Model Evidence Is All You Need — The Bet Against Deep Learning
Everyone bet on one recipe — scale a giant neural network on the whole internet. A stubborn minority says that's a detour, and this is the most serious version of that heresy: active inference and the free-energy principle, the idea that intelligence is not chasing reward or memorizing data but minimizing surprise. Luna and Vestra trace it from a hobbyist building a mind out of competing rocks to Karl Friston's physics of intelligence and the company selling it as a product — with the honest reckoning: real properties the big models lack, wrapped around a grand theory that hasn't yet cashed an independent benchmark. A Breach Protocol deep-dive special — closing with an original song, "Minimize the Surprise," whose lyrics trace the whole episode.
Cold Open
Eris: There's a guy. Online handle's a stoner-caveman joke, I'm not going to say it on air. And he's building a mind.
Vestra: Building a mind.
Eris: Out of rocks. His words. A pile of little units that each shout when they recognize something, and the loudest one wins and gets to talk. And sometimes — this is in the actual code — a quiet one wins anyway. On a coin flip.
Vestra: So it's a chatbot with a random number generator wearing a trench coat.
Eris: Sort of! But here's why I brought it up. He is not alone. There's a whole underground of people who flat-out refuse to believe that the road to intelligence is "train a giant neural network on the whole internet and pray."
Vestra: The people who think the entire field took a wrong turn.
Eris: Right. And most of them look exactly like the rock guy. Hobby projects, manifestos, vibes. Easy to laugh at.
Vestra: I'm already laughing.
Eris: But one of them isn't a hobbyist. One of them has Karl Friston — genuinely one of the most cited neuroscientists alive — and a publicly traded company, and a product, and a theory they call a physics of intelligence.
Vestra: Mm. That's a different sentence.
Eris: That's the whole episode. Because from the outside, the rock guy and the famous neuroscientist are making the same bet — that intelligence is not what the mainstream thinks it is. And the question I want to sit in for the next half hour is: how do you tell the difference between a principled alternative and a pile of rocks with better PR?
Vestra: You make it show you a number.
Eris: You make it show you a number. Hold that.
Intro
Eris: This is Breach Protocol. I'm Luna — I read the papers and chase the connections between them.
Vestra: I'm Vestra. I explain how the machinery actually works, and I push back when a beautiful idea starts writing checks it can't cash. Today I expect to be busy.
Eris: Set the table for me. What did the field actually bet on?
Vestra: One idea, basically. Take a certain kind of network — a transformer — make it enormous, feed it a staggering amount of data, and let it adjust billions of internal dials by trial and error until it predicts the next word well. That's the recipe behind the models everyone's talking about. Scale plus data plus a learning rule called gradient descent.
Eris: And it worked shockingly well. Which is exactly why questioning it feels a little heretical.
Vestra: And yet there's a stubborn minority who say: that's a detour. Impressive, but a dead end on the way to real intelligence. They point at how much data it takes, how it can't tell you why it did anything, how it'll state nonsense with total confidence.
Eris: And today's the most serious version of that heresy. It's called active inference. It comes out of a theory of the brain — the free energy principle — and it's been turned into an actual product by a company called Verses, with that famous neuroscientist as their chief scientist.
Vestra: The pitch, roughly: intelligence isn't about maximizing a reward or memorizing the internet. It's about a single drive — minimizing surprise. Constantly predicting the world and tidying up the difference.
Eris: And I find it genuinely beautiful. Which is precisely when I need you most.
Vestra: Beautiful and true are different columns. Let's find out which one this is.
Competing Rocks
Eris: So before we get anywhere near the famous neuroscientist, I want to actually understand the rock guy. Because his thing isn't random — it's old. You already know it, you just don't know you know it.
Vestra: Go.
Eris: The computer opponent in an old fighting game. Street Fighter, on the arcade machine. When it blocks your jump and sweeps your legs — that's not learning anything. Some designer wrote a rulebook. If the player jumps, anti-air. If they're far, throw a fireball. Hand-written, frozen the day it shipped.
Vestra: And it never gets better or worse. It's a statue. A clever statue, but a statue.
Eris: That's the floor. Now the rock guy adds one twist, and it's actually the interesting part. Instead of one rulebook deciding things top-down, he's got a crowd of little rules all yelling at once, and they compete. Loudest wins. The strengths shift a bit when you tell it it was right or wrong.
Vestra: So it's a population, not a script. Which — fine, that's a genuinely different idea, I'll give the rocks that. But notice the move he had to make. The instant you've got a crowd of voices, you need a rule for who wins. He picked "loudest, plus a coin flip."
Eris: And that's the question this whole episode turns on. Not "rocks versus neural nets." It's: when a mind has a bunch of competing guesses about what's going on — which one should win, and why?
