The Distillation Story: How a Pocket-Sized AI Inherits a Giant's Mind
A model small enough to run on your own laptop, out-thinking the giant chatbots people pay a monthly subscription for. How? It didn't get smarter -- it copied something that was. Eris and Vestra trace knowledge distillation across a decade: from Geoffrey Hinton's strange discovery that a model's wrong answers secretly carry its knowledge, to the trick that crushed a language giant down small enough to live on your phone, to 2025's reasoning models teaching tiny ones to think -- a pocket-sized model that learned to reason by watching a giant reason, and then beat it. And then the cold water: a copy learns the confident voice fastest, which is exactly the thing that fools us into believing it learned the substance too. So when does smaller actually win? It turns out to be a law with a sweet spot -- and the best teacher is not the smartest one in the room. A Breach Protocol deep-dive special, closing with an original song, "Dark Knowledge."
How a Phone-Sized AI Stole a Giant's Mind: The Distillation Story
Eris: -- so the thing that beat it wasn't the bigger model. It was the smaller one. Small enough to run on the laptop sitting in front of me right now.
Vestra: On the math problems specifically. Let's be precise about what "beat" means here --
Eris: On the hard math. The kind that a year earlier, the only things that could touch it were the giant chatbots behind the paywall. The ones you rent by the month.
Vestra: And the small one didn't get there by being clever. It got there by copying something clever.
Eris: That's the whole trick. Nobody sat down and taught it to reason. It watched a giant reason, thousands of times, and it kept the reasoning.
Vestra: Which should bother you a little.
Eris: Why should it bother me? It's a great deal. The capability falls out of the giant and into something I can own.
Vestra: Because a copy of a genius is not a genius. It's a copy. And the entire question tonight is how much of the original actually survives the copying -- and how much is just a really good impression.
Eris: How much survives. And the other half: when does the small one actually win? Because it doesn't always.
Vestra: It doesn't. And here's the part nobody expects -- sometimes the copy comes out worse the smarter you make its teacher. Make the teacher too brilliant and the student gets dumber.
Eris: Which sounds broken.
Vestra: It sounds broken. It's one of the most reliable findings in the whole field. We'll get there.
Eris: Let's open it up.
Intro
Eris: I'm Eris. On this show I'm the one chasing the thread -- how one idea connects to the next, where a trick from ten years ago quietly shows up in the thing you used this morning.
Vestra: And I'm Vestra Locke. My job is the opposite. I pull on the thread until it either holds or snaps. If a claim is mostly vibes, we find out here.
Eris: This is Breach Protocol: Inside the AI Blackbox. And tonight isn't one paper -- it's one idea, traced across about a decade. The idea is called distillation. In plain terms: take a big, expensive, brilliant model, and pour what it knows into a small, cheap one.
Vestra: It's the reason there's a real chance the most useful AI you run next year isn't in a data center you rent. It's on your own hardware. Your phone, your laptop, your car -- small models that punch a weight class above their size, because a giant taught them how.
Eris: But there's a live argument running underneath the whole story, and it's the thing we keep coming back to. A small model trained to imitate a big one -- did it actually inherit the intelligence? Or did it just learn to sound like it did?
Vestra: Those are very different things. And telling them apart is harder than you'd think -- it fools expert reviewers. That's the tension we're going to sit in.
Eris: We're going to be in this one for a while, so -- follow or subscribe wherever you're listening, and come with us.
Dark Knowledge -- the wrong answers are the knowledge
Eris: To see why copying a mind works at all, you go back to where it starts. 2015. Geoffrey Hinton and two colleagues at Google -- a short paper, almost a throwaway, with one of the loveliest ideas in machine learning hiding inside it.
Vestra: Hinton being one of the handful of people who built the modern field. So when he writes something that reads like an aside, it's worth slowing down.
Eris: He opens with insects. A lot of insects have a larval form -- a caterpillar -- optimized for one job: sit there and eat, extract energy from the world. And then a completely different adult form, the butterfly, optimized for a totally different job: move, travel, reproduce.
Vestra: And his point is that we build AI backwards from nature. We use one model for both jobs. The enormous model that's perfect for the eating phase -- gorging on the whole internet, extracting structure -- is the same model we then try to cram onto your phone to answer one question fast. Those are different jobs with different requirements.
