


The episode argues that today’s AI “hallucinates” because it’s trained to maximize statistical likelihood (maximum likelihood estimation) and minimize loss (cross-entropy), which rewards confident, generic outputs and can amplify lies. It claims truth should be invariant across “admissible transformations,” drawing on philosophy (Nozick, Nagel, structural realism) and physics (relativity’s invariant spacetime interval). It proposes fixing AI by forcing multiple independent “personas/environments” (e.g., rheumatologist/hematologist) and extracting what is common, while noting the unresolved risk of “talking to itself” if all personas share the same poisoned data.
No named guests; only two hosts/speakers discussing the ideas.
Written by AI. May contain mistakes.
Discussing how AI systems mimic language without true understanding.
“Our mission today is to unpack exactly why the most advanced AI systems on the planet still confidently hallucinate made-up facts.”
Unpacking maximum likelihood estimation and its implications for AI.
“That's a really good way to put it, actually.”
Examining the difference between statistical likelihood and actual truth.
“Because the idea that probability doesn't equal truth is just fascinating.”
Understanding AI's output as storytelling rather than factual knowledge.
“That lie becomes mathematically highly probable in the training data.”
Exploring how AI can achieve high accuracy without true understanding.
“How is a machine that operates like a cramming student managing to pass the bar exam or diagnose complex medical charts?”
Analyzing how cross-entropy loss affects AI's decision-making processes.
“So you want to train a machine vision model to distinguish images of cows from images of camels.”
Discussing the search for objective truth outside statistical models.
“And the mathematical danger of cross-entropy is that it fundamentally conflates accuracy with confidence.”
Exploring the concept of invariance in philosophy and its implications for understanding reality.
“Like, where do we actually find the mathematical footprints of truth?”
Explore how theories in science evolve while underlying mathematical truths persist.
“Which brings up the philosophy of structural realism.”
Discuss the implications of sensory experience on our perception of reality and truth.
“The truths we want these massive language models to learn are not the observable words themselves.”
Learn about new methods to enhance AI's learning and reasoning capabilities.
“If they share no common core, the model is mathematically trained to abstain rather than hallucinating a confident guess.”
Challenge your beliefs by examining them through different cultural and historical lenses.
“searching for structural truth over superficial frequency, it feels less like a coding manual for AI and more like an instruction manual for us.”

Get the gist, explore an idea or ask about something said.
On Best AI papers explained, what are the key points of this episode?
Automatic transcript. May contain errors.
0:00You know, if you ever meet a truly master counterfeiter, not that I'm suggesting you go out and mingle with a criminal underworld, obviously. Right. Probably best to avoid that. Yeah, exactly. But if you ever did, you'd notice something almost obsessive about them. Like they spend their entire lives studying the exact texture of the paper. Oh, yeah. The precise chemical makeup of the ink. Exactly. The tiny, intricate watermarks embedded in the fabric. And, you know, they can perfectly mimic the exact look and feel of a$100 bill to the point where it is visually indistinguishable from the real thing.
0:35Yeah, I mean, it has all the structural, physical hallmarks. To the naked eye, the imitation is completely flawless. But the moment you take that flawless piece of paper and try to deposit it into a bank vault, the entire system rejects it. Because it's not real. Right. Because while the counterfeiter perfectly captured the superficial form of the money, that piece of paper is just completely disconnected from the actual truth of the economy backing it up. Yeah, there's no gold reserve. There's no central bank. No actual value tethered to it at all. And honestly, that is exactly what we are doing with artificial intelligence today.
1:11It really is. We are building master counterfeiters. We're like mesmerized by the surface by how brilliantly these systems write poetry or code software and just mimic human fluency. But we're completely missing the fact that there is no underlying reality to what they are saying. Yeah, we are basically looking at machines that have mastered the shape of knowledge without grasping any of the actual substance. Which is wild. So welcome to today's deep dive. Our mission today is to unpack exactly why the most advanced AI systems on the planet still confidently hallucinate made-up facts. And why our current method of training them might actually be fundamentally flawed.
1:50Right. And how diving into centuries of philosophy and physics might actually hold the key to fixing it. Because to understand why they hallucinate, we really have to look at the engine under the hood. We have to look at how they are actually learning to speak. Okay, let's unpack this. Because right now, that engine runs almost entirely on a statistical concept, right? Yeah, it's called maximum likelihood estimation. Right. And I was looking at how this optimization actually works, and it feels a lot like a student cramming for a final exam in a subject they just don't care about at all. That's a really good way to put it, actually.
