As I understand it, death is suboptimal and the process of dying seems like a pretty miserable experience for everyone involved. So recently, I’ve been thinking about ways we might be able to avoid that whole thing. Specifically, how we can achieve immortality.
One proposal you’ll hear is digital immortality through a virtual clone: a software replica of your mind that lives on after you die. To me, this misses the point of immortality, which is that you live. Instead of creating a copy of your brain which lives on in your place, we should understand how we as individuals might live forever without a dramatic disruption to our experience of being alive. The transition should be so gradual you barely notice it happening.
This post has two broad parts:
In the first, I discuss the nature of consciousness and what it means to be immortal. Further, why any procedure to achieve immortality must respect the constraint of continuity of consciousness.
In the second, I build up the proposal that we should use neural networks as the core technology of human simulation and that we can transition from a biological brain made of cells to a silicon brain made of a neural network mesh.
Everything that follows is speculative.
First and foremost, I think of immortality as the indefinite continuation of a specific instance of personal consciousness.
I understand consciousness as the ongoing process that produces subjective experience, the feeling of being you. I believe, and assume in this post, that consciousness follows a functional model defined by its function (what it does) rather than the biological substrate it runs on, in line with leading theories of consciousness. Even if your brain were made of metal instead of cells, as long as the underlying computation is identical, then it makes no difference to the state of consciousness.
Under this framing, the indefinite preservation of the human body and the physical brain is not a precondition for immortality. Of course, the fact that our consciousness is currently deeply coupled with our body and brain creates important constraints.
Additionally, a copy of one’s consciousness is not immortality from the perspective of the original. A copy’s existence or non-existence has no bearing on another instance’s function or experience. What matters is the currently existing instance of consciousness. If that ceases to exist, then so do you.
Many believe otherwise. Some hold that identity is shared more broadly across conscious beings, while others point to sleep and anesthesia as examples of consciousness ceasing without being “death”. Although I subscribe to a physicalist view and would claim that the continuity of the underlying physical process is sufficient to preserve the original instance, this is contested and a thorny topic. Instead of getting hung up here, I’ll proceed under the assumption that a copy is a distinct instance from the original and take a conservative view on what might disrupt consciousness.
As a final note on this topic, I also believe that it’s ok to accept a tradeoff in fidelity to the personality and memories of the original in the transition to an indefinitely persistent form of consciousness. Procedures for this transition should do their best to avoid a reduction in the quality of consciousness, but ultimately, a degraded experience of consciousness is preferable to death. This is at least partially mitigated by the likelihood that digital forms of consciousness will be much easier to modify and improve than human brains. So even if the procedures available in your lifetime lead to a version of you that is substantially less intellectually capable, doesn’t retain memories, or has a dramatically different personality, future advancements might allow you to restore full fidelity or even enhance the experience of existence beyond the human baseline.
The actual transition of consciousness is tricky.
From the outside, we can’t directly inspect the internal experience of being conscious. The entity after transition might behave identically to the original while lacking the internal process of consciousness (a philosophical zombie). Alternatively, the original instance of consciousness might not have persisted through the procedure, resulting in a fresh copy of the original which is conscious in its own right but not the original. Since we can't verify success directly, the safest approach is one where each step is small enough that failure is bounded. Our best bet is gradual replacement uploading, where we replace the brain one component at a time with a silicon substitute that performs the same function.
Here’s the case for its practical feasibility. Consciousness appears to be an aggregate property of many distributed components, governed less by any individual neuron than by the collective firing patterns of neural ensembles. If that's true, then there likely exists some granularity of replacement, at which swapping a biological component for a functionally equivalent silicon one does not interrupt consciousness. To this point, hippocampal prostheses supplement parts of the brain’s functionality with a computer running a relatively simple algorithm. Although the gap between supplementing a region and fully replacing one is large, it’s tentative evidence that silicon can stand in for biological neural circuitry.
