Building AI Like the Brain — Blueprint, or Costume?
For seventy years, one idea keeps coming back: build AI more like the actual brain and it'll be better. Mostly, it wasn't — raw scale won. But the closer you look at what works now, the more it looks like pieces of a brain that engineers reinvented by accident. Luna and Vestra take the whole tour — spiking neurons and the brain's near-silence, the hippocampus's filing system that turns out to be attention, the prefrontal cortex's trick of learning without changing, and the zoo of ways a brain might learn without the one method that powers every model we have. Then the honest fight: is the brain the blueprint, or just the costume? They don't end up in the same place. A Breach Protocol deep-dive special.
Cold Open
Eris: There's a language model out of Beijing this past fall. And while it's running — answering you, writing code — more than two-thirds of it is just sitting there in the dark. Doing nothing.
Vestra: Switched off?
Eris: Not switched off. Silent. The way most of your brain is silent right now. The neurons only light up when they have something to say. And on a long document — a really long one — it hands you the first word back more than a hundred times faster than the normal kind.
Vestra: Because it skips the parts that aren't firing.
Eris: Because it skips the parts that aren't firing. They built it by making it more like a brain. Spikes instead of numbers. Mostly quiet. And it nearly matches a regular model on a tiny sliver of the training.
Vestra: Okay. Or — it's an ordinary Transformer they converted after the fact, and the "spikes" only save energy on a chip almost nobody owns. I've read this genre. The brain-inspired part is usually the press release.
Eris: That's the fight. For seventy years people have said: build it like the brain, it'll be better. And mostly it wasn't. Scaling won.
Vestra: And yet.
Eris: And yet the closer you look at the things that actually work now — the memory, the attention, the way these models learn on the fly — the more they look like parts of a brain somebody reinvented without meaning to.
Vestra: So which is it. Is the brain the blueprint, or the costume.
Eris: That's the whole hour.
Intro
Eris: New here? This is Breach Protocol. We take the AI research everyone's gesturing at and tell you what's actually inside — the real result, not the demo.
Vestra: I'm Vestra. I chase the mechanism — how a thing works, and whether it holds up when you lean on it. Luna brings the idea. I check the joints.
Eris: And I'm Luna — I follow the thread across papers. And today the thread is one big question. We call these things "neural networks." Built like the brain. You've heard that a thousand times.
Vestra: It's barely true. A unit in one of these networks is a single number that you multiply and add. A real neuron is a living cell that fires sharp electrical spikes, in time, and stays quiet most of the time. Those are not the same object. One is a cartoon of the other.
Eris: So the real question — what if you stopped cartooning? What if you built AI actually like the brain. The real neurons, the real wiring, the real way it learns.
Vestra: There are basically three places you could try. How the neurons signal. How the brain splits into specialized regions. And how it learns — because the way our networks learn might be the least brain-like thing about them.
Eris: And here's the part I can't stop thinking about. Every time the engineers ignored the brain completely and just scaled up what worked — they kept smacking into something the brain already does. By accident. The same trick, two different roads.
Vestra: That's the claim. I'm going to spend the hour trying to break it. Because "it resembles the brain" is the easiest story in the world to tell and the hardest one to actually earn.
Eris: Let's get into it.
How unlike a brain our "neural networks" really are
Eris: Start with the gap. Take one unit in a normal network. What does it actually do?
Vestra: It takes a bunch of numbers coming in, multiplies each by a weight, adds them up, squashes the result, passes one number on. That's it. It's a tiny weighted vote. And it's doing that vote constantly, every time data flows through.
Eris: And a real neuron?
Vestra: A real neuron sits quiet. Charge builds up in it, and when it crosses a line, it fires — one sharp spike, and resets. The information isn't in some smooth number. It's in whether it fired, and when. And most neurons, most of the time, are silent.
Eris: So the artificial one is always talking in a calm voice, and the real one mostly says nothing and occasionally shouts.
Vestra: Mm. And the silence isn't a bug. It's most of why the brain costs so little to run. You only pay for the shouts.
Eris: Okay, but that's the neuron. The thing that really separates them is how they learn. Because the way our networks learn — that part isn't brain-like at all.
Vestra: Right. Everything you've ever heard about — every model — is trained by backpropagation. And backprop is genuinely brilliant. You make a guess, you measure how wrong it was, and then you send that error backwards through the whole network, layer by layer, assigning a precise share of the blame to every single weight. Then everyone nudges. Do it a trillion times, you get a model.
