Interview with Gérard Biau: AI as a Scientific Assistant: Institutions, Acceleration, and the Limits of Prediction

AI as a Scientific Assistant: Institutions, Acceleration, and the Limits of Prediction

An Interview with Gérard Biau

August 18, 2026

This interview was originally conducted on May 15, 2025.

Gérard Biau is a professor at the Probability, Statistics and Modeling Laboratory (LPSM) of Sorbonne University. He is a senior member of the Institut Universitaire de France and was elected to the French Académie des sciences in 2024. At the time of the interview, Gérard Biau also held the position of director at the Sorbonne Center for Artificial Intelligence (SCAI).

Statistical learning is not merely one branch of contemporary AI. For Gérard Biau, it is a connective language across the sciences. In this interview he discusses how AI is reshaping research practices, institutional responsibilities, and scientific judgment. Moving between his work at SCAI, the PostGenAI@Paris cluster, and the Académie, he describes the infrastructures through which AI enters universities, industry, and public debate. A central tension runs through the conversation: AI can act as an extraordinarily productive scientific assistant, yet its speed and predictive orientation may narrow research horizons and erode foundational expertise. Biau therefore stresses the need to preserve mathematical depth, causal reasoning, and the time required for speculative work. As one of HiAICS’s closest partners, he has accompanied the project over several years.

Andreas Sudmann: As a new member of the Académie des sciences, how do you see yourself contributing to its role in advising public institutions and society on the rapidly evolving field of AI? Are there particular committees or working groups in which you expect to apply your expertise in applied statistics?

Gérard Biau: In France, the Académie des sciences has an important role in public debates about science. Its members are regularly asked to offer their views, participate in roundtables, and contribute to reports or task forces on particular topics. This is part of the role of an academician, and I am increasingly being asked to give talks, courses, or shorter interventions in companies and at universities on matters of scientific importance. Given my expertise, these requests naturally concern AI. I therefore expect this part of my work to grow, although at this early stage I cannot yet point to one specific AI committee in which all of it will take place. The requests come through several channels and concern different audiences. At the same time, an academician has to remain somewhat above the immediate controversies. One has to weigh different positions and speak carefully, because in a sense one is speaking on behalf of science. This does not mean avoiding difficult judgments. It means distinguishing a scientific assessment from a polemical intervention and making the uncertainties visible. In France, the Académie is seen as an institution that carries a certain wisdom. That creates a responsibility to avoid presenting issues as simply black or white. It requires diplomacy. Consider, for example, the environmental impact of AI, which is discussed frequently in panels, companies, and public debates. It would be too simple to say that AI consumes energy, harms the climate, and should therefore be stopped. AI is not carbon-neutral, but significant efforts are being made to develop more frugal systems with lower energy consumption, and AI can also help address important problems related to climate change. Both parts belong in the same account. An academician should provide an honest and scientifically balanced picture, even when the public debate pushes strongly toward a simple position for or against a technology.

Andreas Sudmann: How do you see your expertise in applied statistics informing your work within the Académie? In particular, how might it help initiate interdisciplinary dialogues or research programs that use AI to address complex scientific problems in fields such as climatology?

Gérard Biau: It is useful to return to the circumstances of my election. The Académie is divided into nine sections, including mathematics, biology, and the section of mechanical and computer sciences to which I belong. Traditionally, each section elected two new members in the relevant election cycle. The procedure has now partly changed. In alternating years, the Académie can elect researchers in precisely defined thematic areas in order to shape its scientific policy and ensure that major areas of contemporary science are properly represented. This procedure was used for the first time in 2024. Rather than filling places only through the established sectional routine, the Académie first identified fields that required stronger representation. One of the designated fields was statistics, which is why I was elected. Other new members were selected for areas such as cybersecurity and quantum physics. The choice of these themes is itself a statement about the present and future of science. It shows that the Académie wants to remain connected to the evolution of science and to the fields that are shaping it today. My role is therefore partly to represent expertise in statistics. That matters because statistics is not only a field in its own right; it is also a common language across the sciences. It is present in physics, climatology, economics, and many other domains. Researchers in these fields ask different questions and use different vocabularies, but statistical concepts often allow them to communicate. I can help facilitate connections among these disciplines through a statistical framework. Statistics is also at the heart of AI, while AI is now present in nearly every scientific field. I therefore expect colleagues from different sections to ask me to explain what AI is and is not, where the field is moving, which advances matter, and which methodological errors should be avoided. This will not be a one-directional form of instruction. The questions raised by other disciplines will also change how statisticians understand the uses and limits of AI. This interdisciplinary circulation of knowledge within the Académie will be an important part of my role.

