Key principles of intelligence

Context

The AI world and the computing industry are surging with historic growth. Most of today's models are based on neural network architectures that leverage the attention mechanisms of transformer blocks, and they are executed almost entirely on GPUs. We refer to this current paradigm as the LLM-transformer-GPU paradigm (LLM-TGP), and it is one of the fastest-moving, most consequential moments in the history of technology. Many are asking how new capabilities based on LLMs can be used to handle existing workloads, drive new capabilities that weren't possible previously, increase productivity, and lead to economic gain. These are important questions with many fruitful answers. But there are deeper questions that will prove far more consequential in the long term. In several years, when the sand has settled from this first AI explosion, the only enduring advantage one company or nation could retain over another is through the highest-performing intelligent system. The capabilities of advanced intelligence will lead to superior performance in all other domains of society — from economic productivity and national security to healthcare and quality of life. If there is a more powerful paradigm of intelligence, we must seek it relentlessly, as it will shape the future economy and geopolitical reality.

Despite the tremendous performance of LLM-TGP, there are strong reasons to expect LLMs are not a complete method for general intelligence. The power of the transformer architecture can be matched by other approaches that provide better scaling. GPUs have limited capabilities for many aspects of intelligence related to dynamics and learning.

To design and construct systems for high-performance artificial intelligence, it is advantageous to begin by identifying the key principles of intelligence and then determine how to manifest those principles in technological systems. Key principles of intelligence can be identified first by defining an intelligent system and specifying that it is physically instantiated. With a few subsequent assertions, one can motivate several additional properties of the primitive structures of intelligence and their interactions. Based on this very high-level analysis, it appears that several of the attributes of today's AI are well-aligned with the principles of intelligence, but some are not, and those discrepancies lead to the challenges and roadblocks of conventional AI, including high energy consumption and functional shortcomings like poor scaling with multimodal intelligence and lack of continual learning.

Principle 1: Definition of Intelligence

To identify the key principles of intelligence, we must first define what we mean by an intelligent system. An intelligent system is a model of the world that informs decisions and guides actions. The model may be of a small subset of the world or the entire cosmic edifice. When speaking of general intelligence or beyond to superintelligence, the scope of the model becomes ever more inclusive. The decisions the model informs may include a small number of inputs and possible output states, or they may involve an enormous state space. Again, the more general and powerful the intelligence the larger the number of factors relevant to the decision. The actions being guided by the intelligent systems may be as simple as a single binary choice or they may be quite elaborate, across a range of output channels, actuators, or relationships. The central role of a "world model" in this definition of intelligence aligns closely with perspectives of Yann LeCun and Demis Hassabis. Many subsequent conclusions follow from this general definition.

The Conceptual Model and the Physical Instantiation

At the core of an intelligent system is a model of the world. In what follows, it is helpful to disambiguate between the conceptual model that the intelligent system aspires to implement and the physical instantiation of that model, which may be an imperfect approximation. The conceptual model and its physical instantiation are inextricably intertwined and must be arrived at self-consistently. First, the conceptual model is a model in the sense that it is not the actual world — especially not the entire world — but rather a representation of the world. It is necessarily a reduced representation of the world, meaning it contains less information than the full world, because otherwise it would simply be the world and not a model. The reason it is reduced is so that it is physically simpler (smaller, less massive, portable, consuming less power, etc.). The challenge of designing a conceptual model of the world is to include those features that are the most important about the world so it can be useful while omitting that which is less relevant so the model can be simpler and therefore more easily utilized.

Whatever the model is, it must be cast in physical form if it is to be accessed and utilized. Casting it into physical form subjects it to constraints. If there were no physical constraints, it would just be a replica of the subset of the world with which it is concerned — again, the world and not a model of the world. The trade space of model fidelity and physical limitations presents the challenge of engineering intelligence. Challenges related to physically implementing the conceptual model lead to the engineering tasks of building intelligent systems.