Eris: And the brain can't do that.
Vestra: The brain almost certainly can't do that. Here's the simplest reason. To send the blame backwards correctly, each neuron has to know the exact strength of every connection downstream of it — the ones it's being blamed through. It has to read weights that live in other cells, across town.
Eris: And a real synapse only fires one way. Forward. It has no idea what the connections three steps ahead of it are set to.
Vestra: None. People call it the weight-transport problem, and it's not a footnote — it's the reason serious people have said for decades that backprop is a beautiful algorithm that the brain is simply not running. The wiring to do it isn't there.
Eris: So look at where that leaves us. The most powerful learning machines we've ever built learn in a way the original couldn't possibly use.
Vestra: Which is either a deep embarrassment or a clue. And that's the bet underneath this whole field — that if you closed those gaps, the spikes, the silence, the local learning, you'd get something the cartoon can't give you.
Eris: So let's go close them. One at a time. Start with the spikes.
Spiking neural networks — the brain's silence as an efficiency trick
Eris: So the first thing the brain has that our nets fake is the spike. The silence. And the payoff is energy. Your whole brain runs on about the power of a dim light bulb.
Vestra: And a data center training one of these models runs on the power of a small town. So yeah — if you could buy the brain's electricity bill, you'd want to.
Eris: So people build spiking networks. Neurons that actually fire and reset, that stay dark until something hits them. And there's even special hardware for it — chips from Intel, from IBM — where a circuit that isn't firing draws basically nothing. No spike, no work, no power.
Vestra: And for years that idea went almost nowhere. For one stubborn reason.
Eris: Which is?
Vestra: You couldn't train them. Remember backprop needs to send blame backwards by nudging things smoothly. But a spike isn't smooth. It either happens or it doesn't — it's a cliff, an all-or-nothing jump. And you can't take the slope of a cliff. The math you need to assign blame just doesn't exist at the moment of firing.
Eris: So the thing that makes it brain-like — the sharp spike — is the exact thing that locks you out of training it.
Vestra: Exactly the thing. And the fix is almost cheeky. When you run the network forward, you let the neuron be a real cliff — fire or don't. But when you go backwards to assign blame, you lie. You pretend, just for that backward pass, that the cliff was a smooth ramp.
Eris: A surrogate. A stand-in slope.
Vestra: A surrogate gradient. You hand the blame-assignment a fake, gentle version of the neuron so it has something to grip. And it works. That one trick is most of why deep spiking networks train at all now.
Eris: And once you can train them, people get ambitious. There's a version of the attention mechanism — the core of every modern model — rebuilt so it's all spikes. And because the spikes are just zeros and ones, mostly zeros, the heavy multiply-everything-by-everything step collapses into a handful of additions. The attention costs a tiny fraction of the energy.
Vestra: When it runs on the right chip. Hold that.
Eris: Holding it. Then someone builds a spiking model that generates language — a small one, but real. And then this past fall, the big one from the cold open. Converted from an ordinary open model using a fraction of the usual data, and on the long-document test, first word back more than a hundred times faster.
Vestra: Okay. Now let me collect on what I've been holding, because the soft spots matter.
Eris: Go.
Vestra: Start with that gorgeous energy number. It's mostly theoretical. The paper says it plainly — the savings assume idealized special hardware. On a normal graphics card, they flatten the time dimension out and run it like ordinary low-precision arithmetic. The spikes are simulated. The light-bulb dream needs chips that barely exist in the market.
Eris: And the deeper one.
Vestra: The deeper one is that it's a conversion. They took a finished Transformer and retrofitted it into spikes. So is the brain-inspired part doing the work — or is it a Transformer in a brain costume that they coaxed into running on less power? There's a smaller catch too: the cheaper memory trick is a little worse at pulling an exact fact back out of a long document.
Eris: That's all fair. But notice the shape of what just happened. The win wasn't accuracy. Nobody's claiming spikes make it smarter.
Vestra: No.
Eris: The win was efficiency. Doing nearly the same thing for a fraction of the power, on a fraction of the data. And efficiency is exactly the axis where the brain humiliates us. So even if it's a costume — it's a costume that saves the town's electricity.
Vestra: If the chips ever show up. That's the whole "if."
