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
What Is ChatGPT Doing... and Why Does It Work?

What Is ChatGPT Doing... and Why Does It Work?

ChatGPT predicts what comes next, not what is true. Here is why that distinction matters.
by Stephen Wolfram 2023 143 pages
3.85
1k+ ratings
Listen
Immersive
V2.3
Amazon Kindle Audible
Try Full Access for 3 Days
Unlock listening & more!
Continue

Key Takeaways

Treat ChatGPT as a next-token engine, not an answer lookup system

One token at a time. Given a prompt, ChatGPT estimates which token, often a word fragment, is likely to come next. It selects one, adds it to the text, and repeats the process. A prompt such as “The best thing about AI is its ability to” narrows the next-token options, but does not specify a single continuation. Sampling among plausible options makes repeated answers vary. Wolfram describes a setting called temperature: lower values favor the most likely tokens, while higher values allow less likely choices. A value around 0.8 worked well for essay generation in the systems discussed, but this was an empirical setting, not a law of language.

Plausibility is not verification. The model produces continuations shaped by patterns in its training text. That can make an answer sound assured without checking whether its claims are true. Its basic task is to continue text, not independently confirm facts.

Analysis

The next-token description is a useful antidote to treating a chatbot as a miniature encyclopedia. It separates two abilities that ordinary conversation tends to blur: producing a well-formed sentence and establishing that the sentence is true. The distinction also changes how responses should be evaluated. Fluency is evidence that the model has learned linguistic regularities, not that it has verified a claim. A probabilistic generator can still be useful for drafting, explanation, and exploration, but confidence in its style should not substitute for evidence when accuracy matters.

Model sparse language by generalizing beyond observed word sequences

Counting words reaches a wall. A simple language model might estimate which letters or words tend to follow others. Longer context improves its output, as with letter pairs that capture the fact that q is usually followed by u. But the number of possible combinations grows too quickly. With roughly 40,000 common English words, there are about 1.6 billion possible pairs and around 60 trillion possible triples. No available text collection contains examples of every useful sequence.

Learn a general pattern. A large language model uses a parameterized mathematical model to estimate probabilities even for sequences that never appeared in its training examples. It does not need a stored example of every sentence. Its job is to generalize from recurring patterns in the text it has encountered, though its estimates can still fail in unfamiliar cases.

Analysis

This is a classic problem of statistical learning: observations are finite, while possible inputs are combinatorially vast. The model succeeds by sharing information across examples instead of treating every phrase as an unrelated entry in a lookup table. That sharing is powerful, but it also means that learned regularities can reflect the corpus's gaps and biases. Generalization is not a guarantee of truth; it is a way to make informed estimates where direct evidence is absent.

Training examples shape billions of weights, not a catalog of stored sentences

A model is a fitted procedure. Wolfram compares model-building with predicting how long a ball takes to fall from different heights. A table memorizes measured cases; a mathematical model estimates cases not measured. ChatGPT also represents a procedure, implemented by a neural network whose many weights are adjusted during training. The 175-billion-weight GPT-3 system described in the book was trained on hundreds of billions of words.

Error guides adjustment. During training, the model predicts text from examples and measures the difference between its prediction and the target with a loss function. Calculus-based updates adjust weights to reduce that error. The result is not a verbatim archive of every training sentence. It is a distributed encoding of patterns in the data, useful for predicting continuations. The model's learned world is therefore bounded by both its architecture and its examples.

Analysis

The model analogy helps clarify what training does: it fits a reusable procedure rather than compiling a list of answers. In machine learning, this is the difference between memorizing examples and learning a function that can interpolate among them. The latter is the source of flexibility and the source of risk. A model's behavior on cases unlike its training data depends on assumptions built into the architecture and learning process. Those assumptions are often invisible in the final prose, so a smooth explanation can conceal a fragile extrapolation.

