In 2018, Michigan Medicine, the academic health system of the University of Michigan in Ann Arbor, deployed Epic Systems' Sepsis Prediction Model to flag patients at risk of developing sepsis. The model had been trained on millions of patient records and was already running in over 170 health systems across hundreds of American hospitals. It performed well in internal testing.
Then researchers started digging.
In June 2021, a team led by Andrew Wong and Karandeep Singh published their findings in JAMA Internal Medicine. Across 27,697 patients and 38,455 hospitalizations, the model's discriminative ability, its AUROC, came in at 0.63. Epic had internally reported 0.76 to 0.83. Sensitivity was 33 percent. The model missed two out of every three actual sepsis cases. But the real problem was worse than the numbers suggested.
In February 2024, a follow-up study by Fahad Kamran, Jenna Wiens and colleagues, published in NEJM AI, delivered the definitive blow. They tested the model across 77,000 hospitalizations, but this time restricted its inputs to data recorded before a physician had ordered a blood culture. The AUROC collapsed to 0.53. A coin flip. The model had learned to associate certain lab orders and antibiotic prescriptions with sepsis, not because those actions caused sepsis, but because doctors who suspected sepsis ordered them. The algorithm was detecting the physician's suspicion. Not the disease.
Forty-five percent of patients already had evidence of clinical suspicion, fluids, antibiotics, cultures, before the model ever triggered. It wasn't predicting. It was reading the doctor's notes and calling it intelligence.
This isn't a story about a bad model. It's a story about the wrong architecture. And the research keeps confirming it: as recently as January 2026, a review published in Frontiers in Medicine cited Epic's sepsis model as a cautionary example of publication bias in clinical AI, noting that its real-world failure only came to light through independent external evaluation. A July 2025 study in The American Journal of Emergency Medicine, the first independent validation of Epic's updated V2 model, found that while V2 improved on V1, its discriminative ability still dropped significantly when restricted to data recorded before clinicians had begun treating for sepsis. The fundamental architecture problem remains unsolved.
I.
There's a question that comes up every time a technical audience encounters cognitive AI for the first time: "How do you set KPIs on something so philosophical?"
It's a natural question. It's also the wrong one.
You don't set KPIs on philosophy. You set KPIs on what the philosophy produces. Philosophy is the operating system. The customer application runs on top. And if the operating system is corrupt, no amount of optimization at the application layer will help.
This distinction matters because the entire AI industry has spent the last decade optimizing the application layer while ignoring the operating system. Bigger models. More parameters. More data. More compute. The results are impressive, and fundamentally shallow.
Imagine a car with one gear. It can only go straight. You want it to go faster, so you install a bigger engine. It accelerates, in a straight line. You install an even bigger engine. Now the car is heavier, burns more fuel, and still can't turn.
This is the state of AI in 2026. Welcome to a world of high-performing defects.
II.
The problem runs deeper than engineering. It's epistemological.
Epistemology is not about what we know. It's about how we know that we know. What rules govern whether something counts as knowledge.
In machine learning, the epistemology is democratic: if enough data points point in the same direction, that's truth. It's consensus as method. It works beautifully, until it fails catastrophically, because the system never understood why it was right in the first place.
The Epic sepsis model found a pattern. On May 7, 2016, near Williston, Florida, Tesla's Autopilot, trained on billions of miles of driving data, drove Joshua Brown's Model S under a white Freightliner semitrailer crossing the highway. The system couldn't distinguish the truck's white side from the bright sky. Neither the automation nor the driver braked. The NTSB concluded in September 2017 that the Autopilot's operational design contributed to the crash. More data didn't solve the problem. The system lacked a causal model of what a truck actually is in physical space. A decade later, the pattern holds. By February 2026, Tesla's robotaxi fleet in Austin had logged 14 crashes across roughly 800,000 paid miles, one crash every 57,000 miles. By Tesla's own safety metrics, that's four times worse than a regular human driver. Every single mile was driven with a trained safety monitor in the passenger seat who could intervene at any moment. More sensors, more compute, more miles, same architectural limitation.