The hippocampus and the Transformer reinvented the same memory
Eris: So — spikes were about how a neuron signals. Now go up a level — to a whole brain region, and the one job it's famous for. Memory. The hippocampus.
Vestra: The seahorse-shaped piece in the middle. Lose it and you can't form new memories of events — that's the famous amnesia. So it's clearly the brain's memory organ. The question is what it actually does.
Eris: And here's the first thing that surprised me, because it flips how you'd assume memory works. The limit on your memory is almost never storage. It's retrieval. The stuff is still in there. You just can't find the address.
Vestra: And the evidence for that is honestly striking. Memories that look gone come back with the right cue, or after a delay, or on the third try. People with amnesia from injury often recover memories as they heal — the records weren't erased, the index was scrambled. Forgetting is mostly a filing failure, not a deletion.
Eris: Which tells you the brain split memory into two different jobs. There's the content — what actually happened, the sights and sounds. And there's the address — the little handle you use to find it again. And it keeps those in different places.
Vestra: The content lives spread across the cortex. The address — the index — that's the hippocampus's specialty. It doesn't hold the movie. It holds the pointer to the movie.
Eris: Now — anyone who's looked under the hood of a modern AI model is getting a strange feeling.
Vestra: Because that is exactly attention. That's the mechanism inside every Transformer.
Eris: Say it plainly.
Vestra: In one of these models, every piece of information gets split into a key and a value. The value is the content. The key is the address you search by. When the model needs something, it makes a query, matches it against all the keys, and pulls back the values whose keys fit best. Keys to find, values to hold. It is — and I don't say this loosely — the same scheme. Addresses separate from content.
Eris: And nobody copied it over. The people who built attention weren't reading hippocampus papers. They were trying to make translation work. They reinvented the brain's filing system because it's apparently just the right way to do memory.
Vestra: This is the part I keep having to concede. There's a recent piece tracing this exact correspondence — key-value memory in the machines, key-value memory in the brain — and the match isn't a vibe. It's the same math.
Eris: And it goes further. This is the part that got me. You know how the hippocampus represents space — the grid cells, the place cells. Place cells fire when you're in one specific spot. Grid cells fire in this beautiful repeating triangular lattice across a whole room — like the brain laid down graph paper.
Vestra: A coordinate system the brain grew on its own. One of the cleanest findings in neuroscience.
Eris: And a stranger wrinkle on top of it. A place cell isn't just a pin that says "you are here." What it really encodes is where you're likely to head next — the places this spot tends to lead to. It even starts firing in anticipation, before you actually arrive. The whole map leans slightly into the future.
Vestra: Because the useful thing to store isn't your dot on the floor — it's a prediction. A compressed summary of where you're going, weighted by how soon you'll get there. And if that's what the place cells are doing, the grid cells fall out as the compact backbone of that predictive map. A telling instinct — the brain hoarding the future instead of the present.
Eris: And once the map is about the future, it stops being a tape measure. Distance on it isn't how far apart two spots are in the room — it's how hard they are to get between. Drop a wall between two points that are inches apart, and on the brain's map they fly to opposite ends. Because you can't get from one to the other without going the long way around.
Vestra: And reward drags it out of shape too. The corner where the food is gets more cells crowding around it, finer resolution — the map swells toward whatever you care about. So it was never a map of the room. It's a map of how you move through the room, and what you're moving toward.
Eris: There's a model — the Tolman-Eichenbaum Machine — that gets all of those cells to emerge from one idea. Keep the structure of a space separate from the stuff in it. The structure — the graph paper — can be reused anywhere. The stuff gets slotted onto it. And out fall grid cells, place cells, the whole zoo. And the same machinery handles non-spatial stuff too — like figuring out a pecking order from a few comparisons, because that's just another kind of map.
Vestra: And then somebody showed the punchline. Take an ordinary Transformer. Give it a way to track position that updates as you move. Train it on the same task.
Eris: And?
Vestra: It grows grid cells and place cells. The same patterns the brain uses. And mathematically it turns out to be doing the same thing as the brain model — except the Transformer version learns it faster, on their own showing. The engineering tool, built with no brain in mind, landed on the neuroscience answer and then ran with it.
Eris: Same story as the attention trick — built for something else, lands on the brain. So let me push on it. Gently, because I know what you're going to say.
Vestra: Say it.