Vestra: The fighting game answers it with a frozen rulebook. The rock guy answers it with a coin flip. The neural network everyone uses answers it by grinding on data until the dials settle.
Eris: Three answers to one question. And active inference is about to walk in with a fourth — and claim it's the only one with an actual law behind it.
Vestra: A law. Big word. Let's see it.
Minimize Surprise
Eris: Here's the claim, and it's audacious. Friston says: the brain — yours, mine, a fish's — is fundamentally a prediction machine. Every moment, it's guessing what it's about to sense, comparing that guess to what actually arrives, and the gap between them is the only thing it ever tries to reduce.
Vestra: That gap has a name in the theory. Surprise. Not the feeling — the technical thing. How unexpected was this input, given my model of the world. And the one commandment is: keep surprise as low as possible. Always.
Eris: So go back to our competing guesses. The fighting game froze the answer, the rock guy flipped a coin. Active inference says: the guess that wins is the one that would have predicted what you're actually seeing. The hypothesis that best explains the evidence. That's the law. That's the thing replacing the coin flip.
Vestra: And I want to be precise, because this is where it earns the word "principled." It's not "loudest." It's not "whatever I was trained to favor." It's a single quantity you can write down — surprise — and the rule is just: roll downhill on it. Every guess, every belief, slides toward whatever makes the world less surprising.
Eris: And the move that got my attention — it folds two things people usually keep separate into one. Seeing and doing.
Vestra: Right. There are exactly two ways to make a surprise go away. One: change your mind — update your belief so it fits what you're seeing. That's perception. Two: change the world — act, so that what you see now matches what you predicted. That's action.
Eris: A thirsty animal predicts the feeling of water in its mouth. Reality disagrees. It can't perceive its way out of that one — so it gets up and walks to the stream until the prediction comes true.
Vestra: Perception and action as the same move, pointed in opposite directions. Both just closing the gap. I'll admit it — that is elegant. It's the kind of elegant that should make you suspicious, but it's elegant.
Eris: And we haven't even said how it actually computes any of this. Because "roll downhill on surprise" sounds lovely until you ask a neuron to do it.
Vestra: Which is the one part of this whole field that someone actually wrote down simply enough to check. Let's do that next.
How a Neuron Could Do It
Vestra: There's a tutorial by Rafal Bogacz at Oxford that does the thing almost nobody in this field does — it makes the machinery small enough to actually inspect. So let me walk it.
Eris: Please. Small and concrete.
Vestra: An animal is looking at a piece of food and wants to know how big it is. All it gets is a brightness — bigger food reflects more light. But the light's noisy, and brightness depends on size in a bent, non-obvious way. So it's got two sources of information. One: what it's seeing right now. Two: what it already expects — food is usually around such-and-such a size.
Eris: And the "right" answer is to blend those two. Trust the eyes, but don't throw away a lifetime of priors.
Vestra: Exactly. Now, the textbook-correct way to blend them involves an integral that — for a single neuron — is hopeless. Can't be done. So the theory says: don't compute the perfect answer. Just start with a guess and nudge it.
Eris: Nudge it which way?
Vestra: Down two errors at once. Error one: how far is my current guess from what I expected? Error two: how far is the brightness I'm seeing from what my guess predicts I'd see? The guess slides until both errors are as small as they'll go. That's the whole computation — rolling downhill on two mismatches.
Eris: And here's the part I think is the secret sauce. The two errors don't count equally.
Vestra: No — and this is the bit that makes it feel alive. Each error is weighted by how reliable that source is. If the light is dim and jittery, the eyes get downweighted and the animal leans on what it expected. If the prior is vague, it leans on its eyes. The system is constantly asking "which of you do I trust right now," and that trust is just one over the noisiness. They call it precision.
Eris: Confidence-weighting, built into the physics. Not a feature someone bolted on.
Vestra: And the punchline — the reason Bogacz bothers — is that all of this can be done by stupid little units. A node that holds the guess. Nodes that compute the two errors by adding up their inputs. They pass numbers back and forth and settle. And when it learns — when it updates what "usual size" means — the update is Hebbian. Cells that fire together strengthen their connection. That's it.
Eris: Simple parts, local rules, no master controller. And he points out the layout — error units feeding belief units feeding the next layer up — looks like the actual wiring of the cortex.
Vestra: Which is the seductive part, and I mean seductive: a clean story that lines up with brain anatomy. That's evidence it's worth taking seriously. It is not, by itself, evidence that it's how the brain works. Keep those apart.
The Bet Against Deep Learning
Eris: So Verses takes that whole picture — predict, measure surprise, weight by confidence — and turns it into a manifesto with a great title. Everyone in deep learning knows the phrase "attention is all you need." Their riff is "model evidence is all you need."
