On Uploaded Intelligence

Introduction

Much of the ambition behind artificial intelligence is to make machines more capable. The most successful recent approach has been to teach them through the artifacts of human thought: writing, demonstrations, and judgments about good answers (e.g. RLHF).

But this invites a different (abeit rather creepy) possibility. Instead of learning from the traces human minds leave behind, perhaps we could bring the minds themselves into the digital world.

In the animated series Pantheon, a teenage girl discovers that the stranger she has been messaging online is her late father. An experimental brain scan has allowed his consciousness to run on computers.

Initially, uploading seems like the ultimate extension of learning from humans. Instead of supplying more examples of someone's reasoning, we would reproduce the process that generates them. However, the fundamental objective is different - we would be trying to preserve a particular person, not merely learn from their behavior.

In a world where artificial intelligence eventually surpasses unaided humans across much of intellectual life, could changing the medium in which we think allow people to keep pace? Or would the process produce increasingly capable replicas while leaving the original people behind?

It helps to start with the physical processes behind memory and thought.

Memory Mechanisms

The recent 2026 Nobel Prize in Physiology or Medicine awarded Karl Deisseroth, Peter Hegemann, and Georg Nagel for their discoveries behind optogenetics ^[1], which allows researchers to influence selected cells using light-sensitive proteins. In neuroscience, this makes it possible to intervene in neural activity and investigate its contribution to behavior.

In a 2012 experiment, Liu et al. ^[2] labeled neurons active while mice acquired a fear-associated memory. Later stimulating those neurons with light induced freezing behavior associated with recall.

These experiments demonstrate that physical changes can affect how memories are recalled. Rather than treating memory as an inaccessible property of the mind, we can investigate mechanisms that contribute to it. If the mechanisms could be understood well enough, perhaps their relevant effects could be reproduced elsewhere.

But triggering a memory is not the same as extracting it. In these experiments, the surrounding brain continued to participate in the response. The stimulated cells were part of a collective functioning system, not a self-contained memory file.

Moreover, behavioral mechanisms may extend beyond the brain. In a 2023 experiment, increasing heart rate through optogenetic stimulation heightened anxiety-like behavior in mice in risky contexts. The effect depended on interactions between bodily activity, the brain, and the environment ^[3].

For uploaded intelligence, this suggests that reproducing cognition may require more than copying neural connections. Some bodily signals and sensory interactions may contribute to human behavioral mechanisms.

This suggests that human cognition could be reproduced digitally by preserving the essential physical processes of the brain and body, without replicating every molecular detail.

Human Compression

In a sense, this resembles the idea of compression in computer science, with an unusually difficult definition of success.

A representation can omit enormous amounts of information with a narrow scope; a description sufficient to predict a person's lunch order would be inadequate for continuing their research or maintaining friendships. Preserving broadly human abilities would also be different from preserving human individuality. A digital system might possess the same cognitive abilities as a human without preserving that person's memories, personality, or experiences.

This difference helps us characterize how much information is required to represent a person. The answer depends on which aspects of the original mind we want to preserve.

Consider a simple model with a finite distribution of possible minds and a fixed set of situations, each producing a deterministic response. Group together minds that respond identically in every permitted situation. If there are $N$ distinct representations, a fixed-length encoding needs at least $$B \geq \lceil \log_2 N \rceil$$ bits to distinguish them.

However, it remains a difficult problem to decide which situations describe the model sufficiently.

A representation might reproduce someone's professional expertise while losing the reasons they chose that profession. It might handle nearly every ordinary conversation convincingly, but fail to recognize a close friend. Such a failure could barely affect a reward function while changing something central to the person's life.

A compact representation of a person does not necessarily mean that recreating them requires little information. For example, if a computer already stored a complete simulation of every human, a simple identification number would be enough to retrieve any individual. The number itself contains almost no information about the person; the complexity resides with the model behind it. Similarly, a short description of someone's memories and personality might allow an AI to imitate them, but that does not mean the description alone is sufficient to reproduce their mind.

Nor would preserving a static snapshot be sufficient. An upload would need to encounter unfamiliar situations, form memories, and revise its beliefs. An immutable system would preserve current opinions and behavior while losing the ability to reconsider them.

