What this covers
Adam Brown and Dwarkesh Patel explore how much of what appears to be fixed cosmic destiny—entropy's inevitable triumph, the speed of light, the energy available to civilizations—may actually be engineerable or contingent. Brown argues that deep principles in physics simultaneously impose constraints and reveal counterintuitive freedoms. He draws on vacuum decay, the holographic principle, and thermodynamic bounds to show that sufficiently advanced descendants might escape what we regard as cosmic fate. The conversation moves between rigorous physics and long-term futures: what bubble universes could mean for civilizations, whether the cosmological constant might be manipulable, and how extracting energy from matter—from chemical reactions through nuclear fission to black hole mining—each reveals a layer of what's physically possible.
The breadth spans fundamental cosmology and engineering. Brown explains why quantum fluctuations seeded all structure in the universe, walks through Hawking radiation and the Bekenstein-Hawking bound as the most reliable fact we have about quantum gravity, and discusses AdS/CFT as the isomorphism underpinning modern theoretical work. He then turns to physical limits as roadmaps: why Landauer's principle sets a minimum energy cost for computation, how a space elevator's rope faces the same tensile-strength problem as black hole mining, and the counterintuitive claim that near-saturation of such bounds often signals where deeper understanding is needed. The discussion ventures into philosophy of physics—how representation matters in breakthroughs, whether LLMs might invent notations humans cannot—and ends on tensions between different formulations of quantum mechanics and their implications for how rational agents should weigh futures across many-worlds branches.
Brown argues that much of what seems like fixed cosmic destiny—the heat death, the laws of physics, the energy available to civilizations—is actually engineerable or contingent, and that the same principles (vacuum decay, holography, thermodynamic bounds) reveal both deep constraints and surprising freedoms for sufficiently advanced descendants.
- The cosmological constant may be manipulable, letting descendants escape heat death by triggering controlled vacuum decay
- The holographic principle shows information scales with area not volume, our most confident fact about quantum gravity
- Physical bounds (MC squared, speed of light, tensile strength) tend to be saturable and constrain but also define what future engineering can achieve
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.
Beyond the MC squared limit on energy extractable from matter, future civilizations face Landauer's limit: a given amount of energy has bounded usefulness because computations make errors and there is a fixed cost—a bit of free energy—to correct each error; perfectly error-free computers are very unlikely because the cosmological-constant-driven minimum background temperature of the universe will inevitably leak in and cause errors.
“there is a fixed cost of a bit, basically a bit of free energy, in order to correct those errors”
Often in the history of physics a breakthrough is just how you think about something—the representation you use; Einstein's summation convention and Penrose's notation for spacetimes are examples of representations that made certain things clear, analogous to new experimental techniques, and better LLMs may invent superior representations suited to themselves even if not to us.
“Often in the history of physics, a breakthrough is just how you think about it, what representation you do.”
The most underappreciated discovery in cosmology is that all structure in the universe—galaxies, you, me—originated from quantum fluctuations: just after the Big Bang the universe was homogeneous to one part in 10^5 or 10^6, and those tiny quantum-seeded density fluctuations grew via gravitational collapse (denser regions pulling in more matter) into all observed structure.
“those inhomogeneities were seeded by quantum fluctuations. When the universe, just after the Big Bang, was considerably smaller than it is today, the effects of quantum mechanics were correspondingly more important.”
To extract nearly all the MC squared of matter you need a process that destroys baryon number (start with protons/neutrons, end with none, releasing all energy as radiation); electromagnetic, strong, and weak forces cannot do this—only gravity can—so a small black hole can in principle process protons and neutrons into photons, gravitons, and neutrinos, enabling power plants approaching 100% of true mass-energy efficiency.
“The only force of nature that will do that, with a small caveat, the only force of nature that we know that will do that is the gravitational interaction.”
