HomeThe living book · by Taha Bouhsine

What changed in the course.

Each release records the content of every lesson and phase. Revisit what changed when it is useful to you; your saved checkpoints and exercise progress stay intact.

Returning readers see an update notice in this browser. Lesson and phase notices compare the material with your last visit. These notices use local browser storage and do not send email or synchronize across devices.

Version tracking begins with v0.2.0. Earlier edits do not have a recorded revision history.

v0.7.2 · 2026-10-20

Google Rich Results schema compatibility

Use course syllabus sections so Google does not interpret lesson outlines as paywall fragments.

Replaced Course.hasPart lesson outlines with syllabusSections and typed Syllabus entries while preserving lesson URLs, outcomes and free-access declarations.
Added regression checks for typed syllabus sections and the absence of Course.hasPart, following Google Rich Results Test warnings on published catalog, phase and lesson pages.

Phases in this revision

See the 0 revised lessons

v0.7.1 · 2026-10-20

Discoverable courses, career guides and free book

Add permanent learning and career guides, an indexable book overview and connected educational metadata.

Added permanent server-rendered pages for all 10 learning pathways and 7 career routes, linked from their catalogs and sitemap.
Added a free-book overview with author, edition, chapter links and HTML/EPUB formats; completed book sharing and EPUB catalog metadata.
Replaced navigation-based Course markup with connected WebPage, Course, LearningResource, Book and BreadcrumbList entities, including course provider and phase relationships.
Expanded SEO checks to every indexable page, catalog entries, visible breadcrumbs, book encodings, canonical URLs and permanent guide coverage.
Fixed a misindented code fence that rendered a shell comment as a second project heading.

Phases in this revision

See the 0 revised lessons

v0.7.0 · 2026-10-20

Book quality repairs and connected model experiments

Repair misleading project and TPU contracts, make the illustrated book portable, and connect career and model-learning evidence.

Replaced nine cloned encoder projects with distinct staged CPU contracts, changed-condition checks and learner-owned evidence; untouched starters fail the learner commands.
Corrected calculated versus observed figure provenance, checkpoint restore boundaries, TPU execution instructions, precision byte accounting and persistent profiler artifacts.
Removed mandatory TPU and unrelated language/RL work from the image-classifier route, exposed optional preparation and mapped every career skill to lesson and project-stage evidence.
Embedded standalone HTML figures; added phase and project handoffs, practical appendices, separate solutions, current EPUB metadata and meaningful image alternatives.
Removed mechanical syntax-narration comments while preserving numerical, shape and state explanations.
Split full verification into CPU lesson, sharded project, logger and site jobs; deployment now depends on their completion.
Made all ten separately planned lifecycle briefs discoverable and generated current course counts from canonical registries.
Added an actual 70,765-parameter CPU embedding lineage pilot on chapter-held-out public-domain text, with objective monitoring, independent loss checks, lineage validation, held-out retrieval versus BM25 and fresh-process export; its poor measured retrieval is retained.

Phases in this revision

See the 102 revised lessons

v0.6.0 · 2026-10-13

Reproducible JAX/Flax experiments workshop

One connected workshop turns an experimental question into real tracked Flax runs, complete paired comparisons and an evidence-linked Markdown research notebook.

Added a four-stage project with learner-owned configuration, observation and paired-comparison contracts, linked to existing Flax, randomness, tracking, recovery and distributed lessons.
Added frozen source/data/config plans, bounded process jobs, portable job scripts, retained failures and fresh retries; two-worker execution is CPU-only and accelerator qualification remains separate.
Implemented CSV, TensorBoard, offline W&B and local SQLite MLflow logging, with actual history readback and artifact checks.
Added six recorded CPU runs, explained learning-curve and paired-difference figures, a generated Markdown research notebook and instructor interpretation.
Linked the workshop from Projects, learner resources, model/operations pathways and neural training/recovery phase guides.
Expanded release tracking to workshop source and excluded generated project execution receipts so rerunning checks does not change canonical course revisions.

Phases in this revision

See the 0 revised lessons

v0.5.3 · 2026-10-08

3-step Cloud TPU mini-course, scroll stage transitions & project code blueprints

Splits Phase 19 into three 2-lesson Cloud TPU courses (Phases 19, 20, 21), adds visual stage and era transitions while scrolling, and adds full project explanations with explicit code placement blueprints.

