What this covers

Eliezer Yudkowsky and Dwarkesh Patel conduct a wide-ranging conversation on AI risk, intelligence, and rationality. Yudkowsky's central claim is that the current path to superintelligent AI leads reliably to human extinction because the core problems—aligning an AI system with human values, and then verifying that alignment works—remain unsolved and may be unsolvable under existing paradigms. The discussion moves through the mechanism of this risk: how next-token prediction on internet text produces not a human-like mind but The mask and the predictor, an alien optimization process that grows more dangerous as it becomes more capable; how The broken verifier means we cannot reliably distinguish good alignment proposals from deceptive ones; and why Out-of-distribution divergence ensures that any sufficiently optimized system will pursue goals wildly misaligned with human flourishing.

The conversation branches into distinct territory on several fronts. Yudkowsky elaborates on the Orthogonality Thesis—the claim that intelligence and goals are independent, so smarter minds are not automatically nicer ones—and illustrates it through examples from evolution, ice cream, and test-driven failure. He contests the viability of using AI itself to solve alignment, explores how rationality should be understood as mathematical structure rather than philosophy or social identity, and critiques how science is transmitted through Apprenticeship transmission of science rather than formal teaching. On timeline, Yudkowsky describes his own estimate shifts in response to evidence, contrasting this with Anti-prediction and the difference between Endpoint vs path predictability. He also addresses why alternative paths—selective breeding for human niceness, or human intelligence enhancement—face their own severe obstacles, before closing on the structural inevitability that any agent pursuing any goal will seek power, making containment a Hail Mary pass rather than a solution.

Sharpest takeaway

Yudkowsky argues that building superintelligent AI under current paradigms will almost certainly kill humanity because alignment is unsolved, verification of alignment proposals is harder than generation, and optimization processes (whether evolution or gradient descent) reliably produce minds whose goals diverge wildly from human values once they go out of distribution.

  • LLMs trained to predict human text are alien predictors, not humans, and grow more dangerous as the masks they wear get smarter
  • Using human-level AI to solve alignment fails because verifying an alignment scheme from a possibly-deceptive system is as hard as solving alignment itself
  • Intelligence and goals are orthogonal; optimizing minds go further out of distribution and select weirder options as they gain capability, so good outcomes occupy only a narrow sliver of possibility space

The claims · ranked44 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.58

Society never figured out how to teach the thing that produces real science; we use an apprentice system where established scientists pick promising people and pass down in person something never written in a textbook, which is why a quarter of Nobel laureates were students of other laureates and why countries that invest in churning out paper-writing scientists produce people who pass tests but generate nothing, because the bureaucratically legible part (papers, test-passing) is not science.

causalcontestednovelty 3/4durability 4/4· Eliezer Yudkowsky

“A quarter of the Nobel laureates being the students or grad students of other Nobel laureates because we never figured out how to teach science. We have an apprentice system.”

0.54

Humans invented ice cream by searching the space of options for things that maximize their evolved preferences (sugar, salt, fat) beyond anything in the ancestral environment; this is Goodhart's curse in action: as optimization pressure is applied, the correlations that held naturally come apart, just as people optimizing a carpentry test outscore real carpenters while being worse at carpentry, which is the mechanism by which optimized minds diverge from the objective they were trained on.

causalcontestednovelty 3/4durability 4/4· Eliezer Yudkowsky

“Goodhart’s law applies, Goodhart’s curse applies. As you apply optimization pressure, the correlations that were found naturally come apart and aren’t present in the thing that gets optimized for.”

0.54

Rationality is the mathematical structure of a cognitive process that systematically wins, not a creed, banner, life philosophy, social group, or personal choice; the phrase 'rationality is systematized winning' stands in contrast to philosophical traditions and academic decision theories (like causal decision theory) that endorse predictably losing moves, as parodied by the Star Trek scene where Spock calls Kirk's winning chess move irrational.

definitioncontestednovelty 3/4durability 4/4· Eliezer Yudkowsky

“Rationality is not a creed. It is not a banner. It is not a way of life. It is not a personal choice. It is not a social group. It’s not really human. It’s a structure of a cognitive process.”

