Adam Brown is a theoretical physicist at Stanford and founder of Google DeepMind’s Blueshift team, which works on AI reasoning. He discusses the ultimate fate of the universe, the possibility of engineering vacuum decay to change the laws of physics, the holographic principle and black hole information, the role of AI in physics research, and his personal experiences hitchhiking and studying the Nagasaki bombing.
The ultimate fate of the universe and vacuum decay
The universe’s fate has shifted dramatically over the past century: from a static universe, to an expanding one, to the 1990s discovery of accelerated expansion driven by dark energy (the cosmological constant).
If the cosmological constant is truly constant, distant galaxies beyond ~12 billion light-years are receding faster than light can reach them, meaning a finite amount of free energy is accessible — leading inevitably to heat death.
However, the cosmological constant may not be constant. In many extensions of known physics, it can take different values in different “vacua” — local minima in a higher-dimensional energy landscape.
Our universe may sit in a local minimum, not the true lowest vacuum. Transitions to lower-energy vacua are possible, either spontaneously via quantum fluctuation or deliberately engineered.
Vacuum decay as engineering: Distant descendants could, in principle, trigger a vacuum decay to a state with a lower (or zero) cosmological constant, avoiding heat death. This would involve creating a bubble of new vacuum that expands at the speed of light, engulfing everything in the future light cone.
The energy required is surprisingly small — less than the rest mass energy of a human body — but the control problem is extreme: concentrating it precisely to avoid forming a black hole or landing in an inhospitable vacuum.
The risk is enormous: most vacua are completely inhospitable to life, changing constants like the electromagnetic force or strong nuclear force. The engineering must target only the cosmological constant.
Conservation of energy is not global in general relativity: Energy is only locally conserved. In an expanding universe, energy is not conserved globally, so spawning a bubble universe does not violate energy conservation.
Bubble universes in the past: Our own Big Bang may have been a vacuum decay event from a higher-energy parent vacuum. Adam estimates ~50% probability for this.
Governance implications: If individuals or AIs can trigger vacuum decay, the negative externality is total — destruction of the entire future light cone. This implies a need for strong governance structures, not libertarian freedom.
Why is our universe the way it is?
The laws of physics appear finely tuned for complexity and life — chemistry, structure at all scales from viruses to galaxies. This is surprising if laws of physics are random.
The anthropic principle is the standard explanation: we observe these constants because only such values permit observers. This requires some mechanism for constants to vary across a multiverse (bubble universes, simulations, etc.).
Some physicists argue life might be more adaptable than we think (the “puddle in a hole” analogy), but extreme values — like a cosmological constant that rips everything apart every microsecond — seem genuinely incompatible with any intelligence.
Most underappreciated cosmological discovery: The quantum origin of all cosmic structure. Tiny quantum fluctuations during inflation, imprinted as 1-in-100,000 density variations in the cosmic microwave background, seeded every galaxy, star, and planet through gravitational growth. This was confirmed by precision measurements of CMB anisotropies in the 2000s.
AI and physics research
Einstein-level AGI as a benchmark: Adam suggests the terminal step for AI would be inventing general relativity from pre-1900 physics — starting from thought experiments and conceptual inconsistencies, tracing them to a unified theory. He estimates this could happen within ~10 years.
Current LLMs excel at undergraduate-level physics but have not produced conceptual breakthroughs. They are strong interpolators, and the level of abstraction keeps rising.
AI advantages: They have read vastly more than any human, can be queried as non-judgmental tutors, and can debug misunderstandings. Adam uses them to learn advanced topics outside his specialty (e.g., squeezed light at LIGO).
Evaluating AI on physics exams: On Adam’s graduate general relativity final from Stanford, LLMs went from zero (3 years ago) to weak-student level (1 year ago) to essentially acing it (recently).
Physics problems have two stages: Translating a word problem into math, then solving the math. LLMs are particularly strong at the first stage.
Positive transfer: Improving at one domain (e.g., general relativity) improves performance at others (e.g., coding), suggesting general reasoning gains.
Using astronomical data: Efforts like Shirley Ho’s at Flatiron are feeding observatory data directly into transformers, hoping to find patterns no human could detect.
The beauty problem: Revolutionary theories often win not because they fit data better (epicycles fit data as well as heliocentrism initially) but because they are more beautiful. Teaching LLMs aesthetic judgment in physics remains an open challenge.
Physics stagnation and experimental frontiers
Particle physics has stagnated because the Standard Model (developed in the 1970s–80s) was too successful — it predicted everything seen at colliders, leaving no clear direction for new discoveries.
The LHC found the Higgs (expected) but nothing unexpected (no supersymmetry, no extra dimensions). Building larger colliders is prohibitively expensive (CERN costs tens of billions).
Better bets: Smaller, cheaper experiments looking in unexpected places. BICEP (tens of millions of dollars at the South Pole) searched for primordial gravitational waves in the CMB — a revolutionary signal if found.
Cosmic archaeology as particle physics: The Big Bang itself was the highest-energy collider. By studying its aftermath (CMB, gravitational waves), we can probe energies inaccessible to any terrestrial experiment.
Information from the distant past: The universe was opaque to light before ~300,000 years after the Big Bang, but anisotropies in the CMB allow confident inference of earlier conditions. Information-theoretically, nothing from the past is inaccessible in principle.
Mining black holes
Hawking radiation: Black holes slowly evaporate via quantum effects. A solar-mass black hole has a temperature of nanokelvins and would take ~10^55 times the current age of the universe to evaporate completely.
Mining proposal: Reach a mechanical claw just outside the event horizon to grab Hawking radiation before it falls back in, speeding up evaporation.