Andreas Sudmann: Within the Académie’s institutional framework, where do you expect discussions of ethical guidelines, reproducibility, and best practices for AI in scientific research to take place? How can the Académie address these issues, including the implications of the European AI Act, while preserving rigor and transparency in data-driven research?

Gérard Biau: To understand this, one first has to understand how the Académie works. My section is called Sciences mécaniques et informatiques, and it includes most of the applied mathematicians. There is a separate mathematics section that is primarily devoted to pure mathematics. Within each section, we conduct the traditional work of a learned society: we examine candidates for prizes, prepare reports when requested, and organize events. Sections are in this respect small scientific societies within the larger institution. There is also work at the level of the Académie as a whole. The general assembly is historically called the comité secret, although it is not secret at all, and it meets roughly every two months. Physicists, mathematicians, biologists, physicians, and researchers from other fields all participate. The assembly therefore does not usually conduct highly specialized thematic work. It discusses broader questions, including the ethics of science, data practices, reproducibility, and changes in how research is conducted. Only two months ago, the assembly decided to re-establish a committee on scientific ethics. Such a committee had existed before but was no longer active. There is now a clear need for it again, and I expect to be asked to contribute opinions or to help hear from relevant actors on questions concerning ethics and scientific practice. AI regulation and the AI Act make this responsibility especially demanding because AI, statistics, and data are involved in almost every field. No one specialist can master the legal, technical, ethical, and disciplinary dimensions alone. The Académie must bring the relevant perspectives together and listen carefully before it issues advice. Whenever an AI question arises, it turns to the small number of members with relevant expertise, which also creates a difficult position for those members. One joins the Académie partly to learn about fields beyond one’s own, but one can quickly become occupied with the domain one already represents. I have only been a member for four months, so this is still an expectation rather than a settled routine, but I anticipate being asked to contribute frequently.

Andreas Sudmann: Given the Académie’s long-term strategic role in French and international science, how can it most effectively help shape research agendas and the responsible use of AI? What perspective from applied statistics do you hope to bring to its strategic deliberations?

Gérard Biau: The Académie des sciences is sometimes described in France as the parliament of science. Parliament makes laws, which is not the role of the Académie, but the Académie has considerable scientific authority. When a major issue arises, the government can ask for its opinion. If the government were considering whether AI should be taught throughout the French education system, perhaps even from a very early age, it could ask the Académie to assess the current state of AI education and recommend a direction. The request would not simply be whether AI is good or bad. It would concern the existing evidence, the level at which different concepts can reasonably be introduced, and the competencies teachers and students would require. The Académie can also act on its own initiative. If it considers a topic important, it can decide to prepare a report without waiting for a government request. It recently produced a report on research evaluation in France, for example. It could likewise decide that the environmental impact of AI requires a clearer scientific assessment and mobilize academicians to examine the evidence. In this way, it can both respond to a public authority and identify a question that public authorities have not yet formulated adequately. Reports from the Académie are public and carry substantial weight. They are treated as high-level reference documents. They do not become law and should not be accepted as a kind of infallible scientific bible, but they can establish a direction for future policy and make clear where evidence is strong or still uncertain. In that sense, the Académie can influence the country’s scientific trajectory. This is not unique to France. A report from the National Academy of Sciences in the United States also has significant authority. Applied statistics contributes to such deliberations by providing the methods needed to evaluate evidence across domains, compare heterogeneous studies, and clarify what conclusions data can and cannot support.

Andreas Sudmann: At the same time, you are director of SCAI. How does leading a major AI research center shape what you can contribute to the Académie, and how might your work within the Académie in turn benefit SCAI’s strategy and mission?

Gérard Biau: It is important to bring practical experience into the Académie. At SCAI, we encounter the concrete difficulties of disseminating AI, teaching it, and deploying projects. We see which operations work well and which are difficult to implement. Everything is concrete there: courses need instructors, collaborations need institutional arrangements, and technical systems have to function in actual research settings. In that sense, we are on the front line. We can report not only what a technology promises in principle but what happens when different scientific communities try to use it. We also have an excellent view of developments in France and abroad because we maintain close connections across Europe, including Germany, as well as with the United States and China, and because our students come from many parts of the world. SCAI can therefore provide the Académie with information from the field and a clear picture of what is happening in practice. This kind of feedback is important for an institution whose discussions necessarily take place at a high level. Conversely, my role in the Académie helps me perceive emerging scientific priorities. One quickly senses which topics are becoming central and how researchers from very different fields frame them. The thematic areas chosen for the 2024 elections are themselves an indication of what the Académie considers important for the science of tomorrow. They provide information about long-term priorities before those priorities are fully translated into programs. As director of SCAI, I gain a better view of the research directions that matter nationally and of how France may seek to develop them in the near future. This can inform our choices about partnerships, teaching, and the fields in which SCAI should build capacity. The relationship is therefore reciprocal: SCAI supplies practical knowledge, while the Académie offers a broader strategic horizon.