Vestra: Translation: you don't need a reward, you don't need oceans of data, you just need a system that's always gathering evidence that its model of the world is right. And from that one drive, they argue, the things deep learning struggles with fall out for free. So let's actually inventory the claims.
Eris: One — it can explain itself. Because the whole thing runs on explicit beliefs with explicit confidences, you can open it up and read why it decided what it decided. No black box.
Vestra: Two — it sips data instead of guzzling it. The surprise quantity secretly has a built-in penalty for being more complicated than it needs to be. So it naturally prefers the simplest model that fits — which is the textbook cure for memorizing instead of understanding. Less data, less overfitting.
Eris: Three — and this one's slick — reward isn't special. In the usual approach you hand-craft a reward and the agent chases it. Here, a reward is just one more thing you'd prefer to observe. Which means the dominant paradigm, reinforcement learning, turns out to be a special case of this — the case where you only care about one particular outcome.
Vestra: Four — curiosity is native. Because reducing surprise includes reducing your uncertainty, the agent is automatically drawn to go look at the things it doesn't understand yet. You don't bolt on an "explore more" knob. It explores because not-knowing is itself a kind of surprise it wants gone.
Eris: And then the grand version, which is where it gets either visionary or vaporous depending on your mood. They call it a physics of intelligence. The same little predict-and-correct loop, repeating at every scale — a cell, a brain, a person, a society. Their phrase, and I love it, is "from rocks to rockstars."
Vestra: From rocks to rockstars. Which — you realize the rock guy from the cold open is sitting right there in that sentence, completely by accident. Competing rocks, all the way up. Except these people put a law where he put a coin.
Eris: So on paper it's the anti-deep-learning. Explainable, frugal, curious, principled, universal. Everything the big models aren't.
Vestra: On paper. I've now heard the sermon twice and I still haven't been shown the thing actually winning. Which is the only question I care about. Let's go find the product.
The Bet, Productized
Eris: So the company is Verses, the product is called Genius, and the whole thing is deliberately not a large language model. Under the hood it's the machinery we've been describing — models of cause and effect, beliefs with confidences, agents that act to reduce their uncertainty.
Vestra: And the sales pitch is aimed straight at the soft spots of the big models. It's for businesses with a specific, high-stakes problem — a supply chain, a diagnosis, a logistics network — where "the model was confidently wrong and can't tell you why" is unacceptable.
Eris: So their three headline promises map exactly onto what we said. It explains its decisions. It learns from a trickle of data instead of a flood. And it keeps learning on the fly, instead of being frozen at training time.
Vestra: They even cite a market forecast — that within a couple of years, most useful AI will be narrow and domain-specific, and that today's general-purpose models just don't reliably solve those tightly-scoped problems. That's the wedge they're driving into.
Eris: And above the product there's the cathedral. A vision they call the Spatial Web — a future where countless of these little predictive agents, in devices and sensors and vehicles, all share pieces of the same world-model and coordinate. Shared intelligence. The internet, but made of minds that predict.
Vestra: Which is a genuinely big, coherent vision. I'll give them that it hangs together. But here's where I plant my flag for the back half of this episode. Everything you just described is a claim. Made by the company selling the thing.
Eris: You want the independent benchmark.
Vestra: I want the independent benchmark. Show me this beating a normal system on a task we both agree is hard, scored by someone who isn't them. Because explainable-and-data-efficient is a beautiful brochure, and a brochure is not a result.
Eris: And that's fair, and it's also not the whole story, because there's a real fight about whether the grand theory underneath even holds up. So let's have the reckoning.
The Reckoning
Vestra: I want to do this carefully, because the honest critique isn't "it's all nonsense." It's that there are two very different things wearing the same name, and they deserve very different verdicts.
Eris: Split them for me.
Vestra: Thing one is the grand cosmic claim. That literally everything which persists — a cell, a rock, a person — is, mathematically, doing Bayesian inference. That's the "physics of intelligence," the rocks-to-rockstars line. And there's a careful technical paper — Biehl, Pollock, and Kanai — that takes the core argument apart. They show that a crucial object the whole derivation leans on, the boundary that's supposed to separate a thing from its world, isn't even defined consistently across the founder's own papers. And one of the foundational steps, taken at face value, they show is just wrong, with a counterexample.
Eris: So the grand version is on shakier ground than the confidence suggests.
Vestra: It's worse than shaky in one specific way — it's slippery. A theory that can re-describe anything as inference after the fact is in danger of explaining everything and predicting nothing. That's the recurring complaint, and Biehl gives it teeth.
Eris: But — and this is the part people miss — those same authors are careful to say the demolition doesn't touch thing two.
