LLMs have mastered the language all of science is written in. Brains run on something older, and uploads would need it.

Two debates keep coming up here: when AGI arrives, and whether we'll eventually upload our brains into machines. They matter more than they seem. Treating AGI as a future event takes the urgency out of what has already been happening to jobs and schools since at least GPT-4. And the uploading debate shapes whether we see ourselves as machines that could simply be copied. I think we're getting both wrong, in opposite directions: in an important sense AGI is already here, and uploading is a much harder problem than scaling. That's the argument in short. The rest of this post is the evidence and my answers to the obvious objections, with sources in the first comment. I'm not a neuroscientist or physicist, so I've tried to keep what's established separate from what I'm proposing. The language layer All of science is recorded in language. Every paper, proof, textbook, patent and manual. Math is notation, and notation is language. Nothing in the scientific record exists outside of it. LLMs operate natively in that medium. That's a categorical shift, not an incremental one: a huge share of tasks whose input and output are language can now be automated. What's left is a gap in reliability and depth, not in access to the medium. However you define AGI, machines have broken into the medium all of our science is recorded in. Arguing about the date AGI arrives distracts from what is already happening to every job, school and institution that runs on language. Code is language too. GPT-4 could already reflect on almost any text you gave it; the newest models write, debug and manage real software projects with less and less human help. The language layer keeps expanding into everything that gets done through text. The usual objection is that LLMs aren't AGI because they can't adapt. Their weights are frozen after training. That's true. But look at brains. For hundreds of millions of years, nervous systems learned, adapted and solved problems with no language at all. A crow, an octopus and a dog are all intelligent without words. Even in humans, plenty of what we know, like how to ride a bike or catch a ball, was never in language to begin with. So adaptivity and language are two different layers. Brains had adaptivity first and added language very late. LLMs got language without the adaptivity. To see what that missing layer is, you have to look at what brains run on underneath language. What brains run on First, what's established. Digital hardware works by suppressing variation. Every transistor is an analog device underneath, with its own slightly different switching point that drifts with heat and age. Engineers build in wide safety margins so none of that matters: above a threshold is 1, below is 0. A clock forces every operation into discrete time steps. Shielding and error correction make sure outside fields can't change the result, and when they do, we call it a bit flip and treat it as a fault. The design goal is that the computation comes out the same no matter what the surrounding universe is doing. Brains do the opposite. A neuron firing is all-or-none, but whether it fires, and exactly when, depends on continuous conditions that keep shifting. Each neuron has its own threshold, and it moves with recent activity. Firing happens in continuous time, and timing carries information: whether one neuron fires a few milliseconds before or after another can decide whether the connection between them strengthens or weakens. This is called spike-timing-dependent plasticity, and it's one of the ways brains adapt. Brains are also physically coupled to the electromagnetic field. Every active neuron writes a small disturbance into it, and when millions act together, the result is strong enough for EEG and MEG to measure from outside the skull. Neurons also read from it. Strong magnetic pulses (TMS) can make them fire, and neighboring neurons subtly influence each other's timing directly through the field, not only through synapses. This is called ephaptic coupling, and in one experiment on slices of mouse brain, a wave of activity even jumped across a surgical cut. The field isn't only made by brains. Earth generates one too, shaped by the sun, and every brain on the planet sits inside it. In a small 2019 Caltech study, not yet independently replicated, rotating an Earth-strength magnetic field around participants changed brainwave activity in 4 of 34 people, even though none of them noticed anything. Vision is the same story at a different frequency: light is part of that field, and the eye reads it directly. Every time you read someone's expression, a signal from their brain reaches yours over that same field, carried by light. My hypothesis There is one electromagnetic field, shared by every brain, the planet and the sun. Brains are open systems coupled to it in both directions, writing to it and reading from it, and that coupling is part of what they are, not noise on top of a computation. Digital systems run on electromagnetism too, and they even leak it: a chip's emissions are distinctive enough that side-channel attacks can pull encryption keys out of them. But the chip never reads those emissions back. They're waste, and the hardware is engineered so the field can't influence the computation. A brain writes and reads, and what it reads shapes its own activity. A computer sits in the same field and its camera reads the same light, but everything gets rounded into discrete values before it is used. To be clear, I'm taking a side here. A simulated calculator really calculates, but a simulated fire doesn't burn anything, and I'm betting the brain is more like the fire. Nothing above proves that; the established facts are consistent with both. My claim is that they make the fire side more plausible than people here usually assume, and that uploading quietly depends on the calculator side being true. What this means for uploading I'm not arguing uploading is impossible. I'm arguing it's a different problem from the one we're currently solving. To upload a brain digitally, you'd have to discretize it: fixed time steps, finite precision, every value rounded. That rounds off the fine continuous variation the brain uses as signal, and more importantly, it turns an open system into a closed one. The result might produce