

Predict a class probability

Train a two-feature probability model, inspect its gradients and compare estimates with outcomes before choosing an action threshold.

Try experiment Read the lesson

Combine many imperfect trees

Fit bootstrap trees, inspect their votes and out-of-bag coverage, and test whether shared mistakes survive a measurement shift.

Try experiment Read the lesson

Correct residual errors with boosting

Fit small trees to the mistakes left by earlier predictions, inspect each correction and choose complexity using separate development evidence.

Try experiment Read the lesson

Find structure without labels

Alternate k-means assignments and center updates, then project onto a principal axis and inspect distinctions that vanish.

Try experiment Read the lesson

Find unusual examples

Fit a reference-distance detector, inspect unseen normal behavior and count false alerts when confirmed problems are rare.

Try experiment Read the lesson

Recommend from incomplete feedback

Complete a small ratings picture with predictions, compare simple and learned recommenders, and inspect cold start and exposure bias.

Try experiment Read the lesson

Validate without reusing the answers

Inspect cross-validation folds, repair duplicate and time leakage, and fit preprocessing inside the training boundary.

Try experiment Read the lesson

Finish a tabular experiment

Turn a real-data comparison into a frozen, reproducible report with error examples and carefully bounded claims.

Try experiment Read the lesson

Generative models beyond text

Understand Neural Networks

Compress and reconstruct

Train a tiny shape autoencoder, inspect its bottleneck and test why reconstruction does not guarantee useful generation.

Try experiment Read the lesson

Generate with a variational autoencoder

Train a tiny probabilistic encoder and decoder, trace a noisy latent sample and compare reconstruction with KL regularization.

Try experiment Read the lesson

Infer what cannot be directly observed

Fit a two-component mixture in separate EM steps and compare exact latent inference with a restricted variational approximation.

Try experiment Read the lesson

Generate by removing noise

Train a small 2D noise predictor, inspect corruption arithmetic and manually step a frozen DDPM sampler.

Try experiment Read the lesson

Generate by learning a flow

Train a velocity field and compare Euler and Heun generation on the diffusion lesson’s four-mode target.

Try experiment Read the lesson

Control the generated result

Train a shape-conditioned flow and compare guidance, validity and coverage at equal sampling work.

Try experiment Read the lesson

Learn with a generator and discriminator

Train an actual tiny GAN, inspect discriminator gradients and compare missing modes with a reproducible diffusion reference.

Try experiment Read the lesson

Judge a generated distribution

Compare fidelity, mode coverage, copying flags, task gates and licensed source receipts instead of trusting one score.

Try experiment Read the lesson

Learning through actions

Build AI Systems

Explore before you decide

Choose between uncertain rewards and discover why trying a new option can improve later decisions.

Try experiment Read the lesson

Learn from delayed rewards

Teach an agent a route, inspect a Q-learning update, and uncover a reward rule that rewards the wrong behavior.

Try experiment Read the lesson

Solve the Bellman equations

Compute values in a small looping decision process, separate policy evaluation from improvement, and compare exact expectations with sampled returns.

Try experiment Read the lesson

What can the agent observe?

Hide a cart’s velocity, compare aliased observations and recover motion from a short exact history.

Try experiment Read the lesson

Copy a demonstrated action

Train a puzzle policy from exact expert targets, inspect legal masks and compare proposals after a non-expert move.

Try experiment Read the lesson

Learn from verified self-play

Generate game experience, check it against exact rules, and use verified outcomes to change future decisions.

Try experiment Read the lesson

Know which policy produced the data

Compare a logged behavior policy with a target policy using trajectory weights, coverage checks and bounded replay.

Try experiment Read the lesson

Stabilize neural Q-learning

Train a small neural action-value model and compare replay sampling with delayed target networks.

Try experiment Read the lesson

Increase the probability of useful actions

Learn a stochastic policy from sampled returns and inspect the variance of actual updates.