The target is an online process, something that retains a meaningful connection to its past while remaining capable of developing its future. Even if we could achieve perfect reconstruction, a more fundamental (perhaps metaphysical) question remains.

Teletransportation Paradox

Suppose we could encode every relevant detail of a human brain (through a pain-free and noninvasive procedure) into a petabyte of information. If that information were to be transferred to a computer capable of reproducing the brain's processes, would the resulting digital mind be the same person?

This illustrates the teletransportation paradox ^[4], explored by philosopher Derek Parfit. Although the replica would be indistinguishable from the original at the moment of reconstruction, it remains unclear whether uploading preserves the subjective experience or merely creates a new consciousness with memories.

Now suppose the procedure leaves the original biological brain intact. The original person and the replica each believe themselves to be the same individual. Yet their experiences would immediately begin to diverge, producing two distinct individuals with a shared past.

Destroying the original after the procedure would not resolve this ambiguity. The testimony of the replica would not establish whether the original consciousness had continued or whether a new consciousness had inherited its memories. From the original person's perspective, there is no obvious way to distinguish these outcomes.

The problem becomes more apparent if the procedure creates multiple replicas. Each would share the same memories and psychological history, but their experiences would diverge immediately after reconstruction. Although all would share psychological continuity with the original person, they could not all be numerically identical to that individual ^[5].

Parfit opines that personal identity might not be what matters most. Perhaps preserving someone's memories, relationships, intentions, and psychological continuity is sufficient for a meaningful form of survival, even if the resulting individual is not numerically identical to the original. This argument might satisfy someone who wants their commitments and projects to continue. It may not satisfy someone who wants to experience that future themselves.

Humans + Machine > Machine?

Uploaded intelligence presents another possibility. If the human mind could operate on digital hardware, it may not be constrained by the biological limitations of the brain.

Brains have limited working memory, operate on relatively slow neural signals, and require biological processes to sustain their activity. On the other hand, artificial intelligence can take advantage of increasingly powerful computational hardware, external memory, and parallel processing.

An uploaded mind could hypothetically operate at different speeds, access external memory, or interact with artificial intelligence through interfaces more direct than language. Multiple copies might explore different approaches to a problem simultaneously, similar to how agents operate now.

These possibilities remain speculative fiction. Nevertheless, the distinction between biological and digital intelligence raises an interesting question about the future of human cognition. If machines eventually become substantially more capable than humans, cognitive augmentation might become the only way to remain intellectually competitive.

Brain-computer interfaces already demonstrate more modest ways of connecting biological cognition to digital systems. For example, experimental devices developed by Neuralink have enabled people with paralysis to control computers or communication using neural signals. A far cry from reproducing consciousness, but they illustrate the possibilities of computational systems extending human capabilities.

One possible future involves a gradual progression from restoring lost functions to augmenting existing ones. External memory, computational assistance, and increasingly sophisticated neural interfaces could expand what people are capable of without requiring their biological minds to be replaced.

Another possibility allows human and artificial intelligence to coexist within digital environments. The two might collaborate, compete, or eventually become difficult to distinguish as their capabilities develop.

Nevertheless, the more extensively we modify a digital mind to improve its capabilities, the less obvious it becomes whether we are enhancing the original person or creating someone fundamentally different.

Conclusion

The prospect of uploaded intelligence offers both a fascinating and unsettling possibility for humanity. In a world increasingly shaped by artificial intelligence, humans may no longer be confined to the cognitive limitations of their biological brains. Thought could become faster, memories could be extended, and entirely new forms of intelligence might emerge through the integration of humans and machines.

Yet this possibility introduces a peculiar dilemma. The more we modify human cognition to keep pace with machines, the less clear it becomes what remains distinctly human. At some point, the boundary between augmentation and replacement may disappear entirely.

Perhaps uploaded intelligence will allow humanity to transcend biological limitations, or perhaps it will only create digital successors that inherit our memories and ambitions. Ultimately, the question is whether humanity can endure in a world of increasingly powerful machines, and what we might have to become in order to survive.