Earlier proposals to mine black holes faster—reaching in with a claw to grab Hawking radiation that doesn't escape—do not work; Brown's paper showed you cannot beat the mass-cubed evaporation timescale because the scoop is like a space elevator requiring a rope whose tensile-strength-to-mass-per-unit-length ratio must be as strong as the laws of nature allow, and even a fundamental string saturating that bound is just strong enough to support its own weight with nothing left for payload.
“the existing proposals did not work. They didn't work to speed it up. And in fact, you can't speed it up.”
The holographic principle states a gravitational theory in n dimensions is equivalent to a non-gravitational theory in n-1 dimensions (because information scales as area, as if a lower-dimensional non-gravitational system); this led to Maldacena's AdS/CFT correspondence—the most cited high-energy paper ever—an exact duality between a flavor of string theory and a non-gravitational boundary theory, valuable because non-gravitational theories are far easier to define and understand.
“a gravitational theory in n dimensions is like a non-gravitational theory in n minus 1 dimensions”
AdS/CFT is not an approximation, analogy, or metaphor but a precise isomorphism—the two theories in different dimensions are exactly equivalent, so as a matter of philosophy both are equally real; complexity need not be conserved across the isomorphism, which lets physicists do 'arbitrage', using whichever description is simpler (sometimes black hole physics simplifies messy quark-gluon plasma calculations at RHIC, sometimes vice versa).
“It's not an analogy, it's not a metaphor, it is not an approximation that is valid in some domain and not another. It really is the case that these two theories are exactly equivalent to each other.”
Unlike a black hole horizon, where everyone agrees on the point of no return, a cosmological horizon is relative to each observer: for any person there is a boundary (around the distance from which you can send information and have it returned) between what you could in principle communicate with and what you never can, so we live on the boundary of infinitely many other universes whose centers are located elsewhere, just as those distant observers live on ours.
“A black hole horizon -- there is a point of no return... And everybody can agree where that is. For cosmology, there is a point of no return, but the point of no return is relative to a given person.”
Theoretical physicists tend to be poorly calibrated and in love with their own theories, making overconfident predictions (e.g. confident supersymmetry predictions before the LHC), but this may be optimal at the ecosystem level: since the remaining good ideas look like bad ideas, irrational optimism is needed to push past initial discouragement—analogous to startups, where proper calibration about success odds might stop you from trying.
“the good ideas that look like good ideas have already been had... Where we're at now is the good ideas that look like bad ideas. So in order to motivate yourself to get over the hump... you need a little bit of irrational optimism”
A recurring pattern in physics is that when you derive a bound but can't think how to get close to it, that's an indication you need to think harder, because often—with enough ingenuity—there is a way to saturate the bound; finding something that nearly saturates a bound is a sign you're on the right lines, reflecting that nothing in physics is ever a coincidence.
“if you have a bound but you can't think how to get close to the bound, that's usually an indication that you need to think closer. Because often these bounds, if you're clever enough, there's a way to get to the bound.”
Chemical reactions extract only about one part in 10 billion of the rest-mass energy because they only address the electromagnetic energy of electrons; nuclear reactions do far better (one part in ~10^3) because they tap the strong/weak interactions between protons and neutrons, but even nuclear processes conserve baryon number, leaving ~99.9% of energy inaccessible.
“You burn gasoline and you extract, as a function of the rest mass of the gasoline that you started with, one part in 10 billion of energy”
According to general relativity, energy is conserved locally but not globally: in an expanding universe energy is simply not conserved at the global level and can appear, which is why spawning a bubble universe does not violate conservation of energy and why an unbounded amount of free energy could be available if the cosmological constant can be manipulated.
“Energy, or the conservation of energy, is not conserved in general relativity. Energy is not conserved. It's conserved locally, at things you can do at a local level.”
Pre-1970s, black holes were thought to be one-way (matter falls in, never comes out), but Hawking and Bekenstein discovered that once quantum mechanics is involved, energy leaks out as Hawking radiation—though extremely slowly: a solar-mass black hole has a temperature measured in nanokelvins and would take about 10^55 times the current age of the universe to give back all its energy unaided.
