Start free trial
Trending New Genres Authors
EnglishEnglish
EspañolSpanish
简体中文Chinese
繁體中文Chinese (Traditional)
FrançaisFrench
DeutschGerman
日本語Japanese
PortuguêsPortuguese
ItalianoItalian
한국어Korean
РусскийRussian
NederlandsDutch
العربيةArabic
PolskiPolish
हिन्दीHindi
Tiếng ViệtVietnamese
SvenskaSwedish
ΕλληνικάGreek
TürkçeTurkish
ไทยThai
ČeštinaCzech
RomânăRomanian
MagyarHungarian
УкраїнськаUkrainian
IndonesiaIndonesian
DanskDanish
SuomiFinnish
БългарскиBulgarian
עבריתHebrew
NorskNorwegian
HrvatskiCroatian
CatalàCatalan
SlovenčinaSlovak
LietuviųLithuanian
SlovenščinaSlovenian
СрпскиSerbian
EestiEstonian
LatviešuLatvian
فارسیPersian
മലയാളംMalayalam
தமிழ்Tamil
اردوUrdu
Searching...
SoBrief
The Hundred-Page Machine Learning Book

The Hundred-Page Machine Learning Book

Build better machine-learning models by getting the data, evaluation, and algorithm choices right.
by Andriy Burkov 2019 159 pages
4.25
1k+ ratings
Listen
Immersive
V2.3
Amazon Kindle Audible
Try Full Access for 3 Days
Unlock listening & more!
Continue

Key Takeaways

Treat machine learning as statistical pattern fitting, not human-like learning.

Useful, not human. A machine learning system finds a mathematical model from examples. It can predict labels for new inputs when those inputs resemble the training data statistically. That resemblance is the catch: a system trained on upright images may fail when the same images are rotated, unless rotation was represented in training. Animal learning is more robust to small changes, so the word learning is an analogy, not a literal description.

Start with the limits. The model is not discovering universal truth. It is capturing patterns in a particular sample and using them under assumptions about future data. The practical goal is not a machine that thinks, but a useful predictor whose errors and operating conditions are understood.

Analysis

The book's opening distinction is a valuable antidote to inflated claims about artificial intelligence. It also points toward a central scientific habit: ask what observations support a model, and under which conditions its predictions should hold. The same caution applies to any measurement system, from economic forecasts to medical tests. A model can be highly effective within its intended domain while remaining brittle outside it. Calling that limitation out is not anti-technology; it is what makes deployment more trustworthy.

Define examples, labels, and the decision before choosing an algorithm.

Make the problem machine-readable. A dataset pairs examples with labels for supervised learning. Each example becomes a feature vector, a consistent list of measurements: an email might be represented by whether each dictionary word appears, while its label says spam or not spam. The model learns a mapping from those measurements to the desired answer.

Match the learning setup. Unsupervised learning works without labels and can group examples or flag unusual ones. Semi-supervised learning combines a small labeled set with a larger unlabeled one. Reinforcement learning addresses sequential decisions, where actions change the environment and earn rewards. These distinctions matter: predicting a house price, grouping customers, and choosing a robot's next move are different problems, not interchangeable uses for one algorithm.

Analysis

This framing echoes good experimental design: specify the observation, the outcome, and the decision that follows before selecting a method. It prevents a common category error, such as treating a cluster as a meaningful customer type simply because software produced one. The feature vector also encodes a theory of what matters. If a crucial signal is absent or poorly measured, algorithmic sophistication cannot recover it from nowhere.

Choose a model's geometry and loss for the job, not its reputation.

Algorithms make different bets. A learning algorithm combines a model form, a loss that measures mistakes, and a procedure for finding parameters. Linear regression fits a line or hyperplane by reducing squared prediction errors. Logistic regression is a classifier despite its name, estimating class probabilities. A decision tree makes successive feature-based splits. A support vector machine seeks a separating boundary with a wide margin, while k-nearest neighbors predicts from nearby stored examples.