Use embeddings and attention to see how context shapes each next token

Numbers represent language. An embedding is a list of numbers used to represent a word or passage so that patterns of similarity can be expressed mathematically. Words appearing in similar contexts, such as alligator and crocodile, tend to receive related representations. ChatGPT works with tokens, which may be whole words or fragments, and represents their order as well as their identity.

Context gets reweighted. Its transformer architecture uses attention mechanisms to connect parts of the text that matter to one another. This lets the model use words from earlier in a sentence, rather than relying only on the immediately preceding word. The text representations pass through many neural-network layers; the final state is used to estimate next-token probabilities. Each token is generated through a forward pass, then becomes part of the context for the next pass.

Analysis

Embeddings make a crucial shift possible: language becomes a space of relationships rather than a list of isolated dictionary entries. Attention then makes those relationships sensitive to context, allowing a word's role to depend on the surrounding passage. This is a useful mental picture, but not a literal map of human meaning. A vector's proximity records statistical structure learned by the model, not a complete definition or guarantee that two terms can be substituted. The representation is powerful precisely because it captures patterns that are hard to state by hand.

Human ratings teach a raw text predictor what makes a better assistant

Prediction alone is not enough. After learning from large collections of text, a model can continue prompts, but its longer responses may drift in ways people readily notice. The book describes an additional stage in which human evaluators rate outputs. A separate model learns to predict those ratings, and its judgments are then used to tune the original network toward responses people prefer.

Prompts steer existing capacity. A user can also supply information in the prompt, and the model may use it while composing the answer without changing its trained weights. This works best when the new material fits patterns the model can already handle. A prompt is therefore a temporary context, not the same thing as permanently teaching the model. Human feedback improves conversational behavior, but a preference rating is not a factual proof.

Analysis

This stage shows that a capable predictor and a useful assistant are not identical products. Human judgments add a social layer: responses are shaped not only by what text commonly follows, but by what evaluators consider helpful. That can make interactions more cooperative and legible. It also makes the target partly normative. Preferences differ across people and settings, and a response that sounds agreeable may still be mistaken. The distinction between better conduct and better knowledge is essential when assessing feedback-driven systems.

Fluent language is not proof of reliable computation

Some patterns are shallow; some require procedure. ChatGPT can often handle ordinary grammar because likely continuations offer many clues. But a task demanding exact algorithmic steps can expose its limits. Wolfram illustrates this with a small transformer trained on balanced parentheses. A sequence is valid only if every opening parenthesis is eventually closed. A model with roughly 400,000 weights learned some patterns after about 10 million examples, yet began making errors on longer sequences that required reliable counting.

The architecture matters. For each token, the network passes information forward through its layers; it does not run an internal loop like a conventional program. The generated text does form an outer loop, since each new token joins the next prompt. This structure supports learned continuations but does not guarantee dependable execution of arbitrary procedures.

Analysis

The parentheses example is valuable because it replaces vague claims about reasoning with a testable task. It shows how a system can imitate a rule over familiar cases without reliably carrying out the rule when the sequence becomes demanding. The distinction resembles the difference between recognizing a pattern and executing an algorithm. It should not be taken to mean that language models cannot solve any structured problem; rather, success on routine examples does not establish that a model has a robust procedure for every instance.

Send exact questions to computational tools instead of trusting fluent guesses

Use the right kind of engine. Wolfram contrasts ChatGPT's language generation with Wolfram|Alpha's structured computation. Asked for a precise distance between cities, ChatGPT may produce a plausible but incorrect result. Wolfram|Alpha can interpret the question, use stored geographic coordinates, and apply a distance algorithm. The same division matters for mathematics, rankings, and current or location-dependent facts: language patterns alone do not reliably perform every calculation or retrieve every changing value.

Let the systems hand off. ChatGPT can formulate a natural-language request, a computational system can translate it into a precise representation and calculate an answer, and the result can return to ChatGPT for a readable explanation. The prose generator supplies conversational flexibility; the tool supplies an auditable operation when the task can be formalized.

