ExpertsFMAdam Brown – A deep but accessible introduction to general relativity
Dwarkesh Podcast · July 10, 2026
5 Selected Lessons:
priority · 2 heuristic · 1 case · 1 prediction · 1
Ranking energy-extraction efficiency by which force you access
Source
Dwarkesh Podcast — Adam Brown – A deep but accessible introduction to general relativity (July 10, 2026) · at 39:25 ▸ · at 1:07:17 ▸
When this applies
when ranking competing methods of extracting value from a resource by what fraction of the resource's total stored potential each method can actually reach
The takeaway
Rank methods of extracting value from a resource by how deep into the resource's total stored potential their underlying mechanism can reach, not by how mature or familiar the method is.
Why it works
Each extraction mechanism can only tap the layer of energy coupled to the force it uses; chemical reactions touch only electromagnetic bonds, nuclear reactions touch only strong/weak binding energy, and only gravity couples to all mass-energy including rest mass itself.
Where it stops holding
This stops working when the scarce resource isn't a layered store of potential (so the 'how deep can you reach' framing doesn't apply) or practical constraints (e.g. engineering feasibility) dominate over theoretical efficiency.
Detail
hierarchy:
- Chemical bonds: ~10^-10 of rest-mass energy extracted
- Fission: ~10^-3; Fusion: ~10^-2 of rest-mass energy extracted
- Gravitational infall to near an event horizon: up to ~100% of rest-mass energy extracted
rationale: 99% of an object's energy sits in the rest mass of its protons and neutrons, untouched by chemistry or nuclear reactions; only gravity, by coupling to all mass-energy, can access nearly all of it, making black holes 'the most efficient possible power plant.'
In their words
If I start off with a mass object of m, I can extract up to quantum fractions essentially, I can extract essentially 100% of the rest mass energy that I've gone in. It is the most efficient possible power plant because by building an apparatus like this, in principle, I could extract 100% of the energy of whatever I started with.
Judge a theory's power by the range of scales/phenomena it unifies with one equation
Source
Dwarkesh Podcast — Adam Brown – A deep but accessible introduction to general relativity (July 10, 2026) · at 27:54 ▸ · at 1:22:13 ▸
When this applies
when assessing how impressive or valuable a unifying model/framework is
The takeaway
Rate the strength of a unifying framework by the number of orders of magnitude and the diversity of phenomena it spans using the same underlying machinery, not by the simplicity of the observation that originally motivated it.
Why it works
A theory built from a narrow, almost trivial starting observation gains credibility and significance in proportion to how many wildly different scales and phenomena it correctly describes without modification.
Where it stops holding
This stops working when apparent 'reach' is actually the result of curve-fitting separate add-on assumptions for each new domain rather than genuine extrapolation of the same core machinery.
Detail
hierarchy:
- Thought experiments (elevators, buckets of water) — the seed idea
- Solar-system-scale predictions (Mercury's orbit, light bending)
- Galactic-scale phenomena (galaxy rotation, black holes)
- Cosmological scale (expansion and fate of the entire universe)
rationale: Brown notes 'it's crazy, you know, huge numbers of order orders of magnitude' that the same equations cover, calling it a mark of the theory's exceptional reach and beauty.
In their words
General relativity does all of that and goes one step further. It describes the motion of apples falling off trees. It describes the motion of Mercury and the planets in the solar system, and it describes the expansion of the entire universe.
Push a formula to its extreme to find where it must break
Source
Dwarkesh Podcast — Adam Brown – A deep but accessible introduction to general relativity (July 10, 2026) · at 36:21 ▸ · at 39:25 ▸ · at 41:50 ▸ · at 44:12 ▸ · at 1:02:58 ▸ · at 1:07:17 ▸
When this applies
when testing whether a formula or model is universally valid, push its inputs toward extreme values and check whether it predicts something physically/logically absurd (like exceeding a conserved total)
The takeaway
Treat a formula's prediction of exceeding an available conserved quantity as proof that unmodeled physics/constraints kick in before that point, and go looking for what they are.
Why it works
A model derived in a limited regime (here, weak gravity) can be extrapolated algebraically past that regime; if the extrapolation implies violating a conservation law, that signals the model is incomplete and new mechanics must take over before the breakdown point.