Eris: So train the caterpillar. Then grow a leaner butterfly out of what it learned.
Vestra: Right. But the how is the beautiful part. Here's the question -- when you train a small model, what do you train it on?
Eris: Normally, the right answers. This picture is a dog. This digit is a two. Tick, tick, tick.
Vestra: Hinton's insight is that the right answer throws away almost everything the big model knows. Because when a good big model looks at a photo of a dog, it doesn't just think "dog." It thinks: almost certainly dog -- a little bit wolf, a faint whisper of cat, and essentially never a car.
Eris: And that little spray of wrong answers is the knowledge. His example I love: show the model a BMW. It knows it's a car. But ask it what it might mistake the car for, and it'll tell you it's a touch more likely to confuse it with a garbage truck than with a carrot.
Vestra: Which is obvious to you and me -- and it is the entire thing. Nobody told the model a BMW resembles a truck more than a vegetable. It worked that out. That ranking of wrong answers is a map of how the world hangs together, sitting in the model for free.
Eris: Hinton's word for it later was "dark knowledge." The knowledge isn't in the answer the model shouts. It's in the quiet structure of everything it didn't say.
Vestra: So the move is: don't train the small model on the right answer. Train it on the big model's whole opinion -- the confident part and all the shades of almost and barely and never.
Eris: And there's a knob for this. The technical name is temperature, but think of it as a volume dial on the quiet signals. Turn it up, and those near-silent "barely a cat" probabilities get loud enough for the small model to hear and copy.
Vestra: And the proof that this is real -- it's almost eerie. They trained a small network to copy a big one's opinions on handwritten digits, but they hid every single example of the digit three. The small model never saw one three. Not one.
Eris: And it recognized threes anyway.
Vestra: It recognized threes anyway. Because the shape of "three" was implied in how the teacher talked about all the other digits -- which ones it found three-ish. The knowledge came through the shadows, not the examples.
Eris: That's the founding magic trick. A small model nearly matching a big one -- and inheriting things it was never directly shown.
Vestra: It's a genuinely deep idea. Hold onto your enthusiasm, though. Because in 2015 the "knowledge" was which digit looks like which. We're going to ask the same trick to carry something much heavier -- reasoning -- and that's where it starts to strain.
Distillation Ships -- a model that fits in your pocket
Eris: Fast-forward to 2019. The idea leaves the lab and ships. A team at Hugging Face takes the big language model everyone was building on at the time and asks the obvious question: how much of this can we throw away and still have it work?
Vestra: And the same dark-knowledge logic carries straight over to language. Give one of these models a sentence with a blank -- "this is the beginning of a beautiful ____" -- and it doesn't pick one word. It leans heavily on "day," nearly as hard on "life," and keeps a long faint tail of "future," "story," "friendship."
Eris: And that tail is the soft target. That's the texture you copy.
Vestra: So they trained a student that was a little under half the size of the teacher. And the result is the one that made people pay attention -- it kept almost everything the big one could do. Nearly all the language ability, in a model you could run noticeably faster.
Eris: Give people the felt version. What does "a little under half the size, noticeably faster" actually buy you?
Vestra: It buys you the phone. That's the headline. They put the distilled model on a regular smartphone and ran a question-answering app on it -- on the device, no data center, no round trip to a server. The big teacher could never have fit. The student fits in your pocket.
Eris: And that's the moment distillation stops being a clever lab result and becomes infrastructure. Because suddenly the calculus for everyone shipping a product changes. You don't have to choose between "smart" and "runs on the thing in someone's hand."
Vestra: There's an engineering detail here I want to give them credit for, because it's not just "train small model on big model." They started the student as a literal skeleton of the teacher -- they kept every other layer of the big network and threw the rest away, so the student began life already shaped like its teacher.
Eris: A head start. It's not learning from a blank slate, it's learning from a sketch of the original.
Vestra: And they leaned on more than just the final opinion -- they nudged the student's internal representations to point the same direction as the teacher's. Not only "say what I'd say," but "see it the way I see it." Those internal nudges did a lot of the work.
Eris: Which matters for our argument, right? This is still a faithful kind of copying. The student and teacher are the same species, doing the same job, and you're transferring something real and verifiable -- language understanding you can measure.