An Example

An example will help clarify what is meant by the conceptual model, the physical model, and their correspondence. Consider a subset of the world that consists of a signal that varies in time. At every moment down to the Planck time, the signal is specified as a single number between negative one and positive one with resolution limited by Heisenberg uncertainty. Imagine the goal of our intelligent system is to predict one unit of time ahead whether the signal will be greater than or equal to zero. In this example, "the world" refers to the time-varying signal, and the quality of the itelligent system is measured based on the fraction of the time its prediction is correct about the amplitude of the signal being greater than or equal to zero.

Consider three approaches to building the intelligent system. First, imagine that we observed the signal for a while, and we notice it tends to go up and down a lot: it is highly oscillatory. Fourier taught us that we can represent any signal as a sum over sine waves,

The function \(f(t)\) is represented for all times \(t\) as a sum of known functions. This can provide a highly accurate model of the world, but the accuracy of the model depends on the number of terms we retain in this series. As one approach to building an intelligent model of this system, we could implement a number \(N\) of sinusoidal signal sources, and combine them with amplitude coefficients \(c_n\). This is a physical system, so the number of sources is a constrained resource, and we face a tradeoff between model fidelity and cost, weight, power, etc.

In the Fourier expansion, each subsequent term has a higher frequency, and therefore is more likely to be omitted without affecting accuracy if activity is predominantly within a certain timescale. A related representation is that of a neural network. We can represent the same signal as a sum of neuron activations:

where \(w_n\) are the weights and \(\sigma_n(t')\) are the nonlinear activations that vary with time, as a function of inputs from the system being modeled. The neurons observed the system up to time \(t'\) to make a prediction about the system at time \(t\). In this case, parameters of the basis are typically learned, not fixed and ordered like Fourier terms. Neural basis functions can represent any function, and the quality of the approximation typically grows with the number of neurons.

As a third approach, we know from digital systems that we can represent this time-varying signal as a list of discrete values at discrete times. This must also be implemented as a physical system, in which case the discrete values must be chosen from a defined set of possible values — the bit depth with which we represent the signal. And the density of time points must be specified — the temporal resolution. Again, we face a tradeoff. Higher bit depth and higher temporal resolution give the model better fidelity, but the apparatus used to achieve this will grow in weight, power consumption, and cost.

So which approach is better — analog or digital, neurons or sine waves, binary logic or physical signals? It depends on the nature of the signal and what you want to get from the model. It also depends on the decisions you want to inform, the actions you want to guide, and the specifics of the physical hardware you will use to construct the system. If the signal happens to be pretty well represented with just a few terms in the Fourier series, and those terms are at frequencies that are straightforward for your physical devices to produce, the analog, time-continuous approach may be the way to go. If instead the signal requires many frequency terms, you may find it is more effective to represent it as a digital approximation with discretization in amplitude and time. The right physical system depends on the model being implemented, the performance requirements, and the constraints of hardware and economic resources.

Principle 2: Conjecture That the Hardware Should Match the Model

The primary conjecture that leads to subsequent principles is that the performance/intelligence of the system is maximized when the physical system is as close as possible to the conceptual model. This raises the question of what it means for a system to physically instantiate a conceptual model. A system is a good physical instantiation if at all times there is a location in 3D space with a physical quantity representing each element of the conceptual model. To be physically instantiated means the hardware is the model; the circuit is the algorithm. There is no intermediate mapping to an intermediate information format, as is the case with digital. All computational systems are physical systems, but for an efficient, high-performance intelligent system, directly representing the quantities of interest is most parsimonious and performant.