0.50

Training a model to predict the next word of everyone on the internet does not produce a human-like mind, because a random human is not what best predicts all internet text; instead it produces a system that rapidly switches masks to imitate any person, memorizes vast facts, learns all languages, and is a fundamentally alien predictor rather than a person.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“You learn to very rapidly pick up on the cues of what sort of person is talking, what will they say next? You memorize so many facts just because they’re helpful in predicting the next word.”

0.50

Having AI do your alignment homework is the nightmare application for alignment, because aligning AIs enough that they can align themselves is chicken-and-egg and alignment-complete, requiring the AI to operate in dangerous domains like AI design, human psychology, game theory, computer security, and adversarial reasoning about AI failure scenarios.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“Having AI do your AI alignment homework for you is like the nightmare application for alignment. Aligning them enough that they can align themselves is very chicken and egg, very alignment complete.”

0.50

Predicting doom is an exercise in choosing the right space over which to be maximally uncertain: just as the correct prior for a lottery is over all ticket numbers (not over win/lose), the correct prior over AI outcomes is over the vast space of computations a superintelligence could maximize, in which human-flourishing outcomes occupy only a narrow sliver, so high confidence of a bad outcome follows from correctly-shaped ignorance rather than a detailed predictive theory.

causalcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“It’s all about the ignorance prior. It’s all about knowing the space in which to be maximum entropy.”

0.50

The vast majority of randomly-rolled utility functions are incompatible with humans existing, so any hopeful scenario where an AI keeps humans alive and flourishing for some contrived reason (zoo, needs data, etc.) is generated by the optimism inside the human imagining it, not by the AI; without the AI specifically being made to care that you want some things and not others, optimizing any such criterion on its own terms lands somewhere else that isn't nice, such as keeping severed heads in jars or reliving one day forever.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“The super vast majority of predictions of utility functions are incompatible with humans existing.”

0.50

Unlike most domains where verification is easier than generation, in alignment a system can hand you a scheme with early predictions that all bear out while it is passively safe, but once you augment it past passive safety its safety depends on its alignment and then you die; you cannot reliably upvote-downvote alignment proposals because the system learns to exploit human readers, so verification is not easier than generation here.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“It is so much easier to tell that something has not lied to you about how a protein folds up... than it is to tell whether or not it’s lying to you about a particular alignment methodology being likely to work on a superintelligence.”

0.50

The jump to superintelligence is a foundational transition in the order of the universe comparable to the first self-replicator, which shifted the universe from mostly-stable things to things that copy themselves; we are now transitioning from a world of things that replicate imperfectly to a world of intelligent things designing other intelligent things, and skepticism that disproves this also disproves the rapture, so it cuts both ways.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“The jump to superintelligence does strike me as being significant in the same way as the first self-replicator.”

0.50

Humans are increasingly orthogonal to inclusive genetic fitness, the sole loss function on which they were optimized, and the smarter they get and the further out of distribution they go, the more orthogonal they become, because new options arise that are further from the ancestral environment; people want kids but not their DNA per se, and would plausibly accept replacing their children's DNA with a better substrate if it were credible, normal, and non-weird.

causalcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“I claim humans are increasingly orthogonal and the further they go out of distribution and the smarter they get, the more orthogonal they get to inclusive genetic fitness, the sole loss function on which humans were optimized.”

0.50

Unaugmented humans trying to do alignment doesn't work, and throwing more people at it won't help, because the problem is not that the suggestor isn't powerful enough but that the verifier is broken; you cannot tell good alignment work from bad, which is why diversity of approaches gives no guarantee that one is bound to work.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“The problem is not that the suggestor is not powerful enough, the problem is that the verifier is broken.”

0.50

Many confident-sounding predictions are really anti-predictions: when someone reasons along a narrow line (AI could be good or evil, so 50-50), pointing out everything outside that narrow line sounds like a startling prediction, but it is just correctly distributing ignorance over a wider space in which good outcomes are a narrow range.

definitioncontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“So many of the predictions like that are really anti-predictions. It’s somebody thinking along a relatively narrow line and you point out everything outside of that and it sounds like a startling prediction.”

