Interactive Labs

Explore the trade-offs behind an ML system. Follow a continuing project, predict what a design choice will change, and use the evidence to explain your recommendation.

Start with your volume

Every volume begins with Lab 00. Choose or restore your project, try the notebook, and see how the chapter labs build toward a final discussion. No previous saved progress is required.

Volume I

Onboarding + 16 chapter labs

Start Lab 00 →

Explore the Volume I route

1. 00 Welcome to your deployment project

   Choose a deployment project and practice predicting, comparing, and saving your work.

2. 01 The system runs. Is it useful?

   Choose the next data, algorithm, or machine intervention using separate evidence of behavior and execution.

3. 02 Where should the work live?

   Defend a conditional placement or an explicit hold by screening locality, capacity, operation, and partition obligations.

4. 03 A late requirement makes earlier work conditional

   Trace a requirement through the lifecycle and plan a justified repair, retest, and return path.

5. 04 Make the evidence trustworthy and deliverable

   Create a versioned data contract that supports independent evaluation and a functioning input path.

6. 05 Open the model and inspect its state

   Explain how topology, gradient dependencies, phase, and batching change work, movement, and live state.

7. 06 Choose structure for the task, not the label

   Choose a task-supported architecture while explaining its information assumptions and dominant resource regime.

8. 07 Make the graph executable

   Choose an execution plan whose visibility, setup costs, backward state, and target support match the work.

9. 08 Train to a target, not to a busy GPU

   Choose or reject a training recipe using state fit, effective batch, numerical progress, and a cause-specific bottleneck diagnosis.

10. 09 Make fewer examples earn their cost

    Choose a selected pool or retain the full pool using learning value, coverage, complete preparation cost, and realized delivery.

11. 10 The checkpoint is not the deployed artifact

    Choose a supported deployment artifact on a quality/resource frontier, or hold conversion when evidence is missing.

12. 11 A faster kernel must survive the whole path

    Choose a legal mapping and resource intervention with counted traffic and an explained application-level consequence.

13. 12 Earn the right to make a performance claim

    Decide which performance and behavioral claims are supported, contradicted, or still unresolved.

14. 13 Serve useful work when demand arrives

    Choose a serving policy with an explicit demand envelope, waiting behavior, state fit, and lost-service accounting.

15. 14 An update breaks the contract without crashing

    Create an evidence-based incident response with replayable lineage, feasible rollout, and tested recovery.

16. 15 Useful for whom, under which obligations?

    Recommend release, restricted use, baseline retention, mitigation, or hold using justified obligations and observed consequences.

17. 16 Defend one coherent deployment decision

    Integrate earlier concepts into a release, restriction, baseline, fallback, or hold decision without inventing a new score.

Volume II

Onboarding + 17 chapter labs

Start Lab 00 →

Explore the Volume II route

1. 00 Choose a fleet project

   Choose a fleet track and rehearse prediction, immediate feedback, saved comparisons, and explicit saving before technical investigations.

2. 01 Introduction

   Defend when fleet coupling is necessary and how scaling, failures, and partitions change useful work.

3. 02 Compute Infrastructure

   Select a workload-matched accelerator, node, and rack envelope that passes capacity, movement, facility, and ownership-cost constraints.

4. 03 Network Fabrics

   Choose a fleet fabric by matching message size, shared cuts, isolation, and failure evidence to the traffic contract.

5. 04 Data Storage

   Place and package data so the complete input path feeds the fleet and checkpoint publication preserves restorable state.

6. 05 Collective Communication

   Choose a semantically correct collective and compare its topology, overlap, and compression costs at a fixed useful-work target.

7. 06 Distributed Training

   Defend a parallel training plan against memory, communication, pipeline, and time-to-target limits.

