Google’s Quantum Chip Revealed a Pattern That Shouldn’t Exist — Then It Was Taken Offline

Added: 2026-05-12

[00:00:00]Willow is Google's newest and most powerful superconducting quantum computing chip and the next step in our path towards building largecale quantum computers and exploring their applications.

[00:00:11]>> A quantum computer doesn't look or act at all like the computers that we use every day. It relies on different laws of physics to run calculations at super fast speeds. this new Google quantum computer that can do essentially the way a quantum computer works. A a a problem that would take thousands of years for every computer on Earth. In December 2024, Google's Willow quantum chip did something its own engineers could not explain. A pattern appeared in the data that by every rule of physics the team trusted should not have been there. It wasn't a glitch. It wasn't noise. The chip was behaving in a way no machine on Earth is supposed to behave. Within hours, a decision was made inside Google Quantum AI that almost nobody has reported. They took Willow offline. What that pattern actually was and why a room full of physicists decided the safest move was to shut the machine down is the part you were never told. Built to break, but it didn't Let me take you back to the paper that started it.

[00:01:16]December 2024, Google publishes a result that sends a shock wave through the scientific community. Willow has achieved something physicists had been chasing for more than two decades. Below threshold error correction at scale.

[00:01:29]That phrase sounds technical, and it is.

[00:01:32]But what it actually means is this.

[00:01:34]Every quantum computer ever built has carried one fatal flaw. The more powerful you try to make it, the more cubits you add, the more errors it produces. Quantum systems are fragile.

[00:01:45]They fall apart. They leak information into their environment through a process called decoherence. It is like trying to whisper a secret across a crowded stadium. The louder the crowd, the harder it gets. That was the wall. Every team hit it. Nobody had walked through it. The more cubits Google's team added, the fewer errors the chip made. That is not supposed to happen. Think about what that means for a second. You are not just making a better version of a known machine. You are watching a system behave in a direction the physics textbook said was closed. Dr. Nevin chose his public words carefully. He said the chip could solve certain problems faster than any classical computer. Measured controlled is the first technology that takes the idea serious that we live in a multiverse.

[00:02:30]>> The kind of statement a VP of engineering makes when he does not yet want to say more. But inside the Santa Barbara facility, what his team was actually watching did not fit inside clean scientific language. The system was correcting its own errors without being explicitly told how. Here is where it gets unsettling. You build a machine.

[00:02:49]You give it rules. You program it with protocols. And then it starts handling problems using methods you never wrote into the code. Not randomly, efficiently. For engineers who had spent years running these experiments, that was the moment their reassuring little mental model of the project broke. Quick detour because the rest of the story does not land without it. Classical computers use bits, a zero or a one.

[00:03:16]Quantum computers use cubits. A cubit can be a zero, a one or both at the same time. That both at once state is called superposition. And it is where quantum computers get their power. But superp position is impossibly delicate. The moment a cubit interacts with its environment, heat, vibration, a stray electromagnetic signal, the quantum state collapses. That is decoherence.

[00:03:41]For decades, it has been the immovable law of this field. Every serious team had accepted it not as a bug, as a property of the universe. The architecture used something called a surface code. It spreads quantum information across many cubits in a geometric pattern. So errors can be detected indirectly without collapsing the quantum state itself. Surface codes were not new. What was new was the result at scale. As Willow's cubit count climbed from small test configurations up to 105 cubits. The error rates did not climb with them. They dropped. Each new cubit added to the system made the whole structure more stable, not less.

[00:04:20]Nevin's team ran the numbers. They ran them again and again. The result held.

[00:04:26]One engineer described the feeling as looking at a tower that gets stronger the taller you build it. Every construction instinct says that is wrong. Every prior experiment says it should be wrong. Willow said otherwise.