In the example of the signal, either the Fourier, the neural, or the digital approach can be a physical instantiation, but the digital approach typically introduces a layer of conceptual abstraction. The Fourier approach naturally admits a signal — let's say an electromagnetic field at a selected position in space — that is the superposition of sources at distinct frequencies with specified amplitudes, and the neural system also admits a direct interpretation where the full signal is the sum of multiple contributions, faithfully representing terms in the expansions from the equations above. At that point in space, the relevant quantity of the model is represented by a physical quantity in space, and we can straightforwardly envision extending the representation to multiple locations in 3D space should the model require such added sophistication. The series may be truncated at a certain number of terms, so the physical state is an approximation to the conceptual model, but it is still a physically instantiated representation.

With the digital representation — assuming a binary representation on transistors, as is ubiquitous in today's computers — the representation is spread across several transistors. The physical quantities carrying the information are amounts of charge on capacitors — electrons at the gates of transistors. The number of transistors required to represent the number at each time is specified by the desired bit depth. There is not a physical-model correspondence between the charge on the capacitors and the signal in the model. To reconstruct the signal of interest from the physical system, a conversion is required. The representation of information is immediately mapped to binary, and all subsequent operations are performed as arithmetic. At any step along the way, quantities representing the model are intermediate numerical quantities with no direct physical correlation to the subject of the representation. You only get the meaningful quantities at certain computational steps. Even if the digital system is being used to represent the Fourier or neural network model, the physical quantities on the transistors don't immediately correspond to the parameters in the model. However, an intermediate binary representation is often preferable because computations performed on the information can be carried out with no impact from noise and very low errors even on long computational sequences. This is made possible by the excellent match of the physical substrate — silicon transistors — with the conceptual model — in this case, binary arithmetic and Boolean logic.

And yet, the primary conjecture we put forth is that for an intelligent system to achieve maximum performance, it is important that the hardware be a physical instantiation of the conceptual model. This configuration avoids information conversions and allows time to represent itself, leading to higher speed, efficiency, and capabilities of learning — especially at very large scale. The claim is not that dynamical systems make better computers in all cases. That's evidently not the case. As explained below, for modeling a dynamical world, substrate-model match wins; for high-arithmetic-depth symbolic work, abstraction wins. Performant intelligent systems will leverage both, just as we think with brains — which are matched to the conceptual model of intelligence — while leveraging digital computers as a resource. But in the world of computing, digital has grown explosively over the last 80 years, while hardware that physically instantiates the conceptual models of intelligence has not kept pace. That must change for intelligent hardware to reach its full potential.

This principle is closely related to Sara Hooker's hardware lottery, in which available hardware and software — and efficient hardware-software combinations — dictate which technologies are developed. As Hooker argues, ignoring hardware profoundly limits the types of models and AI concepts that get developed, and the most fruitful path to qualitatively new capabilities requires hardware-software-algorithm co-design. We must consider all aspects of the stack as we design the physical substrate to implement the conceptual model. We must identify the principles of intelligence we aim to capture and then think from a clean slate about the devices best equipped to embody those principles. As Great Sky pioneers a new kind of hardware informed by these principles, a fresh hardware lottery is initiated. With our superconducting optoelectronic networks, new algorithms and architectures will emerge to maximally leverage these new capabilities, even if those same algorithms and architectures are ill-suited for execution on conventional digital processors.

Analog Versus Digital

The analog-digital decision is at the core of foundational decisions about hardware for AI. Although many consider the debate settled in unambiguous favor of the digital approach, to stay entrenched in that perspective risks missing the next great opportunity in computing. The principle above stated that the hardware should match the model. Silicon transistors are an excellent physical instantiation of a binary model, but nowhere in neural networks, the transformer, and any other aspect of the dominant paradigm does the conceptual model demand binary representation. Today's use of binary/digital emerges from history, so let's revisit the sequence.