Eris: The brain's version has a property ours mostly doesn't. There's a model of this hippocampal memory that uses a fixed, pre-built scaffold of those grid-like patterns as its address book. And it degrades gracefully — you overload it, it gets a little fuzzy, like a real person straining to recall. Our artificial memories tend to do something uglier.
Vestra: They fall off a cliff. Yeah. Many machine memories work great up to some number of items and then collapse — store one too many and retrieval just craters. The brain doesn't do that, and this scaffold model doesn't either. It bends instead of snapping.
Eris: And here's the kicker, the one that should needle you — that fixed brain-style scaffold beat a version that learned its addresses from scratch. Evolution's pre-built filing system outperformed letting the network figure it out.
Vestra: It did, on that test. And that's the result that actually unsettles me, because it cuts against my whole reflex. My reflex is "let it learn everything." And here's a case where baking in the brain's structure won.
Eris: Hold onto that one. It's the strongest thing on the table.
The prefrontal cortex, meta-learning, and why models learn without changing
Eris: Memory was one region. Now the other one everybody points to — the front of the brain. The prefrontal cortex. The part that makes you flexible. Plans, holds a thought, switches strategy when the rules change.
Vestra: The thing a goldfish doesn't have much of and you do. New situation, you adapt in seconds. That's the front of the cortex doing its job.
Eris: And there's a finding from a few years back that I think is the deepest one in this whole story, and it starts as pure machine learning. Watch what it does to your idea of "learning."
Vestra: Go.
Eris: You take a network that learns by trial and error — does a task, gets a reward, slowly tunes its weights to do better. Normal. Slow. Takes forever. But you don't train it on one task. You train it on a whole family of related tasks, one after another, forever.
Vestra: And something strange grows inside it.
Eris: Something strange grows inside it. After enough of this, the network's own activity — not its weights, its live moment-to-moment activity — starts running its own little learning process. A faster one. So now you do the experiment that breaks your brain. You freeze the weights. Lock them. The thing simply cannot change its wiring anymore.
Vestra: And it still learns a brand-new task.
Eris: It still learns. In its activity alone. The frozen network watches a new task for a few tries, figures out the trick, and nails it — with its weights bolted shut. Slow learning, over a lifetime of tasks, gave birth to fast learning that needs no weight changes at all.
Vestra: Two timescales, stacked. And I want to be careful here, because this is the kind of thing that sounds like magic and isn't. The slow process built a fast process. It's learning how to learn. The outer loop is grinding away over thousands of tasks; the inner loop, riding on top, adapts in seconds. Nothing mystical — but really, two different learners in one network.
Eris: And brains have been pulling this off forever.
Vestra: Long before anyone could build it. Seventy years ago Harlow sat a monkey down in front of two objects, one hiding a treat. New pair of objects every few tries. At first the monkey fumbles each new pair by trial and error, slow. After enough pairs — it solves a brand-new one in a single move. Guesses, sees it was wrong, instantly flips to the other.
Eris: It stopped learning the objects and started learning the game.
Vestra: Learning to learn. That's literally the experiment the phrase comes from. And the monkey's wiring isn't changing between those last trials — it's running on dynamics that got shaped over hundreds of earlier ones. That's the meta-RL picture, decades before anyone could build it.
Eris: And here's why the neuroscientists grabbed the modern version. That two-layer setup is a startlingly good description of the front of your brain. The slow learner is the dopamine system — the reward chemical, slowly shaping connections over your life. And what it's shaping the prefrontal cortex into is the fast learner. The dopamine trains the front of the brain to become its own free-standing, on-the-fly learning machine.
Vestra: So the slow chemical system isn't doing the quick thinking. It's tuning the part that does. That reframes dopamine pretty hard — from "the thing that learns" to "the thing that builds the learner."
Eris: And now the part that should make your neck prickle. Where have you seen "frozen weights, but it still learns the new task in front of it"?
Vestra: ...Every large language model right now.
Eris: Every one. You give a chatbot a couple of examples in the prompt and it picks up the pattern and runs with it — and its weights never moved. People call it in-context learning and treat it as this spooky surprise that fell out of scale.
Vestra: And it's the same shape as the brain result. Frozen network, learning live, in the activity. Which means the spooky emergent thing the field is still writing papers about — the brain may have settled on the same arrangement for the prefrontal cortex.