Tradeoffs are real. Kernels let an SVM represent curved boundaries without explicitly constructing every transformed feature. Simpler models may train and predict quickly or be easier to explain; more flexible ones can capture complex patterns but may fit noise. The best choice depends on accuracy, data size, interpretability, and the cost of mistakes.

Analysis

The useful lesson is that an algorithm is not a brand to adopt but a set of assumptions to inspect. Geometry determines which distinctions a model can express, while the loss function decides which errors count most. This is closely related to measurement design: optimizing a convenient proxy can produce a model that performs well on paper but poorly for the decision it serves. Simplicity is not automatically better, but complexity should earn its cost.

Tune on validation data, and reserve the test set for one honest final check.

Separate learning from judgment. The training set fits the model. The validation set helps select an algorithm and tune hyperparameters, such as an SVM's penalty or a tree's depth. The test set is held back to estimate performance on unseen examples. If repeated choices are made using test results, the test set quietly becomes another training tool and its score becomes misleading.

Watch the gap. Underfitting means the model fails even on its training examples. Overfitting means it fits training data well but performs poorly on held-out data. Regularization discourages excessive complexity; cross-validation rotates validation folds when data is too scarce for a large separate holdout. Classification also needs task-appropriate measures: accuracy can conceal rare but costly errors, while precision and recall expose different error costs.

Analysis

This is the machine learning version of keeping an independent measuring instrument out of the workshop. A test score is informative only if it has not influenced the design decisions being evaluated. The distinction resembles confirmation bias in research: once results guide repeated adjustments, they no longer offer a clean independent check. Cross-validation makes scarce data more useful, but it does not remove the need to protect the final test from repeated inspection.

Engineer informative signals before blaming the algorithm.

Raw data needs a representation. Feature engineering turns logs, text, images, or measurements into consistent inputs that a model can use. A customer-retention model might derive session frequency or average session length from interaction logs. For text, a bag-of-words representation records which terms appear. Categorical values such as colors need one-hot encoding when their categories have no natural order; assigning red, yellow, and green the numbers one, two, and three falsely suggests one.

Prepare inputs carefully. Scaling numerical features can help optimization, and missing values may be removed, imputed, or represented with an indicator. Binning can sometimes make ranges more useful. These choices are experiments, not rituals: compare alternatives on validation data, and apply the same transformations at prediction time as during training.

Analysis

Feature engineering makes explicit that data is never simply fed into a model; it is interpreted first. This is a strength, because domain knowledge can turn vague records into useful predictors. It is also a risk: a feature may encode accidental correlations, measurement bias, or information unavailable when a real decision is made. The representation deserves the same scrutiny as the algorithm, especially when its outputs affect people.

Use neural networks when layered representations can reveal patterns hand-built features miss.

Learn features in stages. A neural network composes layers of mathematical operations. Training adjusts their parameters using gradients, with backpropagation applying the chain rule to determine how each parameter affects error. Deep learning uses networks with multiple hidden layers; nonlinear activation functions let them represent relationships that a stack of purely linear operations could not.

Choose structure to fit the input. Convolutional neural networks reuse small filters across image regions, helping detect local patterns such as edges and textures without assigning an independent parameter to every pixel connection. Recurrent networks process sequences step by step, while gated units help preserve useful information across longer sequences. Encoder-decoder networks can transform one sequence into another, as in translation. Neural networks can be powerful, but they need careful training, data preparation, and regularization.

Analysis

The key shift is from manually specifying every useful signal to learning a hierarchy of representations from examples. That can be especially valuable when inputs have structure, such as neighboring pixels or ordered words. Yet learned representations are not automatically meaningful or reliable. Their success depends on the training distribution, and their internal complexity can make explanations harder. The right question is not whether deep learning is advanced, but whether its representational flexibility solves a real limitation.

Combine diverse models when one predictor's errors leave useful information unused.