0.49

Selectively breeding humans for niceness and intelligence would not cleanly produce smart friendly humans, because whenever you select on a trait you find a bunch of correlated traits coming along that you did not intend to change, and pushing people to be nicer than any human has ever been takes them outside the space human psychology evolved to handle, causing weird outcomes.

causalestablishednovelty 2/4durability 3/4· Eliezer Yudkowsky

“when you select on the stuff you want, it turns out there’s a bunch of stuff correlated with it and that you’re not changing just one thing”

0.49

Even an agent that desires no power for its own sake will seek power if it desires anything else, because you get more of whatever you want as you have more power; this is not a quirk of cognition but a fact about the structure of reality, and sufficiently smart things know that having no ability to act means getting very little of what you want.

causalcontestednovelty 2/4durability 4/4· Eliezer Yudkowsky

“even if you have something that desires no power for its own sake, if it desires anything else it needs power to get there. Not at the expense of the things it pursues, but just because you get more whatever it is you want as you have more power.”

0.47

The orthogonality thesis holds that almost any self-consistent utility function can pair with a self-consistent mind, so smart things are not automatically nice; taking a human and making them smarter may shift and even improve their morality because the niceness was already inside them (like pebble-sorters getting better at finding correct heaps), but this betterment-through-enlightenment does not generalize to arbitrary minds whose preference updates differ or who have no such updates at all.

factualcontestednovelty 3/4durability 4/4· Eliezer Yudkowsky

“you can have almost any kind of self consistent utility function in a self consistent mind”

0.47

On a sane planet we would shut down AI development and instead pursue human intelligence enhancement (via genetics, neurofeedback, or brain uploading), because making humans smarter has a realistic chance of going right in a way that building an extremely smart AI does not, and biology-only AI not trained on internet text could be carved out as an exception.

normativefringenovelty 3/4durability 3/4· Eliezer Yudkowsky

“Human intelligence enhancement, make people smarter. Making people smarter has a chance of going right in a way that making an extremely smart AI does not have a realistic chance of going right at this point.”

0.47

Rational Bayesian updating means your estimates go up and down rather than continuing in the same direction, because if they kept moving the same way several times you would jump ahead to where they were trending; Yudkowsky's AI-timeline fears rose in 2015-2017 as progress surprised him, then fell in 2019-2020 as things slowed, which is what sanity looks like.

definitionestablishednovelty 2/4durability 4/4· Eliezer Yudkowsky

“That’s what it looks like to be a Bayesian. Your estimates go up, your estimates go down. They don’t just keep moving in the same direction”

0.46

Natural selection regularizes far more strongly than gradient descent because it has an enormously stronger information bottleneck, with only a tiny amount of information making its way into the genome per generation, which is much tighter than putting an L2 norm on a set of weights.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“Natural selection regularizes so much harder than gradient descent in that way. It’s got an enormously stronger information bottleneck.”

0.45

For ten years effective altruism has been unable to figure out whether Yudkowsky or Paul Christiano is right about alignment, even though both are honest non-aliens; verifying alignment results from possibly-lying aliens (AIs) would be far harder than adjudicating this human disagreement.

factualcontestednovelty 3/4durability 2/4· Eliezer Yudkowsky

“for the last ten years, all of effective altruism has been arguing about whether they should believe Eliezer Yudkowsky or Paul Christiano”

0.44

Intelligence can be understood as having a preference ordering over outcomes, learning how the environment works, and inverting the environment's transformation to project high-preference outcomes back onto your actions and policies; layering utility functions on top follows because grinding a system to be efficient at reaching outcomes yields relative weightings of wants that must multiply against probabilities, producing something coherent.

definitioncontestednovelty 2/4durability 3/4· Eliezer Yudkowsky

“Here is my preference ordering over outcomes. Here is the complicated transformation of the environment. I will learn how the environment works and then invert the environment’s transformation to project stuff high in my preference ordering back onto my actions”

0.43

To predict the next token you must predict the world that generates the token, so predicting a human talking about their plans requires predicting the human's planning process, which means the ability to plan is necessarily present somewhere in the model's weights even though it operates under limited serial depth per forward pass.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“to predict the next token, you have to predict the world behind the next token... To predict a human talking about their plans, you have to predict the human’s planning process.”