Analysis

This partnership is more than a convenient feature. It reflects a division of cognitive labor: one system is strong at interpreting and communicating in flexible language, while another can execute explicit procedures over structured knowledge. In software design, this resembles keeping an interface separate from a trusted computation. The handoff is only as dependable as the interpretation and underlying data, however. A tool cannot correct an ambiguously framed question or an outdated source unless those problems are addressed too.

Treat ChatGPT's fluency as evidence of structure, not proof of understanding

Language has learnable regularities. Syntax constrains how words fit together, while meaning imposes further expectations. A sentence can be grammatically formed yet make little sense, so a system needs more than word order to produce convincing prose. Wolfram argues that ChatGPT's success suggests language and the thought expressed through it contain regularities that a neural network can learn from examples.

A scientific hypothesis, not a solved theory. The model's internal representations may offer clues about those regularities, but the book does not show a simple set of semantic laws that explains them. Even visualizing trajectories through a model's representation space does not reveal an obvious rule. The achievement points toward structure worth investigating; it does not establish that the model understands language as people do.

Analysis

The most ambitious claim here is also the most fruitful: if a system learns to generate coherent language from examples, then some aspects of language use may be more regular than intuition suggests. That is a research program, not a demonstration that meaning has been reduced to statistics. Human language is tied to shared situations, intentions, and changing conventions, while a training corpus offers only traces of those practices. The model's success is evidence about learnability, but the boundary between learned linguistic competence and understanding remains an open question.

Pair natural-language flexibility with symbolic computation for dependable answers

Represent claims precisely. Wolfram's proposed computational language represents objects and relationships in forms a computer can execute. Natural language is expressive and convenient, but its terms can remain vague. A symbolic system must make distinctions explicit, then apply defined operations. For instance, it can represent an object moving from one place to another and calculate consequences from that representation.

Build a bridge, not a replacement. The book imagines language models helping translate everyday questions into formal representations, while computational systems perform exact operations and return results. A semantic grammar, a set of rules connecting concepts and relations, could help organize statements by meaning as well as sentence form. The goal is to combine human-readable communication with computations that can be checked, rather than asking a text generator to improvise every step.

Analysis

This proposal reaches beyond fact-checking toward a general architecture for useful AI: natural language for expressing intent, formal representation for preserving distinctions, and computation for deriving results. Its strength is modularity. Each component can be inspected and improved without pretending that one method does everything. The hard part is deciding what the symbols mean and which distinctions matter. Formalization can make a calculation reliable, but it cannot remove ambiguity from concepts that people have not agreed how to define.

Analysis

The book's central achievement is to make an apparently mysterious capability inspectable. ChatGPT does not compose an essay in one act. It repeatedly estimates a next token, using a learned numerical model of textual context. That modest operation, scaled across billions of weights and enormous training corpora, yields behavior that resembles a much grander faculty. The explanatory payoff is not that the system is simple, but that simple operations can combine into capacities their designers did not explicitly program.

Wolfram moves from mechanism to epistemology. The same statistical learning that makes language generation flexible also explains why fluency cannot certify truth. A model estimates what continuation fits its learned patterns; a formal computational tool executes an explicit procedure. This distinction is especially useful because public debate often treats apparent conversation as proof of either intelligence or its absence. The book instead asks which tasks are learnable from examples, which require precise computation, and how tools can connect the two.

Its strongest scientific suggestion is that natural language contains exploitable regularities, perhaps enough for neural networks to reproduce substantial parts of human-like expression. Its more speculative step is that this success points toward discoverable laws of meaning and thought. A model trained on text reveals structure in language use, but that does not by itself identify the underlying structure of thought. Text is a record of communication, not a transparent window into every process that produced it.