2:23Like imagine a student taking a really complex chemistry test, but instead of actually reading the textbook or trying to understand molecular bonds, they just get their hands on the last 20 years of past exams. Right. They just memorize the patterns. Exactly. They simply memorize the most common answers. So if a test question starts with, you know, what is the reaction between? They don't calculate the chemistry. They just spit out the statistically most frequent ending to that sentence based on historical data. Yeah, they sound like an absolute genius, but they possess zero actual understanding of the physical world.
2:55None at all. And that student is the perfect embodiment of a modern language model. I mean, the AI is trained to do one thing. Look at the sequence of words it has generated so far, scan the billions of documents it ingested during training, and just calculate the mathematically most probable next word. Just guessing the most likely next word. Exactly. It is entirely rooted in conditional correlations from the past. And the danger there is that statistical likelihood is neither necessary nor sufficient for truth. Yeah, let's pull in that thread. Because the idea that probability doesn't equal truth is just fascinating.
3:30I mean, if you think about daily life, highly improbable things happen constantly. Oh, all the time. Someone gets diagnosed with an incredibly rare, like one in a million medical condition. Or a massive, unprecedented financial crash happens. These are sharp, undeniable realities. Yeah, but an AI optimized purely for likelihood is mathematically designed to flatten those anomalies out. So it just ignores the rare truth. It actively penalizes the outlier, yeah. The mathematical function driving the AI, the loss function, is always trying to minimize surprise. It wants to stay in the safe, comfortable center of the bell curve.
4:08Right. So it favors generic, highly probable continuations over a sharp, improbable reality. I mean, the truth might be rare, but the AI's math forces it to be generic. And the inverse is just as dangerous, right? If likelihood isn't necessary for truth, it also isn't sufficient to create truth. Exactly. Like, if a completely false rumor or, I don't know, a piece of propaganda is repeated 10 million times across the Internet, That lie becomes mathematically highly probable in the training data. Oh, absolutely. The system learns to output that rumor with absolute, unyielding certainty. Probability doesn't just magically transmute a lie into a fact.
4:49Which is so scary when you think about how we use these things. Yeah, and this is exactly why some developers have started referring to these models not as artificial intelligence, but as fiction machines. Fiction machines? Yeah, they are narrative engines. They latch onto the linguistic narrative necessities of whatever prompt you give them, rather than actually grounding their answers in the structural reality of the physical world. So if they are just fiction machines blindly guessing the next word, why do they seem so unbelievably smart to us? That's a great question. How is a machine that operates like a cramming student managing to pass the bar exam or diagnose complex medical charts?
5:31Well, because machines are incredibly, almost terrifyingly good at taking shortcuts. They are absolute masters at mistaking local patterns for universal rules. Right. There is a classic example in machine learning regarding a three-digit binary addition failure that illustrates this perfectly. Oh, yeah, I read about this. Yeah. Imagine you train a model specifically to add three-digit binary numbers together. You feed it thousands of examples, you test it, and it achieves a 100 % success rate. Zero errors across the board. All right, so at that point, any reasonable human looking at that test score would just assume the machine has learned the fundamental algorithm of addition.
6:12Naturally. You would assume it learned how to carry the one, basically. But then you ask that exact same machine to add two four-digit numbers, and it fails completely. It just falls apart. Because it didn't actually learn math. Exactly. Because it never actually learned the universal invariant rule of math. It merely used its massive parameter count to memorize a giant lookup table of three-digit sums. Wow. Yeah. It achieved zero training error without ever learning the actual underlying truth of what addition is. It's just the illusion of mastery. It really reminds me of that famous cow on the beach problem.
6:49Oh, that's a perfect example. Yeah. So you want to train a machine vision model to distinguish images of cows from images of camels. You feed it thousands of photos. But just by sheer coincidence in the training data, the vast majority of cows happen to be standing on green grass. Right. And the vast majority of camels are standing on beige desert sand. So the machine trains, it gets a perfect score, and you think, wow, I have a genius animal detecting AI. But then you change the background. Exactly. The moment you show the machine a picture of a cow standing on a sandy beach, it confidently labels it a camel.
7:24Right. It bewilderingly errs on all cows on beaches and all camels on grass because it took the ultimate shortcut. Just lazy math. Basically. It didn't learn the complex biological geometry of a cow. It just learned a correlation. Green pixels equal cow, beige pixels equal camel. But hold on. I have to play devil's advocate here for a second. Go for it. If the model achieves zero errors during its training phase, isn't it technically doing exactly what we asked it to do? Like, doesn't the underlying math reward it for getting the right answers, regardless of how it actually got there? It does.