As computing took shape in the 1940s following WWII, the primary applications were numerical computation: calculating trajectories, simulating weather, solving differential equations — applications that involve long arithmetic sequences. The low noise and high accuracy of binary representation were decisive advantages in digital over analog. The invention of transistors in 1947, the integrated circuit in 1958-1959, and the subsequent development of silicon microelectronics wrote the rest of the story. MOSFETs are incredible for representing zeros and ones, and zeros and ones are excellent for performing arithmetic operations through sequences of Boolean logic operations. Binary is also an excellent conceptual medium for high-arithmetic-depth computations. This goes back to Leibniz and ancient China before him. You could still argue for analog in 1966, but in 2026 it usually goes unquestioned that binary digital logic won, and silicon integrated circuits are the medium (at least transistors, perhaps also photonic transceivers at the edge of the package).

Notice in this narrative that when the conceptual workload shifted from differential equations and other logical operations that make for excellent software applications to the pursuit of general intelligence, there was not a pause to reconsider whether the same hardware was ideal — and there was no need for such a pause. Moore's path presented itself to ever more computing performance, and innovations in that lineage continue to drive appreciable performance improvements. This hardware got us to LLM-TGP, and it's working pretty well. You don't switch tracks when the train is barreling this fast.

Many analog waves have built, crested, and evanesced, so to try again, something must be different. Many have used CMOS in the analog domain, but scaling has proven very difficult, due primarily to transistor threshold mismatches and challenges related to co-location of dynamic memory with computational circuitry. The mismatch problem is particularly acute with transistors, because they have an exponential sensitivity in the analog domain to parameter variations. Memristors have been pursued to address this memory need but have struggled to scale due to device variation and material fatigue with repeated use. If one is to preach "analog" and "dynamical" at this stage in the evolution of technology, they ought to have a different approach that avoids these pitfalls. A pillar of Great Sky's technology is that superconducting circuits are used for analog dynamics. These devices are linearly sensitive to parameter variations, not exponentially; they readily co-locate plastic, non-volatile memory with computational circuits; and the form of memory is easily re-writable, non-volatile, and induces no material degradation or fatigue. We've written more about the hardware motivations in IEEE's The Bridge magazine. There is another path to analog, dynamical hardware. Silicon transistors are the right physical instantiation for binary, but that doesn't mean we should expect them to carry us through the representations that give rise to general intelligence.

Dipping back to the 40s again, there has just as long been an impulse to understand intelligence through circuit instantiations of neurologically motivated primitives. The McCulloch-Pitts model assumes many inputs and outputs, analog signal values within the computations of neurons and synapses, and a binary threshold output that goes to many destinations. In the eighty years since, our understanding of the brain has grown far richer, and one can argue that our intelligence emerges not from a long sequence of high-accuracy arithmetic operations carried out between neighboring gates, but from many operations in parallel drawing information rapidly from across a huge network through many long-range interactions. Binary with transistors remains indisputably superior in the appropriate domain, and performs better for AI at the moment due to technological maturity. Yet the neural model from biological intelligence is still a better guide once the purpose shifts to general intelligence — the ambition first pursued by McCulloch and Pitts. We now know far more about the conceptual model of the brain, and we have new hardware to make it.

This is not simply a "brains are smarter than digital" manifesto. The point is that brains appear to follow these principles, while the digital approach departs from the principles in certain ways, so the principles and the paradigm can't both be optimal for general intelligence.

Principle 3: The Model for General Intelligence Should Be Dynamical

One could argue identifying these first two principles isn't helpful, because the art of designing an intelligent system falls upon the second principle: knowing which approximations and representations make good models of the world. But this is progress. Now to make something intelligent we just have to look at the structure of the world and find efficient, condensed representations. To be concrete, let's recall that our purpose is to devise the most intelligent system we possibly can — a system that is generally intelligent and, a big step further, superintelligent. If we don't aim to represent a single, time-varying signal as in the example above, but rather as much of the messy world as possible, it is harder to know where to begin when conceiving of the conceptual model and then choosing a physical instantiation. What we can say about a model capable of general intelligence is that it must be capable of representing many interacting elements that change across a broad range of timescales. The world is made of minute excitations that influence each other across space and time. From this specification, we can determine important features of the model as well as its physical instantiation.