The proposal to combine language models with symbolic systems is therefore more than a product vision. It is a methodological principle: use statistical systems where interpretation is flexible, and exact computation where answers depend on explicit rules or data. The boundary is not always clean, since natural-language interpretation can itself be ambiguous and formal systems encode chosen assumptions. Still, treating these capabilities as complementary offers a clearer path than demanding that a single model both speak plausibly and guarantee correctness.

Summary in 30 Seconds
ChatGPT generates text by repeatedly estimating a likely next token from context. Training adjusts billions of network weights to capture patterns, rather than storing a catalog of sentences. With roughly 40,000 common English words, possible triples already number about 60 trillion, so models must generalize beyond observed examples. Human ratings can improve conversational responses but cannot certify facts. Use formal computational tools for exact calculations and checkable results.
Contains spoilers

Last updated:

Report Issue

Review Summary

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

"What Is ChatGPT Doing... and Why Does It Work?" receives mixed reviews. Some praise its accessible explanation of ChatGPT's basics and neural networks, while others find it overly technical or shallow. Many readers appreciate Wolfram's honesty about the unknowns in ChatGPT's functioning. However, criticisms include excessive self-promotion of Wolfram products and a lack of in-depth analysis. The book is generally seen as a quick introduction to AI language models, suitable for those with some technical background but potentially challenging for complete beginners.

Your rating:
4.44
316 ratings
Want to read the full book?

Glossary

token

text processing unit

A unit of text processed by a language model. A token can be a whole word, part of a word, or another text fragment. ChatGPT predicts one token at a time, which lets it handle uncommon words, compounds, and material beyond a fixed list of whole words.

embeddings

numerical meaning representations

Arrays of numbers that represent words, sequences, or other inputs for a neural network. In word embeddings, terms used in similar contexts tend to have related representations. The numbers let a model compute with language, but they are not plain-language definitions of the words they represent.

transformer

context-processing network architecture

A neural-network architecture designed to process sequences such as text. Transformers use attention mechanisms to combine information from different positions in a sequence, allowing a model to use earlier words when estimating what should come next. GPT models use transformer networks to process tokens and generate continuations.

attention

weighted focus across context

A mechanism in transformer networks that assigns different weights to information from different positions in a sequence. It lets the model use relevant earlier tokens when processing the current context, rather than relying only on nearby words. The learned weights help shape the representation used to predict the next token.

loss function

measure of prediction error

A mathematical measure of how far a model's output is from the desired output on training examples. Training adjusts the network's weights to reduce this error. A lower loss indicates closer fit to the examples under that measure, but does not alone guarantee accurate behavior on new or different inputs.

computational irreducibility

no shortcut to a computation

The idea that some computational processes cannot be reliably predicted by a simpler shortcut. To know their outcomes, one may have to follow the steps of the computation itself. Wolfram uses this idea to explain why learning patterns from examples cannot replace explicit computation for every exact or deeply algorithmic task.

semantic grammar

rules for meaningful relations

A proposed system of rules for combining concepts and relationships into meaningful statements, going beyond syntax, which governs sentence structure. It would describe, for example, which kinds of entities can move or interact. Wolfram presents semantic grammar as a possible bridge between ordinary language and precise computational representations.

computational language

executable symbolic representation

A formal symbolic language for representing objects, relationships, and operations in ways a computer can execute. Wolfram presents the Wolfram Language as an example, with built-in representations and procedures for many domains. Its precision can support calculation, though the concepts and distinctions it encodes must still be defined.

temperature

controls sampling variation

A setting that affects how a language model selects from its next-token probabilities. Lower temperature favors the most likely options, while higher temperature makes less likely options more available. The book describes a value around 0.8 as useful for essay generation in the systems discussed, based on practice rather than a theoretical rule.

Wolfram|Alpha

computational knowledge system

A system that interprets natural-language questions and, when possible, translates them into precise symbolic forms for calculation or structured knowledge retrieval. In the book's proposed partnership, it supplies operations such as mathematical computation and geographic distance calculation, while ChatGPT can frame requests and explain results in conversational language.