8. 07 Preserve valid work when failures become routine

   Preserve valid work when failures become routine

9. 08 Allocate useful fleet progress under competing claims

   Allocate useful fleet progress under competing claims

10. 09 Recover useful progress from a busy fleet

    Recover useful progress from a busy fleet

11. 10 Serve acceptable requests on time at a sustainable cost

    Serve acceptable requests on time at a sustainable cost

12. 11 Decide whether and how a heterogeneous fleet should learn

    Decide whether and how a heterogeneous fleet should learn

13. 12 Operate and change a coupled model portfolio

    Operate and change a coupled model portfolio

14. 13 Security and Privacy

    Choose controls for an explicit fleet adversary, compare privacy and runtime costs, and bound a release claim with lifecycle evidence.

15. 14 Robust AI

    Diagnose silent failure, bound defense and fallback decisions, replay layered poisoning response, and distinguish label coverage from generated-claim evidence.

16. 15 Sustainable AI

    Compare fleet interventions under explicit useful-work, facility, lifecycle, and demand boundaries.

17. 16 Responsible AI

    Make an accountable release decision from distinct fairness, observation, capacity, and data-governance evidence.

18. 17 Fleet Synthesis

    Replay inherited evidence, diagnose a cross-layer failure, compare equal-budget repairs, and defend a bounded fleet decision.

Volume III

Onboarding + 20 chapter labs

Start Lab 00 →

Explore the Volume III route

1. 00 Choose an agent project

   Choose an agent project and practice using comparisons to explain a systems decision.

2. 01 Introduction

   Define the repair contract and defend how much autonomy its evidence can support.

3. 02 The Foundation Model

   Choose an invocation and output interface that fits the request, preserves status and supports scoped checks.

4. 03 Test-Time Compute

   Choose a bounded deliberation policy, explain what evidence it adds, and defend when to abstain.

5. 04 Context Engineering

   Assemble and audit a bounded, current working set that supports the next repair decision.

6. 05 KV Cache Management

   Choose exact state allocation and waiting policies under a real pool bound, then evaluate any semantic approximation separately.

7. 06 Long-Term Memory

   Choose a bounded memory policy that preserves provenance, honors writes/deletions and supports current repair decisions.

8. 07 Tool Calling

   Defend a tool-call policy that distinguishes structural validity, action-time permission, effect settlement, and bounded observation.

9. 08 Agent Sandboxes

   Choose a containment and reset policy whose permitted reach and carried state match the repair task.

10. 09 Agent Security

    Defend how untrusted content may influence proposals without granting new authority or releasing protected information.

11. 10 The Agent Harness

    Choose a continuation and admission policy that accounts for task budgets, pending liability, and approved authority.

12. 11 Durable Execution

    Defend a restart protocol that reconciles durable intent, actual effects, and valid authority after interruption.

13. 12 Failure Recovery

    Choose a recovery route that contains partial effects and preserves a defensible task outcome.

14. 13 Agent Evaluation

    Defend what a matched evaluation establishes about task success, repeated reliability, and grader error.

15. 14 Operating Agents

    Choose what evidence to retain and whether the candidate warrants a controlled release, restriction, or hold.

16. 15 Trajectory Curation

    Defend which observed failures warrant intervention and which trace records may legitimately enter a learning dataset.

17. 16 Trajectory Fine-Tuning

    Choose where a behavior belongs and construct a faithful, correctly weighted training record.

18. 17 Reinforcement Learning from Verifiable Rewards

    Defend a verifiable-reward learning proposal against proxy rewards, credit errors, invalid comparisons, and rollout limits.

19. 18 Multi-Agent Coordination

    Choose whether delegation improves the task relative to a capable controller under the same permissions and budget.

20. 19 Agent Economics

    Defend an architecture using complete cohort accounting, useful throughput, conserved budget liabilities, and accepted outcomes.

21. 20 Conclusion

    Defend a scoped agent-system dossier by combining prior evidence, unresolved obligations, and explicit reversal conditions.

Volume IV

Onboarding + 18 chapter labs

Start Lab 00 →

Explore the Volume IV route

1. 00 Start the physical AI investigation

   Learn how to investigate tradeoffs without confusing a model of a machine with evidence from a physical machine.

