AI for Smarties: A Conversation with Bertrand Meyer
Q1. You open AI for Smarties by acknowledging that you kept meeting highly educated people who were fascinated by AI but, in your words, “had no clue of what’s inside” — unable to answer what a neuron in a neural network actually computes. That observation is striking coming from someone who created Eiffel, wrote Object-Oriented Software Construction, and won the ACM Software System Award. What specifically convinced you that this book needed to exist — and what is the most important thing that someone who uses AI every day but cannot answer that basic question is getting wrong as a result of that gap?
You are right, this book was out of my comfort zone! My previous work and in particular my previous books have been about software engineering, programming methodology, formal methods, OO technology (of course), programming languages. For AI, I started out as a user, and was struck by the lack of clear explanations of what is going on under the hood.
The situation is the same for many people, including many engineers and others who have considerable expertise in other disciplines. It is dangerous to be using such a powerful technology and not to understand its true potential and its risks. This situation leaves an opening for unscrupulous people with an agenda. You see exaggerations both ways: those who warn that Artificial General Intelligence is around the corner and will turn us all into slaves, those who state that AI is pretty dumb and gets everything wrong, and everything in-between.
I think the strength of my book is that I do not have an agenda, except to explain. I want my readers to master the technology so as not to be mastered by it, or by people using it. I take neither a messianic nor an apocalyptic attitude, but stick to concepts and facts.
As the preface states, I do not claim that reading the book qualifies you to apply for the position of chief scientist at Open AI. Not quite yet. It is an introductory text. But you will understand what today’s AI is, how it works, what it does, what it does not do, what effect it is having and will have on jobs, careers and education, and how to organize your own strategy in its inevitable presence.
Q2. Your book does not shy away from the mathematics — you dedicate an entire chapter to “the math you need” and make the case that understanding AI properly requires engaging with it. That is a deliberate and somewhat countercultural choice at a moment when most AI books for general audiences either hide the math entirely or treat it as intimidating decoration. What is actually lost when someone tries to understand AI without the mathematics — and what can they see clearly once they have it that was invisible before?
That’s a really important point. First I should point out that the math, while serious, is not PhD-level stuff. I start from high-school-level concepts and explain all the basics, such as what is a matrix, what is a derivative and so on. I have always been surprised to see people who are scared of mathematics while they master very intricate concepts in other disciplines. There is really nothing difficult there. If you go through that reminder chapter then you can understand the techniques of Modern-AI and why they have made such incredible breakthroughs possible.
Why should people go through those explanations? They simply gain, at relatively little effort (again, it’s not that hard), an incredible edge over those who are just passive users. Let me give just three examples.
Example 1: if you ask an LLM the same question twice, you typically get different answers. What is the story here? Is the technology non-deterministic? Well, actually, not in any actual sense. If you have learned the basic mathematical techniques, you will know the reasons for this phenomenon, how to cope with it, and even how to take advantage of it.
Example 2: AI tools – not just LLMs – make occasional mistakes. By now many people know the term “hallucination”. But the reasons for this phenomenon remain mysterious to them. In reality, there are precise causes for hallucination; if you know them you can be much more effective in using AI results, knowing when to trust them and when not.
Example 3: we have heard about artificial intelligence for decades. Some of the debates seem eternal. But something changed fundamentally in the past few years. Most people do not realize that earlier techniques, what the book calls “Old-AI”, essentially failed, and that the techniques that actually succeed today, “Modern-AI”, are radically different. Not just better ideas: a complete reversal of the approach and rejection of earlier ones. The book explains this 180-degree turn; if you do not realize what happened you just cannot be in control of your use of AI. That involves understanding the math.
Q3. You describe the explanations of transformers in the existing literature as “often nebulous, complicated and confusing” — a pointed criticism from someone who has spent a career thinking about how to explain complex technical concepts clearly. What is the specific conceptual gap in how transformers are typically explained that produces that confusion — and what is the insight that, once grasped, makes the architecture genuinely comprehensible rather than merely memorizable?