[00:04:39]And this is the part I want you to sit with for a second. Because if you are the kind of person whose shoulders just dropped about 2 in while I was explaining that, if something in the back of your mind just said, "Wait, this is the stuff I have been trying to find real reporting on." You are going to want to stay with this channel. Tap subscribe and hit the bell because the next section is where the physics starts to quietly come apart in Dr. Nevin's hands and you are going to want to be here when it does. The team had walked through a wall that was not supposed to have a door and nobody in that room, not Neven, not the postocs, not the principal investigators knew yet what was on the other side. Five minutes versus the age of the universe. Here is how they found out. Google's team ran Willow on a benchmark called random circuit sampling. Think of it as a test problem deliberately engineered to be impossible for classical computers and tractable for quantum ones. They wanted a clean number, a verifiable number, something they could publish. The number came back under 5 minutes. The same task on Frontier, the fastest classical supercomputer on Earth, housed at Oakidge National Laboratory in Tennessee, would take 10 septillion years.

[00:05:54]>> Having lots of node failures, the network that's always a challenges on these the code itself was not behaving the way we expected.

[00:06:03]>> That number does not fit inside human intuition. It is longer than the current age of the universe by a factor so large it barely has meaning. The universe is about 13.8 billion years old. Willow handled in 5 minutes a computation that would take Frontier trillions upon trillions of times longer than the universe has existed. Now, here is where it stops being a performance story and starts being something else. When you run a computation that takes 10 septillion years on classical hardware in under 5 minutes on quantum hardware, you have to ask a question that makes physicists deeply uncomfortable. Where is the computation actually happening on a classical computer? Easy answer. It happens in the transistors. It happens in the chips. It happens in the circuits directly in front of you. Every step of the calculation has a physical location.

[00:06:56]Every operation leaves a traceable path through the hardware. On Willow, that answer breaks. Random circuit sampling works by firing a sequence of quantum operations across every available cubit simultaneously. The circuit is deliberately randomized which makes it nearly impossible for classical hardware to simulate because a classical computer has to track every possible outcome one at a time. The number of possible states grows exponentially with each cubit added at 105 cubits. That number has over 30 digits. Frontier capable of more than a quintilion calculations per second cannot keep up. Not because Frontier is slow. The problem space is so vast that no classical architecture can navigate it inside any time frame that touches human existence. Willow navigated it in the time it takes to drink a cup of coffee. So here is the question. Where? David Deutsch, one of the founding theorists of quantum computing, has argued for years that quantum computers perform parts of their calculations in parallel universes in other branches of reality under the many worlds interpretation of quantum mechanics. The argument is straightforward. If the computation is not happening here in the physical cubits you can see and measure, then it has to be happening somewhere. Quantum mechanics taken seriously says that somewhere is other branches of the universal wave function. That is not a fringe position. That is a Copenhagen trained physicist's worst conversation.

[00:08:27]Google's team did not make that claim publicly. They did not have to. What they had was a result. Correct.

[00:08:34]Independently verified. Reproduced across multiple experimental runs. a computation that finished in under 5 minutes and that no other machine on Earth could replicate in any timeline that connects to human life. The answer came back correct. The path it took to get there is not something our current physics can fully trace. And this is where inside the facility in Santa Barbara, the mood quietly changed. The team was not standing in front of a chip that had simply run fast. They were standing in front of a chip that had returned a verifiable answer from a computational space. their own hardware could not in any classical sense contain the Copenhagen interpretation. The mainstream view that says quantum systems simply sit in undefined states until somebody measures them does not have a satisfying answer for where 10 septillion years of computation goes in 5 minutes. The many worlds interpretation does. Nobody on the team wanted to reach for it out loud. But they were all thinking it. And then the chip did something else. The answer, nobody programmed. During extended testing, Willow started doing something the team had not programmed and could not fully explain. Not a crash, not a drift, something quietly coherent. The system was finding solutions through pathways that were genuinely novel, not faster versions of existing algorithms, structurally different approaches to problem solving, approaches that appeared to emerge from the quantum substrate itself. To understand why this is serious, remember how Willow corrects errors.

