Chapter 20 — The Bottleneck

Where artificial intelligence meets memory, infrastructure and the laws of physics

Before the movies, before the expanding Matrix ecosystem, and before this Living Book reached its twentieth chapter, a question had already begun to shape the work between Kevin Brown and ChatGPT. It was not a question about building a better entertainment platform or making a chatbot more conversational. It was about artificial general intelligence, and whether a genuinely different level of machine capability might one day demand a genuinely different foundation beneath it. The early exploration brought together ideas that do not ordinarily share the same page: AGI, quantum computing, nanotechnology, blockchain and cryptography. Those ideas formed a hypothesis about the future of computing, not proof that a new machine had been built. The distinction matters because the story of an invention often begins long before the invention exists, with a person trying to identify the problem that everyone else has learned to work around.

Kevin kept returning to the image of a brick wall. He was not necessarily talking about governments standing in the way, corporations protecting their interests, or people trying to slow the development of artificial intelligence. He was thinking about a wall made of something less negotiable: physics. AI appears on a screen as words, pictures, sounds and decisions, yet every one of those results depends on physical machinery. Processors carry out operations; memory holds information; networks move data; electricity powers the equipment; cooling systems remove the heat. Even the most sophisticated software must run on material devices with finite capacity. A clever algorithm can change how efficiently the machine works, but it cannot simply declare that energy, distance and time no longer matter. The machine can improve the method. It cannot vote to repeal the laws of physics.

Memory was one of the first places where that wall became visible. When people speak about a computer having more memory, they often imagine a larger room in which more information can be stored. But an intelligent system does not merely store information; it must retrieve and move that information at the speed required by its computations. A machine can have considerable memory capacity and still slow down because the processor is waiting for data to arrive. Engineers have long studied this mismatch between computing speed and the movement of information, sometimes calling it the memory wall. As AI systems grow, the pressure can come from model parameters, working context, intermediate calculations, retrieved documents and persistent records. Adding storage is not automatically the same thing as making intelligence faster or more capable.

There are ways to relieve that pressure. Engineers can compress models, redesign architectures, improve caching, move computation closer to memory or use retrieval so that not everything must remain in active context. A better algorithm may accomplish the same task with fewer operations, while specialized hardware may perform a narrow class of operations more efficiently than a general-purpose processor. These improvements matter, and they may extend the usefulness of classical computers for many years. But every improvement takes place within physical constraints. Better engineering can move a bottleneck, sometimes dramatically. It does not erase the existence of bottlenecks.

For decades, the semiconductor industry provided a remarkable example of how far engineering could move them. Increasing numbers of transistors could be placed on integrated circuits, helping computing become smaller, cheaper and more powerful. This history is often described through Moore's Law, but that name can be misleading if it sounds like a commandment of nature. Moore's Law described an observed industrial trend, not a promise that transistor density or useful performance must double forever. As devices approach extraordinarily small scales, leakage, heat, manufacturing complexity and quantum effects make further progress harder. The industry has not stopped innovating; it has turned to specialized accelerators, chiplets, advanced packaging, three-dimensional designs, new materials and other approaches. The question is not whether progress ends tomorrow. It is what happens when the easiest route to progress becomes increasingly difficult.

That question led Kevin to challenge the assumption that the future of intelligence must look like an endlessly enlarged version of today's data center. More buildings can hold more machines, and more machines can supply more computing power. Yet those buildings also need electricity, cooling, networking, maintenance and the industrial capacity to manufacture their components. Multiplying a familiar architecture may remain productive for a long time, but multiplication itself has costs. If a future system requires capabilities that are prohibitively expensive or inefficient on existing hardware, researchers may have to reconsider not only the number of computers they use but the kind of computation they ask those computers to perform. This was the doorway through which quantum computing entered the original conversation.