Try experiment Read the lesson

Use a critic to guide learning

Train separate policy and value heads, trace their coupled updates, and audit a biased bootstrap.

Try experiment Read the lesson

Limit a policy update

Compare raw and clipped probability-ratio objectives on the same frozen batch, and test the limits of clipping.

Try experiment Read the lesson

Trade off reward and exploration with entropy

Train a bounded continuous actor and twin critics, and compare task reward with action-distribution entropy.

Try experiment Read the lesson

Learn without new interactions

Fit policies from a fixed log, trace unsupported action values, and test a conservative objective against behavior cloning.

Try experiment Read the lesson

Evaluate learning fairly

Compare complete seed cohorts, data and fitting budgets, frozen checkpoints and task success under a misleading reward bonus.

Try experiment Read the lesson

Planning and reasoning

Build AI Systems

Reason with procedures and search

Compare a local shortcut with a procedure that explores alternatives and checks its result.

Try experiment Read the lesson

Use an informed search

Compare uniform-cost search and A* on a weighted map, audit estimates and reopen a location when a cheaper route arrives.

Try experiment Read the lesson

Solve a constraint problem

Build a small timetable, remove unsupported choices and compare backtracking with constraint propagation.

Try experiment Read the lesson

Update beliefs from evidence

Use sensor frequencies to distinguish detection from posterior probability and test how base rates and repeated evidence change a belief.

Try experiment Read the lesson

Factor a probabilistic model

Build an eight-row joint distribution from local tables, test conditional independence, and separate observing a variable from intervening on it.

Try experiment Read the lesson

Infer state across time

Track a hidden moving object by alternating motion predictions and noisy sensor updates, while keeping a distribution over possible positions.

Try experiment Read the lesson

Approximate a belief with particles

Propagate, weight and resample possible positions. Compare a seeded particle filter with exact inference and diagnose lost diversity.

Try experiment Read the lesson

Reason about an opponent

Play a small stochastic game, reveal its search tree and compare adversarial planning, probabilistic planning and a policy learned from sampled returns.

Try experiment Read the lesson

Ask what would change under intervention

Change a causal graph, compare observed and randomized outcomes, and discover why a population effect does not determine an individual counterfactual.

Try experiment Read the lesson

Predict what an action will change

Learn transitions from observations, compare imagined and actual routes, and test when replanning helps.

Try experiment Read the lesson

Refine an idea over several passes

Watch distance information spread through a grid and discover why stable internal states can still be wrong.

Try experiment Read the lesson

Test the parts of a planning system

Compare direct action, search, and refinement under a shared planning budget, then test unfamiliar layouts.

Try experiment Read the lesson

Learning without losing old skills

Build AI Systems

Build a trustworthy practice loop

Generate examples, filter them with a verifier, train a candidate, and protect earlier skills before promotion.

Try experiment Read the lesson

Learn something new without losing the old

Train a shared neural network on two tasks, measure forgetting, and compare replay with a parameter-change penalty.

Try experiment Read the lesson

Learn a better way to adapt

Optimize a starting weight for rapid adaptation, then evaluate adapted copies on a task excluded from meta-training.

Try experiment Read the lesson

Connecting tools and specialists

Build AI Systems

Define an assistant’s job

Specify an assistant’s inputs, scope, permissions, evidence and stopping rules before choosing a model or adding tools.

Try experiment Read the lesson

Call a tool, then check the evidence

Propose structured calls, inspect independent runtime checks, and recover from simulated failures.

Try experiment Read the lesson

Help specialists understand each other

Compare prose, structured records, and vectors; diagnose lost information and mismatched coordinate meanings.

Try experiment Read the lesson

Learn a shared representation contract

Train an adapter from paired vectors, expose missing information, and propagate messages through typed relations.

Try experiment Read the lesson

Choose where a connected system learns

Compare frozen, adapter-only, partial, and joint training while checking an existing consumer for regressions.

Try experiment Read the lesson

Keep documents from becoming instructions

Trace document-induced proposals and enforce a task-specific runtime boundary without confusing permission with correctness.

