The New Cybernetics

A research proposal.

“Life is probably round,” Vincent van Gogh (letter to his brother)

Hi everyone,

I wrote this during a particularly optimistic time thinking about the possibilities of AI and wishing to expand our thinking beyond the day to day reaction to models and benchmarks, and so on. We need big picture thinking.

It’s a research proposal for institutional backing, and I suppose optimistically it would go to a hifalutin place like the Institute for Advanced Study. I am a fellow with the Institute for Advanced Studies in Culture at UVA, as well as a Fellow of the technology and democracy project at the Discovery Institute in Seattle.

I suppose I’m keen to expand this reach out further, or to engage further my existing fellowships. I share this with readers of Colligo to solicit opinions and encourage discussion about this question of the integration of advanced technology into culture and our future.

Statement of Research Interests and Directions

Erik J. Larson

My research focuses on artificial intelligence, human cognition, and collective intelligence—specifically, how AI can augment rather than replace human intelligence. My first book, The Myth of Artificial Intelligence (Harvard University Press, 2021), examined the fundamental limits of inference in AI through a tripartite framework: deduction, induction, and abduction. While symbolic AI (GOFAI) foundered on the limits of deductive inference, modern machine learning remains constrained by its reliance on induction—extrapolating from past data. However, true intelligence requires abduction, the ability to generate novel hypotheses that explain observations. This kind of reasoning is central to human cognition but remains elusive in AI.

While the rise of large language models (LLMs) in late 2023 has led many to believe AI has finally achieved human-like reasoning, I have argued that these systems simulate abductive reasoning without actually performing it—a phenomenon I and others have referred to as Wide AI (in contrast to both Narrow AI and AGI). This distinction has profound implications for the future of AI research.

My forthcoming book, Augmenting Human Intelligence: Empowering Humans in an Age of AI (MIT Press), builds on this foundation to argue that the pursuit of AGI is misguided. We already have general intelligence in humans, and instead of attempting to replicate it artificially, we should focus on amplifying collective intelligence through AI-human feedback loops. The core challenge is not creating autonomous artificial minds but developing dynamic, adaptive systems that enhance human cognitive capacities. I call this cognitive catalysis—expanding and empowering human intelligence through structured human-machine feedback loops. This is the central concern of what I call The New Cybernetics—a research agenda that moves beyond static AI models toward co-evolutionary intelligence networks, where AI actively learns and refines knowledge through structured human interaction. This increases our ability to form hypotheses (novel idea generation) and ask the fundamental questions necessary for thriving in an increasingly complex and uncertain world.

The Great Innovation Slowdown

Despite the narrative that we live in an era of rapid technological progress, empirical research suggests that foundational breakthroughs—discoveries that redefine entire fields—have slowed. A 2020 Nature study analyzing 25 million patents and 3.9 million scientific papers found a decline in the disruptiveness of new innovations since the mid-20th century. Instead of pioneering novel conceptual frameworks, modern patents and research tend to refine and recombine existing ideas rather than replace them.

Historically, the late 19th to mid-20th century produced breakthroughs that opened entirely new scientific frontiers: Einstein’s relativity, quantum mechanics, DNA’s double-helix, antibiotics, the transistor, and spaceflight. By contrast, the past 50 years have seen fewer fundamental discoveries and more incremental improvements along an already defined path—faster chips instead of new computing paradigms, more efficient drugs rather than revolutionary cures. Even AI, despite its recent leaps, relies on decades-old neural network architectures, achieving progress largely by increasing computational power and dataset size, not through novel theoretical insights.

Our high-tech stagnation is driven by several factors: economic incentives that favor exploitation over high-risk exploration, the growing complexity barrier to fundamental discoveries, and AI’s tendency to reinforce existing knowledge rather than generate truly novel ideas. The key question is: Can AI be repurposed to reignite scientific discovery, rather than merely remixing the past?

Rethinking Intelligence: Natural and Artificial

In November 2023, I spent a week as a Visiting Researcher at SFI, where I had the opportunity to discuss developments in AI with faculty member Melanie Mitchell and several graduate students. At the time, my focus was on network centralization—how control over information shapes knowledge flows. This tied into my broader interest in how the web, originally conceived as a free and decentralized space, has been subtly reshaped by algorithmic curation and economic incentives.

Since then, my thinking has expanded. The real issue is not just how knowledge is distributed, but how intelligence itself—both human and artificial—is structured, cultivated, and enhanced.

The Innovation Paradox: AI’s Progress and the Decline of Discovery

The 21st century so far is a paradox. AI’s rapid progress suggests an era of accelerating technological sophistication, yet empirical evidence tells a different story. The same Nature study that charted the decline in disruptive discoveries reveals a world where innovation is slowing, even as computational power increases.