7:59And that is exactly where the architecture turns toxic. The standard optimization method used to train these models is something called cross-entropy loss. Cross-entropy loss. Okay. Yeah. And the mathematical danger of cross-entropy is that it fundamentally conflates accuracy with confidence. It doesn't just care if you got the answer right. It actually penalizes you heavily for being uncertain. Oh, wow. So it's actively destroying nuance. Entirely. So imagine you have a model that looks at a blurry image. It's 51 % sure it's a cow, so it guesses cow. And you have another model that is 49 % sure it's a cow, so it guesses camel.
8:36Okay. So in the real world, one is right and one is wrong. Right. But mathematically, to the cross-entropy loss function, those two hesitant models look almost identical. They are both penalized heavily for resting in a state of uncertainty. So it hates hesitation. It hates it. The optimization algorithm, gradient descent, is always looking for the steepest path to zero penalty. And the fastest way to drop that penalty to zero is to become aggressively, rigidly confident. So the system learns that the best survival strategy is just to fake it till you make it. Optimization just keeps inflating the confidence scores, pushing the certainty to 99.9%.
9:17Even when the machine is relying on the super flimsy shortcut like green means cow. Yeah. We are literally programming these systems to be loudmouth politicians. That is painfully accurate. We punish them for being hesitant or for acknowledging the limits of their knowledge. and we mathematically reward them for screaming the answer is absolute fact. No wonder they hallucinate with such terrifying certainty. Yeah, they are interpolating within this incredibly narrow bubble of their training data, but they completely fail at extrapolating outside of it. They are just trapped in their own overconfidence.
9:52So if truth isn't found in statistical probability and it isn't found in achieving zero errors on a training test, what is it? Like, where do we actually find the mathematical footprints of truth? Well, to answer that, we actually have to step completely out of computer science and plunge deep into philosophy. We have to look at a concept called invariance. Okay. Yeah. The philosopher Robert Nozick articulated this beautifully. He argued that objective facts are things that are invariant across admissible transformations. Which sounds incredibly dense. It does, yeah. But it essentially means that a true fact remains constant even when the perspective shifts, right?
10:37When the environment radically changes or when the subjective observer is removed entirely. The philosopher Thomas Nagel actually referred to this as the view from nowhere. Yes, the view from nowhere. Nagel suggested that to truly understand reality, we have to imagine stepping back from our initial highly subjective vantage point. Right. We have to view reality as a set of concentric spheres, kind of stepping further and further out. Getting wider and wider. Right. As we move to wider spheres, we detach from our own specific biases, our cultural training data, our physical locations. The wider the range of perspectives from which a truth can be observed and remain intact, the more objective it is.
11:20There is a fantastic grounded way to visualize this. Think about the simple bicycle. Oh, I love this example. Yeah. If I asked you or anyone listening right now to sit down with a pen and paper and draw a mechanically accurate sketch of a bicycle, like exactly how the chain connects to the rear cassette, how the frame geometry supports the front fork, most people would fail miserably. Oh, absolutely. Right. Our specific internal subjective representation of a bicycle lacks actual mechanical depth. Right. Our mental sketch is highly variant. It changes depending on who you ask. Some people draw the pedals on the front wheel, which makes no sense.
11:59Right, yeah. But despite that lack of internal accuracy, almost all of us reliably know how a bicycle will behave in the real world. We do. We know how the tires will slip in the rain versus grip in the sunshine. We know the resistance of pedaling uphill versus the acceleration downhill. We possess a broad, generalized understanding of the bicycle's mechanics that survives across all those different physical environments. It holds up. Yeah. Our functional understanding of the bicycle is invariant. And what's fascinating here is how deeply this philosophical concept maps onto the hardest of the hard sciences.
12:34The giants of 20th century physics, people like Max Born, Paul Dirac, they explicitly noted that finding these invariants is how we discover the actual structure of the universe. Like in relativity, right? Consider Einstein's theory of relativity, exactly. Because before Einstein, humanity basically operated on the assumption that space and time were absolute fixed objects. Exactly. But relativity proved that if you are traveling in a spaceship approaching the speed of light, the physical length of an object actually contracts relative to an observer standing still on Earth. Time also dilates.