The 2017 Transformer paper is beautiful, a true milestone, but completely incomprehensible, right from the first few paragraphs, to anyone not deeply versed into the AI literature – countless papers and books – that preceded it. If you are trying to understand the concepts on your own, most of the explanations you find simply expand on that paper, without explaining what the technology is trying to do and what is new about it. One of the key issues, whose resolution led to the extraordinary success of LLMs, was to keep a sufficient window into past interactions, without consuming unrealistically large computing resources. The “attention” mechanism, extended to “self-attention”, was a brilliant idea that made this goal reachable. But people need to understand the details. For example, why earlier approaches, such as convolution networks (which the book also explains), did not suffice. Or why the basic ideas of “queries” and “keys”, which are intuitive, need to be complemented by “values”.
Q4. Your book includes two chapters on education — one on self-learning with AI and one on educational institutions — and you describe education as perhaps the area of human activity most affected by modern AI. You also make a distinction that seems important: that AI tools for personal self-learning are a different question from AI in formal education, and that there is no suggestion that AI removes the need for a teacher in standard education. What is the most important thing that educational institutions are currently getting wrong about AI — and what would a genuinely well-designed institutional response to AI look like, rather than the panicked or dismissive responses you observe in practice?
Yes, education is an inevitable topic if you are trying to assess the effects of AI, and the two topics are different.
For self-learning, AI can be a fantastic tool, but it needs to be used well; the book describes, based on my own experience, a precise learning strategy based on the combination of AI with other techniques, some of them traditional, such as books and articles, the others IT-based. I have not seen anything similar elsewhere. If we are to benefit from AI resources, we need to have clear guidance and a systematic process.
For education in an institutional context, the book’s basic lesson is to be very careful. I can claim some experience here. One of the characteristic of research academics is that they usually teach and have strong views on educational matters. The other is that few of them condescend to publish on educational issues, in actual education-research venues with high standards such as ACM SIGCSE or ITiCSE. In my work at ETH Zurich I took that step: I taught introductory programming 13 years in a row, teaching altogether some 6000 students, and published on this experience and other educational matters, such as distributed, multi-university courses, in those top education conferences and journals. I learned a few things in the process, including that it is very easy to have “great” educational ideas and to convince yourself that the whole planet should use them. “I tried it, the ideas sound great, the students loved the result!” Good educational publications will reject that kind of submission. They rightly demand solid empirical evidence that the techniques work in a general context. In the two millennia of educational experience leading to the current educational system, only a few new ideas have actually survived the clash with reality. I am extremely wary of people who claim that we should completely rethink education based on AI. Being an enthusiast about AI, as I am, does not mean removing, all of a sudden, all your conceptual and methodological guardrails.
AI as it exists today works best for people who have acquired a sound intellectual discipline (the ones whom AI helps in self-learning). The role of schools is to teach that discipline. It has by no means been demonstrated that AI is better at that than traditional student-teacher interactions. Or even that it helps at all. We have little solid empirical evidence; some early enthusiastic studies were later debunked and forced to undergo retraction. Education is a long-term endeavor; beware of promises of false solutions. Education is also a delicate business, dealing with one of the most precious and fragile systems in existence: the malleable minds of young people. It calls for analysis and care, not precipitation and overhaul.
We do need to teach AI itself, as a topic. But here too I see the naivety of people, often not educators themselves, who have all these revolutionary ideas about education. Very nice, but as every professional educator knows, the problem is not (or not just) to introduce a new topic: since the curriculum is a fixed pie, the problem is what to remove. This is a difficult discussion.
Q5. You dedicate significant space in the book to what you call the risks of AI — hallucinations, IP looting, the AGI debate, and what you memorably describe as “good form covering bad substance.” That last phrase is particularly sharp. Can you explain what you mean by it — where specifically do you see AI producing outputs that have the appearance of quality and reasoning but are covering something hollow underneath — and what practical habits of mind should someone develop to recognize and defend against that failure mode in their own AI-assisted work?
One of the most blatant risks, already identified by people such as Terence Tao, is reflected in the title of one of the book’s sections in the risks-of-AI chapter: “elegant form legitimizing bad substance”. One of the biggest successes of Modern-AI is that it has mastered human language, for decades the holy grail of AI research. Today’s tools can produce impeccable texts. That successes affects our long-established modes of assessing quality. In the past, we were used to assuming that if a text was written well its content was also of decent quality, simply because it is reasonable to expect that if the author devoted attention to form he must have devoted comparable effort to substance as well. Conversely, one of the ways to detect a spam message, for example, was bad spelling and phrasing. With Modern-AI this assumption breaks down. You can have perfect-looking junk! In spite of the proverb, we used to have some justification when “judging a book by its cover”. That is no longer the case. The “cover” can be beautiful and the “book” total slop.