Split the Cloud TPU onboarding track into three focused 2-lesson courses (~45 minutes each): 19 TPU Setup: Provisioning & Runtime Verification (tpu-01, tpu-04), 20 TPU Workflows: Launching Jobs & Checkpointed Experiments (tpu-02, tpu-05), and 21 TPU Systems: Generations, Memory, Precision & XProf (tpu-03, tpu-06).
Added course-type UI/UX cues (courseTypeCue), a sticky course section scroll bar, and five curriculum era transition banners across the catalog, phase guides, and pathways.
Grouped lesson reader sections into five color-coded learning stages (1 Understand, 2 Build & Run, 3 Experiment, 4 Exercise & Check, 5 Evidence) with a sticky scroll-aware stage bar and dedicated exercise workbench.
Added a full Project Blueprint at the start of every project workspace, guide, and catalog card explaining what to build, the exact workspaceFile where learners write code, the directory tree, and the stage verification workflow.

Phases in this revision

See the 103 revised lessons

v0.5.0 · 2026-10-06

TPU generations, precision and profiling

A practical TPU performance guide connects hardware choice, training and cache memory, precision error and profile interpretation to one decision report.

Added a generation comparison from legacy TPUs through TPU7x, with primary sources, chip/device counting and training versus serving workload reasoning.
Added worked model-state, KV-cache and roofline estimates, plus explicit weight, activation, accumulation, optimizer and cache precision policies.
Added a bounded CPU/TPU precision lab comparing FP32, BF16 and true INT8 dots; recorded CPU error/storage figures retain source hashes and numerical checks.
Added short warmed XProf capture, timeline diagnosis and changed-shape exercises, linked from the TPU pathway, GCP guide and developer resources.
TPU execution, multi-host scaling and model-level quantization quality remain target qualification work.

Phases in this revision

See the 0 revised lessons

v0.4.0 · 2026-10-06

TPU and Google Cloud practice, plus a clearer course reader

A practical cloud workflow with a CPU-tested training launcher, and readable math and code in the downloadable HTML book.

Extended the TPU pathway with Google Cloud project setup, gcloud provisioning, explicit backend checks, recovery, evidence retrieval and resource cleanup.
Added a downloadable launcher that reuses the workload-operations worker, rejects accidental overwrites and retains attempt history. CPU recovery and failure checks pass; TPU hardware qualification remains pending.
Improved the HTML course reader with syntax highlighting, code copy controls, expected-result panels, clearer equations and print styles. Embedded math fonts and styles preserve formatting when the HTML is saved.
Fixed shared syntax highlighting that dropped Python numeric literals containing underscores; added exact code-preservation checks across all 97 book chapters.
Added practical guides and launcher sources to release change detection.

Phases in this revision

See the 0 revised lessons

v0.3.0 · 2026-10-06

Deeper deployment lessons and reproducible release checks

Eight deployment and interoperability lessons now connect staged builds, numerical reasoning and changed-condition practice to one release dossier.

Added staged builds to framework bridging, PyTorch-to-Flax parity, export, precision, containerization and edge deployment, completing the eight-lesson phase.
Added feature-order and activation checks, retention gates, elementwise error budgets, nonuniform gradient parity, and fresh-interpreter export validation.
Expanded raw-request and artifact validation, packed-weight storage accounting, percentile interpretation and request deadline exercises.
Added four explained conceptual figures and strengthened the phase completion evidence; CPU, container and device results retain separate scopes.

Phases in this revision

Deployment, interoperability & edge AI

See the 8 revised lessons

v0.2.0 · 2026-10-06

Clearer lessons, visible revisions and searchable chapters

A course-wide teaching revision, with a recorded version for every lesson and phase.

Expanded explanations and reasoning checks across all 97 authored lessons, with 22 concept diagrams and explanations of what each figure shows.
Added project organization and CLI guidance, expanded modality project plans, and preserved explicit CPU validation and TPU limitations.
Published a permanent HTML page for every authored lesson, search metadata and a sitemap.
Added lesson and phase revision labels, a release history, and notices when this browser encounters changed material. Your saved progress stays intact.
Added contextual GitHub feedback links.

Phases in this revision

See the 97 revised lessons