“Once quantum mechanics is involved, in fact, energy, even without you doing anything, starts to leave black holes. The problem... is that it leaves black holes extremely slowly.”
A space elevator's hardest problem is the rope supporting its own weight—each segment must hold all the rope below it, so tension and thickness grow upward; a steel rope would need to be thicker than the Earth by geostationary orbit, but carbon nanotubes are strong enough to just barely work around Earth with only a couple of thickness doublings, while being totally inadequate (off by ~10^12) for the strength-to-weight ratio black hole mining demands.
“each bit of the rope needs to support not only its own weight but also the weight of all of the rope beneath it. So the tension that you require keeps getting more and more”
The fact we are most confident about regarding quantum gravity is that the maximum information (entropy) storable in a region equals its surface area divided by 4G·ℏ (Bekenstein-Hawking), with black holes saturating this bound—a profoundly surprising result because all non-gravitational physics suggests information should scale with volume, and the naive volume-scaling argument fails because any sufficiently dense pile of hard drives collapses into a black hole long before violating the bound.
“the amount of information you can store, measured in bits, the entropy of that region, is given by the area of that region divided by G, Newton's constant, and H bar, Planck's constant”
There was a natural fission reaction in Earth's distant past—a seam discovered in Africa where a fission reaction happened spontaneously due to a higher fraction of fissile isotopes—which did not explode but did the same thing that happens in modern nuclear power plants.
“They've discovered a seam in Africa where it looks like there was a fission reaction that naturally happened. It didn't explode, but it did do the same thing that happens in our nuclear power plants.”
Because colliders probe small scales by high energy (Heisenberg's uncertainty: high energy means short resolution) but face unfriendly cost scaling laws, an alternative is to study the Big Bang itself—the universe played backwards contracts into an enormous explosion 13.8 billion years ago—so by closely examining its aftermath we can probe energies inaccessible to any feasible collider, at the cost of not being able to rerun it.
“there was a ginormous explosion that happened, which was the Big Bang... that's a very big particle collider indeed. And so by just examining very closely the Big Bang and its aftermath, we're able to hopefully probe some of these quantities”
Despite being a tremendous conceptual breakthrough and existence proof of a well-defined theory of quantum gravity, AdS/CFT has a negative cosmological constant while our universe has a positive one, so it doesn't describe our universe; formulating a holographic theory for a positive-cosmological-constant (de Sitter) universe is hard partly because, as far back as Archimedes, you need a fixed external point to observe from, which a positively-curved expanding universe with finite entropy denies you.
“in AdS/CFT, there is a negative cosmological constant in the gravitational theory, and our universe, as we discussed before, has a positive cosmological constant”
In many natural extensions of known physics the cosmological constant is not truly fixed but corresponds to a local minimum (vacuum) in a higher-dimensional energy landscape; distant descendants could escape heat death by deliberately triggering a controlled vacuum decay to transition into a vacuum with a lower cosmological constant, and they would be strongly incentivized to do so.
“you have what's called different vacuum, different regions of parameter space that you can transition between, in which the cosmological constant can take different values”
Almost certainly anything humans can do can happen without humans, so vacuum decay—if possible at all—will definitely happen in our future due to a quantum fluctuation; according to quantum mechanics, if a process is permitted it must eventually happen, though it can take exponentially long, which is why descendants might prefer to take fate into their own hands.
“Almost certainly anything that humans can do can happen without humans.”
The many-worlds interpretation should be the default expectation given quantum mechanics, but you cannot dismiss existential risk by saying some branch will flourish anyway: Born's rule says total utility is the linear sum across branches weighted by the square of the amplitude, so the rational move is to make our own branch as good as possible—though Brown concedes a non-linear 'portfolio' utility intuition inconsistent with Born's rule remains conceivable.