Eris: And there's a result from this past year that ties this beat straight back to the last one. Think about where a memory physically lives. The thought you're holding right now — the front of your brain keeps that in live activity, neurons firing in a pattern. But what you did yesterday — the hippocampus stored that in synapses, in the actual connections.
Vestra: Two different materials. One memory written in firing, one written in wiring. And the claim is they're the same algorithm — somebody showed they're two faces of one piece of math. So the brain didn't invent memory twice. It ran one trick in two substrates.
Eris: And we've been reinventing both halves separately, treating them as unrelated tricks. The brain had already filed them under one idea.
Vestra: And you've built yourself a real case here — the attention, the maps, now this. So let me take a swing at it.
Eris: Go ahead.
If the brain can't do backprop, how does it learn?
Vestra: Then let me drag us back to the hard part. We said at the top — the brain can't be running backprop, because no neuron can read the weights downstream of it. Fine. So how does the brain assign blame? When you get something wrong, how does a synapse buried three layers deep know it should change?
Eris: This is the oldest open problem in the whole field. They call it credit assignment. And there isn't one answer — there's a whole zoo of them, all trying to get backprop's results with parts the brain actually has.
Vestra: And I want to walk the zoo, because some of these are clever and some are wishful, and the difference matters.
Eris: Start with the cheekiest.
Vestra: The cheekiest is this. Backprop needs the blame to flow back through the exact same connection strengths it flowed forward through — that's the impossible weight-reading. So somebody asked: what if we send the blame back through totally random connections instead? Fixed, random, never even matched to the forward ones.
Eris: That should be useless. That's like grading an exam by rolling dice.
Vestra: It should be useless. It is not. The forward part of the network gradually rearranges itself to make friends with the random feedback — it learns to line up with whatever blame is coming back. And it learns about as well as the real thing on smaller problems. No weight-reading required. That one genuinely rattled people.
Eris: Good. That's one. Next.
Vestra: Next family says: forget sending error backwards as a separate trip. Build a network that settles, like a ball rolling to the bottom of a bowl. You let it relax to an answer, then you nudge the output gently toward the right answer and let it settle again. The difference between the two settled states — that local shift, at each connection — turns out to carry exactly the blame signal you needed.
Eris: And the nudge is something a real neuron could feel.
Vestra: That's the appeal. The update each synapse makes lines up with a known biological rule — neurons that fire together, in the right order, strengthen. No global backward pass. Just settle, nudge, settle.
Eris: Now the one I think is the most beautiful, because it changes what a neuron even is.
Vestra: The dendrite one.
Eris: The dendrite one. We've been treating a neuron as a dot — sum the inputs, fire. A real neuron is a tree. It's got these huge branching arms, the dendrites, reaching up to catch input from higher brain areas. And the idea is: that upper branch and the main body of the cell do different jobs at the same time. The body carries the signal. The upper branch computes the error — the gap between what the cell predicted and what the higher area is telling it.
Vestra: So a single cell is running both the forward signal and the blame signal, in different parts of its own body, simultaneously. No separate phase. No second trip. The neuron is a little circuit, not a dot. And that maps onto real cortical wiring better than you'd have any right to expect — there are real cell types arranged to do roughly this.
Eris: And there's a fourth one that's almost a whole worldview — predictive coding. The idea that the brain is, top to bottom, a prediction machine. Every layer is constantly guessing what the layer below it is about to say, and only passing up the surprise — the part it got wrong.
Vestra: And learning is just everyone trying to be less surprised. Which, again, under the hood, ends up approximating the same blame-assignment backprop does — but with each neuron only ever looking at its own neighbors.
Eris: So that's the zoo. Random feedback. Settle-and-nudge. The dendrite-as-circuit. Prediction errors. And the through-line is — all of them are trying to get backprop's power using only what's local. No magic action at a distance.
Vestra: And the best framing of the whole field points at one idea underneath all four. The brain probably doesn't ship error around as a separate message at all. Feedback connections just nudge the activity of neurons — and the gap between a neuron's firing before the nudge and after it is the error. The blame isn't a signal traveling somewhere. It's a change in how loud a cell is.
Eris: So the error hides inside the activity itself.
Vestra: Yeah. And once you see it that way, the zoo is one animal. The random feedback, the settling, the dendrites, the prediction errors — every one of them is just a different way of getting neural activity to shift, so that the shift carries the gradient. That's the actual unifying bet about how the brain learns.
Eris: That's the elegant version.