brain-like output, the way an LLM produces human-like language, but it would be a model of a brain running inside a sealed box. The most developed upload proposals would also destroy the original brain in order to scan it. That's usually debated as an identity question: is a perfect copy still you? My point is different and more physical. On this view, the copy wouldn't be the same kind of system at all. It would be a closed model of an open system, so a destructive upload trades the original for a description of it. Gradual replacement, swapping neurons one at a time for artificial ones, doesn't escape this either. This isn't the Ship of Theseus, where every new plank is as good as the old one. A real neuron doesn't only pass signals through synapses; it also writes to and reads from the field around it. A replacement that only copies the synaptic behavior would cut that coupling, one neuron at a time. The question isn't whether it's still the same ship. It's whether the new planks still touch the water. Energy is a clue here. A brain runs on about 20 watts. In 2013, one of the world's fastest supercomputers needed 40 minutes to simulate a single second of activity in a network the size of 1% of a brain, and that network was randomly wired, with no field effects at all. Feynman made a similar point about physics: to simulate nature efficiently, you need a machine that works like nature. The cheapest way to run a brain may simply be to be one. So if uploading is possible, I suspect it runs through analog neuromorphic hardware built to couple to the field the way neurons do, not through bigger digital models. Objections I expect "Just simulate the field too." You can, and we agree on more than it seems: the brain's activity includes field effects, and a simulation could model them in as much detail as you like. The disagreement is whether a model of that coupling is the coupling. A simulated brain would be coupled to a simulated field, a description of the field rather than the field itself. Hooking it up to sensors doesn't close the gap, since sensors turn the field into a stream of rounded numbers, and nothing the upload does writes back into the field in a way that shapes its own activity. If you believe a perfect model of a process is that process, the way a simulated calculator really calculates (functionalism), that's a coherent position. But it's an assumption, not a result, and it's where the real disagreement lives: is a brain more like a calculator or more like a fire? "Analog systems aren't infinitely precise either. Below the noise floor there's no information." True. The point isn't infinite precision; it's what the variation is connected to. A simulation can add noise, even physical noise from a hardware random generator, but that noise isn't coupled to what it stands in for. It doesn't come from the neuron's neighbors, its own field or its surroundings. A brain's variation does. "Everything is coupled to the EM field, including rocks." Right, so coupling alone isn't enough. The claim is that brains use the coupling, since it shapes when neurons fire and how connections change, while digital systems are engineered to make it irrelevant. Coupling is necessary in this view, not sufficient. "This is just the old 'brain as a radio' idea." That idea, going back to William James in 1898, says the brain receives something non-physical from outside. This is different in two ways: the field is physical and measurable, and the coupling runs both ways. Brains write to the field, which EEG and MEG record, and read from it, which ephaptic coupling and TMS show. "Isn't this just Penrose and Hameroff's quantum microtubules?" No. Everything here is classical electromagnetism, measurable at body temperature, and the argument doesn't need quantum effects. If their Orch-OR theory turns out to be right, it would only widen the gap between a brain and a digital copy, since quantum systems are exponentially harder to simulate on classical computers. But this argument stands or falls without it. "People are fine inside Faraday cages and MRI scanners." Shielding blocks outside radiation, but the coupling between neurons inside the brain continues regardless. Static magnetic fields like Earth's pass straight through a Faraday cage; blocking them takes special magnetic alloys like mu-metal. "AI doesn't need to copy the brain, just like planes don't flap their wings." Agreed, and that's my point about LLMs: they reached the language layer without copying brains at all. But uploading is different. It doesn't aim for intelligence in general; it aims to continue one specific brain, and for that, how the original works is exactly what matters. You can build a plane without copying a bird, but you can't upload a bird that way. "LLMs do adapt, through in-context learning and fine-tuning." In-context learning disappears when the conversation ends, and weight updates happen offline in separate training runs. A deployed model doesn't rewire itself from experience while it runs. A brain never stops. "LLMs still make mistakes, so language isn't solved." Agreed, which is why I said a huge share can be automated, not that language is solved. I'd argue reliability is an engineering problem, though that's debated. The barrier of the medium itself is gone. What would change my mind The coupling itself is established: it has been measured in rodent brain tissue, weak fields shift the timing of neurons in awake monkeys, and the physics is the same in human brains. What's open is whether it does work that matters. If replacing neurons in a living network with artificial units that copy their synaptic behavior, but don't emit or respond to fields, leaves the network's timing and synchrony unchanged, then the coupling is a side effect rather than part of what the brain does, and the core of this argument fails. Summary The language layer is where all of science is recorded. LLMs have already crossed into it, and we should stop debating whether that counts and deal with what it means. The field layer is what brains ran on long before language: continuous, adaptive and coupled in both directions to everything around it. Digital systems are built to keep it from mattering. I think that's the gap between an LLM and a brain, and between a digital copy and the original. One is built to resist the physical world in order to stay stable. The other is part of it.