The first step to digital immortality is simulating the human brain, because before we go wild on uploading real people to the cloud, we first need to know that we can actually run a human brain on a computer. The technological progression here starts with whole brain emulation (WBE), the replication of the neuronal state of the brain in software.
Current approaches to WBE involve mapping out an organism’s brain via an imaging technique (e.g. electron microscopy) then using a neuron model to run a digital replica of it. As a concrete example of WBE, researchers at Eon Systems recently released work where they took a computational model of a fly’s brain and integrated it into a physics-based simulation of a fly body. The groundwork involved cutting a fly’s brain into ultra-thin slices and creating a map of every neuron and synapse. This map was then implemented in a neuron model and integrated into a virtual environment where sensory data from the environment come into the simulated brain and motor commands come out to be executed by the simulated fly body.
Simulating a human brain requires basically the same type of work, but at a scale of roughly 106 more neurons and synapses. For an overview of what’s necessary to get to human WBE and what timelines might look like, it’s worth a skim through this survey. As a ballpark, the author estimates “$5-50B over 10-25 years” to complete human WBE, requiring innovations in imaging and simulation — ultimately challenging but doable, especially considering probable technological acceleration from the use of AI research tools.
Once we’ve hit this milestone, once we’re able to create a faithful model of a human brain from neuronal imaging, there’s still some way to go before we can make an individual person immortal. That’s because imaging a human brain at the level of granularity of individual neurons and synapses necessary for WBE requires techniques that are destructive to the brain. In other words, we can’t simulate an individual via WBE without destroying their brain first.
So the crux of our technical problem is that we need to be able to create a full neuronal simulation of an individual while restricted to non-destructive brain recording techniques like implanted brain-computer interfaces (BCI). To give some sense of numbers here, there are about 1011 neurons and 1014 synapses in the brain while modern BCI like Neuralink is at ~103 electrodes. Although BCI is increasing electrode count quickly, even if we get to 107 electrodes within 40 years (assuming doubling every 3 years), we need to be able to take 107 data points from an individual and extrapolate that to 1014.
We clearly need a powerful generalized representation of the human brain to bridge the gap. We can’t rely on individual data alone to determine neuronal state. We need a strong prior for how human brains are structured and behave, on top of which we can leverage individual data to specialize to a specific person’s brain.
This is where foundation models and the neural network algorithms powering modern AI come in. My proposal is that we can simulate an individual’s brain non-destructively through a multimodal autoregressive neural network model. We first train a foundation model to learn that generalized representation of human brains, then we fine-tune it to simulate the individual. Under this regime, the place of WBE is in creating synthetic training data for these foundation models.
The foundation model is trained on a broad slice of human population. It takes in current neuronal state, sensory data like video and audio, perhaps also language, haptic, hormonal and proprioceptive inputs. It outputs neuronal state one timestep into the future. The training data is drawn from two broad sources: “life logging” living people and synthetic data from putting human brain emulation in a virtual environment.
Life logging has been floated before e.g. in this LessWrong post and this Reddit discussion. It involves real-time recording of an individual’s sensory perspective and neuronal state throughout a broad sample of the individual’s daily life — at a minimum smart glasses to record video plus audio and implanted BCI to record neuronal state. As a scenario, we might imagine 100,000 people, who collectively cover demographic attributes spanning the range of humanity, wearing recording devices 24/7 for years at a time. We would get to study how the brain reacts to a wide variety of situations and even how it changes over time at different stages of life.
Separately, we can generate an enormous amount of synthetic data from running WBEs in a variety of virtual environments, which could be physics based simulations of the world. As an example, we might create an environment where the individual crosses a street and study how the brain processes information, plans, and makes decisions. The neural network is meant to learn a generalized, distilled representation of human brains from the upstream WBE neuron model by studying its outputs.