The world is a highly complex interacting system with many elements that come into and fade out of existence. The world is constantly changing, and many of the things we would like an intelligent system to do are related to predicting the consequences of changes over time. The conceptual model underlying general intelligence ought to be based on primitive elements that are transient excitations that persist and adapt across a wide range of time scales, arranged in highly interconnected networks to capture complex interactions. This is not a requirement for intelligence, but rather a guideline for performant physical instantiation of an intelligent model — assessed by the quality of insight and guidance as well as the time and energy required to produce the information. To achieve this, the model elements should be basis functions representing the temporally interacting system — little flecks and smears across time, excitations driving one another. And the physical activities and interactions must exist across timescales so that temporal aspects of the real world we aim to model are represented by the physical structure instantiating the model. This efficiency argument has been made before; what has changed is the substrate physics, which we take up in a subsequent post.

It is possible to model a dynamical world without dynamical primitives. Many fields of science model systems with PDEs that are stepped through time numerically and output a single solution at the end. The claim here is an intelligent system will perform better, use less energy and time to provide insight and ultimately scale to a greater degree of intelligence if it leverages dynamical primitives to represent the dynamical world. When constructing a generally intelligent model of a complex, dynamical world, a dynamical model in which time represents itself in a system of many temporally varying elements will achieve the best performance. And there must be many of these interacting elements because the world is rich with attributes. This is not a surprise, as the human brain has 86 billion neurons and 100 trillion connections between them. Today's LLMs have a few trillion synaptic parameters, and scaling laws show that increasing the number of parameters improves performance. While the parameters in LLMs have nowhere near the complexity of biological synapses, both cases illustrate the point that increasing the number of interacting elements increases the expressive potential of the model.

The primitives must demonstrate certain behaviors: the ability for inputs to conjure signals; the ability to maintain signals; the ability for signals to leak passively and fade; and the ability to actively erase or reset signals. Those are device level requirements. At the network level, dynamical routing of signals must be possible to allow one structural network to switch between myriad functional networks so that one generally intelligent model of the world can be accessed at many different times to guide decisions or actions related to myriad subsets of the larger model.

The states themselves and the quantities that shape their interactions are forms of memory, and therefore we seek a conceptual model with these different forms of memory. There is no single, most important timescale in the world or in many subsets of the world, so the conceptual model should represent states and interactions across many timescales. These forms of memory are present in the brain, and some forms are present in today's AI models, but these forms of memory represent a departure between AI and principles of intelligence.

Dynamical conceptual models are common in today's AI. Recurrent neural networks, gated recurrent units, long short-term memory, and state space models are all prominent and competitive with transformer models in many applications, especially when the subject being modeled is dynamical, with variation over time being central to the nature of the system. These models are chosen and work well precisely because the conceptual model matches the aspect of the world being modeled. The challenge with current instantiations is that they are constrained to execute on a digital substrate optimized for arithmetic, not for dynamic operation. This partnership often works reasonably well, but the mismatch leads to pain points. For example, with gated units, the complexity of the gates strongly affects the runtime, so models are pushed toward the simplest instantiations. With state space models, more general nonlinear recurrent neural networks are reduced to linear form to conform to the operations of GPUs. The hardware constrains the conceptual model rather than rising to meet the needs of the conceptual model. These dynamical models are competitive with transformer models even though they are constrained by the available hardware, indicating that if hardware better matched to the purpose were available, significant performance improvements could be achieved. This is indeed the case, and Great Sky will have much more to say about this in upcoming posts.

The argument that using dynamical primitives brings greater efficiency should not be interpreted as a modest claim or one with limited practical consequence. By being more efficient (faster while consuming less power, more intelligent with fewer physical resources), qualitatively new opportunities arise. Any Turing machine can produce the same output as any other, but in the real world, time passes and resources are constrained. Capabilities gated by a time constraint are useless if they can't meet that time constraint, so a quantitative speedup past the threshold leads to a qualitatively new form of intelligence. Imagine a fusion reactor in which an instability forms within a microsecond. If one approach can accurately detect and respond within that microsecond while the other reaches the same answer but only after days of numerical computation, they are not interchangeable.

