5.0 out of 5 stars Nice conceptualization!

T.G. · October 28, 2025

Someone with a physics and mathematics background will readily embrace the LLM Concepts and probabilities. As the learning data base grows so shall imaging,languages,new theories, drug design.... grow!

4.0 out of 5 stars Very good overview of how LLMs work.

M.K. · November 11, 2024

I really enjoyed reading this book. I know there is a free blog post with the same content, but books are easier for me to digest than blogs when it comes to long texts. I definitely understand the technology better now.Some cons:* Kindle version is not great. It's ok to read it on smartphone but not on Kindle. Some images are difficult to see.* Transformers could be explained better* Author tries to plug his product, Wolfram|Alpha, a little bit too much.

5.0 out of 5 stars Authoritative description of how this programming actually works

J. · July 5, 2023

This is the best explanation of how GPT software/neural nets actually find language-based answers. It's a must read for anyone (high school or college students, their parents, physicians or post docs) who intends to use GPT in any field. I think about this program completely differently now that I've read this book.Short, well-illustrated, and always to the point, you'll be led by the hand through every level of development and function in an easy-to-understand format that won't require a degree in computer programming to understand. AI is here, and this is a must read for this year for everyone!!!

3.0 out of 5 stars Advertisement

S. · April 16, 2024

The first half of the book gives some basic idea of how neuro network and ChatGPT work, which is great for AI beginners. The second half raises some questions and has discussions which are interesting. How to solve the problem? Use the software developed by the author. So the second part is more like an advertisement. Given that the software is still under development, at least I am not interested in it.

4.0 out of 5 stars ChatGPT is Horrible at Math; Pair it with Wolfram Alpha!

Q.B. · May 5, 2023

First of all, as a true techie, I love Wolfram's work, not the physics, but his elegant, refreshing and bold contribution to mathematics and creating tools to bring that complexity to everybody. His first go at it, Mathematica put so many modes of computation on one's desktop as a standalone. This includes ultra high precision (seemingly un-limited digits), way powerful symbolic calculations including complex algebra expression factoring, equation solving, graphing, and full differential and integral calculus, and to top it off, rule-based logic and far more.You have to wait until the later chapters for it, but after explaining the brain-dead simplicity of neural network "learning", he figuratively destroys ChatGPT, dismembering it, and thrusting its flaws at the reader's face for all to see. I found it shocking, but most fascinating, as he describes why fixing ChatGPT's Achilles heel will be near impossible and certainly not worth the effort.ChatGPT which works on such a brute force simple single principle, neural net weightings, can never be a match for real intelligence crafted into a product. I love geniuses who later in life wax philosophical (If you do, check out Irwin Schrodinger's "What is Life" and his [20 years later] "My View of the World". Steven Wolfram is definitely a computational genius, with a divine gift of designing, creating, and polishing the highest mathematical tools, and then giving them to the world. As such he is entitled to wax philosophical, too.As a religious Jew, and I think anyone religious would agree, no amount of appending the most fitting next nucleotide into a string of DNA will produce something as complex as a housefly, all the more so a human - sorry Darwin. All the top NASA engineers and scientists can not come close to the fly's taking and landing ability on any angled surface, its ability to evade capture by a human being by quickly seeing and evaluating trajectories 100's of times per second, its ability see in a myriad of directions simultaneously, its getting its energy from simple waste materials, and its manufacturing ability to replicate thousands of itself from those same materials in record time.Steven Wolfram has thus done humanity a great service. Fear no longer. This wave of AI is brain dead and will not take over the world, not with its lack of the most elementary mathematical, or even arithmetical ability.As I thank Steven Wolfram for his perfectionist products, and enjoy reading his books large and small, I hope he will understand that to my mind, he has made it clear that, in the most similar way, his taking down of ChatGPT by exposing its limitations of its simple technique of choosing the best 'atom' to append, he has, in one glorious stroke, also eviscerated Darwin and exposed the extreme stupidity of believing that random mutations plus survival of the fittest (picking the most fitting next DNA 'letter') can ever compete with an alternative such as truly thoughtful and intelligent crafting and designing.Wolfram refreshingly blurts out the truth. I agree and will close, sharing that expecting Darwin's theory to produce life is now as foolish as expecting ChatGPT to calculate numbers, to take one example, to calculate Pi to even a few dozen places, a task Wolfram's tools now can all do in a split second.To also put it bluntly, after reading "What Is ChatGPT doing ... and Why Does It Work?" , knee jerk use of science or technology to defend atheism, is no longer an effective option.