“Born's rule, which is one of the foundational rules in quantum mechanics, tells you how much to care about each branch. You don't care about them equally.”
New revolutionary theories are not chosen merely by best matching data with loss minimization; the heliocentric model with epicycles fit the data no better than geocentrism, yet won because it was more beautiful, and beauty has been a reliable guide in the history of science—so an AI optimizing only for data consistency would end up with epicycles, meaning we would have to teach LLMs beauty.
“if you optimize only for being consistent with the data, you're going to end up with epicycles. You're not going to end up with some beautiful new conceptual thing.”
A new vacuum bubble tends to expand at the speed of light, so it is not a box you keep in a lab to do useful work; it engulfs your entire future light cone, meaning you must engineer the new vacuum to remain hospitable for life rather than treating it as a contained energy source.
“the thing about these bubbles is that they tend to expand at the speed of light. So even if you start off outside, you're probably going to end up inside them in short order unless you run away very quickly.”
LLMs appear to be interpolators, but the level of abstraction at which they interpolate keeps rising, so from a sufficiently elevated point of view inventing general relativity from Newtonian physics is just interpolation at a grandiose level of abstraction—which tells us something about the nature of human intelligence as much as about language models.
“it certainly looks like these LLMs and these AI systems in some sense are just interpolators, but the level of abstraction at which they're interpolating keeps going up and up and up”
Particle physics has slowed not from dysfunction but because the Standard Model written in the 1970s-80s was too good—it predicts everything coming out of any particle accelerator we can afford to build—so the field 'won', and getting too good makes it hard to know where to push next; CERN at tens of billions of dollars actually costs more than big AI model training runs, showing academics pooling money can outspend big labs.
“we wrote down theories in the 1970s, and those theories were—it's called the Standard Model—and those theories were too good, in the sense that we won”
There is nuclear insubordination in both directions: the 'good kind' where operators refuse to launch despite orders or sightings (like Petrov declining to report a false attack warning), and the Nagasaki kind where a crew may have dropped a weapon against direct orders.
“there's nuclear insubordination in both directions”
The quantum-mechanical multiverse and the cosmological multiverse become totally intermeshed when you write down a theory containing both, because the formation of bubble universes is itself often a quantum process, leaving you in superpositions over a bubble existing and not existing; the base reality is plausibly the wave function over all the bubbles in the inflaton field, but counting in positive-cosmological-constant bubble universes remains an open problem.
“what gives rise to bubble universes is often quantum processes. So often you end up in superpositions over there being a bubble universe and there not being a bubble universe”
Brown's research into Nagasaki suggests the bombing was highly suspect: the crew had strict orders to bomb only with visual confirmation, missed the target by far more than visual bombing would (suggesting illicit radar bombing), and had decided in the cockpit to drop the bomb no matter what because fuel shortfalls meant they'd otherwise lose the only existing weapon in the ocean—so possibly 50% of all nuclear weapons ever dropped in combat were dropped against direct orders.
“is it possible that 50% of all of the nuclear weapons ever dropped in combat were, in fact, dropped against direct orders?”
In the 1990s we discovered the universe is not just expanding but expanding faster and faster due to dark energy (the cosmological constant), which drags distant galaxies away from us so rapidly that anything beyond about a dozen billion light years can never be reached even by near-light-speed probes, meaning there is a finite amount of free energy in our future and we are doomed to heat death if the constant is truly constant.
“In possibly the worst day in human history, sometime in the 1990s, we discovered that in fact, not only is the universe expanding, it's expanding faster and faster and faster.”
Most physicists would be very surprised if faster-than-light communication turned out possible; Brown puts greater than 90% chance—98% if pressed—that in 100 years we are still limited by the speed of light, though over a million-year horizon the question itself may dissolve into a larger one we'd understand differently.
“there is a greater than 90% chance that in 100 years, we are still limited by the speed of light. There's a 98% chance, if you make me be precise.”