Andreas Sudmann: Our previous interview took place in May 2024, shortly before Sorbonne University secured funding for the PostGenAI@Paris cluster, in which SCAI plays a central role. How has that success changed SCAI’s role, scale, and priorities?

Gérard Biau: I remember telling you during our last conversation that SCAI had reached the end of a cycle. Xavier Fresquet and I had started from scratch. At the beginning of the French national AI strategy, Sorbonne University decided to build its own center rather than apply to become one of the national institutes. We were therefore part of the broader strategy without being fully integrated into its initial institutional structure. We were neither completely outside it nor equipped with the same status and resources as the national centers. Nevertheless, SCAI continued to exist and gradually became a recognized actor. Over time, we built courses, projects, and collaborations, but we eventually reached the limit of what we could achieve within the traditional university environment. We had done almost everything that was possible at that scale. If we wanted to expand and accelerate, we needed a substantial new impetus. I said at the time that winning the cluster would open a new phase, while losing it would have made SCAI’s survival very difficult. The competition was therefore an important institutional threshold, not simply another funding application. We succeeded, and that has given SCAI a new start. We now have more than twenty collaborators, stronger relationships with companies, and growing activities at Sorbonne University Abu Dhabi that are connected to the Paris cluster. Success also makes further cooperation easier because external partners see that the model has institutional support. Our aim is to serve as a bridge between the university and the economic world. Germany has a more established culture of cooperation between universities and companies, visible for instance in the Fraunhofer institutes. In France, universities and companies have traditionally been more separate, yet progress in AI requires them to work together. The two sides cannot remain isolated when research, infrastructures, and applications develop so rapidly. The cluster makes this cooperation easier because companies increasingly recognize SCAI as an important point of access to the university and trust us as an intermediary. Within the university, colleagues also see more clearly what we contribute. We have demonstrated that the approach can work, so there is less reluctance and we are granted more autonomy. SCAI has therefore entered a new phase. The immediate task is to stabilize the model and scale it over the cluster’s five-year funding period. After that, we may discover that another step, perhaps even a larger institutional structure, is necessary. For now, however, the priority is to consolidate what we have built and make this new scale sustainable.

Andreas Sudmann: The cluster uses the term post-generative AI for systems that may operate more autonomously and contextually than current generative models. How might this shift alter research methods and scientific discovery across the cluster’s three main areas of disruptive technologies, future health, and resilient societies?

Gérard Biau: This connects directly to our earlier discussion about mathematics. Generative AI is already here, and the technology will continue to improve through agents and related developments. There is an important technological side to post-generative AI, but the more consequential question may be how these systems enter everyday life, professional work, teaching, administration, and scientific research. Post-generative AI refers to the stage at which generative systems cease to be isolated tools and become embedded in these practices. The issue is therefore how generative AI lands in institutions and changes ordinary activity. In research, we already have, and will increasingly have, super-assistants that know a great deal, compute rapidly, propose ideas that we still have to evaluate, and make connections across fields. They do not remove the need for scientific judgment. Their proposals have to be graded, corrected, and placed in a disciplinary context. Yet they can accelerate the exploratory work that precedes a result. Interdisciplinary work could become easier. I know very little about quantum physics, for example, but I could ask an agent whether a statistical idea resembles something in quantum physics and whether the fields are using different vocabularies for related concepts. The system might reveal a connection that would otherwise take much longer to identify. Even when the analogy is imperfect, it can tell the researcher where to begin reading or whom to contact. Routine data processing will also become much faster. A researcher can ask for a summary of a data set, a first visualization, or potentially interesting patterns without first writing every Python program by hand. These preliminary tasks can be tedious, although they are necessary for seeing what a data set might contain. AI compresses that interval dramatically. It will certainly contribute to remarkable discoveries, especially in areas such as medicine and healthcare. Yet the deeper transformation concerns how research itself is conducted. Until now, we have generally programmed a computer, run an experiment, inspected the results, changed the program, and eventually written a paper. The computer executed a task that the researcher had specified. AI-enabled computers will increasingly become active laboratory assistants that participate in this iterative process, answer questions, and suggest the next operation. In that limited but important sense, they will do research with us. Researchers in mathematics, physics, biology, and other fields already sense this shift. Many come to SCAI because they know something significant is happening but are unsure how to respond, which tools to trust, or how to integrate them into their methods. SCAI’s role is to provide appropriate tools, explain their limits, and accompany scientific communities through this transition.