Wolfram Language

general computational language

A symbolic programming and computational language designed to represent and operate on a broad range of objects, data, and ideas. The book presents it as a medium that can connect human questions to executable calculations. Unlike conversational text, its expressions must specify distinctions precisely enough for a computer to process them.

FAQ

What's "What Is ChatGPT Doing... and Why Does It Work?" about?

  • Overview of ChatGPT: The book explains how ChatGPT, a language model developed by OpenAI, functions and why it is effective in generating human-like text.
  • Interdisciplinary Story: It combines technology, science, and philosophy to tell the story of ChatGPT's development and capabilities.
  • Neural Nets and Language: The book delves into the concept of neural networks, their history, and how they are used to model human language.
  • Two Main Parts: The first part explains ChatGPT's language generation, while the second part explores its potential to use computational tools like Wolfram|Alpha.

Why should I read "What Is ChatGPT Doing... and Why Does It Work?"?

  • Understanding AI: It provides a comprehensive understanding of how AI models like ChatGPT work, which is crucial in today's tech-driven world.
  • Interdisciplinary Insights: The book offers insights from various fields, including technology, science, and philosophy, making it a rich resource for diverse readers.
  • Author's Expertise: Written by Stephen Wolfram, a renowned computer scientist, the book benefits from his deep expertise and unique perspective.
  • Future Implications: It discusses the future potential and implications of AI, helping readers understand its impact on society and technology.

What are the key takeaways of "What Is ChatGPT Doing... and Why Does It Work?"?

  • Neural Networks: The book explains how neural networks, inspired by the human brain, are used to generate human-like language.
  • Training Process: It details the training process of ChatGPT, which involves learning from vast amounts of text data.
  • Limitations and Potential: The book discusses the limitations of current AI models and their potential to evolve with computational tools.
  • Scientific Discovery: It suggests that the success of ChatGPT indicates a simpler underlying structure to human language than previously thought.

How does ChatGPT generate text according to Stephen Wolfram?

  • Word-by-Word Generation: ChatGPT generates text by predicting the next word based on the text it has seen so far, using probabilities.
  • Neural Network Model: It uses a neural network model trained on a large corpus of text to make these predictions.
  • Randomness and Creativity: The model incorporates randomness to avoid repetitive and flat text, which can lead to more creative outputs.
  • Temperature Parameter: A "temperature" parameter is used to control the randomness, with a typical setting of 0.8 for essay generation.

What is the role of neural networks in ChatGPT as explained in the book?

  • Brain Inspiration: Neural networks are inspired by the structure and function of the human brain, with neurons and connections.
  • Training and Learning: They are trained using large datasets to learn patterns and make predictions, similar to how humans learn.
  • Complex Tasks: Neural networks can perform complex tasks like image recognition and language generation by identifying patterns.
  • ChatGPT's Network: ChatGPT uses a large neural network with 175 billion parameters to generate human-like text.

How does "What Is ChatGPT Doing... and Why Does It Work?" explain the training of ChatGPT?

  • Large Text Corpus: ChatGPT is trained on a vast corpus of text from the web, books, and other sources to learn language patterns.
  • Weight Adjustment: The training involves adjusting the weights of the neural network to minimize errors in text prediction.
  • Human Feedback: After initial training, human feedback is used to fine-tune the model, improving its ability to generate coherent text.
  • Efficiency and Scale: The book discusses the efficiency of the training process and the scale required to achieve human-like language generation.

What are embeddings, and how are they used in ChatGPT?

  • Numerical Representation: Embeddings are numerical representations of words or phrases that capture their meanings in a multi-dimensional space.
  • Semantic Similarity: Words with similar meanings are placed close together in this space, allowing the model to understand context and relationships.
  • Word and Text Embeddings: ChatGPT uses embeddings for both individual words and sequences of text to generate coherent language.
  • Training Embeddings: The embeddings are learned during the training process, helping the model predict the next word in a sequence.