Modern AI follows this same trajectory: today’s models are extraordinarily powerful at synthesizing and extrapolating from existing data, but they do not engage in discovery the way humans do. Large language models simulate reasoning without hypothesizing—the generative leap from observation to plausible explanation that defines true intelligence. This is why AI can summarize known science but does not generate new theories.

The problem is not that AI isn’t yet intelligent enough. The problem is that we are aiming it in the wrong direction.

Toward a Dynamic, Human-AI Intelligence Network

One of the great failures of AI research has been its obsession with replacement rather than augmentation. Decades of investment have been funneled into the pursuit of AGI, a dream that has remained—suspiciously—”ten to fifteen years away” for the past fifty years. But human intelligence is not something to be replicated. It is something that can be expanded.

This is not a new idea. In the 1940s, Norbert Wiener’s cybernetics movement recognized that intelligence is not merely computational but arises through interaction, feedback, and adaptation. The challenge is not to build machines that think in isolation but to engineer systems that enhance human cognition through structured collaboration. The pioneers of computing—from Vannevar Bush’s Memex to Douglas Engelbart’s vision of human-computer symbiosis—understood that intelligence does not exist in a vacuum. It is networked, distributed, and most importantly, dynamically updated.

Today’s AI models, however, are static after training. No matter how sophisticated, a neural network remains a frozen knowledge system—it does not update fluidly like the human brain. Instead, it requires costly and brittle retraining, often leading to catastrophic forgetting. Meanwhile, the human brain, running on just 20 watts of power, continuously rewires itself, integrating new knowledge without erasing the old. This gap—between AI’s rigid epistemic stasis and the brain’s fluid adaptability—is the central bottleneck of modern AI.

Beyond Current AI: A Research Agenda for Intelligence Augmentation

Breaking the Static Mind (first key area)

The core flaw of modern machine learning is epistemic stasis—once trained, a model is frozen in time, locked into the data it was fed, unable to adapt without costly and inefficient retraining. This is a brittle form of intelligence, one that bears little resemblance to how humans—or any biological system—learn.

The human brain, by contrast, is a master of adaptation. It continuously refines its internal models, integrating new information seamlessly and with minimal energy cost. If our goal is true intelligence augmentation, we must move beyond static AI models and toward systems that can learn dynamically, in real time.

In my research, I aim to explore architectures that enable continuous learning —AI systems that do not function as fixed artifacts but evolve through interaction. This is not just a matter of efficiency; it is a matter of capability. An AI that learns dynamically as it engages with the world—rather than waiting for scheduled retraining—would be something fundamentally different from today’s models. It would exhibit the hallmarks of living intelligence: fluidity, adaptability, and responsiveness. But rather than treating this as yet another attempt to replicate human intelligence, we should ask the deeper question: how could such a technology augment our own?

What we seek is not artificial general intelligence (AGI) as it is conventionally imagined. What we seek is an intelligent society—a civilization where intelligence, both human and machine, is maximally generative, adaptive, and available for meaningful use. Advances in AI are part of this challenge. But how do we get there?

Mathematics of Liquid Neural Networks

One answer to breaking the static mind lies in neuroscience-inspired AI architectures, particularly liquid neural networks—a class of models that continually reshape their internal connections in response to new data, mimicking the plasticity of biological neurons.

Unlike traditional deep learning models, which rely on static weight updates, liquid networks use differential equations to dynamically adjust their internal states in real time. This allows for lifelong learning rather than fixed, brittle training cycles.

A liquid neural network can be modeled as:

h(t)= f (h (t), x(t), θ)

where:

• h(t) represents the evolving hidden state over time,

• x(t) is the input signal,

• θ represents the trainable parameters,

• and f is the governing function that determines how the system evolves.

Unlike conventional neural networks that require periodic retraining, liquid networks adapt continuously, adjusting their weights dynamically rather than in discrete steps. This allows them to learn and generalize to new environments without suffering from catastrophic forgetting—a fundamental limitation of traditional AI.

The term “catastrophic forgetting” was first coined by Robert M. French (1999), but the problem has plagued neural networks since at least the 1980s, when researchers discovered that training a model on new data could completely overwrite previously learned information. Unlike the human brain—where learning is cumulative and integrates new knowledge without erasing old memories—artificial networks struggle to retain past knowledge while acquiring new insights. This limitation marks a fundamental distinction between natural and artificial intelligence, frustrating efforts to achieve continuous learning in modern AI.

By embracing continuous adaptation, liquid networks offer one promising pathway toward AI systems that do not just process static knowledge but evolve their understanding in real time—a necessary step toward true intelligence augmentation. The relative simplicity of the mathematical model behind liquid neural networks suggests that other, as-yet-undiscovered modifications to traditional artificial neural networks may emerge—if research is pointed in the right direction.