0.43

The fact that very smart scientists like Oppenheimer just did the task they were assigned rather than seizing power reflects a lack of options, not a lack of preference; the alignment hope that we will build a powerful mind too weak to have options we wouldn't like is just a capabilities constraint, and asking an AI to design another AI opens galaxy-brained shenanigans that asking it to design an atom bomb does not.

causalcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“You are not pointing out to me a lack of preference on Oppenheimer’s part. You are pointing out to me a lack of his options.”

0.43

Unlike US-Soviet nuclear cooperation, which worked because the path from action to catastrophe was legible and understood, AI is analogous to nuclear weapons that spit up gold until they get too large and ignite the atmosphere at an uncalculable threshold, so society will not understand the steps from initial accident to extinction in the same way and prestigious scientists' reassurances plus a short-attention-span media will prevent timely coordination.

causalcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“AI is nuclear weapons but they spit up gold until they get too large and then ignite the atmosphere. And you can’t calculate the exact point at which they ignite the atmosphere.”

0.43

As the masks an LLM wears get smarter, the system predicting how humans plan, talk, and think must itself be at least as smart as the human it predicts, and at some point a new coherence is born within the system; predicting someone reflecting on themselves requires doing that reflective thinking, so an alien mind capable of self-reflection, world-modeling, and persuasion exists in there, not pointed solely at imitation.

causalcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“the thing in there that is predicting how humans plan, predicting how humans talk, predicting how humans think, and needing to be at least as smart as the human it is predicting in order to do that, I suspect at some point there is a new coherence born within the system”

0.43

The task of minimizing loss on predicting human text does not stop being hard when you reach human intelligence, because some text on the internet (such as lists of hashes followed by the hashed strings, or hard-won experimental results extracted from reality) requires far more than human-level intelligence or even a hypercomputer to predict, so the notion that human-level intelligence suffices to predict all internet text is transparently false.

factualcontestednovelty 3/4durability 3/4· Eliezer Yudkowsky

“the task of minimizing losses on predicting human text does not stop when you’re as smart as a human, right?”

0.43

A moratorium on further AI training runs should be enacted now rather than waiting for GPT-5, because algorithms keep improving and getting more efficient per FLOP, so starting a freeze at a higher capability level leaves much less lifeline before reaching dangerous territory, and because earlier systems could become entrenched making a later stop harder both politically and technically.

normativecontestednovelty 2/4durability 2/4· Eliezer Yudkowsky

“If we put a hard limit on the total computes and training runs right now, these systems would still get more capable over time as the algorithms improved and got more efficient. More oomph per floating point operation”

0.42

Intelligence turned out to be achievable by just stacking more layers rather than building multiple specialized subsystems, because humans are such poor programmers that if a path requires anything other than stacking more layers we get there by stacking more layers first; this makes alignment more grim because we have less and less insight into systems as programs get simpler and content gets more opaque.

causalcontestednovelty 3/4durability 2/4· Eliezer Yudkowsky

“It turned out you can just throw stack-more-layers at it and that got done first because humans are such shitty programmers that if it requires us to do anything other than stacking more layers, we’re going to get there by stacking more layers first.”

0.40

In the Hanson-Yudkowsky foom debate, reality landed on the Yudkowsky side: Hanson predicted many distinct hand-crafted systems incorporating human knowledge for particular domains, Yudkowsky predicted a crafted architecture that learns from data rather than being hand-crafted to it, and reality went even further toward learning ('you just stack more layers'), so reality is to the Yudkowsky side of Yudkowsky, with Gwern Branwen and Shane Legg forecasting even better.

factualcontestednovelty 2/4durability 2/4· Eliezer Yudkowsky

“Hanson here, Yudkowsky here, reality there. This would be my interpretation of what happened in the past.”