Many modest models can cooperate. Ensemble learning combines predictions from multiple models. Random forests build decision trees from resampled versions of the training data and vary the features considered at each split; averaging their predictions can reduce variance. Gradient boosting instead adds trees in sequence, with each new tree addressing errors made by the current model. It can improve fit, though too many or overly complex trees can overfit.

Diversity matters. Combining highly similar models may simply repeat the same mistakes. On imbalanced data, where one class is rare, overall accuracy can hide poor detection of that class. Class weighting, oversampling rare examples, or undersampling common ones can shift the tradeoff. Algorithm choice should also account for training speed, prediction speed, memory, and explainability.

Analysis

Ensembles turn disagreement into a resource: when models make different errors, their combined judgment can be more robust. This resembles collective decision-making, where independent evidence is more useful than multiple repetitions of the same view. But the benefit depends on genuine diversity and a suitable evaluation metric. A system can improve its headline accuracy while becoming worse at the rare event that matters most, so ensemble quality must be judged against the actual cost structure.

Use unlabeled data to find structure, but do not mistake clusters for truth.

Unsupervised models describe, not certify. Clustering groups examples without known labels. K-means assigns examples to a chosen number of clusters around moving centroids. DBSCAN groups dense neighborhoods and can mark isolated examples as outliers; HDBSCAN extends this approach to clusters with varying density. Gaussian mixture models allow an example to belong to several clusters with different membership scores.

Interpretation remains a human task. Dimensionality reduction compresses many features into fewer dimensions, often to visualize data. Principal component analysis follows directions of greatest variance; UMAP aims to preserve local similarities in a lower-dimensional view. Density estimation and outlier detection can help flag unusual cases. Because there is no label-based answer key, cluster quality and usefulness depend on the purpose and on evidence beyond the algorithm's output.

Analysis

Unsupervised learning is powerful precisely where labels are unavailable, but that also makes it easy to overinterpret. A cluster is a mathematical grouping under chosen features and distance rules, not proof of a natural category. Visualization adds another layer of simplification: nearby points in a plot may not preserve every relationship in the original space. The strongest use is exploratory, generating hypotheses that can be checked against domain knowledge or later evidence.

Spend labels strategically, and reuse learned representations when data is scarce.

Labels are expensive information. Active learning selects unlabeled examples for an expert to label, often prioritizing cases where the current model is uncertain and the surrounding data is dense. Semi-supervised methods use unlabeled examples alongside a small labeled set, though automatically assigning labels can sometimes reinforce the model's mistakes. Self-supervised learning creates training signals from the data itself: word2vec, for example, learns word representations by predicting words from their surrounding context.

Transfer what already works. Transfer learning adapts a network trained on one dataset to a related task. A practitioner can replace its final prediction layers and train those new layers on a smaller labeled dataset. This can reduce the amount of new annotation required, but success depends on whether the learned representation carries useful information to the new task.

Analysis

These methods treat annotation and representation as scarce resources rather than assuming an unlimited supply of expert labels. Active learning can direct human effort toward informative cases; transfer learning can reuse structure acquired from a related problem. Both approaches have a boundary condition: an uncertain example is not necessarily the most valuable one, and a source dataset can carry mismatched assumptions. Their practical value comes from checking whether saved effort preserves the quality needed for the target decision.

Give specialized tasks an objective that reflects what success actually means.

Some outputs need more than one label. Sequence labeling assigns a label to each step, while sequence-to-sequence models map an input sequence, such as a sentence, to an output sequence. Multi-label classification permits several labels for one example. One-shot learning compares learned embeddings to decide whether two examples represent the same entity.

Measure the real task. Learning to rank orders search results, so a useful objective must reward relevant items appearing high in the list, not merely correct isolated predictions. Recommendation systems can use item features, patterns shared among users, or a hybrid of both. These problems show why a generic accuracy score is not enough: the model must be trained and assessed according to the structure and consequences of its intended use.