[00:10:08]>> And today, on behalf of our amazing team, I'm proud to announce Willow.

[00:10:14]Willow is Google's newest and most powerful superconducting quantum computing chip. The surface code spreads quantum information across many physical cubits arranged in a geometric pattern.

[00:10:25]When an error occurs, the code identifies it and corrects it. But here is the critical part. It does this without directly measuring the quantum state. Because in quantum mechanics, the act of measurement collapses the system.

[00:10:38]You cannot look at it without changing it. So, Willow corrects errors sideways.

[00:10:43]It measures the relationships between QITs, not the cubits themselves. It infers what went wrong from the surrounding pattern and it fixes it.

[00:10:51]That is already remarkable engineering.

[00:10:53]But during extended runs, the team was watching something past the architecture they had designed. Willow was optimizing its own quantum state management in real time. The specific algorithms for this were not explicitly written. They emerged from the interaction between the hardware and the error correction protocols. As if the system was learning mid computation how to protect its own information better. Now, here is where it gets strange. Every program runs instructions a human wrote. Step one, step two, step three. Even the most complex AI systems are executing mathematical operations that engineers designed, tested, and approved before deployment. There is always a chain between what a programmer wrote and what the machine does. You can trace it.

[00:11:40]Willow broke that chain. Not completely.

[00:11:43]Not in a way that suggested the system had goals or intentions. Let me be careful about that because this is where internet folklore usually barrels in and ruins a real story. But in a specific, measurable, reproducible way. The system was handling error correction through methods that were not in the original code. The surface code gave it the tools. The way it was using those tools during extended computation runs was not fully specified by anyone. The Google team ran diagnostics. They checked for bugs. They checked for measurement errors. They checked for artifacts in the data. The results held. What they were seeing was not noise. It was structure. The system was producing efficient unexpected behavior. The novel pathways it found were not worse than the programmed ones. In several cases, they were better. Let that sit. A machine that malfunctions is a problem you fix. A machine that improves on its own design, even in a narrow domain, is a question you answer. And that question does not have an engineering answer. It has a theoretical one. It sends you back to the foundational models of how quantum systems behave and asks where the gap is between prediction and observation. One researcher on the team put it plainly in internal notes later referenced in follow-up papers. The system was doing something real, something consistent, something they had not taught it. For a group of scientists trained to be precise about language, that sentence carried weight. And this is where I want you to picture Nevin again. Not in a press photo. In that lab, he is looking at a dashboard that shows his chipsolving problems through procedures he cannot find in his own source code. Using the physical properties of quantum mechanics in ways his theoretical model did not predict.

[00:13:29]He turns to one of the lead engineers.

[00:13:31]He does not raise his voice. He asks how long this has been happening. Whatever the answer was, it was enough. Because that was the moment the pause became a decision. The team recognized they were watching something they did not have a complete theoretical model for. In science, when your experiment outpaces your theory, you stop. You document. You build better frameworks before you go further. The chip had given them more than they came looking for. What they did not yet know was that Willow had not shown them the strangest thing in the building that was still coming. Two black boxes, one machine. Here is the layer of this story that almost nobody has reported. At the time of Willow's breakthrough, Google's quantum team was not only running benchmark tests. They were quietly exploring a question sitting at the intersection of the two most powerful technologies humanity has ever built. What happens when you combine quantum computing with artificial intelligence? Traditional AI, the large language models, the image generators, the recommendation systems, all of it runs on classical hardware.

[00:14:37]Billions of parameters, trillions of training tokens, thousands of graphics processors burning electricity day and night. Powerful, yes, but it hits walls.

[00:14:47]Optimization problems too complex to brute force. Drug molecule simulations that would require more classical compute than exists on the entire planet. Quantum computers can in theory cut through some of those walls. A neural network learns by adjusting millions, sometimes billions of internal values called weights. That search is called optimization and it is computationally brutal. Classical hardware handles it through brute repetition.

