1.

ARGUMENT

1.

A Nobel laureate bets his reputation on a single conjecture: any pattern found in nature can be efficiently modeled by a classical computer — no quantum machine required. If Demis Hassabis is right, protein folding, fluid dynamics, weather, and consciousness itself are all just search problems waiting for the right model. The proof is already arriving faster than the theory.

2.

Nature isn't random — selection pressure makes it learnable

• Proteins fold in milliseconds despite 10^300 possible configurations; Go has 10^170 possible positions — both dwarf the atoms in the universe, yet both are solvable
• Hassabis calls this "survival of the stablest": mountains shaped by weathering, planetary orbits carved by gravity, stable elements forged by nuclear selection — all non-random patterns left behind by processes that ran billions of times
• The conjecture's core logic: if a system survived selection, it has structure; if it has structure, a neural network can follow the gradient to learn it

3.

AlphaGo and AlphaFold prove the paradigm on two radically different domains

• AlphaGo modeled Go's dynamics, layered Monte Carlo Tree Search on top, and produced Move 37 — a strategy no human had ever played in thousands of years of the game's history
• AlphaFold modeled protein structure space and solved computational protein folding, earning Hassabis and John Jumper the Nobel Prize — directly vindicating Christian Anfinsen's 1972 conjecture
• AlphaFold 3 extended to protein-protein, protein-RNA, and protein-DNA interactions; AlphaGenome maps single genetic mutations to function; AlphaProof earned a silver medal at the Math Olympiad — each a new domain, same paradigm

4.

Veo learned physics from YouTube — no equations, no embodiment, no physics engine

• Google's video generation model renders liquids through hydraulic presses, specular lighting, and material dynamics that Hassabis spent years painstakingly hand-coding as a game developer in the 1990s
• Neuroscience predicted embodied interaction was necessary for understanding intuitive physics — Hassabis himself believed this five to ten years ago; passive observation alone proved sufficient
• WeatherNext outperforms traditional fluid dynamics supercomputers on cyclone path prediction, suggesting that even Navier-Stokes dynamics — "traditionally thought of as very, very difficult intractable problems" — sit on a learnable lower-dimensional manifold

5.

The boundary of what classical systems can model may be a new complexity class

• Hassabis frames P=NP as a physics question: if information is "the most fundamental unit of the universe, more fundamental than energy and matter," then what a Turing machine can efficiently compute defines the structure of reality itself
• Factoring large numbers — uniform, patternless — may require quantum computers; chaotic systems with sensitive initial conditions are "right on the boundary"; but everything shaped by selection likely falls inside the class
• No formal definition exists yet — Hassabis is working on it "in my few moments of spare time" — but the empirical evidence is outrunning the theory: proteins, Go, weather, genetics, and mathematics all yielded to the same approach

6.

AGI on classical hardware is the conjecture's ultimate test — and Hassabis gives it 50% odds by 2030

• His AGI definition is deliberately high-bar: match every cognitive function of the brain with consistency across all domains, not the "jagged intelligence" of current systems that ace coding but stumble on common sense
• The lighthouse test: invent a physics conjecture worthy of Terence Tao's attention, or create a game as deep and elegant as Go — "not just come up with move 37, a new strategy, but can it invent a game"
• Hassabis concedes current systems "definitely can't" generate worthy conjectures and puts the odds at "50/50 whether new things are needed or whether the scaling of the existing stuff is gonna be enough" — half the probability space includes a world where the paradigm plateaus

7.

The conjecture is unfalsifiable — and unfalsifiable claims aren't science

• "Any pattern found in nature" defines the conjecture's scope by what exists in nature, a boundary that shifts every time we discover something new; Hassabis himself cannot draw the line between natural and abstract systems, admitting chaotic and emergent systems are "right on the boundary" and that he "doesn't know" which side they fall on
• Every success — AlphaFold, AlphaGo, WeatherNext — is claimed as evidence for the conjecture, while every failure can be reclassified as "not a natural system" or "not yet attempted with the right model"; a hypothesis that absorbs all outcomes and excludes none is not a hypothesis but a narrative
• The P=NP connection, the conjecture's theoretical anchor, is a hobby project with no formal definition, no published paper, and no peer review — it is a Nobel laureate's cocktail-napkin intuition dressed in the language of complexity theory, and treating it as more than that mistakes ambition for proof

8.

Science has always advanced on conjectures that couldn't be falsified — until they could

• Both sides assume a conjecture must be falsifiable at the moment of its formulation to be scientifically valuable; but Anfinsen's 1972 thermodynamic hypothesis about protein folding was equally unfalsifiable when stated — it took fifty years and AlphaFold to vindicate it, and it won two Nobel Prizes in the process
• The history of physics is littered with productive conjectures that outran their formalisms: the atomic hypothesis was "not even wrong" for a century before Einstein's Brownian motion paper; continental drift was ridiculed for fifty years before plate tectonics provided the mechanism; the Church-Turing thesis remains unproven yet defines all of computer science
• What matters is not whether Hassabis can draw the boundary today, but whether the conjecture generates experiments that split the hypothesis space — and it already has, across six domains in a decade, each time surprising experts who predicted failure