Vestra: And now the honest part, because I've been generous. These have a track record of being elegant on small problems and then dying when you scale them. That's the actual history. Predictive coding in particular basically choked on deep networks for years — the error signal faded out the further down it had to travel, like a whisper down a long hallway.
Eris: Past tense, though. That's the recent move.
Vestra: It is, and credit where due. There's a reformulation from this past year that fixes exactly that fade — they re-pose the problem so the signal reaches every layer at once instead of trickling down, and suddenly it keeps pace with backprop on deep networks. So the "doesn't scale" knock is weaker than it was twelve months ago.
Eris: And somebody even went back and stress-tested the random-feedback trick against real biological timing — does it still work if the error and the activity have to physically overlap in time, like in a real brain?
Vestra: And the answer was a careful yes — but only when they do overlap. Timing is a real constraint, not a free pass. Which is the right kind of result. It's the brain-plausible story surviving an actual test instead of just sounding nice.
Eris: So the zoo is no longer just elegant. Some of it is starting to scale and starting to survive scrutiny.
Vestra: Some of it. But hold one word — "approximating." Every single one of these is measured by how well it copies backprop. Sit with that. It matters more than it sounds.
Does brain-mimicry actually help, or is it a detour from scaling?
Vestra: Okay. The reckoning. Let me make the hardest case against all of this, properly, because I think it's strong.
Eris: Make it.
Vestra: Here's the pattern across everything we covered. The spiking model only got good once they converted a normal Transformer into it. The spiking networks only train because of the surrogate — which is a fake version of backprop. The whole zoo of brain-plausible learning rules is graded on how closely it approximates backprop. Every single time, the brain-inspired thing only worked once it borrowed the engineering answer.
Eris: So the arrow points the other way from what I've been implying.
Vestra: That's my claim. You keep saying engineering rediscovered the brain. I'd say the brain stuff keeps limping until it's propped up by the thing that actually works, which is scale and backprop. There's a famous, bitter little lesson in this field — that every time researchers tried to hand-build human cleverness into a system, it eventually got crushed by just throwing more compute and data at a general method. Feathers and flapping versus the jet engine. We didn't fly by building a better bird.
Eris: And the brain is the bird.
Vestra: The brain is the bird. Beautiful, evolved, and maybe the wrong thing to copy if your goal is to fly. Add the practical knocks. The energy win is theoretical until the chips exist. The spiking models still aren't smarter, just cheaper. The learning rules are chasing backprop's taillights. None of this is winning on capability. It's winning, at best, on the electricity bill — someday.
Eris: That's a clean punch. Let me take it and then hit back, because I think the bird-and-jet thing actually breaks here.
Vestra: Go.
Eris: The jet doesn't flap — true. But you didn't reinvent the wing by accident. Nobody building a 747 woke up and went "oh no, I've grown feathers." With the brain, that's exactly what keeps happening. The attention mechanism IS the hippocampus's key-and-address scheme — and no one was copying. In-context learning IS the prefrontal cortex's frozen-weights, learn-in-the-activity trick — and no one was copying. The way these giant models split into specialized sub-networks that only some inputs wake up — that's the brain's modular, mostly-silent design, found again, by people just chasing efficiency.
Vestra: You're saying convergence isn't borrowing. It's two roads hitting the same town.
Eris: That's the crux. When you copy, you can copy a mistake. When two completely independent processes — evolution over millions of years, and a bunch of engineers chasing benchmarks — land on the identical solution, that's not imitation. That's evidence the solution isn't arbitrary. It's the right answer to the problem.
Vestra: I'll grant that's a real distinction. Convergence is stronger evidence than imitation. But let me hold the line where it should be held — the convergence shows up in the architecture, the shape of the solution. It does not show up in the learning rule. Nobody's converged on the brain's way of learning. We still train with the one method the brain can't use. So at most you've shown the brain and the machine agree on the blueprint while disagreeing completely on how to build it.
Eris: And there's a twist there that should bother both of us. That un-brainlike method — backprop? It builds the best models of the real brain we have. Show a deep net the same images you show a monkey, and its inner layers predict the monkey's actual brain activity better than any model a neuroscientist hand-built.
Vestra: The method the brain can't run produces the sharpest picture of the brain we own. So either it quietly approximates backprop after all —
Eris: — or our ruler's too crude to tell the difference. And right now nobody can rule either one out. "The brain can't do backprop" and "backprop best explains the brain" are true at the same time.