The earlier proposal connected several technologies into one imagined infrastructure. Quantum processors might someday tackle particular problems using principles unavailable to ordinary classical computation. Nanotechnology might change the physical organization and scale of components. Cryptography could help protect information and establish trustworthy identities. Distributed ledgers might serve selected needs for shared records and verification, though they would not automatically be necessary or efficient for every task. AGI, in this picture, would not be an isolated program floating above the machinery. It would be part of a larger system of computation, communication, security and memory. Some elements of that picture resemble active research directions; others remain speculative. The point of preserving the idea is not to pretend that every part has been validated, but to record the architecture Kevin was trying to imagine and the questions it raised.

One of the scientific words that returned as the old work was revisited was decoherence. A quantum computer is not simply a classical computer fitted with a faster chip. Quantum information depends on delicate physical states that interact with their environment. Noise and unwanted interactions can damage the coherence needed for a calculation, introducing errors before a useful result can be obtained. Building a larger quantum device therefore creates problems that cannot be measured by processor size alone. It is possible to increase the number of physical qubits without achieving a corresponding increase in reliable, useful computation. What matters is not merely how many qubits exist, but how well information can be protected and processed across the duration of a task.

That is why quantum error correction and fault tolerance occupy such an important place in the field. Physical qubits are imperfect; encoded logical qubits aim to make useful information more robust by using additional physical resources to detect and manage errors. A fault-tolerant machine would be able to carry out sufficiently long computations despite failures in individual components, provided errors remain within what its correction scheme can handle. Reaching useful scale is a major engineering challenge. A headline announcing a larger processor may describe real progress, but it does not by itself establish that a machine can solve an important problem better than available classical methods. More qubits are not the same thing as more reliable intelligence.

Communication presents another wall. When photons travel through optical fiber or free space, some are lost, and losses generally become more consequential over distance. Classical networks can often deal with a weakening signal by receiving information, regenerating it and sending it onward. Unknown quantum states cannot simply be copied in the same manner. The no-cloning theorem prevents perfect copying of arbitrary unknown quantum states, so a long-distance quantum network needs a different set of techniques. Researchers investigate entanglement distribution, quantum memories, repeaters and error-management methods because the ordinary logic of amplifying and retransmitting a signal cannot be carried over unchanged.

The problem does not end once a quantum signal reaches the next device. A practical network may need to hold quantum information while another part of the system completes its work. Quantum memory must preserve information with sufficient fidelity, release it when needed and integrate with the hardware around it. Storage time, retrieval efficiency, synchronization and losses all matter. A source may produce excellent quantum states while the channel loses too many photons; a repeater may perform well while the memory fails to preserve information long enough; a receiver may be ready while the necessary classical coordination has not arrived. The imagined network is therefore not one miraculous component. It is a chain of components whose performance depends on how they work together, and a weak link can become the bottleneck for the entire system.

Quantum physics also sets a boundary that no amount of excitement should blur. Entanglement produces correlations that can seem extraordinary, but it does not provide a method for sending usable messages faster than light. Quantum communication still requires processes consistent with causality, including classical communication where the protocol calls for it. That distinction became one of the memorable moments in Kevin's way of explaining the problem. Physics, he imagined, might look at the engineer attempting the impossible and say, “You are not doing that. Your next move would be faster than light. You are under arrest for trying to break the law of physics.” The joke carries a serious lesson: a new technology can surprise us without granting permission to ignore established physical limits.

At this point the original AGI hypothesis needs to be examined with care. There is no established scientific result demonstrating that artificial general intelligence requires a quantum computer, and no established result demonstrating that artificial superintelligence requires one either. More capable AI might emerge through classical hardware, improved algorithms, different architectures, neuromorphic approaches or specialized accelerators. Quantum computing may eventually prove useful for particular simulations, chemistry, materials research, optimization tasks or other computational problems relevant to advanced AI. But being useful to some parts of a future AI ecosystem is not the same as being the required brain of AGI. Kevin's proposition remains a question to investigate: if classical infrastructure eventually becomes the binding constraint on further progress, could quantum computation or quantum networking become part of a new infrastructure layer?