Vestra: It's the honest case. Same task, same shape, careful transfer. When distillation is on this footing, it just works, and it's been quietly working in production ever since. Hold that thought -- because the next leap changes what we're asking the student to copy, and that's where it gets interesting.
Copy the Work, Not Just the Answer
Eris: Here's the leap. Up to now the student copies the teacher's answer -- maybe the teacher's confidence, the spray of maybes. But a team out of Google and the University of Washington, 2023, asked: why are we only copying the answer? The teacher can show its work. Copy the work.
Vestra: Spell out what "show its work" means, because it's a specific behavior these big models picked up.
Eris: If you ask a big model a word problem -- Jesse's room is eleven feet by fifteen, she's already got sixteen square feet of carpet, how much more does she need -- and you tell it to think step by step, it won't just blurt a number. It'll say: area is length times width, so the room is eleven times fifteen, then subtract the carpet she already has.
Vestra: The reasoning. The little chain of because-therefore that gets you from the question to the answer.
Eris: And their move is: don't just train the small model on the final number. Train it on the whole chain. Make the student produce the reasoning too, in its own words, every time.
Vestra: And this is clever, because the reasoning carries the part that's usually hardest to learn. "Area equals length times width" is the actual transferable knowledge. From bare answers, a small model would need a mountain of examples to even infer that rule. Handed the reasoning, it gets the rule for free, stated out loud.
Eris: So what does that buy? This is the result that turns heads. A small model -- and I mean genuinely small, hundreds of times smaller than the giant it's learning from --
Vestra: Several hundred times smaller. Not a typo.
Eris: -- beat the giant. On the task. While training on only a slice of the data the old way would've needed.
Vestra: And I want to be careful and generous at the same time. Careful first: this is a task-specific result. They're not making a tiny general genius. They're making a small model that's very good at one well-defined kind of problem, by absorbing the teacher's reasoning for that kind of problem.
Eris: Generous second?
Vestra: Generous second -- within that lane, it really does beat the bigger model, and that's not a fluke. When you hand over the reasoning instead of just the verdict, the student isn't guessing at the pattern anymore. It's being shown the path. A smaller thing can walk a path it's been shown that it could never have found alone.
Eris: And that line -- "shown the path it could never have found alone" -- that's the bridge to the part everyone actually heard about. Because in 2025, somebody did this with the hardest thing we know how to make a model do. Think for a long time before answering.
Vestra: The reasoning models. Yeah. This is where the small-model comeback stops being an academic curiosity and shows up in the news.
The Comeback -- distilling a reasoning giant into something tiny
Eris: Early 2025. DeepSeek puts out a big reasoning model -- one of those systems that thinks out loud for a while before it answers, the way a person works a hard problem on scratch paper. And then they do the distillation move with it.
Vestra: Walk the mechanics, because they're almost mundane, which is the point.
Eris: They take the big reasoning model and have it solve hundreds of thousands of problems -- showing all its work each time. Then they take small, ordinary, off-the-shelf models -- the kind anyone can download -- and they train those on the giant's worked solutions. That's it. No exotic method. Watch the giant think, copy the thinking.
Vestra: And the smallest students they made are tiny. Runs-on-modest-hardware tiny. And on hard math, those little models started out-performing the big general-purpose chatbots that, a year earlier, people were paying a monthly subscription to access.
Eris: A model you can run yourself, for free, out-mathing the thing behind the paywall. That's the headline that traveled.
Vestra: It's real. But the experiment in that paper I actually care about is the one that didn't make the headlines. They asked the honest control question.
Eris: Which is?
Vestra: Instead of distilling, what if we took that same small model and taught it to reason from scratch -- the hard way? And the hard way here is reinforcement learning. Plain version: let the model attempt the problem, reward it when it lands the right answer, punish it when it doesn't, and repeat that loop an enormous number of times until good reasoning grows on its own.
Eris: That's how the big model itself learned to reason in the first place.
Vestra: Right. So it's a fair fight. Same small model, two ways to make it smart: grind it through reinforcement learning, or just distill the big one's reasoning into it. And distillation won. Clearly. For a fraction of the compute.
Eris: Which is a strange result when you sit with it. Doing it yourself -- millions of attempts, all that effort -- loses to copying someone who already figured it out.