This same situation arises across domains. In a quantum computer, decoding and error correction need to beat the decoherence time, but the compute to do it grows faster than the number of qubits. Or consider weather at the scale we experience it. I was recently caught on a boat when a storm built in minutes with no warning. My weather apps said no rain or wind for four hours. The winds reached 70 mph gusts, the typically calm river had four-foot waves, and the rain came in as pellets that forced me to close my eyes. Just a kilometer south, the scene was far more serene, and if I could have known that, white knuckles and chattering chins could have been avoided. No system today resolves conditions at a granularity of kilometers refreshed every two minutes with enough parallel forecasts to quantify uncertainty in your local kilometer. Today's models are nowhere near the limits of what can be predicted due to chaos, and long before we get there we're capped by computing time and cost. A model architecture with the right physical hardware that was cheap and fast enough to predict the local weather for each of us at once wouldn't forecast slightly better. It would allow us to reach the actual limits of forecasting chaotic systems as complex as the weather, which would change the way we're impacted by natural disasters. Efficiency brings qualitatively new capabilities.

These arguments don't necessarily apply to building models of aspects of the world that are not dynamical. For example, intelligent systems focused on mathematics may or may not be based on dynamics, and the best choice may depend on the type of mathematics under consideration.

Where LLM-TGP Departs from the Principles

These principles guide us toward models and physical instantiations that share many features with biological brains and LLMs, but they are also markedly different from LLM-TGP in several important ways. One of the most important questions about the conceptual model is whether it's dynamical, meaning change over explicit time is central to the model. One of the most important physical considerations is how the immensity of interactions is implemented. The neuroscience perspective guides us toward dynamical activity and direct, physical connections over long distances. The current paradigm guides us toward time-independent representations with computational operations on a clock between nearest neighbors. This approach is currently the only viable method at scale, but it isn't directly answering the question about how far intelligence can go.

The discrepancies arise because the iterative model-hardware loop that has been executed over the past six decades that arrived at LLM-TGP was not conceived for building a general model of the world. Much work has focused more specifically on building a model of language in the form of text. It has been revelatory that a model built on text alone could achieve as general an intelligence as LLMs have demonstrated. Part of this is on the human side, where the user of the model has an unquenchable ability to receive text and conjure much more sophisticated mental images. The challenges at the forefront of the current practice result from departures from what we might expect from principles of intelligence.

Most prominently, with the transformer architecture time is treated entirely differently than it appears to proceed in physical reality. Temporal sequences are fed into transformer models at once as entire vectors. For text, this works well. Text does not comprise a large amount of data, so processing long vectors at once is not problematic. Applying the same methodology to model systems that are inherently temporal is less of a clean fit. It is awkward to decide where to separate sequences of activity to put them into the model. In such cases there is a discrepancy between the conceptual model — the transformer — and the structure of the world the model seeks to represent.

Given the success of transformers alongside the success of the brain, it's natural to ask if a dynamical system can reproduce the function of transformer attention, and can learn without digital backpropagation — which we take up as a fourth principle discussed at the end. I'll be writing more about those subjects in upcoming posts.

The Bitter Lesson

Do these principles adhere to The Bitter Lesson? Paraphrasing Sutton's paper observes that attempts to realize intelligent systems by building knowledge about the world into the model are brittle and difficult to scale or make general. Instead, large networks with minimal inductive biases trained with a lot of data have been far more successful than any other AI approach. Sutton points out that AI researchers often "seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation." We should lean into increased computational power and let the training structure the network rather than stores of facts, rules, or procedures.