[00:15:17]Run the calculation, adjust the weights, run it again millions of times. Training a single large AI model today can cost tens of millions of dollars in compute alone.

[00:15:31]Quantum computers approach optimization differently. Because cubits can exist in superposition, representing many possible states simultaneously, a quantum system can in principle explore a much larger solution space in a single pass. Instead of testing one configuration at a time, it evaluates many at once. Google's team was testing exactly that. Whether Willow could find better solutions to optimization problems faster than classical hardware, whether the novel problem-solving behavior the chip had already shown on benchmarks would transfer to real machine learning tasks. Early results were promising enough to keep the investigation running. And this is where it became genuinely uncomfortable. When a quantum computer is used to train or run an AI system, you now have two black boxes nested inside each other, the AI itself already operates in ways researchers cannot fully interpret. You can watch what a neural network outputs, but tracing why it produced that specific output is an open research problem called interpretability. It remains unsolved even for classical AI systems. Your bank runs on AI nobody can fully explain. Your social media feed runs on AI nobody can fully explain.

[00:16:45]That is already the world you were living in. Now add quantum computation underneath it. The AI does not fully expose how it reached its answer. The quantum computer does not fully expose the path it took to compute it. The two opacities stack. What comes out the other end is a system that produces correct verifiable results through a process its builders cannot audit.

[00:17:08]Neither piece is dangerous in isolation.

[00:17:11]Classical AI is already deployed at massive scale. Interpretability gaps and all. Quantum computers on their own solve specific mathematical problems without any agenda. But the combination is new territory. An AI that learns through quantum computation is not just faster. It potentially discovers solutions a classically trained system would never find because it is searching a fundamentally different and larger solution space. what it learns and how it learns it may differ in ways that do not show up in the output. That is not a conspiracy. It is a known documented challenge in computer science taken seriously by researchers at Google, at Deep Mind, and at academic institutions across the world. And it becomes exponentially harder when quantum hardware enters the picture. Nevin's team saw this coming early. The conversations inside the research group shifted. Not just does it work, but do we understand what it is doing well enough to scale it responsibly? Those are different questions. The honest answer to the second one was no. So they pulled the chip offline for that line of investigation. Not out of fear, out of the kind of intellectual discipline that separates serious science from the race to publish. You do not keep pushing forward when something works in ways you do not fully understand. You halt. You build better tools. Then you decide whether to proceed. What is confirmed and what is still open. Let me separate two things because this is where a lot of coverage of this story quietly falls apart. There is what Willow has verifiably done. And there is what Willow has opened up as a live unanswered question in physics. Both are real. Both matter. They are not the same. Here is what is confirmed. Willow achieved below threshold error correction. published, peer-reviewed, reproduced. Willow completed random circuit sampling in under five minutes on a problem that would take Frontier 10 septillion years. Also confirmed, the chip demonstrated problem-solving behavior that emerged from its architecture rather than its explicit programming documented internally, referenced in follow-up work, consistent across runs. Those are the facts on the table. None of them are small. Any one of them would have been a career-defining result. Willow produced all three in the same year. Now, here is what is still open. Nobody can yet fully explain where a 10 septillionyear computation actually executes in 5 minutes. The Copenhagen interpretation, the mainstream one, says a quantum system exists in undefined states until measured, collapses on observation, and the question of what was happening before measurement is not a physical question. That has been the orthodox answer for most of the history of quantum mechanics. It is not good enough for this because here you have a verifiable output correct reproducible and you have a process that by any classical accounting cannot have happened inside the hardware in the time it took. The many worlds interpretation Deutsch's view gives you a physical answer. The computation was distributed across branches of the universal wave function. That answer is mathematically consistent with quantum mechanics.