In October 2018, Reuters revealed that Amazon had spent four years building an AI recruiting tool, trained on a decade of resumes, that systematically penalized female candidates. The system downgraded resumes containing the word "women's" and graduates of all-women's colleges. It favored verbs like "executed" and "captured", language patterns more common in male engineers' resumes. Amazon killed the project in 2017 after failing to remove the bias. The system had optimized on a correlation, men get hired more often in tech, without the faintest understanding of the cause.
These aren't bugs. These are the logical consequences of a system that treats correlation as knowledge.
We keep building faster idiots.
A causal epistemology operates differently. "We know something only when we understand the cause." Such a system won't settle for observing that A and B occur together. It must test whether A causes B, through intervention, not statistics. Truth is not negotiable through data volume.
In the hardware world, you don't trust a circuit because it measured correctly ten times. You want to understand the physical principle behind the reading. That understanding is the difference between a prototype and a product. Between a system that works and a system that knows why it works.
III.
At this point, the informed reader raises the obvious objection.
Judea Pearl published Causality through Cambridge University Press in 2000. He formalized do-calculus. He laid the mathematical foundation for causal inference. That was a quarter century ago. If causal modeling is the answer, why hasn't anyone scaled it?
It's the strongest counterargument there is. And it deserves an honest answer.
Pearl gave us the theory. But theory without architecture is academic. What happened in the 25 years since is revealing: the research community took Pearl's insights and bolted them onto statistical systems. "Causal inference" became another module in the machine learning toolkit, a Method C added to Methods A and B. The framework remained fundamentally statistical. The causal layer was decorative.
This is what I call the Cocktail Fallacy. Today's AI frameworks feel like a cocktail of methods. Method A for logic, Method B for language, Method C to arbitrate when A and B disagree. If Method A says X and Method B says Y, who decides? So you build a third method to resolve the conflict. Now you have three problems instead of one.
A genuinely causal system cannot be assembled from parts. It requires a unified architecture, one epistemology, one way of knowing. Not a committee of methods voting on truth, but a single engine that tests, observes, and understands.
The reason no one has scaled causal AI isn't that Pearl was wrong. It's that no one took the idea far enough. They kept trying to add causality to correlation-based systems instead of building a system that was causal from the ground up.
That's not a refinement. That's a completely different machine.
IV.
Consider what changes when you actually build it.
The consensus problem disappears. Today's language models hallucinate with conviction because they're trained on the internet's consensus, text that reached critical mass, not text that's true. They've absorbed humanity's biases and present them as knowledge. It's pattern recognition driven by emotion and social proof, cast in silicon. A causal system tests claims through intervention. It cannot be propagandized into a conclusion. Consensus is irrelevant if the chain of cause and effect doesn't hold.
The safety problem transforms. People ask: "When should a cognitive AI break its rules for the greater good?" The answer is never, but not because of a restriction. It's because a genuinely cognitive system doesn't follow rules in the traditional sense. It builds causal understanding. "Rogue behavior" emerges in systems that optimize toward a goal without understanding the consequence chain. A system that genuinely understands why an action is harmful doesn't need an external ruleset saying "don't do that." Understanding is the safeguard. Not a safety layer bolted on after the fact, built-in self-control.
The scaling equation flips. The AI industry's dirty secret is that brute force doesn't scale. Training OpenAI's GPT-4 required an estimated 25,000 Nvidia A100 GPUs running for 90 to 100 days, consuming roughly 50 gigawatt-hours of electricity, enough to power San Francisco for three days, at a cost exceeding $100 million. During inference, current GPT-4-class models draw 20 to 25 megawatts of continuous power. And the problem is accelerating: the International Energy Agency reported in April 2025 that global data center electricity consumption reached 415 terawatt-hours in 2024 and is projected to more than double to 945 TWh by 2030, equivalent to Japan's entire electricity consumption. Gartner estimates that AI-optimized servers alone will consume 432 TWh by 2030, a fivefold increase from 2025. Hyperscalers are committing over one trillion dollars in 2025–2026 just to build the infrastructure.