There is a tradeoff between the two types of data. Life logging captures high fidelity environmental data but coarse granularity neuronal data. Life logging will also involve capturing a large amount of “boring” data; daily life can be repetitive and most of the recorded data will contain little novel information. Synthetic data captures high fidelity neuronal data but lower fidelity environmental data.
Life logging is crucial, synthetic data is highly beneficial but not strictly necessary. A WBE model might be more complete in its modeling of the brain’s machinery, but it’ll always have gaps to the ground truth of how the human brain behaves in real situations, which we can get through life logging, even if it’s a noisy picture. Further, life logging can capture changes in the brain’s behavior over time, something a static model of neurons can’t do.
Once we’ve taken all this data and trained the foundation model with it, we then fine-tune it for an individual who wants their brain simulated. This individual might life log their daily life for a year and for special sessions designed to surface as many memories and unique aspects of their personality as possible. This leads to another reason why life logging data in pre-training is so important; otherwise the individual’s life log data is out of training distribution and the model will be forced to generalize to a domain it hasn’t seen before.
But this whole train of thought raises the question: even if we can build a foundation model that produces the right neuronal state outputs for an individual, is such a model conscious? The claim that it is conscious has two components: a continued commitment to the functional model of consciousness and an empirical claim about the level of abstraction at which consciousness operates.
Our best guess from empirical research is that consciousness can be characterized by neuronal activity, without needing to descend to molecular-level biology. Some hold that consciousness-relevant computation happens below the level of neuronal state, at the level of dendritic dynamics, neuromodulation, or even quantum effects in microtubules. If sub-neuron dynamics turn out to matter, the approach can adapt by operating at a finer grain, at the cost of significantly harder data collection and increased computational requirements.
But if neuronal state is causally sufficient for consciousness and the functional model holds, then consciousness is a function determined by neuronal state. Further, we can treat this function like a black box. As long as any given neuronal state produces the same conscious experience, it doesn’t matter whether the function is in reality a biological brain or a neural network. The implication is that we can change both the underlying hardware (cells → silicon) and the computation being performed (biological neurons → foundation model) while preserving consciousness, which lives at a higher level of abstraction than either the hardware or the computation underneath.
A natural question is whether the prediction error of the neural network will interfere with the quality or presence of consciousness in the neural network simulation. Although prediction error is a significant concern, two thoughts. First, neurons are by nature stochastic so consciousness has some tolerance for noise in neuronal state; the open empirical question is whether the model stays within those bounds. Second, prediction error leading to an altered quality of consciousness is an acceptable tradeoff against non-existence.
At last, we have the foundation for immortality. The final missing piece is being able to facilitate a continuous transition of consciousness from a biological brain to a silicon one. The ideal scenario is gradual replacement: switching over the brain in chunks small enough such that no individual chunk is load-bearing for consciousness.
A monolithic neural network is opaque. Its internal representations don’t map cleanly onto parts of the brain so there’s no principled way to swap in a small piece of it while leaving the rest of the brain biological. To this end, we might train many small models instead. First, we subdivide the brain into small chunks such that no individual piece is load-bearing for consciousness in the sense that its temporary loss wouldn’t interrupt consciousness. Then for each chunk, we train a model whose inputs and outputs match the chunk’s interface. This gives us a procedure to replace the brain piecemeal with a mesh of interconnected neural networks.
The process of training these small models might look pretty similar to the proposal for training the larger monolithic model. Life logging combined with synthetic WBE data trains the subregion foundation model; the individual’s life-logging fine-tunes it.
Although I won’t make a concrete decision on what specific subdivision makes sense, we can discuss some factors to consider. The smaller the chunks, the more confident we should be that they aren’t load-bearing for consciousness. On the other hand, from the perspective of reducing engineering scope by avoiding an excessive number of small pieces, we want chunks large enough to keep the total count manageable. We intuitively want to choose chunks along functional units and natural boundaries. Chunks that cross natural boundaries are harder to characterize and verify so it makes more sense to take advantage of pre-existing natural divisions. The right granularity and level of abstraction of these subregions then depends on the specific region of the brain — we might divide the hippocampus into subfields, the cortex into cortical areas, and the thalamus into nuclei.
