The Future of Whole Brain Emulation

Long a controversial staple of science fiction, emerging tools and recent breakthroughs have neuroscientists pursuing the concept of uploading consciousness under a less fictional name: whole brain emulation.

Where We’re At
Since 2008, neural recording capabilities have improved by roughly two orders of magnitude1. Calcium imaging can now capture nearly 80% of neurons in a larval zebrafish brain and about a million cortical neurons (~1.4% of total neurons)2 in mice. High-density probes like Neuropixels record from thousands of neurons simultaneously¹. This would have seemed impossible a generation ago.

In parallel, researchers have produced complete synaptic-resolution maps of brains across various organisms. Called connectomes, these comprehensive neural wiring diagrams have been completed for C. elegans worms, fruit flies, and larval zebrafish. The cost per reconstructed neuron has plummeted from roughly $16,500 to about $100¹.

The convergence of improved recordings, mapped connective structures, and improved computational power has enabled proper brain emulation of simpler organisms. In C. elegans (about 300 neurons), closed-loop simulations now reproduce basic behaviors¹. Zebrafish (100K neurons) simulations accurately replicate eye-movement dynamics¹. Fruit fly (140K neurons) models can predict neural responses during feeding and grooming¹. We’ve been able to simulate these organisms reasonably well, given the small number of neurons required to reconstruct their entire nervous systems.

Even crude reconstructions of these smaller-scale organisms represent real progress. The obvious question is whether this approach can scale to more complex organisms and, eventually, to humans.

What’s Still Missing
To reach human-scale emulation, we need to solve a set of separate, interlocking problems.

The Scale Problem:
The most advanced mouse “connectome” captures about 120K neurons out of the ~70 million neurons in a mouse’s brain¹. Imaging a mouse brain at sufficient resolution would generate an exabyte of data; a human brain, up to 2.8 zettabytes¹. That’s roughly 0.5% of the entirety of the internet, just to capture a single brain.

The Human Recording Problem:
For humans, EEG remains the standard for neural recordings, but it’s a blunt instrument. It captures only aggregate electrical activity from millions of neurons via surface electrodes, offering little insight into individual neuronal dynamics. EEG cannot reliably distinguish neighboring circuits, let alone individual cells, and because EEG uses surface electrodes, its imaging depth is on the scale of single millimeters3, obscuring activity of deeper structures.

The other side of human state-of-the-art imaging is functional imaging, like functional magnetic resonance imaging (fMRI) and functional near infrared spectroscopy (fNIRS). They measure blood flow changes as a proxy for neural activity4^,5. While useful in some contexts, this proxy is flawed, as all of them are. Blood flow responses vary across brain regions and individuals, change with age and health, and unfold on timescales of seconds, which are much slower than neural signaling.

The Resolution Problem:
Even setting aside scale, our recording paradigm captures only part of the picture. Invasive techniques typically focus on action potentials, the moments when a neuron fires, while ignoring subthreshold activity. This binarization treats neurons as simple on/off switches. Membrane voltage fluctuates continuously as it integrates thousands of inputs, and these dynamics are computationally significant. Subthreshold activity can synchronize distant neural populations, shape learning, and encode information independently of firing.

All of that only accounts for neurons. The brain contains a roughly comparable number of glial cells, astrocytes, microglia, and oligodendrocytes that actively modulate synaptic transmission, shape learning, and maintain circuit function. A recording paradigm that ignores this entire class of relevant cells misses half of the picture.

The Chemistry Problem:
Reductively, all electrical activity in the brain is the result of chemistry: ion gradients across membranes, neurotransmitter release and binding, and neuromodulators that reconfigure circuit behavior over seconds or hours. To emulate rather than approximate, we would need to capture this chemical foundation.

Real-time chemical sensing in the brain, with spatial and temporal precision, remains an unsolved problem. Even outside the brain, continuous hormone monitoring is still in its infancy.