Our understanding of cosmology keeps changing because the long lever arm of asking about the very distant future causes even small present uncertainties to pan out into enormous differences in the distant future, which is why questions about ultimate fate are so sensitive to small changes in our understanding.
“it's because you have the lever arm, the long lever arm of asking about the very, very distant future that makes even small uncertainties pan out to absolutely ginormous distances in the distant future”
The very early universe was opaque to light because it was a dense plasma that absorbed light, so we cannot directly see anything before about 300,000 years after the Big Bang; nevertheless we can infer a great deal about earlier times from the patterns of anisotropies in the cosmic microwave background, and at the information-theoretic level nothing about the distant past is inaccessible in principle.
“the very early universe, just after the Big Bang, was opaque to light. We can only see light past about 300,000 years after the putative Big Bang.”
For the cosmological multiverse there is no good way to decide the weighting factor for utility across universes—we lack the equivalent of quantum mechanics' Born's rule—so it remains genuinely open whether the right way to aggregate utility is a simple linear sum or some non-linear function, leaving room for intuitions like valuing the finality of total extinction more than proportionally.
“for the cosmological multiverse, there isn't a particularly good way to decide what the weighting factor should be. We don't have the same equivalent of Born's rule in quantum mechanics.”
Running many AI instances in parallel to search for productive theories (e.g. best-of-n over random parts of the current model of the universe) faces a key challenge: how to evaluate whether you ended up with a good theory; tasks are easily parallelized when correctness is self-evident (like problems in NP), but knowing whether a generated theory like special relativity is 'onto a winner' is hard, making evaluation the tricky part of the pipeline.
“how do you evaluate whether you had a good theory at the end? That's going to be the tricky bit.”
There are speculative but not wild theories of the early universe in which our Big Bang was itself a local vacuum decay event, where we started in a higher-cosmological-constant vacuum and transitioned into the bubble in which we live; Brown estimates roughly 50% probability for this picture.
“what we see as the Big Bang was in fact just a sort of local vacuum decay that then gave rise to our the bubble in which we live”
Most possible vacuums are completely inhospitable to all forms of intelligence, so triggering vacuum decay carelessly risks landing in a vacuum where you have changed not just the cosmological constant but constants like the electromagnetic or strong nuclear force, which would be catastrophic; the engineering goal is to target only the cosmological constant and avoid making a black hole.
“most of those vacuums are incredibly inhospitable to life as we know it. In fact, seemingly they're just completely inhospitable to all forms of intelligence.”
The energy required to trigger vacuum decay is probably small—much less in MC squared terms than the energy in your body, though more than current particle colliders produce—so the difficulty is not energy but the control problem of concentrating it into a precisely shaped bubble that expands correctly, lands in the target vacuum, and does not collapse into a black hole.
“The energy requirements are probably pretty small, much more than we can currently make in our particle colliders, but much smaller just in terms of MC squared than the energy in your body.”
The problem with particle colliders is that the expected information is low—we know roughly what we'd get—and they got too expensive; often smaller observational experiments like BICEP (tens of millions, looking for primordial gravitational waves in the CMB at the South Pole) are worth doing precisely because they're cheap, while a $50 billion collider is not the obvious place to spend science money.
“it's not the obvious place to spend $50 billion if you had $50 billion to spend on science”
Because quantum computation grows almost exponentially with qubit count for tasks like factoring (a 200-qubit machine is far more than twice as good as a 100-qubit one), the distant future might feed all free energy into one central quantum computer despite redshift costs; quantum coherence can in principle be shared across the universe via photons that maintain coherence in vacuum, though bandwidth and interconnect constraints (like classical computing) could instead favor distributed entanglement networks.
“a 200-qubit quantum computer is much more than twice as good as a 100-qubit quantum computer for certain tasks”
The anthropic principle confidently explains why we live in a weird corner of the solar system rather than the sun (we'd be dead otherwise), and extends to constants of nature like the proton being slightly more stable than the neutron; if true, this demands the existence of other places in a multiverse where these constants take different values and no one lives, naturally drawing you to variables scanning over space.