5.0 out of 5 stars This is an important book.

T. · March 21, 2024

A most distinguished member of the academy, Steven Wolfram has provided a substantial contribution to the body of thought on AI.As the world is agog with the mystery, potential, risk, and opportunity associated with large language models, Dr. Wolfram provides a clear concise explanation of how LLMs came into being, how they work, the class of problems that they address, the large class of problems that they will not address, and the extent to which they are inexplicable.AI changes everything and Generative AI significantly accelerates that change. Fueled by the advent of LLM’s, AI has moved to the top of the agenda of though leaders in the academy, industry, and government.We are seeing innovative, previously unimaginable breakthrough applications of LLMs introduced daily. The scope of the application of these models to change the way we work, communicate, administer, and entertain is staggeringly large and unknowably vast.As you attempt to parse the cacophony of messaging from the large and growing chorus of technology providers, politicians, media, opportunists, charlatans, fear mongers, and pundits regarding LLMs, Dr. Wolfram ‘s book will provide you a solid grounding of what this stuff really is, what it can do, and – importantly -- what it cannot do.

Tout comprendre du fonctionnement interne de ChatGPT

G. · March 13, 2024

Assez technique mais passionnant pour les aficionados !Avec en prime des réflexions sur ce que tout ça nous enseigne du fctnnmt du cerveau.

Very readable introduction into the inner workings of ChatGPT

L. · April 23, 2023

I love this little book because it is very readable, well organized and easy to understand. As a byproduct you get some basic knowledge of machine learning without the need to dive into the mathematical depths. I can warmly recommend this excellent introduction by Stephen Wolfram. He is a master of explaining things clearly.

Interesting book, ...but not so easy

J.J. · December 27, 2023

This book is written by a person who has been working on AI and has been somewhat surprised by the success of Chatgpt. It is certainly worth reading, but a certain level of expertise on computation theory is necessary to understand it fully. Interesting to see the limitations of Chatgpt and LLM models. I think it is very good, but the title is a little misleading (that is why I do not give 5 *)

"It was a dark and stormy..." What's the next word?

Z.F. · May 11, 2025

Mr. Wolfram is a top expert: A well known mathematician, creator of the "Mathematica" software. This book is a must read for anybody who worries about the future of the world with AI, or who wants to invest in it. I am a retired computer engineer, but even people without technical background can vastly benefit from reading it while passing over more technical parts. LLMs (Large Language Models) are at the heart of the Generative AI. Reading the book let's you understand that LLMs basically given any piece of text estimate the probabilities of next word in that text. This is a great achievement, since number of possibilities is astronomical. But it lets you understand, that a text step-by-step created by the software is kind of average of everything that was ever written on the subject and consumed by chatGPT in the learning phase. To me it means that ChatGPT is a "Rumour Machine" or a plagiarism creator but not from a single source, but from millions of them at the same time. It also means that to me there is no intelligence whatever involved and that for a question it would give a most popular answer instead of a correct answer. Many people do not agree with this estimate - I had many heated discussions on the subject. Read the book and decide for yourself.