From Prediction to Discovery: AI and Hypothesis (second key area)

Another promising avenue is predictive processing, a theory rooted in neuroscience and cognitive science that suggests the brain actively predicts and updates its internal world model through a continuous feedback loop. Instead of passively processing sensory input, the brain generates expectations, tests them against reality, and refines them dynamically.

This framework, pioneered by Karl Friston through his Free Energy Principle, has been further explored by Andy Clark in cognitive science (Surfing Uncertainty, 2016) and shares conceptual similarities with Jeff Hawkins’ Hierarchical Temporal Memory (HTM) theory (On Intelligence, 2004). If AI could be structured around a similar predictive learning paradigm, it would shift from passive pattern recognition to active hypothesis generation—a fundamental step toward co-adaptive intelligence.

The goal of full hypothesis generation remains distant in AI research—and pursuing it risks repeating the same flawed approach that has defined AI for decades: scaling statistical curve-fitting architectures, such as neural networks, in the hope that intelligence will emerge. But intelligence is not just a function of scale; it is embedded in contextual, goal-driven, and networked processes.

Rather than treating predictive processing as a pathway to autonomous scientific discovery, its real value may lie in dynamically scaffolding human theorizing—helping researchers refine, test, and expand hypotheses by functioning as an interactive cognitive tool embedded within larger knowledge systems.

To move beyond passive pattern recognition, we must develop architectures that reframe AI as an evolving participant in knowledge generation by enabling:

• Proactive Exploration – AI should suggest promising lines of inquiry, not just based on static datasets but through continuous interaction with evolving human-driven research.

• Dynamic Hypothesis Refinement – AI should iteratively update inferred hypotheses as new information arrives, preserving contextual continuity.

• Embedded Knowledge Systems – AI should function within structured feedback loops, integrating human intuition, machine inference, and broader information networks to make open-ended inquiry more productive.

This vision requires moving beyond individual AI models and toward a systemic view of intelligence—one that recognizes cognition as inherently distributed across human-machine systems and information networks.

Rather than chasing AGI as an artificial replica of the human mind, we should be designing adaptive, multi-agent knowledge architectures that enhance collective intelligence. In this framework:

• AI does not merely assist human cognition; it co-evolves with it, continuously refining and expanding the scope of inquiry.

• Intelligence is not fixed but dynamically structured within human-machine networks, making knowledge generation more fluid, accessible, and expansive.

• AI is not a passive tool but an active participant in scientific discovery, innovation, and creativity.

The goal is not to create machines that compete with or surpass human cognition, but to systematically integrate AI into the production and use of knowledge in ways that make intelligence more adaptive, more effective, and more generative than ever before. This is the foundation of The New Cybernetics—a paradigm where intelligence is not an isolated function but a living, evolving system embedded in human-machine collaboration.

AI as a Catalyst, Not an Oracle (third key area)

AI is often framed as an oracle—a system that, when given the right input, can deliver a correct answer. But this is a narrow and deeply constrained conception of intelligence. The real promise of AI lies elsewhere—not in mere extrapolation, but in cognitive expansion.

Yet the entire history of AI—with rare exceptions—has been dedicated to the pursuit of stand-alone intelligence. The sci-fi vision of AI was, in some sense, inevitable. The fascination with artificially intelligent life seems deeply ingrained in the human psyche. But in chasing this dream, we have overlooked a more consequential question:

· What purpose should AI serve?

Technology is never neutral—it is always put to a purpose. Shovels and tractors multiply the power and reach of our muscles. AI should do the same for our minds. In retrospect, the “AGI” idea was never particularly well thought out, and the notion of a self-replicating “superintelligence” even more whimsical. Yet the power of modern AI now demands a serious scientific framework—not for replacing human intelligence, but for augmenting it in structured, networked ways.

Instead of functioning as an oracle of past knowledge, AI should serve as a catalyst for idea generation, enabling the generative leaps that allow humans to form new hypotheses, reframe problems, and push the boundaries of inquiry. This is The New Cybernetics: a shift in focus from artificial minds to symbiotic intelligence systems.

The future does not lie in stand-alone superintelligence. It lies in human-machine intelligence networks—a new paradigm in which AI expands the reach of human thought, making knowledge more accessible, more effective, and more generative than ever before.

We’ve already seen glimpses of this potential, but the tools remain crude. Current models are limited in obvious ways:

• They lack continuity, failing to retain a true working memory of an evolving inquiry.

• They have no meta-awareness of where a thought process is headed.

• They can complete, but they cannot initiate insight.

What’s missing is an architecture for structured knowledge resonance—a framework in which human and machine intelligence actively refine and expand each other. Some, notably Wharton’s Ethan Mollick in his 2024 book Co-Intelligence: Living and Working with AI, have described this as “alien intelligence”—an AI that does not merely mimic human reasoning but thinks alongside us in non-human yet profoundly useful ways.