Andreas Sudmann: PostGenAI@Paris also places strong emphasis on industrial partnerships and the transfer of expertise into fields such as law and economics. How are its 21 Collaborative Acceleration Programmes structured to integrate academic research, industrial innovation, teaching, and wider societal concerns?

Gérard Biau: This integration is built into the design of the cluster. Each of the 21 Collaborative Acceleration Programmes, or CAPs, is a major research program with both an academic and an industrial component. A proposal could not be approved unless the researchers had established a strong industrial relationship in advance and the industrial partner was already involved. It was not enough to promise that a company might be found later. We wanted the CAPs to be oriented toward cooperation with industry from the outset, so the scientific questions would develop in contact with concrete needs while remaining genuine research questions. We also required them to include teaching. The principal investigators committed both to collaborating with companies and to translating the research conducted in their programs into educational formats. This means that a CAP is expected to connect research, innovation, and the formation of students rather than treating them as separate activities. The research topics are developed together with industrial partners while also feeding directly into courses and training. The structure is intended to make transfer continuous instead of postponing it until after a research project has ended.

Andreas Sudmann: A core objective of PostGenAI@Paris is to train substantially more students in AI and build expertise across disciplines. Which pedagogical approaches and competencies does the cluster prioritize so that the next generation of researchers can use advanced AI systems effectively and responsibly?

Gérard Biau: PostGenAI@Paris certainly trains the next generation of AI makers, including computer scientists and mathematicians who will develop new systems. The larger challenge, however, is to provide every student with an education in AI. We call this AI + X. AI should become part of programs in physics, chemistry, medicine, the humanities, and other fields. This does not mean inserting the same computer-science course into every curriculum. The relevant examples, depth, and methods have to be connected to the discipline in which students will use them. The difficulty is that the necessary teachers cannot simply be produced overnight, while demand is growing everywhere. If every bachelor’s and master’s program requires AI teaching, the number of instructors needed is enormous. PostGenAI and SCAI therefore have to find scalable solutions that support departments rather than leaving each of them to invent a curriculum alone. We organize summer schools and hackathons, develop teaching resources through SCAI Education, and train instructors who can bring these materials into their own fields. The aim is not to turn a philosophy student into an AI specialist. Yet whatever profession that student later enters, AI will be present, so some basic understanding and acculturation are necessary. Students should know what these systems do, what they do not do, and where critical judgment remains indispensable. In January, for example, Xavier Fresquet organized an intensive week for students from the Faculty of Arts and Humanities. They began from scratch with an introduction to what AI is and is not. They then worked with computers and Python to understand concretely what a machine-learning model does rather than encountering it only as an abstract idea. The program also covered prompting, generative AI, agentic AI, and current tools. In one week, the students did not become specialists, but they acquired a durable orientation that will remain useful throughout their working lives. That is exactly what post-generative AI education should accomplish. It combines conceptual understanding, some direct technical experience, and the ability to assess tools in context. We need comparable formats across the university.

Andreas Sudmann: Recent debate has focused heavily on generative AI and large language models. From your perspective as a mathematician and specialist in applied statistics, could this intense focus risk obscuring other forms of AI or narrowing our understanding of how AI is transforming scientific research?