Once we’ve subdivided the brain and trained all our small models, we need to figure out how to get the brain to interface with each model. Because even a tiny amount of round trip network latency is incompatible with the frequencies we want to run inference at (e.g. 1,000 Hz), this looks like some kind of implanted neural prosthesis running inference on-device.
Then it’s time to switch over. Switching over an individual chunk could look like a blue-green deployment: the prosthesis runs in parallel with its biological subregion with both connected to surrounding neurons, neuronal traffic is gradually switched over to the prosthesis, and the switchover can be rolled back if something is wrong. If everything goes right with the neural prosthesis now bearing the entire load of traffic, we might decide to discard the biological subregion.
We repeat this process until all the pieces of the brain are switched over and we are left with a mesh of neural networks. This mesh still communicates with the rest of the body via the same network of nerves as before so we haven’t completely replaced the individual’s dependence on biological components. But the consciousness, the core part of one’s identity, will now run in silicon.
Because the disruption from any individual chunk is bounded by that chunk’s contribution to consciousness and we subdivided the brain specifically such that no chunk is load-bearing, the worst-case disruption should not break the continuity of consciousness. If we’ve done our job well, we’ll have preserved the continuity and quality of consciousness throughout the entire process, even as the underlying hardware and computation of consciousness change dramatically. One caveat: there may be some discrepancies between the behavior of the neural network mesh and the original brain. As I argued earlier, this is a tradeoff worth accepting. Continued existence, even as a degraded version of the original, is preferable to non-existence, and digital consciousness is easier to refine than its biological counterpart.
The final result is a person that should behave, think, and feel like who they were before. For the time being, their new silicon brain is still hooked up to a biological human body. But because software is endlessly reproducible, testable, and easy to modify, the range of ways they might alter themselves becomes essentially unbounded. They are immortal and the world is their playground.
Everything below is order-of-magnitude with the goal of doing a sanity check and gaining intuition.
The human brain has roughly 1014 synapses. At ~10 parameters per synapse — enough for connection weight plus state for short-term plasticity, neuromodulation, and dynamics — that gives ~1015 parameters.
At 106 electrodes per person sampling at 103 Hz, treating each ~1ms timestep as one multimodal token (joint electrode state plus concurrent sensory input), one person produces ~3×1010 tokens per year.
A Chinchilla-optimal run for a 1015 parameter model wants ~2×1016 tokens, or ~700,000 person-years of recording. 100,000 people for 3 years yields ~1016 tokens — enough to train, but on the data-limited side of optimal. Closing the gap requires either a much larger life-logging program or synthetic data from WBE, which is part of why the WBE pathway matters.
For memory, training a dense model at 16 bytes per parameter, requires the equivalent ~200,000 H100s to hold the model. Then for compute, using the standard 6 × params × tokens approximation, training on 1016 tokens at 1015 parameters costs ~6×1031 FLOPs. Assuming performance of 1e15 FLOP/s for an H100 on bf16, this is 9,500 for that cluster of 200,000 H100s, so compute looks to be the bottleneck.
But projecting out a 2x improvement in FLOP/s every 2 years per chip gets us to reasonable numbers pretty quickly — after 30 years of progress along this trend, 9,500 years shrinks to 3.5 months.
Once the foundation model exists, fine-tuning specializes it to one individual — closer to LoRA-style adaptation than continued pretraining. About 1 year of dense recording gives 3×1010 tokens (0.0003% of pretraining), or ~2×1026 FLOPs. On the same 100k-H100 cluster, that's ~20 days of wall time.
Nothing here is obviously impossible, but the recording side is the real constraint. Compute trends largely take care of themselves on a 30–50 year timeframe; the question is whether BCI hardware gets there in time.
Originally posted on X.