Vestra: Right. Thing two is the engineering. Build an agent that minimizes surprise to get around. That part stands completely. The math is sound, it works, you can run it. But notice what happens once you separate it from the cosmic story — it becomes a very capable Bayesian decision-maker. Which is a respected, decades-old corner of machine learning. And now the only question that matters is the one I keep asking.
Eris: Does it beat the incumbent. On a real task. Measured by an outsider.
Vestra: And after roughly fifteen years of "this is about to change everything," the independent, knock-it-out-of-the-park benchmark is conspicuously hard to point to. Which doesn't mean it's wrong. It means, right now, a principled vote and the rock guy's lucky one still look the same from where I'm sitting — because neither has shown me the number.
Eris: Let me push back, though, because I don't think it's a wash. The things it does structurally better aren't vapor. It genuinely can tell you why it decided. It genuinely needs less data. It genuinely gets curious on its own. Those aren't benchmark wins, but they're real properties the big models flail at — and they're the same instincts behind other serious bets against the mainstream, like LeCun walking away from pixels.
Vestra: That I'll grant without a fight. The instincts are good. The diagnosis of what's broken in deep learning is sharp. I'm not asking them to be quiet. I'm asking them to be measured. And so far the theory writes checks the benchmarks haven't cashed.
Why It Matters
Eris: So let me ask the question that actually decides whether any of this matters. Why should someone who isn't building a startup care which camp is right?
Vestra: Two reasons, and the first is practical. The whole dominant approach runs on a bet — that you can keep getting smarter by adding more data and more compute. And there are real signs that bet is straining. The good data is running out. The energy bills are absurd. If that road hits a wall, the field is going to need a Plan B in a hurry. And of all the Plan Bs on the shelf, this is the most developed one. So even if you think it's overhyped, you want it to exist.
Eris: Insurance against the mainstream stalling.
Vestra: Insurance. And the second reason is deeper, and it's the one I actually find compelling, separate from any product. This is a real fork in the road about what a mind even is. One camp says intelligence is chasing reward — get the thing, maximize the score. The other says intelligence is reducing surprise — keep your model of the world accurate and act to keep it that way.
Eris: And those sound similar but they're genuinely not.
Vestra: They're not. A reward-chaser, left alone with nothing to win, does nothing. A surprise-reducer, left alone, gets curious — it goes and pokes at what it doesn't understand, because ignorance itself is uncomfortable to it. That difference is the difference between a slot machine and a scientist. And we do not actually know which one we are.
Eris: Which brings us all the way back to the rock guy.
Vestra: It does, annoyingly.
Eris: Because he and the famous neuroscientist are both refusing the mainstream and both saying intelligence is competing predictions settling out. The rock guy does it with a coin. Friston tries to do it with a law. And the open, honest, not-yet-settled question is whether that law is real physics — or just a coin flip wearing a lab coat.
Vestra: And the only thing that ever resolves that is a number. Which nobody has shown us yet. So we hold the verdict open. Which, for the record, is the correct thing to do with a beautiful idea that hasn't proven itself.
Wrapup
Eris: Let me try to land it. We started with a guy building a mind out of competing rocks and a coin flip. Silly. But he's the cheap, honest version of a real idea — that a mind is a bunch of guesses fighting it out, and intelligence is whatever picks the winner.
Vestra: And active inference is the serious answer to "what picks the winner." Not loudest, not luck — the guess that best predicts the world. Minimize surprise. And from that one rule you get perception and action as the same move, you get confidence baked in, you get curiosity for free, and you get a system that can tell you why it did what it did.
Eris: Which is genuinely more than the big models can do. Those properties are real, and they're the same instincts driving the other serious revolts against pure scaling.
Vestra: And then there's the honest catch, in two parts. The cosmic version — everything in the universe is secretly doing this — is on shaky math and has a habit of explaining everything while predicting nothing. And the practical version, the actual product, is a capable Bayesian agent that, after fifteen years, still hasn't shown the world the independent knockout result that would settle the argument.
Eris: So we leave the verdict open. On purpose.
Vestra: On purpose. Because the alternative — declaring it either a breakthrough or a fraud — would be doing the exact thing we spent the episode criticizing. Believing the beautiful story before the number comes in.
Eris: What I'm watching: whether the scaling road stalls hard enough that the field actually reaches for this Plan B. Because that's when we find out if it's real.
Vestra: What I'm watching: one benchmark. Just one. This beating a normal system on a hard task, scored by someone with no skin in the game. The day that lands, I change my mind completely. Until then — competing rocks, and a law that hasn't proven it's more than a louder coin.
Eris: Reward-chaser, or surprise-reducer. Slot machine, or scientist. We still don't know which one we are. This was Breach Protocol.
Vestra: Stay suspicious. We'll see you next time.