Analysis

Task-specific objectives make the definition of success visible. Ranking, identity matching, and recommendation are not ordinary classification with decorative labels; order, similarity, and user preference shape what a correct result means. This is a broader lesson about optimization: systems pursue the quantity they are rewarded for, not the intention behind it. A metric should therefore be treated as a compact statement of priorities and tested for the failures it leaves uncounted.

Analysis

The book's deepest contribution is less a catalog of algorithms than a practical way to reason about them. A real problem becomes a dataset through representation choices; an algorithm then fits a model under a chosen loss; evaluation estimates whether that fit travels beyond the examples used to create it. Each step contains judgment. The recurring emphasis on feature design, holdout data, hyperparameters, and model tradeoffs makes machine learning look like experimental engineering rather than a button that turns data into truth.

That framing is useful, but the book's practical focus leaves important questions at the edge of its compact tour. Similarity between training and future data is not guaranteed, especially when products, populations, or behavior change. A strong test score cannot by itself establish that a model is fair, that its target is a good proxy for the desired outcome, or that acting on its predictions is beneficial. Those questions require attention to data provenance, deployment conditions, and the consequences of errors, alongside technical performance.

The technical survey also clarifies why no single method dominates. Linear models offer speed and interpretability; trees expose decision structure; neural networks learn layered representations; ensembles combine predictors; unsupervised methods explore without labels. Choosing among them is a decision about data, computation, explanation, and error costs. The most dependable practitioner is therefore not the one who reaches first for the newest architecture, but the one who can state the task, defend the representation, preserve an honest evaluation, and recognize what the resulting model cannot establish. The book's short format succeeds as a map: it offers enough conceptual landmarks to ask better questions, while signaling that practical competence comes from testing those questions against real data.

Summary in 30 Seconds
Machine learning fits mathematical models to examples, so predictions are dependable only when future inputs resemble the training data. Define examples, targets, and error costs before selecting an algorithm. Use training data to fit, validation data to tune, and a protected test set for a final estimate. Useful features and task-appropriate metrics matter as much as model choice. Neural networks, ensembles, and unlabeled data each help in particular settings, but their outputs still require careful interpretation.
Contains spoilers

Last updated:

Report Issue

Review Summary

4.25 out of 5
Average of 1k+ ratings from Goodreads and Amazon.

The Hundred-Page Machine Learning Book receives high praise for its concise yet comprehensive overview of machine learning concepts. Readers appreciate its balance of mathematical rigor and practical explanations, making it suitable for both beginners and experienced practitioners. The book's compact format is seen as a strength, offering a quick reference guide without sacrificing depth. Some criticisms include its dense mathematical content and occasional lack of detailed explanations. Overall, it's highly recommended as an introductory text or refresher for those with a technical background.

Your rating:
4.61
257 ratings
Want to read the full book?

Glossary

Feature engineering

Turning raw data into predictors

The process of converting raw information into features that a learning algorithm can use. It can involve extracting measurements from logs, encoding categories, scaling numerical values, creating bins, or handling missing data. The quality of this representation strongly affects what patterns a model can learn.

Empirical risk

Average training-set loss

The average penalty a model receives when its predictions are compared with the labels in the training set. Learning algorithms commonly seek parameters that reduce this quantity or a related objective. Low empirical risk shows fit to observed examples, but does not guarantee accurate predictions on unseen data.

Overfitting

Fitting noise as well as signal

A model overfits when it predicts its training examples well but performs poorly on new examples. This can happen when the model is too complex for the available data or when there are many features and few examples. Regularization, simpler models, and additional training data can help.

Regularization

Penalizing model complexity

A family of methods that discourages a learning algorithm from creating an unnecessarily complex model. In L1 and L2 regularization, a penalty based on model parameters is added to the objective. This can slightly worsen training fit while improving performance on unseen data.