Gérard Biau: One possible danger is that AI now allows research to move very quickly. It accelerates discovery. In mathematics, it helps us work with data, test ideas, and implement algorithms much faster. Younger researchers are understandably attracted to this speed because the process can feel almost like a game. A new idea can be translated into code and tested almost immediately. The next generation will integrate this way of working very quickly, and to some extent it has already done so. My doctoral students sometimes arrive with code that has clearly been developed through substantial interaction with AI, and I suspect they increasingly use these systems for mathematics as well. Their supervisors may not always know how extensive that interaction has become. We must therefore make sure that essential competencies are not lost. Using AI effectively requires a strong background in mathematics and computer science. Running Python code is one thing; understanding the code deeply, seeing why it works, and being able to identify a hidden error are different matters. Students must retain the foundations of their disciplines if they are to use AI intelligently rather than simply accept its output. A second danger is the growing pressure toward short-term science. Important research often begins with ideas that appear implausible, fail repeatedly, and may only become productive years later. Some research is not connected to an immediate problem at all, and its significance becomes visible only much later. AI encourages a rapid sequence of ideas, notes, implementations, and papers. That can be productive, but it also makes constant output seem like the natural rhythm of research. We should preserve the time needed to think, take risks, and pursue problems without an immediate application. Science takes time, and it sometimes requires remaining with a question even when nothing publishable appears quickly. A third issue is that current systems still operate largely through correlations. At a high level, they receive inputs and outputs and identify statistical relationships. Science, however, is not only prediction. It also seeks an understanding of phenomena as a whole, a view of the world, and an account of causal mechanisms. Large language models are next-token systems. Their mode of operation can push us toward asking only what comes next and toward organizing reasoning as a sequence of locally plausible steps. Scientific reasoning also requires a broader picture and the question of what causes what. It asks which mechanisms generate the observed regularities and how different parts of a phenomenon belong together. Current models can assist with this work, but their basic operation does not supply that causal or synoptic perspective. We risk losing some of that orientation if LLM-based reasoning becomes the dominant model of thought and if the speed of prediction is confused with scientific understanding.

Andreas Sudmann: Your argument suggests that there is a physical world beyond the data on which large language models operate. Agents are now being connected more directly to environments, yet this connection remains limited and raises questions about how much autonomy they should be granted. How does this problem of embodiment shape your assessment of AI in science?

Gérard Biau: There is a physical world, as well as a world of intuition, and current models do not truly engage with it. Many important intuitions in mathematics and science are connected to the way we interact with the physical world. Results that appear highly abstract can still be guided by spatial, mechanical, or visual experience. I see AI systems as extraordinary instruments of deduction, but they do not really understand the world. Science also means understanding the world in which we live, not only identifying regularities within data. Agents may become more directly connected to environments and may act within them, but the depth of that connection remains unresolved. Giving a system access to sensors and actions does not automatically provide common sense or a human relation to consequences. It also creates a practical question about how much autonomy we are prepared to grant. The same broader problem arises in artistic creation. AI systems can produce impressive pictures, images, and videos, but it is unclear whether they are genuinely groundbreaking. Creativity involves more than improving an existing pattern step by step. It includes a representation of the world and is connected to our physical and human existence. Large language models are embodied in infrastructures, data centers, algorithms, and bodies of data, but this is different from the embodied relation to the world on which human intuition depends. Their material infrastructure matters greatly, especially ecologically, yet it does not by itself solve the epistemic problem of grounding. In mathematics, their next-token operation encourages a mechanical sequence of ideas. One can sense this in the way they reason: one idea follows another, often impressively, while the wider intuition remains less secure. What is often missing is an overarching view, a relation to physics, and an understanding of why a problem matters as a whole.

Andreas Sudmann: Before the interview, you described a recent experience in which a large language model helped you work through an advanced mathematical problem and changed your assessment of its reasoning capabilities. What happened?

Gérard Biau: This experience is very recent. Until a few days ago, I had not considered using ChatGPT for serious mathematics rather than small computations. I decided to try, and I was surprised that we could sustain an interesting interaction. The system functioned like a highly capable assistant. I began with an idea of my own whose notation and significance were still unclear. The idea came from me; the machine did not invent the problem. But it immediately understood the basic proposal and suggested a notation. I rejected the first version because it was not quite what I wanted, and it responded that it now understood the distinction and produced another. We could refine the mathematical language through dialogue. I then asked whether the idea was interesting and how it might relate to the literature. The system identified affinities with several existing approaches while also explaining where the proposal differed and what potential it might have. This did not settle the question, but it gave me a map of possibilities against which I could test my own judgment. We then tried to prove something. Its initial suggestions were false. I pointed out the errors, and it acknowledged them and tried again. At one point it invoked a result that did not exist, so I corrected it again. This is precisely why a strong mathematical background remains necessary: without it, one could easily accept an elegant but invalid argument. Yet the system never became tired or discouraged. It remained fully engaged whatever the hour and generated another approach each time an argument failed. Step by step, we arrived at a proof whose details may still need refinement but whose main arguments appear to be there. The system’s persistence made it possible to explore more variants than I might have pursued alone in the same period. The experience was also intellectually pleasurable, which surprised me. Of course, doing mathematics with colleagues and students is pleasurable as well, and the human relationship is completely different. But I found myself enjoying the exchange with the machine. I knew that it was an algorithmic system based on next-token prediction, yet during the interaction one partly forgets this and enters into the mathematical problem. For a time, it feels less like operating a search tool and more like working with an unusually patient assistant.