Adam’s result: This doesn’t work. The rope required would need a tensile strength-to-weight ratio at the absolute limit allowed by the laws of nature (set by the speed of light). A string from string theory saturates this bound but has zero capacity to carry a payload. The evaporation time still scales as mass cubed — no parametric speedup is possible.
Black hole power plants: Small black holes (high temperature) could convert matter to energy with near-100% efficiency (E=mc²), bypassing the baryon number conservation that limits nuclear reactions to ~0.1% efficiency. You throw matter in, capture Hawking radiation (photons, gravitons, neutrinos) out.
The holographic principle
Bekenstein-Hawking entropy: The information capacity of any region scales with its surface area, not its volume: S = A/(4Gℏ). Black holes maximize information density for a given area.
This is deeply surprising — all of non-gravitational physics suggests information scales with volume. But if you try to pack too much information into a volume, it collapses into a black hole before violating the bound.
Holographic principle: A gravitational theory in n dimensions is equivalent to a non-gravitational theory in n-1 dimensions. Gravity “eats” a dimension’s worth of information.
AdS/CFT correspondence (Maldacena, late 1990s): An exact duality between a gravitational theory in anti-de Sitter space (negative cosmological constant) and a conformal field theory on its boundary. This is the most cited paper in high-energy theoretical physics.
It provides a well-defined theory of quantum gravity in AdS, but does not describe our universe (which has a positive cosmological constant).
Both descriptions are equally real — it’s a precise isomorphism, not an approximation.
De Sitter holography (our universe): An active open problem. Proposals include placing the dual theory on the cosmological horizon or in the infinitely distant future. The lack of a fixed boundary makes this much harder than AdS.
Philosophy of infinities
Quantum many-worlds and cosmological multiverse both imply vast infinities of branching universes. Should this change our decision-making?
Born’s rule says you should weight branches by the amplitude squared, making total utility a linear sum — so you should still try to make your branch as good as possible.
But some intuitions suggest non-linearity: the finality of extinction in all branches feels worse than losing half the population in one branch. For the cosmological multiverse, there is no equivalent of Born’s rule to resolve this.
The two multiverses (quantum and cosmological) intermesh when bubble universes arise from quantum processes, creating superpositions of bubble existence.
Engineering constraints on future civilizations
Speed of light: Adam assigns >90% confidence that this remains a constraint for the next 100 years, ~98% for communication specifically.
Energy extraction: E=mc² is the maximum, achievable in principle with black hole power plants.
Computation limits: Landauer’s limit sets a minimum energy cost per irreversible computation. The cosmological constant creates a minimum background temperature, implying a non-zero error rate for any computer — perpetual error-free computation is impossible.
Trade and organization: Whether future civilizations centralize or distribute computation depends on returns to scale. Quantum computers have super-linear returns for certain tasks (factoring), favoring centralization, but other tasks favor distribution. Photons maintain coherence in vacuum, enabling entanglement networks across cosmic distances.
Redshift as a transport cost: Beaming energy across the universe incurs massive cosmological redshift, making it more efficient to use energy locally rather than transport it.
Hitchhiking
Adam hitchhiked extensively across America and Europe (Oxford to Morocco, Princeton to Stanford, etc.).
Key tips: Stand where drivers can see you and safely stop (stoplights, wide shoulders). Don’t write your exact destination on your sign — write the general direction to avoid discouraging partial rides. If you feel unsafe, claim carsickness to exit.
Human encounters: ~20% of rides involve “one of the craziest things that ever happened.” Truckers are often the most interesting — they’re usually violating company policy, have vast audiobook knowledge, and unprocessed life stories.
One trucker was being scammed by his fiancée through advance-fee fraud; Adam helped him realize it at a truck stop in Nevada.
A state trooper drove 1,000 miles out of his way, ending at his father’s grave in Baton Rouge — a spiritual quest triggered by their conversation.
A Wyoming rancher had never contemplated that stars are like the sun; the fact reoriented his entire worldview.
Safety: Adam felt safer hitchhiking than picking up hitchhikers. There’s a moment of mutual anxiety in the first minute, then typically relaxation.
The Nagasaki bombing
During COVID lockdown, Adam became obsessed with whether the Nagasaki mission followed orders.
The primary target was Kokura, but it was clouded over. Strict orders required visual identification of the target. They went to Nagasaki (secondary target), also clouded over.
The official account claims a miraculous hole in the clouds opened on the third pass. But they missed the aiming point by a large margin — the bomb hit on the other side of a hill.
Multiple accounts from crew members are flatly inconsistent. Some suggest they had decided to drop the bomb by radar regardless of visual confirmation — against direct orders.
Curtis LeMay wanted to court-martial them, but the war ended, making it a PR problem.
Striking implication: 2 of the 2 nuclear weapons ever used in combat (Hiroshima and Nagasaki) may both have involved insubordination — Hiroshima’s crew also reportedly deviated from orders. Nuclear war has also been averted by insubordination (e.g., Petrov not reporting a false alarm).
Adam’s career
He has two careers: theoretical physics (collaborating with Leonard Susskind, working on black hole information, holography) and AI at Google DeepMind (Blueshift team on math and reasoning).
Physics moves slowly with high counterfactual impact per paper; AI moves fast with lower individual counterfactual impact but vastly greater total impact.
He attributes physics stagnation to the field’s own success (the Standard Model cleaned up particle physics) rather than dysfunction, though he acknowledges fads and overconfident predictions (e.g., pre-LHC supersymmetry claims).
Timeline to automation as a physicist: He could imagine full automation within ~5 years, roughly coinciding with ASI.