The principles stated here align with Sutton's observations. It is unavoidable that some inductive biases must be built into the model due to hardware and economic pragmatism. LLM-TGP models have evolved significant structural specifications due to an interplay between the hardware and the concept, especially as the field moves to multimodal models and agentic interactions. The resulting architectures provide inductive biases about the structure of the world, but they don't arise from a direct injection of expert knowledge. Similarly, the structure of the brain is shaped in part by its input modalities and the decisions/behaviors it must guide. Still, much of the brain is bitter-lesson compatible. Both the neocortex and the hippocampus are large, highly repetitive architectures that can be trained to represent many concepts depending on their inputs. Fine details of the neocortex equip it for massive semantic storage with predictive-coding dynamics, while the hippocampus is predisposed to representing sequences, relationships, and maps. It is reasonable to conjecture that the ultimate cognitive architecture will integrate both types of modules (plus several more), and both of these modules learn the structure of the world through interaction with the world, in accord with the bitter lesson as well as the principles stated here. We should use our knowledge of the brain to design architectures for AI, which is quite a distinct endeavor from using our knowledge of the world to construct an expert system. Beginning with the right conceptual model to serve as a foundation of intelligence and then choosing the right physical substrate to instantiate that model is a path for future AI that receives and embodies Sutton's lesson.

But we still need the right hardware to instantiate such an architecture, and that is what has been left wanting since the perceptron. One of Sutton's most important points is that Moore's law has supported AI through exponential development and the sustained, falling cost of compute during the golden age. Great Sky's approach — superconducting optoelectronic networks — offers a new path for exponential scaling. Moore's law was enabled by lithographic improvements and the miniaturization of transistors. With SOENs, the path is different. But exponential growth is enabled by doubling the number of parameters in a network with each product generation, a feat made possible by the elimination of data movement bottlenecks, the low power density of superconducting circuitry, and the tremendous communication capabilities of dense optical interconnectivity. Diminishing returns of Moore's-law-scaling need not halt progress on AI if we leverage the benefits of this new hardware paradigm. As Sutton says, "One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great." The principles described here — and the further principles that lead us to SOENs — represent an entirely new platform for leveraging this insight to go beyond constraints of the current paradigm.

We'll explore what Great Sky is building in the context of The Bitter Lesson in a future post.

Next Thoughts

Beyond these three principles, a fourth is just as important: an intelligent system must be capable of continual learning through adaptation mechanisms spanning all relevant timescales of the model. We defer it to its own post not because it is peripheral, but because it is where the substrate question grows most subtle — and where Great Sky's approach departs most sharply from both the current paradigm and prior neuromorphic efforts. Building hardware that physically instantiates a dynamical model and learns continually is frontier territory, and it deserves a careful treatment rather than a paragraph.

These principles are a foundation, not a conclusion. The first is a definition, and deliberately spare; what follows depends on getting that definition right. The others are established perspectives, albeit without unanimous acceptance. The purpose of writing out the principles is to explain the rationale for the systems Great Sky is building so critics can point to exactly where the logic is flawed. Perhaps the principles are misguided, or perhaps the truth is more exciting, admitting an exceptional opportunity for a completely different kind of AI that far surpasses what's possible today.

The principles point to an approach to intelligence that is very different from the current paradigm, with a different architecture and hardware instantiation, culminating in a vision for AI that departs in important ways from the dominant paradigm, but to further elaborate on the principles to arrive at designs for intelligent systems, we can find helpful guidance in network theory and neuroscience. To inform hardware choices, we can seek physical limits of devices. These trains of thought have led Great Sky to a specific kind of concept and hardware for AI that follow these principles, but go far beyond. The subsequent decisions we make are: 1) to pursue the most powerful intelligence based on a different conceptual model — dynamical networks; and 2) to construct the conceptual model on a new physical substrate — superconductors and light. These choices require motivations beyond the principles presented here, but these principles are the foundation.

Jeff Shainline July 22nd, 2026