What Could Be the Next AI Winter?

Departing from my usual—but highly irregular—tutorials on deep learning concepts, I want to excite your minds with a speculative idea relevant to the subject.

Imagine having at your disposal a polymath, learned in all present human knowledge, who neither tires nor gets distracted, who “understands” your requests—at least in the sense of identifying your goals through conversation, assimilating relevant knowledge, and responding coherently in alignment with your intent—and who does so with lightning speed. Such a polymathic personal assistant is nearly a reality today.

For instance, I recently conversed with ChatGPT and similar large language models (LLMs) about the nature and mechanics of deep learning itself. To my surprise, I was able to grasp at least the elementary aspects of the subject! It was easier than doing my own desk research—almost like having a teacher nearby. Imagine that—learning AI from AIs. What a delightful treat.

Of course, these chat assistants are not perfect. They still make glaring errors in memory, semantics, and reasoning. However, it is easy to envision these flaws diminishing as models and hardware continue to scale.

Given this trajectory, it seems reasonable to assume that generative AI is accelerating our progress toward superintelligence—at least in certain respects. For instance, one can easily imagine that, within a few years, the human-computer interface will transform into something more seamless and intuitive. We will converse with our technological extensions and accomplish complex tasks effortlessly. The "ontological middleware" of today—graphical user interfaces in operating systems and apps—could become obsolete.

Initially, these interactions with technology will involve explicit conversation, with LLMs and other AI models acting as intermediaries—translating the salient expectations in our spoken thoughts into machine-readable instructions via smartwatches or similar devices. Like KITT from Knight Rider in the 1980s, but better. Perhaps, in a few decades, high-tech interfaces with the human nervous system—perhaps even non-invasive ones akin to today's biometric wearables—will eliminate the need to speak at all. Our technology will respond directly to our thoughts and intentions.

But this raises a crucial question: How far can these machines, powered by LLMs and other AI methods, become truly superintelligent—on par with an Einstein, a Dickens, or a Darwin? Can we, as Dario Amodei and others suggest, create a "country of geniuses in a data center"?

I think not. Here’s why.

An advanced LLM will soon possess—or may already possess—a comprehensive "model of the world." That is, it will have constructed a robust ontology of reality: the relationships between objects, the nature of space and time, and even the laws that govern change. However, there is a fundamental limitation: an LLM’s understanding of reality is based entirely on human language. It is a model constrained by the way words and sentences capture the world. You might wonder, “But isn’t language enough?”

Language is a protocol, evolved to convey sufficient information to help instantiate a multifaceted world model in another person’s mind. It triggers spatiotemporal hallucinations—what we call imagination—as well as nonverbal intuitions, such as fear, joy, or suspicion. Language synchronizes and alerts minds, but it does not encompass all aspects of reality in and of itself.

Consider, for instance, an exchange between two great thinkers. Suppose Newton explains gravitation to a young Einstein, saying, "Any body in space pulls another body toward it, with a force that decreases in proportion to the square of the distance between them." He shares that careful measurements confirm this model, save for a few exceptions, which remain puzzling.

Einstein contemplates this for a while. Suddenly, he brightens with a superior smile. A new vision has struck him—a spatiotemporal hallucination—of bodies rolling through curved space toward other bodies. He imagines space itself bending under the influence of mass, with the degree of curvature determined by the mass of the object. He has no proof yet, no calculations worked out, but the thought electrifies him. He bids Newton farewell and rushes off to his study, eager to explore the mathematics.

This, of course, is a whimsical caricature of history, but it illustrates a crucial point: true innovation is not merely a matter of logically manipulating knowledge extracted from conversation to produce a result. Even if one were to assimilate all available human discourse, true intellectual breakthroughs require something beyond language—a leap of speculative imagination, a vivid and nonverbal restructuring of concepts in the mind.

This restructuring of concepts must have a stochastic quality, where new spatiotemporal structures—akin to, but not quite, a 3D movie—are generated and then culled or modified based on available knowledge. This architecture of human thinking is well known. The late Daniel Dennett, for example, used the phrase "fame in the brain" to describe how different, potentially conflicting mental states compete for attention.

One doesn’t have to appeal to Einstein’s mental feats in general relativity to see the role of imagination in innovation. Consider Jesus introducing the notion of forgiveness in an environment where it was neither known nor practiced, or Gandhi proposing passive resistance as a means to defeat oppression—despite it seeming counterintuitive, even reckless, at the time. People were hurt and killed in these movements, and it did not appear to be the prudent course of action. But it worked. And in that instance, at least, the cost was far less than what an all-out rebellion would have entailed.

Yes, in a deep, deterministic sense, every invention is deduced from existing ideas and facts. But lessons from neuroscience suggest that it pays to hallucinate rather than just deduce. Of course, many hallucinations lead nowhere. But history has shown that this method—conjuring and testing novel ideas—has been instrumental in art, science, politics, and metaphysics.

This “evolving hallucination” approach is something that LLMs and their underlying transformer architectures are not designed to accomplish. In fact, as Thomas Wolf pointed out in a more recent article than Amodei’s, they are architected to follow the rules of the game—to diligently conform to the status quo that emerges from the data, so to speak. The more they are trained, the better they become at providing accurate answers—within the current paradigm. They are like a straight-A college student: exceptional at deductive reasoning, recollection, and knowledge retrieval, yet lacking in creative invention—whether mathematical, physical, or literary.

I am by no means suggesting that AI can never evolve into truly innovative machines. Perhaps it is possible to architect a multifaceted AI more complex than a transformer—one with layers or modules that are trained to hallucinate structures resembling objects and actions in spacetime (a true “world model”) and test them against relevant facts, logic, and equations, identified in a more conventional transformer-like fashion. What I am suggesting, however, is that unless and until we do this, we are likely headed for another AI winter.