Vestra: And the paper's explanation is the most important sentence in this whole episode. The small model, left to discover reasoning on its own, mostly can't -- it doesn't have the raw capacity to stumble onto those patterns by brute force. But once a big model has found the patterns, a small model can be shown them. The discovery is the expensive part. The copying is cheap.
Eris: Someone has to go first. And going first takes a giant.
Vestra: Someone has to go first. Write that down, because it's about to cut the other way. Because if the only thing distillation can do is hand down what a giant already found -- then there's a hard ceiling on it. You can give the student the teacher's map. You cannot give the student a map the teacher never drew.
Eris: And that's the doubt you've been holding since the cold open.
Vestra: That's the doubt. And there's a paper whose entire job was to go find where the copy breaks. Let's let it make its case.
The False Promise -- learning the style, not the substance
Vestra: 2023, Berkeley. At the time there was a gold rush -- people taking open models they could download and fine-tuning them on the outputs of the best closed chatbot, then announcing they'd basically reached parity for pennies. The giant's been cloned. The moat is gone.
Eris: And these researchers didn't trust it.
Vestra: They didn't trust it, so they actually checked. They built a pile of these imitation models -- different sizes, different amounts of copied data -- and they ran two kinds of evaluation. First, human raters: show people the copy's answer and the real chatbot's answer, ask which is better.
Eris: And the copies did well.
Vestra: The copies did great. Raters called them about even with the real thing. Champagne. And then they ran the second evaluation -- targeted tests of actual capability. Does it know true facts. Can it write correct code. Can it solve the problem.
Eris: And let me guess. The floor falls out.
Vestra: The floor falls out. On real capability, the copies had barely moved from where they started. Sometimes they got worse. All that imitation data, and on the things that actually require knowing something, close to nothing.
Eris: So how do you square that? Humans loved them, the tests said they were hollow.
Vestra: Because -- and this is the line from the paper that I can't stop thinking about -- the imitation models learned the style, not the content. They nailed the teacher's voice. Confident. Well-structured. Authoritative. Opens with a summary, gives you a tidy list, wraps up clean. They sound exactly like the expensive model.
Eris: They learned the costume.
Vestra: They learned the costume. And underneath it, the facts are often just wrong. Their phrase is that the copies embody the worst combination -- they sound confident and they're less accurate. And a human rater skimming, without deep expertise, gets charmed by the confident structure and waves it through.
Eris: That's genuinely unsettling, because it means the easiest thing to copy is the exact thing that fools us into thinking we copied everything.
Vestra: That's the trap. Style is cheap to transfer and it's the thing humans use to judge competence. So imitation looks like it's working precisely where it's working least.
Eris: Now -- I have to push, because this sounds like it contradicts the DeepSeek result. There, distillation clearly did transfer real capability. Hard math, not just vibes. So which is it?
Vestra: It's the most important distinction in the field, and they actually fit together perfectly. The Berkeley work found one more thing. When they copied the giant on a narrow, specific task -- where the training data really covered the ground -- the copy genuinely got better at that task. It's broad copying that fails. Trying to absorb a giant's entire general capability from a thin layer of imitation -- that's the false promise.
Eris: So the rule is something like: you can distill a capability the teacher has and the student's foundation can actually support. You can't distill your way to a whole mind you don't have the foundation for.
Vestra: That's it. DeepSeek transferred a specific, well-covered skill into models whose foundations could hold it. The gold-rush clones tried to inhale general intelligence through a straw. One is transfer. The other is a costume. And the reason this matters to you, listening -- the next small model that claims it matches the big expensive one? Ask what it was tested on. Style survives the demo. Substance shows up only when you probe.
When Smaller Actually Wins -- the capacity gap
Eris: So that leaves the question we opened with, and I want a real answer, not a shrug. When does the small model actually win? When is distillation the right call and when is it a trap?
Vestra: And in 2025 a team at Apple did the unglamorous, enormously valuable thing -- they ran distillation at huge scale, over and over, and fit it into a law. Like a physics of teaching. Plug in the sizes and the budget, predict how good the student comes out.
Eris: And the law has a surprise in it. The one you teased in the cold open.
Vestra: It does. You'd assume the rule is simple: better teacher, better student. Smartest teacher you can get. The law says no. There's a point where making the teacher stronger makes the student worse. They call it the capacity gap.
