At Rise of AI 2026 in Berlin, Christoph von der Malsburg argues that present AI, impressive as a passive "encyclopedia," remains a black box that learns inefficiently and cannot yet safely act on its own, citing Anthropic's refusal to let a government control weapons with its AI as a sign of how shaky current control is. He proposes animals as the model for future "agentic AI": many are born with most of their behavior already in place, and should be able to learn a new environment the way an employee learns a company, by observing rather than months of training data. He argues biology-based AI is regaining interest, citing ventures founded by Jeff Bezos, Yann LeCun and Jeff McCaleb and new German funding calls, but says none of them yet knows how a squirrel does what it does. Drawing on his own decades of research, he lays out a framework built on structured perception, a "jigsaw puzzle" model of object representation in the visual cortex, mapping between sensory input and abstract categories, and innate behavioral schemata, closing by arguing Europe could out-compete data-hungry American and Chinese AI with an approach that does not depend on mass data.
Key takeaways
- Christoph von der Malsburg calls today's agentic AI "playing with fire," citing Anthropic's refusal to let a government control weapons with its AI as proof control remains shaky.
- He argues agentic AI should learn a new environment the way a new employee learns a company, by observing and asking, rather than through months of pre-training.
- He identifies the binding problem, representing which neurons belong together as one coherent object, as AI's central unsolved representation problem.
- He built a fully neural face-recognition model, following work from two companies he co-founded around 20 years ago, that recognizes faces in about 20 milliseconds.
- He argues Europe could compete in AI through biologically inspired systems that gather data through their own curiosity rather than mass data collection.
Detailed notes
What is wrong with present AI
Von der Malsburg opens with the practical problems of deploying today's AI in a company: it learns inefficiently, requiring laborious new data for every application, and is a black box even to its creators, making its behavioral control unreliable. He distinguishes two paradigms: AI as a passive "encyclopedia" that gives impressive answers but only reacts, and the emerging push toward agentic AI that acts directly on a computer or in the world, which he calls "playing with fire" at its current reliability, citing Anthropic's refusal to let the American government control weapons systems with its AI as evidence of how shaky control is. He also points to self-driving cars, still needing millions of miles of human-driving data, as proof direct AI-world interaction remains unsolved.
Animals as the model for agentic AI
Von der Malsburg argues that agentic AI should be able to "learn on the job" the way a new employee learns a company, by observing and asking questions, rather than through extensive pre-training. He points to animals as proof this is possible: many are born with most of their behavior already built in, such as a pigeon able to fly out of the nest or a squirrel able to judge branch strength, find nuts, avoid predators and build a nest without being taught. He has advocated biology as a model for AI for decades against mainstream resistance, but says the mood is shifting: he cites Jeff Bezos founding Prometheus, Yann LeCun founding AMI, and Jeff McCaleb founding two further companies on AGI and on biology and neuroscience, led by Dileep George and Doris Tsao, each funded at the billion-dollar level, alongside smaller German funding calls he has applied to. He cautions none of these ventures yet knows how a squirrel does what it does, and that the roughly 15 to 20 disciplines studying brain and mind largely fail to talk to each other.
Structured perception and the binding problem
Von der Malsburg argues that raw sensory input is an "ocean of complexity" that never repeats exactly, but that the physical world has deep structure, comparable to how computer graphics generate realistic scenes from underlying rules. Citing the no free lunch theorem, he argues a learning system must be tuned to its application domain rather than seeking one universal optimum. He identifies representing a coherent object as the field's central unsolved problem, contrasting symbolic AI, which uses compact symbols for whole objects, with neural approaches that represent detail at the level of individual neurons but cannot capture which neurons belong together as one object, the binding problem, which he says he first proposed 40 years ago and remains largely unaddressed. He offers a jigsaw-puzzle analogy: pieces of a scene, or "net fragments" of neurons in the visual cortex, only activate when they fit their neighbors, which is how newborns segment an object from its background using continuous motion alone.
Mapping and abstraction
Von der Malsburg describes co-founding two face-recognition companies, one in Bochum and one in Los Angeles, roughly 20 years ago, built on mapping an incoming image onto a stored template, work he says the field later abandoned, though he has since built a fully neural model that recognizes faces in about 20 milliseconds. He connects this to Moravec's paradox, roboticist Hans Moravec's observation that AI can play chess but not replicate what a mouse does escaping into a hole, and argues abstraction, relating the concrete world to categories such as language, is achieved through mapping. He traces "schemata" from Immanuel Kant and Edmund Husserl through to Marvin Minsky's and Roger Schank's frames and scripts, arguing those never solved how to connect abstract descriptions back to real situations, something children do effortlessly, as when a simple drawing is instantly recognized as "papa."
Instinct, imprinting and cooperative networks
Von der Malsburg argues that behavioral control, driven by instincts and drives in the midbrain, is a missing piece agentic AI will also need, since humans are guided by hunger, danger and similar drives far more than commonly assumed. He cites Konrad Lorenz, Nikolaas Tinbergen and E.O. Wilson's descriptions of innate perceptive and action schemata, including Lorenz's gosling imprinting on a wiggled boot mimicking its mother, still used to train geese to follow aircraft on new migration routes, and the innate infant schema for a simplified face, which later develops into recognition of seven distinct facial expressions. He summarizes his own framework, "cooperative net architecture": only a small fraction of the brain's neurons, an estimated one in 1,000, are active at any moment, each held active briefly by roughly two dozen connected neurons, forming a self-consistent pattern across the whole brain through self-organization, comparable to a physicist's energy function.
A path for Europe
Von der Malsburg says his personal strategy has been to wait for the limits of deep-learning and backpropagation-based AI to become visible, and that renewed company founding and government funding calls signal that shift is underway; an earlier attempt of his own to start a company eight years ago, backed by Berlin VC Blue Yard, failed because young engineers at the time wanted only deep learning, whereas he has now assembled a team for a German funding application and is preparing another application to a second German funding program. He argues American and Chinese AI is built on large-scale data collection Europe is reluctant to pursue, and frames biologically inspired AI, which gathers information through curiosity and play rather than mass data, as a technology Europe could use to start over and compete, despite established AI already attracting hundreds of billions in investment against almost none for this approach.
Quotes
The present AI is a completely black box, even to the people who have built it.
Christoph von der Malsburg
None of those companies I've spoken about have a clear idea how a squirrel does it.
Christoph von der Malsburg
As soon as you are hungry, or as soon as you get into danger, you know very well who is calling the shots in your brain.
Christoph von der Malsburg
Questions answered in this talk
Why does Christoph von der Malsburg think today's agentic AI is risky?
According to Christoph von der Malsburg at Rise of AI 2026, present AI is a black box even to its own creators, making its behavioral control unreliable, which he illustrates by citing Anthropic's refusal to let a government control weapons with its AI.
Why does he use animals as a model for AI?
According to Christoph von der Malsburg at Rise of AI 2026, many animals are born already able to perform complex behavior, such as a squirrel judging branch strength or finding nuts, which he argues shows a path to AI that learns on the job rather than needing extensive pre-training data.
What is the binding problem in brain and AI research?
According to Christoph von der Malsburg at Rise of AI 2026, the binding problem is the unsolved question of how the brain represents which individual neurons, or which parts of an image, belong together as a single coherent object, a challenge he says he first proposed around 40 years ago.