2. 01 The Causal Boundary

   Decide what may remain a proposal, what requires independent permission, and which delay reduction preserves elementary stopping clearance.

3. 02 The Physical Body

   Choose a usable operating condition by checking all physical budgets, not by trading one failed budget against another budget’s slack.

4. 03 The Cognitive Brain

   Choose a proposer split and chunk supply that fits computation and evidence validity while keeping model output checkable and nonauthoritative.

5. 04 The Nervous System

   Choose transport and expiry rules whose bounded permission path remains able to stop while the proposer stalls or floods resources.

6. 05 Physical Data

   Choose what to collect, label, retain and claim so that data covers the learner’s recoverable states without hiding physical cost or provenance.

7. 06 Policy Training

   Choose a training regime and decoder whose support and timing fit the latch task, then hand a versioned candidate to evaluation without granting it physical authority.

8. 07 Closed-Loop Evaluation

   Choose a bounded claim and evidence campaign whose population, complete denominator, outcome channels and statistical limits match that claim.

9. 08 Sensor Perception

   Choose sensing and a dated observation contract whose physical quantity, qualified error, age and failure behavior support the intended consumer.

10. 09 Spatial Memory

    Choose what a stored belief can still support, when it expires, and what requires a new observation rather than a more confident estimate.

11. 10 Grounded Intent

    Which target can the permission side grant, for how long, and which requests must be refused before trajectory planning?

12. 11 Trajectory Planning

    How much executable motion should a record contain, and which seam and stop must remain valid when the next record never arrives?

13. 12 Safety Enforcement

    Which commands can pass, which can be minimally corrected, and which state-matched fallback still has valid premises?

14. 13 Workload Placement

    Which coupling must be removed or bounded, and what loaded evidence is required before accepting a placement?

15. 14 Supervisory Intervention

    When can a person take proposal authority on a named channel, and how does the machine protect itself before that person is ready?

16. 15 Adversarial Verification

    Which fault trials can refute the containment claim, what would count as valid evidence, and which conclusions remain out of reach?

17. 16 Deployment Release

    Which configuration and envelope, if any, does the evidence justify, and what must withdraw that permission?

18. 17 Responsible Physical AI

    Who can approve each premise, who bears the remaining exposure and work, and what evidence must prevent an unfair or unaccountable deployment?

19. 18 Engineering Physical AI

    Which claims transfer as reasoning, which need new evidence, and what is the next experiment worth performing?

Labs run as full pages in your browser, with the book’s navigation. The first launch downloads the runtime and dependencies and may take a few minutes.

What you will do

1. Read and predict. Meet the situation, see the learning goals and complete route, and predict a consequence before opening an investigation.
2. Change and compare. Change one condition and inspect the immediate feedback. Explain what improves, what costs more, and what remains uncertain.
3. Capture and discuss. Keep contrasting evidence, defend a recommendation or hold, and say what would change your mind. Download the discussion to submit through your course’s usual system.

The book explains concepts, TinyTorch builds the machinery, these labs explore design trade-offs, and StaffML practices defending decisions. Each chapter lab develops its own concepts within your continuing project.

Saving your work

Capturing keeps a comparison fixed while you continue experimenting. Saving a discussion is an explicit action; later changes can make that saved evidence stale. The Design Ledger can retain saved decisions when browser storage is available. Keep a downloaded report for a portable record, especially when storage is unavailable. Storage does not synchronize across devices or submit work to your instructor.

LLM assistance can help with hypotheses and drafting. Your claims still need support from the experiment, including its assumptions and limits. Calculation Notes explain the governing models when you want to dig deeper.

Local notebooks and hardware

Browser labs need no local installation. To modify a notebook or run it locally, install the repository dependencies:

python3 -m pip install -r labs/requirements.txt
python3 -m pip install -e mlsysim
cd labs
marimo run vol1/lab_00_introduction.py

Install dependencies and launch once while connected before relying on offline use.

For a separate hands-on robotics experience, explore the Volume IV Physical AI Hardware Studio.