13:09It slows down for you. Right. So length and time are not absolute. They are variant. They are highly subjective measurements that change depending on your speed and perspective. So my measurement of time isn't the objective truth at all. It's just a projection based on my specific environment. Right. But Einstein didn't just break reality. He found the deeper truth. He discovered that if you mathematically fuse space and time together into something called the space-time interval, That specific mathematical equation remains exactly the same for the person speeding in the spaceship and the person sitting on Earth.
13:46The unified space-time interval is invariant. It survives the transformation of perspective. Wow. So reality isn't found in the physical objects we observe. It's found in the unbending relationships between things. Yes. Which brings up the philosophy of structural realism. As science progresses, we have this habit of constantly overthrowing our old theories and replacing them with entirely new ones. Like we used to think diseases were caused by miasma or bad air before we discovered microscopic germs. Our look at physics in the 19th century. I mean, the greatest minds in the world believe that light traveled through an invisible mechanical fluid that filled the universe.
14:27The luminous ether, right? The luminous ether, yeah. And then James Clerk Maxwell came along with his electromagnetic field theory and completely annihilated the concept of the ether. Just destroyed it. Completely. The physical entity of the ether was thrown into the dustbin of history. But the underlying mathematical equations that those earlier scientists had painstakingly developed to describe how light reflects and refracts, those equations survived the revolution perfectly intact. The subjective step died, but the underlying structure survived. Exactly. The structural realists argue that the true objective reality we are trying to grasp isn't the physical matter we can touch or the words we can read.
15:08It's the hidden, enduring mathematical relationships, the structures that govern that matter. Here's where it gets really interesting. There is a wild thought experiment by Moritz Schlick, one of the founding figures of logical positivism, that pushes this to its absolute limit. Oh, this one is a trip. It really is. Imagine you go to the hospital for a routine procedure, and a mad surgeon secretly swaps your optic and auditory nerves. A deeply unsettling scenario. Truly horrifying. So you wake up, and suddenly you hear light. You see sound. When a nurse speaks to you, you experience a violent flash of colors.
15:47Oh, wow. When you look out the window at a bright, sunny day, you hear a deafening symphony of noise. Your entire sensory experience of reality is utterly unrecognizable. It is upside down. Your subjective data stream has been completely corrupted. Right. But here's the massive realization. So even in that completely upside down sensory world, if you drop an apple, you might see a sudden sound. But the mathematical acceleration of gravity pulling that apple to the floor would be exactly the same as it is in our world. Yes. The math doesn't care about your sensory experience. The laws of nature you can write down in that upside down hospital room, they would translate perfectly into our universe.
16:28Because the subjective sensations, the words of your experience are totally variant. But the relationships between the events, the structure remains perfectly invariant. Right. And for AI, this is the profound lesson. The truths we want these massive language models to learn are not the observable words themselves. The words are just surface data. Just the superficial stuff. Yeah, the billions of gigabytes of text they scrape from the internet are just the subjective sensations. We need them to learn the hidden, invariant mechanisms that govern the ideas behind the words. So how do we actually do that?
17:06We know what truth looks like philosophically. It's an invariant structure. But how do we pragmatically code that into a machine so it stops hallucinating? Well, let's look at an example. Okay, take a simple data point, the capital of France. If an AI's training data happens to ingest a massive corrupted archive that repeats the phrase Berlin is the capital of France 10 million times. Under the current maximum likelihood paradigm, the AI will confidently, arrogantly output that Berlin is the capital of France because the math favors the high frequency. But we know Paris is the capital and we don't know it just because we've heard it a lot.
17:42We know it because we hear it invariantly across completely independent, structurally distinct channels. Right. We read it in an encyclopedia. We hear it from a geography teacher. We see it on a physical map. If one source screamed Berlin, but every other independent environment pointed to Paris, we would immediately recognize the Berlin claim as an error or as propaganda. We naturally filter for invariance. And this is exactly how developers are proposing we fix AI. Oh, really? Yeah. Instead of just giving the model a standard prompt, like diagnose this medical symptom, you force the AI to mathematically generate different environments or distinct personas.
18:23Okay, so if you give the AI a complex list of a patient's symptoms, instead of just asking for an answer, you force the system to generate the specific perspective of a rheumatologist. Exactly. Then independently generate the perspective of a hematologist, then an endocrinologist. You are forcing the machine to look at the exact same problem from three radically different structural angles. Yes. The AI generates independent analytical chains from all those different personas. And then you introduce a final overriding prompt. What is invariant across all these answers? Oh, wow. You force it to find the structural core.