“one should expect intelligent life to live in this weird place in parameter space”
Our universe is incredibly rich, with interesting structure at every scale from viruses to galaxies, and the laws of physics appear very well attuned for life and chemistry—if you threw a dart in laws-of-physics space, chemistry would likely be impossible in most places—which either demands an intelligent-creator explanation or an anthropic one involving variables that scan across a larger possibility space.
“Everything is just tuned just so that chemistry is possible. Perhaps in most places you would throw the dart in possibility space, chemistry would be impossible.”
A skeptical response to fine-tuning is the puddle analogy: a puddle marvels that its depression seems perfectly shaped for it, when in fact the puddle adapted to the hole; similarly, life is highly adaptable—thriving in undersea vents and harsh environments—so apparent fine-tuning may partly reflect life adapting, though a universe with a cosmological constant ripping it apart every microsecond would still be implausibly hostile.
“a puddle that's sitting in some depression in the ground reflects on how wonderful the universe is, that this depression in the ground seemed to have been made the perfect shape for the puddle to exist”
The economic theory of comparative advantage only applies if not all resources can be transported; in a future where you could disassemble a trading partner and apply everything to the party with absolute advantage, comparative advantage breaks down—but cosmic expansion redshifts energy beamed across the universe, so keeping energy in place may beat transporting it, and whether to concentrate all value on Earth or spread across galaxies depends on unknown returns to scale.
“In economics, the theory of comparative advantage only applies if not all resources can be transported.”
It is interesting to reflect on what aspects of human behavior nature has a hard time doing without us: for example, we make colder things in our laboratories than exist naturally in the universe, though the universe certainly could make anything colder by chance.
“we make colder things in our laboratories than really exist naturally in the universe”
Physics problems, unlike pure math problems, have two components: using physics knowledge to turn a word question into a math question, then solving the math; the distinctively LLM-capable step is the first one, recognizing and formulating the physical problem mathematically.
“One is to sort of take this word question and turn it, using your physics knowledge, into a math question and then solve the math question.”
Although there was long debate in the black hole context about whether information is conserved by black holes, the modern consensus is that it is conserved.
“there's been a lot of debate in the black hole context about whether information is conserved by black holes, but the modern consensus is that it is”
LLMs went from totally useless three years ago to extremely useful for physicists, especially as personal tutors that can explain advanced topics, debug your misunderstandings (telling you not just the right answer but why your wrong answer was wrong), and do semantic literature search—filling a gap because advanced topics often lack good resources, analogous to how chess players improved by training with chess machines off the board.
“they will not only tell you what the right answer is, but debug your understanding on the wrong answer”
LLMs are strong on topics with substantial literature—out of millions of physics papers there might be a thousand on squeezed light at LIGO, enough for the model to explain it well and even correct misunderstandings—but if there are only a handful of papers on a topic, the model is typically not strong at it.
“out of a physics literature of millions of papers, there have got to be at least a thousand on that topic. If there was just a handful of papers on a topic, it's typically not that strong at it.”
There is an effort (Shirley Ho at Flatiron) to feed the pipeline of exabytes of astronomical observatory data into a transformer; the dream is that the model finds patterns no human could ever tease out, dramatically improving the return on extremely expensive observatories—though the patterns it finds might not be interesting to humans.
“There's an effort, Shirley Ho and Flatiron, which is basically that exact plan. They take the pipeline of all of the data that comes out of these astronomical observatories, they plug them into a transformer, and see what happens.”
If individuals can trigger vacuum decay that wipes out the entire future light cone—not just themselves but everyone wishing to participate in it—this is an enormous negative externality that makes libertarian fantasies impossible and demands strong governing structures, especially in the worst case where physics might even incentivize sitting on the edge of instability for a power plant.