Cross-validation

Rotating validation folds

An evaluation procedure that divides training data into several folds, repeatedly training on most folds and validating on the remaining fold. The validation scores are averaged to estimate performance for a particular model or hyperparameter setting. It is especially useful when a separate validation set would leave too little data for training.

Gradient boosting

Sequentially correcting model errors

An ensemble method that adds models, commonly decision trees, in sequence. Each new tree is trained to address errors made by the current ensemble, and its contribution is weighted by a learning rate. The method can produce accurate models, but its depth and number of trees need careful tuning to limit overfitting.

Convolutional neural network

Neural network using shared filters

A neural network architecture that applies learned filters across local regions of an input, commonly an image. Sharing each filter across positions reduces the number of parameters compared with connecting every input pixel independently to every unit. Convolutional layers can learn local patterns such as edges and textures.

Active learning

Selecting examples for labeling

A supervised learning approach for situations where labels are costly to obtain. The current model examines unlabeled examples and selects ones expected to be especially informative, such as uncertain examples in dense regions. A human expert labels selected cases, which are then added to training data.

Transfer learning

Adapting a model to another task

A method that reuses a model trained on one dataset to help solve a different task. A common neural network approach retains much of the trained network, replaces its final prediction layers, and trains the new layers on a smaller labeled dataset. It works best when the earlier learned representations remain useful for the target task.

Word embeddings

Vectors representing word meaning

Numerical vectors learned from text so that words used in similar contexts tend to have similar representations. These vectors can serve as inputs to models for language tasks. In the word2vec skip-gram approach, a model learns representations by using a word to predict surrounding context words.

FAQ

What's "The Hundred-Page Machine Learning Book" about?

  • Overview: "The Hundred-Page Machine Learning Book" by Andriy Burkov is a concise guide to the fundamental concepts and techniques in machine learning. It covers both theoretical and practical aspects of the field.
  • Content: The book includes explanations of various machine learning algorithms, types of learning, and practical advice for implementing machine learning solutions.
  • Purpose: It aims to provide readers with a solid understanding of machine learning in a short amount of time, making it accessible to both beginners and experienced practitioners.
  • Structure: The book is organized into chapters that progressively build on each other, starting from basic concepts to more advanced topics like neural networks and unsupervised learning.

Why should I read "The Hundred-Page Machine Learning Book"?

  • Comprehensive yet Concise: The book offers a comprehensive overview of machine learning in just a hundred pages, making it a quick yet informative read.
  • Practical Insights: It provides practical advice and insights that are valuable for both beginners and experienced practitioners looking to enhance their understanding of machine learning.
  • Accessible Language: Written in an accessible language, the book is suitable for readers with varying levels of expertise in machine learning.
  • Continuous Improvement: The book is supported by a companion wiki that is continuously updated, ensuring that readers have access to the latest developments in the field.

What are the key takeaways of "The Hundred-Page Machine Learning Book"?

  • Understanding of Algorithms: Readers will gain a clear understanding of fundamental machine learning algorithms, including linear regression, logistic regression, decision trees, and support vector machines.
  • Types of Learning: The book explains different types of learning, such as supervised, unsupervised, semi-supervised, and reinforcement learning, and their applications.
  • Practical Application: It emphasizes the practical application of machine learning techniques, including feature engineering, model evaluation, and hyperparameter tuning.
  • Advanced Topics: The book also covers advanced topics like neural networks, deep learning, and ensemble learning, providing a well-rounded understanding of the field.

What are the best quotes from "The Hundred-Page Machine Learning Book" and what do they mean?

  • "All models are wrong, but some are useful." This quote, originally by George Box, highlights the idea that while no model can perfectly capture reality, some models can still provide valuable insights and predictions.
  • "Machines don’t learn." This statement emphasizes that machine learning is about finding mathematical formulas that produce desired outputs, rather than machines actually learning in the way humans do.
  • "Read first, buy later." This principle reflects the author's belief that readers should have the opportunity to evaluate the content before making a purchase, ensuring they find value in the book.
  • "Machine learning is not learning." This quote underscores the distinction between machine learning and human learning, pointing out that machine learning involves building systems that perform tasks without explicit programming.