A high-level introduction to the enigmatic success of ChatGPT

P.H. · April 2, 2023

Stephen Wolfram's latest book is directed at explaining the enigma of why the artificial intelligence program ChatGPT is so successful. Wolfram is ideally placed to provide insights into the remarkable success of ChatGPT. He is the inventor of the symbolic mathematical processing program Mathematica and the technical and commercial brains behind Wolfram Alpha which is a natural language knowledge resource accessible to anyone via the internet. Wolfram has been ruminating on all things computational for over 40 years and 20 years ago he wrote a large book called “A New Kind of Science” in which he systematically delved into the way simple computational rules can produce complex behaviour.It is useful to provide some historical context to the remarkable success of ChatGPT whose engine is GPT- 3. It burst onto the scene in late 2022 and surprised even people who work in the field. For most people it a mysterious thing that is almost magical. At the time of writing there is an even more powerful GPT-4 engine. Such is the success of ChatGPT that some people are now calling for an end to AI because of the as yet unknown social and economic consequences of unconstrained applications of AI in general.It was not always thus, because even ten years ago there was some doubt as to the future of AI. Going back to the days of Marvin Minsky at MIT in the 1960s, his blind PhD student Jim Slagle wrote a symbolic mathematical processing program that could solve MIT level integral calculus questions with a high rate of success. Mathematica itself owes some of its intellectual DNA to what Jim Slagle did back then. During the 1990s and early 2000s researchers developed programs for classifying images, for instance, but by about 2010 no major breakthroughs had been achieved in attaining high levels of accuracy. MIT actually considered closing down its undergraduate AI course around 2010 because the theoretical landscape seemed barren. Then Geoff Hinton and his PhD students from Toronto University blew everyone away in 2012 with the success of their convolutional neural network with fast GPU processing which won the AlexNet image classification competition by a wide margin. Since then, the field has progressed at an astonishing pace in just 10 years. People genuinely want to understand the magic behind ChatGPT, and I think Wolfram does a good job in explaining the technical structure of ChatGPT but he also explains just how much art or what he calls “neural lore” goes into the success of ChatGPT.So, who is Wolfram's audience? Could a non-computer scientist or a non-mathematically trained person get a meaningful understanding of how ChatGPT works from the book? Yes, they can because what Wolfram has done is avoid all the detailed mathematics involved in favour of a high-level description of the processes that guide how the system works. He uses his own Wolfram language to demonstrate fundamental points in a visual way. Long time Mathematica users such as myself are used to seeing these renditions but non-Mathematica users may find them obscure. However, he provides an entirely accessible non mathematical explanation of what is going on. I could not see for instance why someone interested in linguistics wouldn't be able to get a high level feel for what's going on inside ChatGPT.So what is actually going on in ChatGPT? This is the enigma because the structure of the program is at one level very simple. People might reasonably assume that someone has encoded some high-level algorithmic rules for how grammar works into the system. But that is not how it works. Nor does it have billions of texts in memory which it somehow searches through to generate a result. Very simply ChatGPT has no explicit knowledge of the rules of grammar for instance. To understand how GPT works there are several building blocks which involve terms such as embeddings, transformers, attention blocks, activation functions, neural nets, weights and back propagation. All of these things have very precise mathematical structures which Wolfram avoids going into in favour of a high-level discussion which gives you a sense of how they work in actual examples. He also gives a sense of the scale of ChatGPT given that there are around 175 billion weights in the system that conspire miraculously to ultimately give rise to sensible answers. How this is ultimately done in detail is unclear and is an area of active research but in essence vast amounts of data are used to train ChatGPT at a massive cost in terms of computation and finance (Wolfram mentions a billion dollars for the training effort). Conceptually you have a massive linguistic space, and you need some measure of proximity or nearness between words and their meanings. Probabilistic concepts based on empirical data sets are used in the context of determining the contextual proximity of related words. The weights in the model can be tuned to improve the model’s probability of rendering language which is acceptable to humans. Surprisingly in re-training the system you don't have to go into a training algorithm and manually adjust all the weights. At a basic combinatorial level with even a few weights that would be a monumental task but with 175 billion it is impossible. Wolfram explains at a high level how they get around this type of problem.There is some obscure stuff like a reference to Rule 30 which only makes sense to people who have read his book “A New Kind of Science” where he shows that you can create remarkable complexity with very simple rules. For those with a physics/mathematical background Wolfram makes some interesting comments about “trajectories” in these massive spaces and in so doing he seeks to look for a principle like following geodesics to explain the remarkable convergence properties of ChatGPT. 
Wolfram identifies a number of structural weaknesses with ChatGPT the most fundamental of which is that it actually doesn’t do any real-world computation. He gives examples of how ChatGPT could interact with Wolfram Alpha which does have powerful computational capacities. No doubt some will see this as commercial special pleading but it is hardly surprising that a guy who has spent his life building engines like Mathematica and Wolfram Alpha would not want to see them integrated with something like ChatGPT.One thing that has occupied my mind is whether programs like ChatGPT will give rise to a global uniformity of outcome. As time goes on the potential training space will be filled with more and more artefacts of engines like ChatGPT. There is however a very counterintuitive probabilistic result for Bernoulli variables with variable probabilities of success which holds that the variance or volatility is maximised when all the probabilities are the same. As the great probability theorist William Feller pointed out this result implies that “given a certain average quality p of n machines, the output will be least uniform if all machines are equal”. Maybe that can be a chapter in another book.