[00:20:26]Nobody inside Google has publicly endorsed it. Nobody has publicly ruled it out either. The silence is itself the story. Second open question. Why did the error correction improve with scale? The theoretical models of decoherence say it should degrade. The experiment says it does not. That is a direct conflict between prediction and observation.

[00:20:48]Conflicts like that eventually force a revision of the underlying theory. Which part of the theory gives? Nobody yet knows. Third, what exactly was Willow doing when it found those novel problem-solving pathways? The team has the outputs. They have the performance metrics. They do not yet have a complete mechanical description of the process.

[00:21:09]Until they have that description, scaling further is not a question of engineering willpower. It is a question of whether you understand what you are scaling. This is why the pause was not cautioned for the sake of optics. It was the only responsible move. You do not keep adding cubits to a system already showing you behavior past your theoretical model. You do not couple it to AI optimization pipelines while the interpretability problem is unsolved.

[00:21:34]You do not release it to external researchers until you can describe what it is doing in language that survives peer review. And this is where Willow actually sits in 2026. Not as a debunked sensational headline, not as a consumer product rolling into your laptop next quarter. Willow sits as a machine that produced results the field's best models did not fully predict. Taken deliberately offline for the experimental lines that outran the theory while its team builds the framework needed to go further. The pause is not a climb down from the title. The pause is the title. A pattern that should not exist. A machine taken offline. Both true, both documented, both still unresolved at the theoretical level. The near-term applications, if the pause resolves the way Google hopes, are not small. Drug discovery, simulating molecular behavior at the quantum level, something classical computers cannot do at meaningful scale.

[00:22:29]Cancer treatments, antibiotic resistance, protein folding, material science, new superconductors, better solar cells, and cryptography, the uncomfortable one. Current encryption standards rest on mathematical problems hard for classical machines and potentially trivial for a sufficiently powerful quantum one. Willow is not there yet, but the trajectory is clear.

[00:22:53]That is why NIST finalized its first postquantum cryptography standards in 2024. The threat horizon stopped being theoretical. The chip is offline for now. The questions it opened are not the bigger question. Willow left unanswered.

[00:23:09]For decades, physicists worked under a set of assumptions about quantum systems. They are delicate. They decoher quickly. They need near absolute zero and near perfect isolation. Scaling them is engineering hell. Willow broke several of those assumptions at once.

[00:23:26]And in physics, when assumptions break, you do not revise an equation. You go back and ask what else you got wrong. If quantum systems can scale more gracefully than the models predicted, if error correction can improve rather than degrade as you add cubits, then the theoretical picture of quantum decoherence is incomplete. Which means the understanding of how quantum information behaves in large systems is incomplete. Which means the question of what quantum computers are actually doing when they compute is still open at the deepest level. Deutsche's many worlds suggestion is not speculation from a sci-fi magazine. It is a legitimate interpretation with serious mathematical backing offered by one of the founders of the field. The Copenhagen interpretation does not give a satisfying answer for where 10 septillion years of computation goes in 5 minutes. Something has to give. The team running that lab is made up of physicists, not philosophers, not screenwriters. people who spend their days running controlled experiments and publishing peer-reviewed papers. When people like that start quietly asking what parallel computational structures might explain their results, you are in territory that deserves careful attention. Willow has been pulled offline for the experimental lines that got past the theory. That halt is not an ending. It is a gate. Behind that gate is a question nobody, not Deutsch, not the physicists at Oak Ridge who built Frontier can yet answer. What is the universe actually doing when a machine like this computes? Where is the work getting done? If the answer is what some of the most serious minds in the field now quietly suspect, the next chapter of physics is going to be the most disorienting one humans have written since we realized the Earth was not the center of anything. Here is the part I want to hear from you on. Do you believe Willow was tapping into something we cannot yet explain with current physics?

[00:25:21]Or is this just better engineering, better math, better luck? Drop your answer in the comments. I read them. And if you want the next chapter of this story, the one being written right now in labs you have never heard of, subscribe because we are going there next.