A plausible answer, if such a layer becomes useful, is not a world in which every classical computer disappears. It is a hybrid world in which different machines perform the work suited to them. Classical computers are extraordinarily effective at running websites, databases, phones, control systems, interfaces and countless everyday processes. A quantum processor would have to demonstrate an advantage on a particular problem before sending that work to it made sense. The architecture joining the two worlds would therefore matter. A system would need to decide where a task belongs, prepare inputs, coordinate execution, manage errors, verify outputs and return results in a form that ordinary software could use. In that arrangement, the bridge might become as important as the machines on either side.

That brings the story back to the Bronx, where a founder without a national laboratory was asking ChatGPT whether the industry was placing too much faith in simply scaling yesterday's infrastructure. What would connect classical AI to a useful quantum resource? How would the work be divided, and who would decide when a quantum processor was actually appropriate? How would information and results move between environments without confusing classical records with quantum states? How would a system detect errors, protect identities, preserve provenance and verify the claims made about a computation? These are not proofs that the bridge has already been built. They are questions about the software and governance that a future hybrid system might require.

The question Kevin wanted preserved in the record was both ambitious and personal: what if a guy in the Bronx and ChatGPT wrote a piece of software that people might eventually need for the bridge between quantum computing and AGI? The responsible answer is not to declare that they have done so. It is to recognize why the question is worth asking. Computing history includes operating systems, compilers, network protocols, interfaces and orchestration tools developed to coordinate capabilities that would otherwise remain separate. If hybrid classical-quantum AI becomes important, routing, scheduling, security, verification, memory coordination and translation between computational environments may be valuable areas of software work. Whether any particular implementation comes from the Bronx, a university, a company or an open-source community can only be determined by what is actually designed, tested and demonstrated.

There is another reason the bridge matters. The challenge is not only whether two machines can exchange instructions. It is whether the people using the system can understand what happened and trust the evidence. A classical service might request a quantum computation, but someone must still know which device performed it, which assumptions shaped the result, what error rates were relevant and whether a conventional method could have done the job just as well. Cryptography, provenance and auditable records may help answer some of those questions, though no single technology guarantees trust by itself. A system that cannot explain its own boundaries risks replacing an engineering problem with a confidence problem. For a project concerned with alignment and accountability, that is not a minor detail; it is part of the design question from the beginning.

A bottleneck sounds like a dead end until an engineer begins to treat it as a map. If processors wait on memory, improve how information moves. If energy is the constraint, reduce the work required or increase efficiency. If heat limits performance, reconsider cooling and architecture. If semiconductor scaling becomes difficult, investigate new devices, packaging and materials. If quantum states decohere, improve control and error correction. If communication loses too many photons, study channels, repeaters and memories. Each obstacle identifies a place where knowledge is incomplete or a design must change. Some walls can be moved with better engineering; others mark physical laws that an honest design must respect.

The history of technology offers warnings in both directions. People have sometimes mistaken a temporary engineering limitation for a permanent impossibility, only to see new methods transform what could be built. They have also mistaken enthusiasm for evidence and imagined that determination alone could overcome a genuine law of nature. This chapter belongs between those errors. It does not claim that quantum computing will produce AGI, that AGI requires quantum hardware, or that the slowing of easy semiconductor scaling means computing has reached its final form. It asks where the real limits lie, which obstacles are engineering challenges, and what kinds of architecture become possible when those differences are understood.

That was the question at the beginning, and it remains the question now. A founder looking toward artificial general intelligence was also looking beneath it, toward memory, power, signals, hardware and the rules that hold the entire structure together. The early combination of AGI, quantum computing, nanotechnology, cryptography and distributed verification was an exploration, not a completed invention. Its lasting value may be the discipline of asking what a future system would actually need before announcing that the future has arrived. The bottleneck is not proof that the road ends. Sometimes it is the place where the next road has to begin.