Whole-Brain Emulation and Other Paths to AGI
4 min read · updated August 3, 2026
Whole-brain emulation is the proposal to build a mind by copying one rather than by designing one. Its distinguishing feature is that it requires no theory of intelligence at all, which is why it has stayed on the table for decades despite being the most obviously difficult route.
Why this route is interesting
Every other path to general intelligence requires figuring out what intelligence is well enough to build it. Emulation requires only measuring a brain in enough detail and running the measurement forward. Anders Sandberg and Nick Bostrom put it this way in “Whole Brain Emulation: A Roadmap” (2008), the canonical technical treatment: the approach is brute force, and brute force has the property that progress is a function of instrumentation rather than of insight.
That makes it forecastable in a way other routes are not. You cannot put a schedule on having the right idea. You can, in principle, put one on resolution, throughput and compute — which is exactly what the roadmap attempted, and why it remains the reference document even though its projections are old.
What emulation requires
Three things, in order, and each is a separate research programme.
- Scanning. Physically measuring the structure of a brain at whatever resolution turns out to be sufficient. Current methods are destructive: fix the tissue, slice it, image the slices with electron microscopy, reconstruct.
- Interpretation. Turning the images into a model with the parameters a simulation needs — which cells, which connections, with what strengths, with what dynamics.
- Simulation. Running that model fast enough to be useful, which is a compute problem whose size depends entirely on the level of detail the previous step turns out to require.
The roadmap’s central and still-unanswered question is the level of detail. Emulation might be possible at the level of neurons and connection strengths; it might require the internal state of each synapse, or gene expression, or molecular dynamics. Each level up multiplies the scanning resolution and the compute by a large factor, and nobody knows which level suffices. A forecast that does not state its assumed level of detail is not saying anything.
Scanning: the connectomics state
Connectomics has produced genuine, checkable milestones, and they are the right things to track.
The nematode C. elegans — a few hundred neurons — was mapped by White and colleagues in the 1980s, the first complete connectome of any animal. Complete larval and adult connectomes of the fruit fly have since been published, the adult whole-brain reconstruction being a large-scale collaborative effort combining electron microscopy with machine-assisted proofreading. Cubic-millimetre volumes of mammalian cortex have been reconstructed with every neuron and synapse traced.
What is missing is not technique but scale. A human brain is larger than the largest completed reconstruction by many orders of magnitude in volume, and the imaging, storage and reconstruction cost scale with it. The useful thing to watch is not any single announcement but the trend in cost and time per unit volume, which is the quantity that decides whether the gap is closable.
Interpretation: the harder half
The strongest evidence that structure is not enough comes from the simplest case. The C. elegans connectome has been available for decades and the worm’s behaviour still cannot be predicted from it — a sustained open-science effort has worked on exactly this and it remains unfinished. If a complete wiring diagram of a few hundred neurons does not yield a working simulation, a wiring diagram of a human brain will not either.
The gap has identifiable contents:
- Synaptic strength. A connectome records that a connection exists. Function depends on how strong it is, and strength is not straightforwardly readable from morphology in fixed tissue.
- Neuromodulation. Chemicals released diffusely change how whole regions behave without corresponding to any wire. They are invisible to a structural scan.
- Dynamics. Cells differ in their response properties, and those properties are not fully determined by shape or connectivity.
- Non-neuronal contributions. Whether glia and other cells participate in computation is an open question in neuroscience, and the answer changes what has to be modelled.
The concrete missing capability, stated so progress on it is recognisable: a method for recovering functional parameters — not just topology — from a fixed, scanned brain. Nobody has one, and until somebody does, scanning resolution is not the binding constraint.
The other non-LLM routes
| Route | Description |
|---|---|
| neuromorphic hardware | Chips that compute with spiking, event-driven elements rather than dense matrix multiplication. Real energy advantages have been demonstrated on some workloads. Missing: a training method for large spiking networks that is competitive with gradient descent on dense hardware. The obstacle is algorithmic, not fabricational. |
| brain-inspired architectures | Predictive coding, cortical-column models, and other designs that borrow neuroscience principles without copying a specific brain. Missing: a demonstration that any of them scales to the capability level dense transformers already reach. They are not currently competitive at scale, and that is the whole question. |
| neurosymbolic and program synthesis | Combining learned components with explicit symbolic structure, aimed squarely at systematic generalisation and verifiability. Missing: a way to learn the symbolic structure rather than hand-supplying it. Hybrid systems work where the symbolic part is authored; authoring it is the bottleneck. |
| open-ended and evolutionary search | Rather than optimising a fixed objective, generate an expanding space of problems and solutions. Motivated by the observation that the one process known to have produced general intelligence worked this way. Missing: any known way to run it at a tractable cost — evolution's compute budget was the biosphere over billions of years. |
| embodied and developmental learning | Learn as an infant does, through sensorimotor interaction over long horizons. Missing: sample efficiency. Real-world interaction is slow and cannot be parallelised the way text can, which is why the field leans on simulation and why the reality gap matters so much. |
A caution about all five: “current methods will not get there, so one of these must” is a non-sequitur that appears constantly. The possibility that current methods plus continued scale suffice is a live option, and it is the subject of the scaling debate elsewhere in this cluster.
Why the route would matter
If emulation arrived first, the situation would differ from the designed-AI case in ways worth thinking about in advance. An emulated mind would have human-derived motivations, which is a real advantage over specifying values from scratch and does not amount to alignment — humans are not reliably benign, and an emulation of a particular person inherits that person.
It would also raise the moral questions in moral status immediately rather than eventually, since an emulation of a person has whatever claim the person had, and it would raise them in the acute form of copying, pausing and deleting. And it would have unusual economics: a mind that can be copied and run at variable speed makes labour reproducible, which is the assumption behind the most extreme wage predictions discussed in post-AGI economics.
None of this is a forecast that the route will succeed. It is the case for why the requirements are worth knowing: they are the difference between tracking progress and tracking press releases.