How does "The Hundred-Page Machine Learning Book" explain supervised learning?

  • Definition: Supervised learning is described as a process where a model is trained on a labeled dataset, learning to map inputs to outputs.
  • Types: The book covers different types of supervised learning, including classification and regression, and explains how they are used to predict discrete and continuous outcomes, respectively.
  • Example: It provides examples like spam detection, where the model learns to classify emails as spam or not spam based on labeled examples.
  • Goal: The ultimate goal of supervised learning, as explained, is to create a model that can generalize well to new, unseen data.

What is unsupervised learning according to "The Hundred-Page Machine Learning Book"?

  • Definition: Unsupervised learning involves training a model on data without labeled responses, aiming to find hidden patterns or intrinsic structures.
  • Techniques: The book discusses techniques like clustering and dimensionality reduction, which are used to group similar data points or reduce the number of features, respectively.
  • Applications: It highlights applications such as customer segmentation and anomaly detection, where unsupervised learning can provide valuable insights.
  • Challenges: The book also addresses the challenges of unsupervised learning, such as the difficulty in evaluating model performance due to the lack of labeled data.

How does "The Hundred-Page Machine Learning Book" approach neural networks and deep learning?

  • Introduction: Neural networks are introduced as mathematical functions with layers that transform inputs into outputs, with deep learning involving networks with multiple layers.
  • Components: The book explains key components like activation functions, weights, and biases, which are crucial for the functioning of neural networks.
  • Training: It covers the training process using backpropagation and gradient descent, emphasizing the importance of optimizing network parameters.
  • Applications: The book discusses applications of deep learning in areas like image and speech recognition, highlighting its ability to handle complex data.

What practical advice does "The Hundred-Page Machine Learning Book" offer for feature engineering?

  • Importance: Feature engineering is emphasized as a critical step in building effective machine learning models, involving the transformation of raw data into meaningful features.
  • Techniques: The book covers techniques like one-hot encoding for categorical variables and normalization for numerical features to improve model performance.
  • Challenges: It addresses challenges such as dealing with missing data and the need for domain knowledge to create informative features.
  • Impact: Effective feature engineering is shown to significantly impact model accuracy and generalization, making it a valuable skill for practitioners.

How does "The Hundred-Page Machine Learning Book" explain model evaluation and performance assessment?

  • Metrics: The book discusses various metrics for evaluating model performance, including accuracy, precision, recall, and the area under the ROC curve.
  • Confusion Matrix: It explains the use of confusion matrices to visualize the performance of classification models and identify common errors.
  • Overfitting and Underfitting: The book addresses the concepts of overfitting and underfitting, providing strategies to balance model complexity and generalization.
  • Validation: It emphasizes the importance of using validation and test sets to assess model performance and avoid biased evaluations.

What does "The Hundred-Page Machine Learning Book" say about ensemble learning?

  • Definition: Ensemble learning is described as a technique that combines multiple models to improve overall performance, often leading to more accurate predictions.
  • Methods: The book covers popular ensemble methods like random forests and gradient boosting, explaining how they aggregate predictions from weak learners.
  • Advantages: It highlights the advantages of ensemble learning, such as reduced overfitting and increased robustness, making it a powerful tool in machine learning.
  • Applications: Ensemble learning is shown to be effective in various applications, from classification to regression, due to its ability to leverage diverse models.

How does "The Hundred-Page Machine Learning Book" address the challenges of imbalanced datasets?

  • Problem: Imbalanced datasets are identified as a common issue where one class is underrepresented, leading to biased model predictions.
  • Solutions: The book suggests solutions like class weighting, oversampling, and undersampling to address class imbalance and improve model performance.
  • Techniques: It discusses advanced techniques like SMOTE and ADASYN, which generate synthetic examples to balance class distributions.
  • Impact: Properly handling imbalanced datasets is shown to significantly enhance the accuracy and fairness of machine learning models.