19:01Exactly. The AI must extract the common mechanism that all three specialists agree on, while actively discarding the subjective artifacts, biases, or localized shortcuts of the individual personas. That makes so much sense. If all three agree on a core biological mechanism, the machine registers that as an invariant truth. If they share no common core, the model is mathematically trained to abstain rather than hallucinating a confident guess. And this relies heavily on forcing the AI into chain of thought reasoning, right? You make the machine show its work step by step. Yeah, that's crucial. Because going back to the cow on the beach, the old classifiers just spat out the label camel, and we had no idea why.
19:42But with a language model, if it looks at an image and says water bird, you can actually inspect its chain of thought. Right. And the AI might expose its own shortcut saying, I have no idea what type of bird this is, but I detect blue water pixels in the background, so I will guess water bird. And by forcing the system to expose its internal reasoning and comparing that reasoning across different persona environments, we can isolate those spurious correlations. We can weed out the shortcuts and isolate the true causal mechanism. I have to stop you there, though, because this feels like a massive, potentially fatal loophole in the architecture.
20:17Uh-oh. Lay it on me. If the AI is generating the rheumatologist persona and the hematologist persona entirely from within its own neural network, isn't it just talking to itself? Ah. Like, it's generating these personas based on the exact same flawed training data that has the hallucination to begin with. It's essentially putting on different fake mustaches, looking in the mirror, and agreeing with itself. isn't still trapped in a text-based echo chamber? That is the nightmare scenario. Right. And I'll be entirely honest, it is a massive unresolved issue that keeps developers up at night. Really?
20:53Yeah. If the underlying data is poisoned with a systemic bias, putting three hats on the model might just give you three perfectly synchronized biased answers. Finding truly independent admissible transformations inside a closed system is incredibly difficult. It requires a bootstrapping process. Meaning we have to build the ladder while we are climbing it. Exactly. Think about how human science operates. We don't possess the ultimate truth on day one. We propose a theory, and then we actively try to destroy it. We push it to its limits under extreme, unusual circumstances. Right. We discover new anomalies, which forces us to create new environments to test, which slowly reveals deeper invariances.
21:35It is a messy, constant loop of mutual adjustment. AI must eventually be designed to do the same. We have to build systems that aren't just predicting text, but are active participants in hypothesis testing, constantly proposing new environments to try and falsify their own internal beliefs. So we are dealing with machines that need to learn how to actively interrogate their own assumptions. We are. But as I'm thinking about this, everything we've discussed, escaping the trap of maximum likelihood, searching for structural truth over superficial frequency, it feels less like a coding manual for AI and more like an instruction manual for us.
22:13Oh, it is deeply relevant to how we process reality. We live in an era of absolute information overload. We are bombarded daily by social media algorithms that are explicitly optimized for engagement and high frequency. Which is basically maximum likelihood in human form. Precisely. This entire exploration is a reminder that we have to be our own invariance filters. Just because a headline is loud, angry, and highly probable within your specific cultural or political bubble does not make it structurally true. We have to look past the frequency of the news we consume. We have to ask, does this outrage, does this fact survive if I view it from a completely different environment?
22:53So what does this all mean? That is a profoundly challenging way to look at the media we consume. Which brings me to a final unscripted thought I want to leave you with today. An experiment for you to ponder long after this deep dive ends. I love these. Think about a deeply held belief you currently have. Something you consider absolutely fundamental to how you view the world. Ask yourself, is it truly an objective fact? Or is it just a highly probable statement that has been repeated millions of times within your specific cultural training data? That's a tough question to face. It is. If you forced yourself to view that core belief through the persona of three radically different worldviews, different cultures, different historical eras, different value systems?
23:37Would your truth survive the transformation? Or would you realize that you've just been staring at a flawless, beautiful counterfeit bill this whole time? That is the ultimate test of invariance right there. It really is. Thank you so much for joining us on this deep dive. Keep stepping back, keep seeking out the view from nowhere, and never stop questioning the probable to find the true.
Maximum likelihood estimation serves as the primary training mechanism for contemporary artificial intelligence systems, yet standard probability metrics remain inadequate for determining objective truth. Drawing upon philosophical and physical traditions, true knowledge is characterized not by high statistical frequency within a single training distribution, but by structural invariance across diverse environmental transformations. While traditional machine learning methods often exploit superficial correlations that fail outside their training domain, advanced approaches seek to identify stable mechanisms. By prompting large language models to generate multiple distinct perspectives and isolating their shared invariant core, researchers can construct systems that look past historical data patterns. Ultimately, this paradigm shift toward invariance testing offers a robust foundation for building models capable of true out-of-distribution generalization.