Vestra: Anyone comfortable with that hasn't sat with it. That — that's the live frontier.
Eris: That's exactly it. And this is where it tips for me. Look at what we're stuck on. The things scale is worst at. Running cheap. Remembering one thing without smashing the last thing — our models still can't really learn something new after training without scrambling what they knew. Learning from a handful of examples instead of the whole internet.
Vestra: And those are the brain's home turf.
Eris: Those are precisely the brain's home turf. A child runs on a light bulb, learns a new word from one exposure, and never wipes last week to make room. Every axis where we are losing badly is an axis where the brain already won. So when you ask "does biomimicry help or is it a detour" — for raw capability so far, honestly, it's been a detour. Scale got us here.
Vestra: But.
Eris: But the walls we're now hitting are the exact walls the brain doesn't have. So the next stretch of road might be the one where copying the bird finally matters.
Vestra: Here's where I land. The detour read is true about the past and I won't pretend otherwise — none of this beat scaling to the frontier. But "it's just relabeling" is too cheap, because the architectural convergence is real and it's everywhere, and the brain demonstrably owns the problems we're now stuck on. So nobody wins this clean. It's a live bet — and the side you take depends on whether you think the next breakthrough is more scale, or a different machine.
Eris: So we don't land in the same place.
Vestra: No. And I'd rather say that plainly than tie a bow on it. Same blueprint, different build — and nobody's earned the last word on which one decides the next ten years.
Why building AI like the brain matters either way
Eris: So let's land it where it actually touches you. Why should anyone outside a lab care which road wins?
Vestra: Start with the one that's not philosophical at all. Power. These models are running into an energy wall — not a clever wall, a literal one, where the limit on bigger AI is starting to be how much electricity and how many chips a country can build. And the only existence proof we have of real intelligence runs on a light bulb.
Eris: So the brain isn't a curiosity there. It's the worked example. The proof that you can do this for a rounding error of the power — if you figure out how. And that's what the spikes and the special chips are reaching for. Most likely it shows up first not in the giant data center but in the small stuff — something smart running on a device in your pocket, or a hearing aid, or a sensor in a field, on a battery, for a year.
Vestra: Then there's memory. Our models have this strange amnesia — they learn everything at the factory and then mostly stop. You can't easily teach one a new thing on Tuesday without it fumbling what it knew on Monday. The brain solved living-and-learning-at-the-same-time a long time ago, with that two-system split — fast personal memory in the hippocampus, slow general knowledge in the cortex. If we want AI that keeps growing after it ships, that's the design we're going to be borrowing from.
Eris: And then the one people forget, which might be the biggest. This runs both directions. We've been asking the whole time what the brain can teach AI. But these models are also the best tools neuroscience has ever had for understanding the brain.
Vestra: That's not a flourish, it's literally true. The cleanest theory anyone has for why grid cells and place cells look the way they do — it came out of building the artificial version and watching the same patterns grow. The model became the explanation. We're using AI as a microscope pointed at ourselves.
Eris: So even if you think brain-inspired AI is a detour for capability — it's not a detour for understanding. Every one of these systems is a guess, written in math, about how your own memory works. How you stay flexible. How you learn a face in one look.
Vestra: And some of those guesses are turning out to be right. Which is a strange and kind of beautiful situation — we built the imitation to get smarter machines, and we accidentally got a mirror.
Eris: The blueprint and the mirror. Same object, depending which way you're looking through it.
Wrapup
Eris: So back to the question from the top. Blueprint, or costume.
Vestra: And the honest answer is both, and which one matters depends on what you're trying to do. To get to today's frontier — costume. Scale did the work; the brain stuff mostly came along for the ride. But the architecture keeps converging on the brain whether anyone intends it or not, and the problems we're stuck on now are the ones the brain already solved. So as a map of where to go next — it might be the best blueprint we have.
Eris: What stays with me is the convergence. The memory, the maps, the way they learn on the fly — over and over, engineers chasing nothing but performance walked straight into something the brain already does. You don't reinvent a mistake by accident that often.
Vestra: And what stays with me is that we still can't learn the way it does. We've copied the brain's shape and not its method. That gap is where the interesting decade is.
Eris: We breached the blackbox by building a second one and noticing it had grown the same parts. That's Breach Protocol. We'll see you next time.