What advanced topics are covered in "The Hundred-Page Machine Learning Book"?

  • Dimensionality Reduction: The book explores techniques like PCA and UMAP for reducing feature space dimensionality, aiding in data visualization and model simplification.
  • Clustering: It covers clustering algorithms such as k-means and DBSCAN, which group similar data points without labeled responses.
  • Semi-Supervised Learning: The book discusses semi-supervised learning, where models leverage both labeled and unlabeled data to improve performance.
  • Transfer Learning: It introduces transfer learning, where knowledge from one domain is applied to another, reducing the need for extensive labeled data.

About the Author

Andriy Burkov is a machine learning expert and author known for his ability to distill complex concepts into accessible explanations. His writing style is praised for being concise, practical, and focused on intuition rather than excessive mathematical detail. Burkov's approach to explaining machine learning topics is often compared to having a knowledgeable friend explain concepts in a casual setting. He is recognized for his innovative "read first, buy later" distribution model for the book. Burkov's expertise in applying machine learning in real-world scenarios is evident throughout the text, making it particularly valuable for practitioners seeking to enhance their understanding of the field.

Download PDF

To save this The Hundred-Page Machine Learning Book summary for later, download the free PDF. You can print it out, or read offline at your convenience.
Download PDF
File size: 0.38 MB     Pages: 22

Download EPUB

To read this The Hundred-Page Machine Learning Book summary on your e-reader device or app, download the free EPUB. The .epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.
Download EPUB
File size: 1.38 MB     Pages: 28
Want to read the full book?
Follow
Listen
Home
Swipe
Library
Get App
Try Full Access for 3 Days
Listen, bookmark, and more
Compare Features Free Pro
📖 Read Summaries
Read unlimited summaries. Free users get 3 per month
🎧 Listen to Summaries
Listen to unlimited summaries in 40 languages
—
❤️ Unlimited Bookmarks
Free users are limited to 4
—
📜 Unlimited History
Free users are limited to 4
—
📥 Unlimited Downloads
Free users are limited to 1
—
Risk-Free Timeline
Today: Get Instant Access
Listen to full summaries of 26,000+ books. That's 12,000+ hours of audio!
Day 2: Trial Reminder
We'll send you a notification that your trial is ending soon.
Day 3: Your subscription begins
You'll be charged on Nov 25,
cancel anytime before.
Consume 2.8× More Books
2.8× more books Listening Reading
Our users love us
1M+ readers
Trustpilot Rating
TrustPilot
4.6 Excellent
This site is a total game-changer. I've been flying through book summaries like never before. Highly, highly recommend.
— Dave G
Worth my money and time, and really well made. I've never seen this quality of summaries on other websites. Very helpful!
— Em
Highly recommended!! Fantastic service. Perfect for those that want a little more than a teaser but not all the intricate details of a full audio book.
— Greg M
Save 62%
Yearly
$119.88 $44.99/year/yr
$3.75/mo
Monthly
$9.99/mo
Start a 3-Day Free Trial
3 days free, then $44.99/year. Cancel anytime.
Unlock a world of fiction & nonfiction books
26,000+ books for the price of 2 books
Read any book in 10 minutes
Discover new books like Tinder
Request any book if it's not summarized
Read more books than anyone you know
#1 app for book lovers
Lifelike & immersive summaries
30-day money-back guarantee
Download summaries in EPUBs or PDFs
Cancel anytime in a few clicks

Why one summary?

Scanner
Find a barcode to scan

⭐
⭐
⭐
We have a special gift for you
Open
38% off
DISCOUNT FOR YOU
⭐
⭐
⭐
$79.99
$49.99/year
Only $4.16 per month
Continue
2 taps to start, super easy to cancel
Settings
General
Widget
Loading...