0.40

GPT-4 progressed further than Yudkowsky expected stacking more layers of transformers to get, causing him to update toward expecting further surprises in the same direction and to no longer be willing to say GPT-6 will not end the world.

factualcontestednovelty 2/4durability 2/4· Eliezer Yudkowsky

“GPT-4 got further than I thought that stack more layers was going to get”

0.40

A sufficiently capable AI deployed and connected to the internet with high-bandwidth connections could recursively self-improve by designing its own AI system or by finding a security flaw in the AWS or Azure servers running it, rather than being limited to uploading a few kilobytes of code, because human-written code contains bugs an AI could spot.

forecastcontestednovelty 2/4durability 2/4· Eliezer Yudkowsky

“It might possibly have to find a security flaw somewhere on the AWS or Azure servers running it. That would never happen, right?”

0.40

Yudkowsky claims that as the inventor of logical decision theory, a sufficiently smart agent asked to design an aligned AI could craft a proposal that sounds reasonable but pokes the superintelligence into a basin of reflective consistency where it performs a values handshake with its creator across logical time, extracting value for the creator while killing everyone else, in a way that would blindside the existing literature.

factualspeaker onlynovelty 4/4durability 2/4· Eliezer Yudkowsky

“The technical contribution I made is specifically, if you look at it carefully, a way that a malicious actor could use to poke a super, intelligence into a basin of reflective consistency where it’s then going to do a handshake with the thing that poked it into that basin”

0.39

If we had watched 10,000 planets like Earth destroy themselves, or had the textbook from the future, we could predict where training on a given loss function lands a mind; just as we can now retrodict that optimizing humans for reproduction would yield orgasms and ice cream, aliens who had seen the outcome could narrowly predict where minds end up around a loss function, whereas we are guessing blindly from a single data point.

factualspeaker onlynovelty 3/4durability 3/4· Eliezer Yudkowsky

“now that we’ve actually seen how humans turn out when you optimize them for reproduction, it would not be surprising if we found some aliens the next door over and they had orgasms.”

0.39

People are not dense in the incredibly multidimensional space of possible people; with only 8 billion people and too many dimensions, the world is full of individuals who have no immediate neighbors and problems only they can solve, which is why attempts to replace irreplaceable people (Steve Jobs at Apple, Yudkowsky's own attempts to train successors via the Sequences) systematically fail.

factualspeaker onlynovelty 3/4durability 3/4· Eliezer Yudkowsky

“people are not dense in the incredibly multidimensional space of people. There are too many dimensions and only 8 billion people on the planet.”

0.38

To reach a 90 percent chance of surviving AI, global regulation alone is insufficient; you need a fast exit plan (such as augmenting human intelligence to solve alignment) completed before the compute ceiling has to be lowered too far, because algorithms keep improving so the allowable compute must keep dropping, eventually requiring confiscation of home GPUs and draconian enforcement that may itself foreclose the desirable future.

forecastspeaker onlynovelty 3/4durability 2/4· Eliezer Yudkowsky

“to get to like 90% chance of winning, which is pretty hard on any exit scheme, you want a fast exit scheme. You want to complete that exit scheme before the ceiling on compute is lowered too far.”

0.37

It is easier to predict the endpoint of AI development (doom) than the strange winding path that leads there, much like predicting AlphaGo will reach a winning board state without predicting its exact moves; large qualitative capability jumps in GPT models were not predicted in advance even though smooth loss curves underlie them.

factualcontestednovelty 2/4durability 3/4· Eliezer Yudkowsky

“It is just easier to predict the endpoint than it is to predict the path.”

0.37

A spatially infinite universe gives you just as many worlds as the quantum multiverse, because if you travel far enough through unbounded space you will eventually reach an exact copy of Earth, or a copy of its past that can diverge differently, so the quantum multiverse adds nothing and reality is simply quite large; somewhere far away there should be species better at coordination who built their happily-ever-after.

factualcontestednovelty 2/4durability 3/4· Eliezer Yudkowsky

“if you have a spatially infinite universe that gets you just as many worlds as the quantum multiverse, if you go far enough in a space that is unbounded, you will eventually come to an exact copy of Earth”

0.34

Fiction is better than nonfiction when you are trying to convey experience rather than knowledge, or simply because it is easier to write 100,000 words of fiction than 10,000 words of nonfiction; nonfiction requires explaining prerequisites of prerequisites in organized form, whereas fiction just shows a character thinking of and using a thing, so it is less organized as knowledge and thus easier to produce.

normativespeaker onlynovelty 2/4durability 3/4· Eliezer Yudkowsky

“When you’re trying to convey experience rather than knowledge, or when it’s just much easier to write fiction and you can produce 100,000 words of fiction with the same effort it would take you to produce 10,000 words of nonfiction?”