Try experiment Read the lesson

Memory and shared working state

Build AI Systems

Find relevant information

Compare keyword and learned-vector retrieval, change passage boundaries, and inspect what a ranking score can and cannot tell you.

Try experiment Read the lesson

Answer with evidence

Build an extractive answer with precise citations, reject unsupported claims, and make conflicts and missing facts visible.

Try experiment Read the lesson

Share the right information at the right time

Control a three-slot event workspace and compare shared processing with a fast local response loop.

Try experiment Read the lesson

Remember selectively and retrieve evidence

Compare similarity with exact and evidence-aware retrieval, age records, and consolidate a carefully scoped procedure.

Try experiment Read the lesson

Choosing what to do next

Build AI Systems

Decide whether another computation is worth it

Compare acting now with gathering evidence, then follow each observation to a different next action.

Try experiment Read the lesson

Coordinate dependencies and reusable skills

Schedule dependent tasks, repair a stale-file conflict, and execute a bounded verification skill.

Try experiment Read the lesson

Learn from routing evidence without fooling yourself

Expose selection bias, compare specialists on the same contexts, and check confidence against outcomes.

Try experiment Read the lesson

Train and evaluate a small neural router

Train a 6 → 8 → 5 network, inspect its action probabilities, and compare it with a strong rule under the same tool budgets.

Try experiment Read the lesson

Testing whether improvement is real

Build AI Systems

Measure the whole workflow

Run an assistant through a fixed case suite and diagnose retrieval, evidence, tool and runtime failures separately.

Try experiment Read the lesson

Test a proposed system improvement

Expose a hidden regression, set release gates, and promote or restore a version with its evidence attached.

Try experiment Read the lesson

Find out whether improvement generalizes

Select a lucky benchmark winner, remove components, and compare step accuracy with whole-workflow reliability.

Try experiment Read the lesson

Assemble an experimental agent

Build AI Systems

Connect an agent’s decisions, actions, and learning

Trace a proposal through a controller, checked execution, episode memory, and a separately updated world model.

Try experiment Read the lesson

Build an architecture that earns its complexity

Train a policy/value network, compare search and memory, freeze a rule-transfer test, and plan a disciplined research program.

Try experiment Read the lesson

Geometry and control foundations

Build AI Systems

Describe where the robot is

Move a two-link arm, express points in named frames and distinguish ideal reach from verified motion.

Try experiment Read the lesson

Represent motion in three dimensions

Compare fixed and moving axes, matrix and quaternion descriptions, and the limits of a projected 3D picture.

Try experiment Read the lesson

Use a Jacobian

Predict small tool movements, inspect missing directions and compare inverse velocity proposals.

Try experiment Read the lesson

Turn a target into motion

Find alternative target-reaching postures and separate geometric reach, joint limits and obstacle contact.

Try experiment Read the lesson

Correct errors using feedback

Tune an ideal cart controller and inspect overshoot, velocity damping and delayed measurements.

Try experiment Read the lesson

Choose continuous actions

Train a small force controller, then compare continuous outputs, discrete choices, noise and saturation.

Try experiment Read the lesson

Stop when the controller leaves its limits

Inspect runtime rejection, delayed observations and bounded fallback without equating a stopped task with success.

Try experiment Read the lesson

Turn observations into estimates

Distinguish noisy sensor readings, delayed information, filtered estimates and simulator truth.

Try experiment Read the lesson

Fuse noisy measurements

Predict motion, correct with sensors and inspect the uncertainty assumptions of a linear Kalman filter.

Try experiment Read the lesson

Reason about forces and contact

Compare force and position commands, friction limits and the forces needed to hold a load.

Try experiment Read the lesson

Plan a collision-free trajectory

Search joint-space waypoints, try shortcuts and independently verify the full motion.

Try experiment Read the lesson

Enforce operating limits

Keep learned proposals separate from action checks, watchdogs, collision prevention and latched braking.

Try experiment Read the lesson