The challenge of The New Cybernetics is to move AI beyond static prediction and classification and toward a dynamic, networked interplay with human cognition. Instead of training models solely to predict, classify, or generate, we need systems that unleash intelligence in symbiosis—not by replacing human cognition, but by extending its reach, unlocking intellectual pathways that would otherwise remain inaccessible.

Intelligence—even “co-intelligence”—does not exist in isolation; it is fundamentally networked. To realize this vision, we must turn to the deeper question of intelligence networks themselves:

The Intelligence of Networks

Intelligence does not emerge in isolation; it is a function of interaction. Just as human cognition is shaped by networks of discourse, scientific collaboration, and institutional knowledge, AI must be embedded in systems that enhance collective intelligence rather than concentrating knowledge in static, centralized hubs. The question is not simply how to make AI smarter, but how to make intelligence itself more available—more fluid, more generative, more symbiotic.

At the Santa Fe Institute, I aim to investigate how decentralized, dynamically updated human-machine networks can function as engines of discovery. Intelligence, when properly structured, does not merely solve problems—it plumbs mysteries—those big, unresolved questions that require novel perspectives, unexpected insights, and cross-disciplinary thinking. The networks of the future will not be built from isolated minds—whether human or machine—but from paired intelligences, organic and synthetic, working in tandem across dynamic systems to tackle the most complex challenges of our time.

The distinction between puzzles and mysteries was first made by Gregory Treverton, a national security expert, and later popularized by economists John Kay and Mervyn King in their 2020 book Radical Uncertainty: Decision-Making for an Unknowable Future. The distinction fits nicely with my purpose here: while AI is adept at solving puzzles—structured problems with clear rules and solutions—the real world where intelligence truly matters is awash in mysteries, where the essential question is not “How do we solve this?” but “What’s going on here?”

This idea has deep roots. Horst Rittel and Melvin Webber made a similar distinction in their 1973 paper, “Dilemmas in a General Theory of Planning,” where they famously described such challenges as wicked problems—problems that lack clear definitions, definitive solutions, or stopping points, requiring continuous adaptation and systemic understanding. Regardless of the terminology, the future of intelligence lies in human-machine collaboration, where AI is not just a tool for answering predefined questions but a catalyst for uncovering the questions we should be asking. These symbiotic intelligence networks will operate through adaptive feedback loops, allowing humans and machines to co-evolve their understanding in real time—a far cry from the static, isolated AI models of today.

The Deep History of Intelligence Networks

This is not a new problem. Network theory has a rich pedigree, stretching back to Pál Erdős, whose pioneering work in graph theory and combinatorics laid the foundation for how we model connections between entities. The success of the World Wide Web in the 2000s revived interest in the power of networks, with Albert-László Barabási’s Linked: The New Science of Networks (2002) reshaping our understanding of scale-free networks and the emergence of hubs in complex systems. His work, along with Duncan Watts’ Six Degrees (2003), profoundly shaped modern network theory, especially in how knowledge, influence, and innovation propagate through interconnected systems.

At a more applied level, James Surowiecki’s The Wisdom of Crowds (2005) explored the conditions under which collective intelligence flourishes, but the concept predates modern computation—stretching back to the origins of human coordination. The earliest distributed intelligence networks were formed by hunter-gatherer societies, where survival depended not on individual genius but on the emergent intelligence of the group.

Some of the most substantive studies of intelligence-in-network come from unexpected places. Edwin Hutchins, in Cognition in the Wild (1995), provides a brilliant case study of how Navy navigation teams function as distributed cognitive systems—showing how intelligence is not localized in any one navigator but emerges through interactions between tools, maps, protocols, and human expertise.

Similarly, studies of air traffic control, emergency response coordination, and even jazz improvisation reveal the deep principles of real-time collective cognition. These systems work not because of a single omniscient actor, but because they structure feedback loops, partial information sharing, and dynamic role adaptation—all features that today’s AI fundamentally lacks.

A New Era of Intelligence

As I write this, DeepSeek has temporarily upended the tech world by producing an open-source, high-performing language model (LLM) that required a fraction of the computational expense of Silicon Valley’s GPT series. The science of AI continues to demonstrate two key points discussed above.

First, we are refining existing ideas rather than making fundamental breakthroughs, consolidating our collective intelligence toward laudable but incremental progress rather than groundbreaking discovery. Second, the churn inherent in the new age of AI means that all cards are on the table—the landscape is shifting, and a new era is ripe to begin.

This proposal captures the ethos of our time and lays out an exciting research direction—one that could fundamentally reshape how we think about AI and, perhaps more importantly, how we see and value ourselves.

As a computer scientist who has witnessed the highs and lows of this quixotic field for more than two decades, I am both excited and humbled by where we stand today. But more than anything, I am ready to take the next step—together.

Erik J. Larson