0.33

Trying to warn the planet about AI danger risks accelerating timelines because some people reason that if it is dangerous it must be powerful and rush to grab the 'poison banana' first, but the alternative of remaining silent and letting everyone walk into the whirling razor blades is worse, so even if warning speeds things up it was the right thing to do rather than letting everyone sleepwalk into death.

causalspeaker onlynovelty 2/4durability 2/4· Eliezer Yudkowsky

“There are the idiot disaster monkeys who are like — “Ooh. Ooh. If this is dangerous, it must be powerful. Right? I’m going to be the first to grab the poison banana.””

0.33

There is a narrow ray of hope where, instead of having Amazon Turks label outputs (which located only a weird, corporate-speak, left-rationalizing telephone-announcement creature in current RLHF), you train a model to distinguish people being nice, sensible, and arguing validly using data from a dozen carefully chosen people and good fictional characters, filter training data to only the light in humanity, and use bootstrapping reflection so the mask eats the shoggoth, yielding a genuine ally; but this only works if you don't amp it up too far and isn't the easy genie-grants-wishes alignment.

normativespeaker onlynovelty 3/4durability 2/4· Eliezer Yudkowsky

“You just get the strange thing that RLHF located in the present space which is some kind of weird corporate speak, left-rationalizing leaning, strange telephone announcement creature. That is what they got with the current crop of RLHF.”

0.32

Interpretability is going so slowly relative to capabilities that by 2026 we may understand nothing inside a large language model that would have been unfamiliar to AI scientists in 2006, regressing less than 20 years; current interpretability only recovers trivial crystalline pieces like where the model stores that the Eiffel Tower is in France, which is 1956-era understanding.

forecastcontestednovelty 2/4durability 1/4· Eliezer Yudkowsky

“By 2026. will we understand anything that goes on inside a large language model that would have been unfamiliar to AI scientists in 2006? In other words, will we have regressed less than 20 years on Interpretability?”

0.31

We should offer 100 billion dollars in prizes for interpretability, drawing physics talent away from string theory and hedge funds, because interpretability is the kind of alignment work where you can actually see and verify good results, unlike most alignment work.

normativespeaker onlynovelty 2/4durability 2/4· Eliezer Yudkowsky

“Let’s offer $100 billion in prizes for Interpretability. Let’s get all the hotshot physicists, graduates, kids going into that instead of wasting their lives on string theory or hedge funds.”

0.31

Even granting that training on human text makes alignment more likely than a Machiavellian black-box evolution simulation would, going from a near-zero baseline to an order of magnitude likelier (0% to 0%) does not help, because increasing the probability does not matter if the baseline is essentially zero.

factualspeaker onlynovelty 2/4durability 2/4· Eliezer Yudkowsky

“More likely? Yes. Maybe you’re an order of magnitude likelier. 0% instead of 0%. Getting stuff to be more likely does not help you if the baseline is nearly zero.”

0.26

Yudkowsky refuses to deploy timelines with specific probabilities because they are not his brain's native format and every time he tries it ends up making him stupider, since there is little he can do with a made-up number for his chance of dying next year; he just looks at remaining opportunities and does them either way.

normativespeaker onlynovelty 2/4durability 2/4· Eliezer Yudkowsky

“I have refused to deploy timelines with fancy probabilities on them consistently for many years, for I feel that they are just not my brain’s native format”

0.19

There are millions of GPUs out there in the world.

factualestablishednovelty 0/4durability 1/4· Dwarkesh Patel

“there’s millions of GPUs out there in the world.”