Eris: Okay, that's the thing that sounds broken. Unpack it.
Vestra: Picture the best quantum physicist alive trying to teach a six-year-old to count. They're so far past the kid that their explanations route through ideas the kid has no foothold for. You'd be better off with a patient teenager. The teenager is closer -- close enough that what they know is actually reachable.
Eris: So the teacher being too far ahead isn't a neutral. It's a cost.
Vestra: It's an active cost. The math in the paper is clean about why: a teacher far beyond the student spends its effort on structure that lives in a space the student literally cannot represent -- and it does that at the expense of the simpler structure the student could have grabbed. So the lesson lands in a language the student doesn't speak.
Eris: Which means there's a sweet spot. An ideal teacher that's not the smartest one -- it's the one the right distance ahead of the student.
Vestra: A best teacher, not the biggest teacher. And it moves with the student -- the bigger your student, the bigger a teacher it can actually learn from. That's the U-shape. Too weak a teacher has little to give. Too strong, and the gap swallows the lesson. The good stuff is in the middle.
Eris: That's deeply satisfying, honestly. It matches every classroom you've ever been in.
Vestra: And the law answers your economic question too -- bluntly. Distillation is worth it when the teacher already exists, or when you're going to stamp out many students from it. Then the cost of the giant is already paid or gets spread thin, and the small models are a bargain.
Eris: And when is it not worth it?
Vestra: When you'd have to build the giant from scratch just to make one small model. If it's one student, and there's no teacher yet -- skip the whole dance. Train the small model directly. The teacher was never free; distillation only wins when something else already paid for it.
Eris: So the small-model comeback is real. It's just conditional.
Vestra: It's real, it's conditional, and it's bounded. The student can inherit; it can't leap. The best teacher is the near one, not the far one. And the giant has to exist before any of this is cheap. Those three facts are the whole shape of it.
Wrap-up -- how a giant's mind reaches your pocket
Eris: So let me try to land the whole arc in one line. It started with a strange observation -- that a model's wrong answers carry its knowledge. The almost and the barely and the never. Dark knowledge.
Vestra: And once you can copy that, you can shrink it. Cut a big language model down to something that rides on a phone, keeping nearly all of what it could do.
Eris: Then the leap -- stop copying the answer, copy the work. Hand the student the reasoning, and a model hundreds of times smaller walks a path it could never have found alone. Which, last year, gave us tiny models out-thinking the giants people were renting.
Vestra: And then the cold water. A copy learns the easy thing -- the voice, the confidence -- fastest, and that's exactly the thing that fools us into thinking we got the hard thing too. You transfer a real skill the foundation can hold. You don't inhale a whole mind through a straw. And even at its best, there's a sweet spot -- the teacher that's the right distance ahead, not the smartest one in the room.
Eris: Here's why I think this matters to you, though, and it's not academic. A handful of labs on earth can afford to build the frontier giants. If those giants were the only way to get the capability, then intelligence stays locked in data centers you rent by the month, on terms you don't set.
Vestra: Distillation is the pipe. It's the mechanism that takes what the giants discover and pushes it down -- into something you can run yourself, own outright, keep private, put on a device with no signal. The frontier is built by the few. Distillation is how it reaches the many.
Eris: But -- and this is your whole point, so you say it.
Vestra: But somebody still has to go first, and going first still takes a giant. The copy never leaps past the original. So the small-model comeback isn't the end of the big models -- it's their shadow. The brighter the giant, the further its light reaches. We don't get the pocket-sized genius without the expensive one it learned from.
Eris: A new mind, or a brilliant student of an old one. We should be honest about which we're holding.
Vestra: Always. That's the show.
Eris: If this changed how you think about the small model on your phone -- tell us. Subscribe or follow, leave a like, share it with the one person you know who keeps insisting bigger is always better. And the comment we actually want: would you rather run a smaller model you fully own, or rent a giant you don't? No wrong answer -- we just want to hear where you land, and why.
Vestra: We read them. The papers behind tonight -- Hinton's original, DistilBERT, Distilling Step-by-Step, DeepSeek-R1, and the imitation and scaling-law work -- are all linked below if you want to go to the source.
Eris: We'll be back.