Andreas Sudmann: Could you briefly describe the mathematical problem you explored with the model?

Gérard Biau: I had an idea for a neural-network architecture inspired by attention models. I asked whether a network could be constructed in this particular way. The system grasped the intuition quickly and helped me formulate the architecture more clearly. I then asked whether the proposal was genuinely different from conventional transformers or attention models. It argued that it was and explained why. The useful part was not merely the affirmation. It pointed to structural features that could be compared with the literature. In particular, it noticed that the matrices involved were rank-one matrices with a special structure and suggested that this was the interesting feature of the idea. I had not initially expressed the point in those terms. The observation helped identify what distinguished the proposal from existing architectures and why the difference might matter. We could then ask what kinds of patterns in the data such a structure might capture and what mathematical properties could potentially be proved. I supplied the questions and the central idea, but the system provided answers that pushed the inquiry forward and gave me new formulations to evaluate. It did not replace the mathematical work, and its mistakes remained serious, but it helped organize that work. In this sense, it was doing mathematics with me. We should still remember that it is a machine, but science has always depended on instruments. Surgeons work with knives, farmers with tractors, and researchers with computers for writing, calculation, simulation, and communication. The relevant question is what an instrument allows us to perceive or do, and what competencies are needed to control it. AI introduces another way of doing science. The experience convinced me that the change is not limited to automating routine calculations. It can enter the formation and testing of ideas themselves. We are entering a fascinating period in which research practices may change profoundly.

Andreas Sudmann: Looking further ahead, which emerging challenges do you expect to become decisive as AI continues to transform the sciences? Which problems require scientific, institutional, or conceptual responses now to prevent them from becoming major obstacles later?

Gérard Biau: The first challenge I encounter repeatedly in courses with students and professionals is the environmental question. AI has a real environmental impact, and people increasingly want to know how much electricity and water is consumed when they use a large language model. AI requires GPUs and large computing infrastructures, which consume substantial energy. GPU farms also create relatively few jobs once they are operating, so their local economic contribution may be limited in relation to their resource demands. The environmental cost of AI will therefore become increasingly critical. Science can help mitigate these effects through more efficient hardware, smaller models, and more frugal methods, but technical improvement will not make the political and ethical question disappear. The issue may also change how younger generations relate to AI. Students are often much more anxious about the environment than people of my generation and can take more radical positions. They may eventually decide that certain forms of AI are too damaging to the planet and refuse to use them. The future of AI adoption is not predetermined. A second major direction concerns robotics and autonomous systems that enter daily life. The question is how the knowledge represented in current systems can be translated into concrete actions. We are still far from reliable machines that can plan and carry out ordinary household tasks such as cleaning dishes, but agentic AI already performs actions, answers messages, and organizes schedules with growing autonomy. Useful robots would need to anticipate, plan, respond to changing environments, and operate safely around people. Transferring current capabilities into action in the physical world will be a central research problem. A third challenge follows from the limitations of today’s LLMs. They are enormous models with billions of parameters, require substantial energy to train, cannot plan reliably, lack common sense, and contain no genuine physics. They rely heavily on brute force. A child does not need to see billions of images to distinguish a cat from a dog, while a machine may require immense quantities of data. Human learning is much more economical and much more closely connected to action in the world. We therefore need forms of AI that are more closely connected to the real world and to physical knowledge. Such an approach is not yet available at scale, but it could become a genuine game changer. The future of AI is not necessarily synonymous with large language models. We may see a different paradigm emerge. Finally, education faces a tsunami. Students of all ages already have access to AI, so schools and universities must decide how these systems should be used in courses and examinations, how AI itself should be taught, and what knowledge still matters. These decisions have to be made quickly because the tools are already present. A few years ago, we said that everyone should learn Python. Today, AI can produce Python code for almost anyone, which forces us to reconsider the purpose and depth of programming education. It may still be essential to understand code even if writing every line manually becomes less important. The larger task is to prepare students for a world that will change extremely quickly, give them tools we cannot yet imagine, and enable them to move flexibly between domains. That flexibility must not come at the cost of disciplinary foundations and critical judgment.