“in a world with big negative externalities, libertarian fantasies can't really happen. It has pretty good big government governance implications”
Generally there is positive transfer between domains in AI models: making a model better at one thing tends to make it better across all domains, though if you over-optimize for one narrow task you eventually hit a Pareto frontier and start degrading other metrics.
“Generally, you see positive transfer between domains. So if you make them better at one thing, they become better at another thing across all domains.”
In physics, if you didn't write a paper no one might write it for years or ever, partly because the field is so specialized (only 100-300 high-energy theoretical physicists worldwide); in computer science, if you didn't do it someone else would soon, because thousands work on the same problems—so physics has higher counterfactual impact per person while AI has vastly larger absolute world-changing impact.
“In physics, you write some papers, and you're like, had I not written that paper, no one would have written that paper for years or ever, perhaps. Computer science doesn't feel like that.”
A practical hitchhiking tactic is to write the direction you're going rather than an exact destination, because people heading partway will still stop (whereas they'd skip you thinking they aren't going all the way), and it makes declining an unwanted short ride easier; the carsick excuse works to exit an unsafe ride since even serial killers don't want vomit in their car.
“don't write your exact destination on your sign. Write the sort of direction in which you're going.”
The very last thing AI systems will be able to do is, given the laws of physics as understood at the turn of the last century, invent general relativity from scratch—and once an AI can do that, there won't be much else left to do as far as humans are concerned, perhaps in 10 years, at which point AI will have fully encompassed human intelligence.
“maybe the very last thing that these systems will be able to do, these LLMs will be able to do, is, given the laws of physics as we understood them at the turn of the last century, invent general relativity from that”
The biggest hitchhiking tip is counterparty modeling: imagine the driver picking you up needs time to see you, evaluate you, decide, and safely stop—so stand somewhere visible, ideally at a stoplight, where there's a safe place to pull over; a good spot can be worth a thousand miles since there is roughly a 1000x variance in hitchhiking spot quality.
“They need time to see you, to evaluate you and to decide they're going to pick you up and then to safely stop.”
Scott Aaronson posted that GPT-4 got a B or an A-minus on his intro to quantum computing class, evidence that models perform at a respectable graduate-adjacent level on technical coursework.
“Scott Aaronson recently posted about the fact that GPT-4 got a B or an A- on his intro to quantum computing class”
Hitchhiking selects for more open and risk-tolerant people, with about 20% who always pick up hitchhikers, and provides a high-variance cross-section—from idle millionaires to ex-prisoners to truckers—where drivers, especially lonely truckers with excess mental capacity and untold stories, frequently confide things they say they've never told anyone, making it a mutually beneficial emotional exchange.
“there's definitely a selection bias in who picks you up. They tend to be more open and more risk tolerant.”
Hitchhiking used to be very common in the US until some mass murderers targeting hitchhikers drove its popularity down; Brown argues this fear may be partly irrational because, given how few hitchhikers now exist, cruising for them is not a high-expected-value strategy for a would-be murderer—though he admits this could be pure cope.
“there were some mass murderers who drove the popularity down by targeting hitchhikers”
Many tasks Brown performed as a physicist are already automated at some level, but full automation—being totally out of the picture and no longer necessary—is probably close to ASI-complete, so his timeline equals whatever the timeline for artificial superintelligence is; he can imagine a scenario where it's five years.
“Until I am totally out of the picture and no longer necessary... That's probably pretty close to ASI complete. So whatever your timeline for ASI is.”
Brown's private benchmark—giving his Stanford graduate general relativity exam to LLMs—went from a score of zero three years ago, to weak-student-level a year ago, to essentially acing the test in the last couple of months, demonstrating extremely rapid capability growth on a hard graduate physics evaluation.
“Three years ago, zero. A year ago, they were doing pretty well, maybe a weak student, but in the distribution. And now they essentially ace the test.”