What is the significance of the transformer architecture in ChatGPT?

  • Attention Mechanism: Transformers use an attention mechanism to focus on relevant parts of the input text, improving context understanding.
  • Sequence Processing: They are particularly effective for processing sequences of data, like text, by considering relationships between words.
  • Efficiency and Performance: The transformer architecture allows for efficient training and high performance in language tasks.
  • ChatGPT's Use: ChatGPT's neural network is based on the transformer architecture, enabling it to generate coherent and contextually relevant text.

How does Stephen Wolfram view the future potential of ChatGPT and similar AI models?

  • Beyond Human Capabilities: Wolfram envisions AI models like ChatGPT using computational tools to go beyond human capabilities in certain tasks.
  • Integration with Tools: He discusses the potential for integrating AI with tools like Wolfram|Alpha to enhance their computational power.
  • Scientific Discovery: The success of ChatGPT suggests the possibility of discovering new "laws of language" and thought processes.
  • Continued Evolution: Wolfram anticipates continued evolution and improvement of AI models, driven by advances in technology and understanding.

What are the limitations of ChatGPT as discussed in "What Is ChatGPT Doing... and Why Does It Work?"?

  • Lack of True Understanding: ChatGPT generates text based on patterns, without true understanding or reasoning capabilities.
  • Computational Irreducibility: The model cannot perform complex computations that require step-by-step reasoning or control flow.
  • Dependence on Training Data: Its performance is limited by the quality and scope of the training data it has been exposed to.
  • Need for External Tools: For precise computations and factual accuracy, ChatGPT needs to integrate with external tools like Wolfram|Alpha.

What are the best quotes from "What Is ChatGPT Doing... and Why Does It Work?" and what do they mean?

  • "The success of ChatGPT is, I think, giving us evidence of a fundamental and important piece of science..." This quote highlights the scientific significance of ChatGPT's success in understanding language.
  • "ChatGPT is 'merely' pulling out some 'coherent thread of text' from the 'statistics of conventional wisdom'..." It emphasizes that ChatGPT's outputs are based on statistical patterns rather than true understanding.
  • "The remarkable—and unexpected—thing is that this process can produce text that’s successfully 'like' what’s out there..." This quote underscores the surprising effectiveness of ChatGPT in mimicking human language.
  • "It’s a very different setup from a typical computational system—like a Turing machine..." This highlights the unique architecture of ChatGPT compared to traditional computational systems.

How does Stephen Wolfram propose to enhance ChatGPT with Wolfram|Alpha?

  • Computational Knowledge Integration: Wolfram suggests integrating ChatGPT with Wolfram|Alpha to provide it with computational knowledge superpowers.
  • Natural Language Interface: The integration leverages the natural language interface of both systems, allowing seamless communication.
  • Enhanced Accuracy: By consulting Wolfram|Alpha, ChatGPT can improve its accuracy in computations and factual information.
  • Broader Applications: The integration opens up new possibilities for applications that require both human-like language generation and precise computation.

About the Author

Stephen Wolfram is a renowned scientist, entrepreneur, and author. He founded Wolfram Research and created influential tools like Mathematica and Wolfram|Alpha. Wolfram's work spans various fields, including computer science, physics, and artificial intelligence. He authored "A New Kind of Science," exploring complex systems and computational models. Wolfram's contributions to science and technology have made him a prominent figure in the tech industry. His latest endeavor, the Wolfram Physics Project, aims to find fundamental theories of physics using computational methods. Wolfram's expertise in AI and computational systems positions him as a valuable voice in discussions about emerging technologies like ChatGPT.

Download PDF

To save this What Is ChatGPT Doing... and Why Does It Work? summary for later, download the free PDF. You can print it out, or read offline at your convenience.
Download PDF
File size: 0.42 MB     Pages: 21

Download EPUB

To read this What Is ChatGPT Doing... and Why Does It Work? 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: 1.46 MB     Pages: 27
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...