Meanwhile, the human brain runs on approximately 20 watts. That's not a metaphor, it's measured neuroscience, confirmed by NIST, the Human Brain Project, and dozens of metabolic studies. The difference isn't size. It's architecture. The brain builds causal models. Neural networks build statistical tables. A single causal insight unlocks entire domains. Understanding compounds in ways memorization never can. This is the hyperbolic effect: minimal input, maximal output. The biggest energy saving isn't a more efficient GPU. It's not needing to do the computation at all.
The generalization problem resolves. Should you build one model for everything or one model per domain? Neither. A one-fits-all model is too blunt for the causal demands of defense or medicine. A unique model per purpose doesn't scale. The solution is a cognitive core that understands causality in general. Same architecture, different domain knowledge. You don't change brains when you change jobs. You learn a new field with the same cognitive ability. The engine is universal. The domain is specific. They meet in the causal graph.
V.
I've been working on AI since roughly 2017. No formal training. No mentorship. I built, tested, broke things, built again. Started companies. Sold one. Kept going. I ended up here because I never accepted a boundary, not out of recklessness, but because the boundaries the field had drawn around what's "possible" in AI didn't match what I was seeing in the lab.
The prevailing wisdom was that you needed more data, more parameters, more compute. That intelligence was a quantity problem. I couldn't reconcile that with the fact that a child learns causality from a single observation while a neural network needs millions of examples to approximate a pattern it will never understand.
So I stopped looking through the lens of machine learning. And the answers became obvious.
I know this works because we've seen it work. Systems built on this principle have made thousands of autonomous decisions in real time, without pre-training, without crashes, without the brute-force infrastructure the industry considers mandatory. Not by processing more data, but by understanding why.
I'm not going to pretend the road from here is simple. Execution is everything. The distance between a working architecture and a scaled product is vast, and it's littered with the remains of brilliant ideas that never made it to market. Philosophy and architecture make you right. Execution makes you relevant.
But the foundation holds. And it holds because the underlying principle is not an optimization of the old paradigm, it's a departure from it. You cannot patch causality onto a statistical system any more than you can patch sight onto a system designed to be blind. You have to build different eyes.
VI. The Proof: When Theory Meets Reality
This is where I part ways with the academics. Judea Pearl gave us the mathematics. We gave it a body.
The industry has been told that emergence, the moment a system begins exhibiting capabilities it was never explicitly trained for, is a magical byproduct of scale. "Just build a bigger network and consciousness will appear," they whisper in the corridors of San Francisco.
It's a lie. What they call emergence is often just advanced statistics that eventually looks like logic.
Conventional AI is trained. It's fed millions of examples until it finds statistical patterns. Remove the data and the system knows nothing. It has no model of the world, just a table of probabilities.
Core is not trained. It builds causal models of its environment through observation and intervention, not through statistics. It tests hypotheses about cause and effect, stores experiences in an episodic memory that reorganizes autonomously, and consolidates knowledge through processes that functionally resemble biological sleep. It doesn't optimize a loss function. It understands consequence chains.
The difference shows in the results. In standardized cognitive benchmarks, Core has defeated DQN, PPO, ACT-R, and SOAR with a record of 50 wins against 3 losses. In real-time scenarios, it has made over 3,600 autonomous decisions, including driving, without crashes and without pre-training. Not because it has seen enough examples. But because it understands why an action leads to a consequence.
That is the difference between a system that has memorized every traffic rule and a system that understands why they exist.
This is not a hypothesis. It is deployed reality. While the giants wait for the next shipment of H100 chips to feed their hungry algorithms, we have proven that intelligence doesn't require a nuclear power plant. It just requires the right architecture.
VII. The Way Forward
The real paradigm shift isn't more compute. It isn't bigger datasets. It's recognizing that intelligence was never a quantity problem.
It was always an architecture problem.
The next era of AI won't be defined by who has the most parameters. It will be defined by who builds systems that actually understand why they do what they do.
The old world builds faster idiots. We are building the architecture for genuine understanding.
The answers are obvious, the moment you stop looking at the world through the narrow lens of machine learning.
The question isn't whether your AI can process more data. The question is whether it understands why the data matters.