One level up
Stress-test any model at its extremes before trusting its predictions inside the normal operating range.
Where it stops holding
This stops working when the model's underlying assumptions already hold at the extreme being tested, so there is nothing new to find.
Detail
rule: Extrapolate a limited-regime formula to its extreme and check for violations of a conserved quantity.
threshold: Once the formula predicts extracting more energy than the object's full rest-mass energy (fraction > 1), that is the signal something must give.
rationale: Full general relativity resolves the paradox by making the gravitational force diverge at a finite radius (forming a black hole) before the Newtonian formula's absurd prediction is reached.
In their words
it sure looks like you could get more than 100%, the fraction with bigger than one. You can get more than 100% of the mass of your brick back by lowering it down to the surface of this object. And that feels wrong.
AI will likely become a superhuman explainer, not just a superhuman prover
Source
Dwarkesh Podcast — Adam Brown – A deep but accessible introduction to general relativity (July 10, 2026) · at 1:35:10 ▸
When this applies
when a powerful automated system could in principle produce correct-but-incomprehensible outputs, and you must forecast whether it will instead be optimized to make its outputs human-understandable
The takeaway
Don't assume a powerful automated capability will default to the least interpretable form of output; check whether early evidence shows it naturally producing human-legible intermediate forms that humans can then extend.
Why it works
A system trained to be superhuman at a task is not automatically optimized only for raw correctness; if comprehensibility is also a target (or an emergent byproduct of doggedly searching for human-legible paths), the system can convert hard-to-understand results into understandable ones.
Where it stops holding
This stops working if future systems are optimized purely for proof validity/speed with no incentive toward interpretability, or if the early positive evidence (one case) doesn't generalize.
Detail
approach: Brown predicts LLMs will act as 'superhuman explainers' as well as 'superhuman provers,' converting hard-to-understand proofs into human-comprehensible ideas that humans can then redeploy.
signals:
- an Erdos conjecture was proved informally (not just as inscrutable Lean code) by an LLM-derived idea
- human mathematicians were then able to fully understand that idea and use it to prove new, different theorems
rationale: Brown finds this early empirical pattern 'pretty supportive' of the more optimistic vision, though he explicitly flags it as early-days evidence from a small number of cases.
In their words
I think that is a possible future, but I actually don't find that to be a very likely future. Because as well as being superhuman provers, we also expect these large language models to be superhuman explains.
Even the theory's inventor can misunderstand his own exact solution for decades
Source
Dwarkesh Podcast — Adam Brown – A deep but accessible introduction to general relativity (July 10, 2026) · at 31:46 ▸ · at 1:18:56 ▸ · at 1:22:13 ▸
When this applies
when a correct formal/mathematical result sits inside your own framework, available immediately, but the community (including its author) misreads its implications for a long time
The takeaway
Deriving a correct result is not the same as understanding it; expect genuine comprehension to lag the derivation, sometimes by decades, and to require both further theoretical generalization and outside empirical confirmation.
Why it works
A formally correct solution can be interpreted through incorrect physical intuitions (e.g., 'bouncing off' the horizon) until later theoretical work (proving the solution's generic applicability) and independent empirical confirmation force a correct reading.
One level up
Don't assume the originator of a model has correctly grasped all of its implications just because they wrote it down first.
Where it stops holding
This stops working when a result's implications are simple enough to be grasped immediately, or when no further theoretical/empirical work is needed to validate the interpretation.
Detail
situation: Schwarzschild derived the exact black-hole solution to Einstein's equations almost immediately, but it was called a 'mathematical monstrosity' and even Einstein wrote confused, incorrect things about the event horizon.
decision: The field required Penrose's (and later Hawking's) theoretical proof that black hole formation is generic, plus decades of astronomical observation (galactic center star orbits, LIGO gravitational waves, Event Horizon Telescope) before consensus formed.
outcome: Black holes are now considered extremely well-established, confirmed by multiple independent lines of theory and evidence.
lesson: A correct formal result can remain misunderstood by its own originators for decades; robust consensus required both generalizing theoretical proof and multiple independent empirical confirmations, not derivation alone.
In their words
People wrote down wrong things for about half a century about what this meant. And perhaps the worst offender was was Einstein, who got extremely confused about it, particularly confused about what I will describe as the event horizon