To take just another one of the examples you cite from the risks-of-AI chapter: the massive intellectual property looting performed by LLMs when processing the Web is an ongoing scandal of planetary proportions. As you may know, I have been involved in an attempt, Recognyze AI, to correct that situation. AI companies are going to have to come clean, or the backlash from society and governments will be huge.
Q6. You invented Eiffel, wrote what became a classic textbook in object-oriented software construction, contributed foundational ideas to software engineering across five decades, and now — at a moment when the field you helped shape is being transformed by AI — you have written a book explaining that transformation to the educated non-specialist. That is a remarkable arc. Here is the personal question: when you look at what large language models and modern AI systems actually do, as someone who understands them deeply from the inside, what is your honest emotional response — is it wonder, unease, satisfaction, alarm, or something harder to name — and has writing this book changed how you feel about where the technology is taking the discipline you devoted your life to?
“Wonder, unease, satisfaction, alarm” – a bit of all of them. More on the wonder side: I keep being awed by the possibilities. Unease or alarm arise when I think of intellectual professions that may be at risk; translators, for example. It is disturbing to think that in such endeavors intelligence, thoughtfulness and experience are no longer a guarantee of success. Also, I experience a bit of puzzlement when I look at the techniques, which are more “flies” than “bees”, referring to a text by the Belgian poet Maeterlinck which I cite in the book. I got into computer science through AI as a student, first in France then at Stanford, and left it for software engineering; coming back now, to a field based on entirely different ideas, is somewhat of a shock, as the techniques we thought would work did not, and the other way around.
You mention my work in software engineering. Programming has fundamentally changed, in just a couple of years, because of AI. What strikes me is how wrong current conventional wisdom is here in dismissing the importance of the role of programmers. That role is more essential than ever, particularly the need for good programmers. As I detail in the book, the demise of programmers has been announced many times before, but this advance is different from previous ones: it’s not like a more high-level programming language or design formalism which will generated code (such as machine code) that most programmers do not even understand. There is no way you can use AI-based development (“vibe coding”, although the expression is going out of favor as quickly as in came into use) without understanding the code that it produces. The book distinguishes between two kinds of technology advances: L-advances, leveling everyone to the level of the best professionals; E-advances, enhancing the advantage of the best. AI can be either, depending on the field, but in software development it is definitely E-type. There will be little room left for incompetent programmers and outsourcing.
There also remains the problem of everything else in software engineering. Coding, after all, is some 10-20% of the effort in large industry projects. We are only at the beginning of what AI can do for requirements, design, testing, proofs, evolution… At the end of this month we have a workshop, called FISEE, in Villebrumier near Toulouse on AI’s role in teaching software engineering, with the emphasis not on whining (“AI is replacing programmers!”) but on successful experiences. In the presentations and discussions I hope to learn more about what the future has in store for us.
Qx. Anything else you wish to add?
Perhaps I should mention that since the book was first circulated, several academics told me that universities should offer a general undergraduate course on AI, including for non-computer-science and non-engineering majors, and that AI for Smarties would be an excellent textbook for such a course. As a result of this observation I included all the exercises in the text itself; previously, they were on the site. The book now includes over 180 exercises.
More generally, the basic ideas of AI are intellectually exciting and understandable by anyone with a high-school education. They are as valuable and enjoyable to learn as, say, Shakespeare’s plays, the Impressionists’ paintings or Newton’s mechanics. They should be part of our canon now; with this book I am doing my bit to help bring them there.
……………………………………………………………………..
Bertrand Meyer boasts an illustrious career as a distinguished Professor of Software Engineering (emeritus) at ETH Zurich and Chief Technology Officer of Eiffel Software in Santa Barbara. His extensive publications, including 13 single-author books and over 300 refereed articles, have solidified his reputation as a prolific contributor to the field. Meyer’s exceptional contributions have earned him prestigious awards like the ACM Software System Award and IEEE Harlan Mills Software Engineering Prize, further cementing his status as a respected figure in industry and academia.
Resources