Each of these problems demands a different approach. Progress on one doesn’t guarantee progress on another. Still, even solving a single sub-problem could be transformational for neuroscience and medicine. We don’t need full emulation to reap enormous benefits along the way.

Prediction vs. Emulation
With enough data, we can reasonably predict targets like gross recordings and labeled behaviors. But this doesn’t capture the architecture of human neural circuits firing. Prediction learns relationships. Emulation captures and simulates fundamental structure and behavior as an emergent property.

Whole brain emulation requires reproducing the internal causal mechanisms of neural circuits, leading to the emergent property of organism behavior, and, maybe, even consciousness. All proper frontier science at its core borders on philosophy. Predictive models depend on subjectively defined labels and objectives, requiring us to decide what the system is meant to predict or optimize. At sufficient resolution, only reconstructing the underlying architecture should allow behavior, and potentially consciousness, to emerge as a property of the system itself.

This leads to a deeper question. If a brain were measured perfectly, would it be perfectly predictable? And if so, what exactly would that prediction represent? Answering yes assumes that the human mind can be fully reduced to numbers multiplied in silicon.

I’ll speculate here: Neural activity is electrical, emergent from chemistry. Chemical interactions are governed by electron interactions. And electrons, at their foundation, are quantum objects, existing in probability distributions until forced into definite states by interaction. We can’t observe these states without altering them. This might not be a limitation of our instruments, but of the universe itself. So what if the thing that makes me, me, is entangled in the way quantum states resolve, moment by moment, across a lifetime of neural computation? While it’s a stretch, what if identity is even partially encoded in these quantum events? Emulation would face an even greater wall than basic technological advancement. You can’t copy what you can’t measure. And you can’t measure what observation destroys.

Where Progress Can Come From
Given these challenges, I see two paths that feel particularly promising.

The first sidesteps the issue of translatability between organisms by working directly with human tissue. Organoids, ex vivo slice cultures, and organ-on-chip platforms offer something animal models never could: actual human neurons with human genetics, in controllable, experimental systems. While not emulations, they’re the real substrate, grown outside the body. They won’t replicate cognition or behavior, but for questions of drug response, toxicity, and cellular mechanism, they may prove far more translatable than any mouse model ever has. The technology is still maturing, with challenges around reproducibility, vascularization, and functional maturity. Of course, a few thousand neurons in a dish is a far cry from 86 billion in a skull, but the trajectory is promising. If we can build reliable, scalable platforms for testing interventions on human neural tissue, we can shortcut the translation problem that has plagued neuroscience-derived therapeutics for decades.

The second is better, bigger, data for more useful predictions. Improved electrode arrays, optical or mechanical recording methods, and novel biosensors capable of tracking neurochemistry in real time could dramatically expand what we can capture. With larger, richer datasets, prediction becomes a data science problem, linking interventions to outcomes without needing to characterize the causal mechanisms or machinery in between. We don’t need to understand every synapse or connection to predict that a compound will be toxic, or that a stimulation pattern will alleviate symptoms. The brain can remain a mostly unsolved problem, so long as its inputs and outputs are reliable. We rarely understand mechanisms completely; we learn what works and refine from there.

I’ve come to think we need more neuroengineering than neuroscience. A scientist seeks to understand. An engineer seeks to solve. While fundamental understanding of human brain function is a ways away, with better tools and problems properly reframed, we can make significant progress.

Emulation of smaller-scale organisms represents an exceptional feat of neuroscience, a testament to decades of painstaking research. The path to human-scale emulation runs through solving the challenges I’ve outlined above: scale, human-compatible recording, and resolution. Perhaps I’m overestimating some of these barriers. Maybe there are approaches I haven’t considered that make certain challenges more tractable than they appear.

If you’re building tools that address these gaps, whether in imaging, biosensing, tissue engineering, computational approaches, or any other fields I might have missed, let’s talk!