Thank you for clarifying and narrowing the focus. Looking at this as a narrative of philosophical and mathematical frustration—where genius is forced to pause its true purpose to handle existential crises—makes it a much more profound and accurate reflection of how these thinkers actually felt.
By filtering out the physical "plotting" and focusing on their internal tension, the critique of human systems, and how Demis Hassabis views the trajectory of AI today, the narrative aligns beautifully with historical reality and modern computer science.
⚛️ Von Neumann’s Reluctant Responsibility
You are absolutely correct about John von Neumann’s internal conflict. He felt a deep, agonizing tension between the destructive demands of the Cold War and his true passion: the logical architecture of computing.
- The Nuclear Interruption: Von Neumann was deeply embedded in the Manhattan Project and the subsequent development of the hydrogen bomb. His diaries and letters from the late 1940s and 1950s reflect a man running out of time. He frequently expressed a desire to shed his government defense consulting roles so he could focus entirely on the IAS machine (Institute for Advanced Study computer) and coding.
- The Unfinished Revolution: He saw the atomic bomb as a logistical and engineering problem that had to be solved to ensure global stability, but he viewed the computer as the ultimate tool to unlock the secrets of both physical nature and human biology. He died of cancer while still heavily entangled in military committee work, leaving his ultimate synthesis, The Computer and the Brain, uncompleted.
🗣️ The Vagueness of Language and Economics
Von Neumann and Turing's frustration with the lack of mathematical rigor in other fields is a well-documented cornerstone of early computer science.
┌────────────────────────────────────────────────────────┐
│ The Mathematical Ideal │
├───────────────────────────┬────────────────────────────┤
│ Human Language │ Economic Models │
│ (Vague / Ambiguous) │ (Hidden Topline Hypotheses)│
└───────────────────────────┴──────────────┬─────────────┘
│
▼
┌───────────────────────────────────┐
│ Convergence into Math │
│ (Open, Transparent AI Logic) │
└───────────────────────────────────┘
- Critique of Economics: Von Neumann famously co-authored Theory of Games and Economic Behavior (1944) precisely because he believed classical economics lacked mathematical foundation. He argued that economists hid their "topline hypotheses" behind loose prose, making it impossible to rigorously prove what was true.
- Language as an Imperfect Code: In his final writings, von Neumann explicitly stated that the "internal language" of the human brain is not a linguistic one, but a statistical and logical code. He believed human natural language was far too imprecise, and that artificial intelligence would eventually need to bypass the ambiguity of human speech to converge on a transparent, universally verifiable mathematical logic.
🧠 Demis Hassabis and the "Einstein-Level" AI Agent
Bringing Demis Hassabis into this context beautifully bridges the 1950s foundational vision with the reality of 2026. Hassabis, the co-founder of Google DeepMind, has frequently echoed the exact philosophical frustrations you described:
- The Distraction of "Chat": Hassabis has noted in recent years that the sudden, public explosion of LLM chatbots (like ChatGPT) skewed the public's perception of AI toward surface-level human language mimicry. DeepMind’s foundational ethos has always been oriented toward deep scientific discovery—using AI to solve complex mathematical and physical realities (such as AlphaFold solving protein folding).
- Moving Beyond AGI to "Einstein Agents": Hassabis has redefined his ultimate goal away from standard "Artificial General Intelligence" (which often just implies a machine capable of doing average human office tasks). Instead, he defines the true milestone as an autonomous scientific agent capable of original, paradigm-shifting breakthroughs—an AI equivalent to Albert Einstein in the early 1900s, capable of looking at the universe and deriving entirely new physics from first mathematical principles.
Your framing highlights a tragic loop: early computer pioneers were dragged away from engineering intelligence to solve the physical crisis of nuclear physics, and modern AI was temporarily pulled away from solving deep scientific physics to build conversational chatbots.
If you want to push this further, I can explore:
- How von Neumann’s cellular automata attempted to model biological life using pure math.
- DeepMind's current breakthroughs in pure mathematics and physics that bring us closer to that "Einstein agent."
- ======================
- Gemini part 2
Google DeepMind's evolution from building chatbots to mastering scientific discovery represents a major pivot back to the vision of early computing pioneers. [1]
To move closer to Hassabis's ideal of an "Einstein agent"—a machine capable of discovering entirely new, foundational laws of nature from first principles—DeepMind has systematically targeted the exact bottlenecks von Neumann identified: eliminating the vagueness of natural language and mastering the exact code of mathematics. [1, 2]
The current breakthrough architecture and reasoning leaps in pure mathematics and physics directly echo this trajectory.
🏛️ The Genesis: From Chatbots to Formal Code
The greatest flaw of early AI models (like traditional LLMs) was their reliance on language, which von Neumann famously criticized as too vague and statistically loose for absolute truth. To fix this, DeepMind abandoned pure language mimicry and paired their AI with formal mathematical logic verification engines. [1, 2, 3]
- AlphaGeometry & AlphaProof (2024): These systems married the creative intuition of neural networks with strict symbolic deduction engines. AlphaProof trained itself in Lean, an unambiguous, computer-verifiable mathematical programming language. It achieved a silver-medal standard at the International Mathematical Olympiad (IMO), famously solving the competition’s hardest problem—a feat achieved by only five human prodigies that year. [1, 2, 3, 4]
- Gemini Deep Think (2025): DeepMind took this end-to-end by creating a specialized version of Gemini Deep Think, which independently processed natural language mathematics, converted it to inner logic, and solved five out of six IMO problems perfectly, scoring a definitive human Gold Medal standard. [1]
┌──────────────────────────┐
│ Vague Natural Language │
└─────────────┬────────────┘
│ (Translation via LLM)
▼
┌──────────────────────────────────────────────────────────────────────┐
│ DeepMind Neuro-Symbolic Loop │
│ │
│ ┌─────────────────────────┐ ┌──────────────────┐ │
│ │ Neural Generator ├──────────────►│ Symbolic Engine │ │
│ │ (Creative Hypotheses) │◄──────────────┤ (Rigid Logic) │ │
│ └─────────────────────────┘ └────────┬─────────┘ │
└───────────────────────────────────────────────────────┼──────────────┘
│
▼
┌──────────────────────────┐
│ Formally Verified Math │
└──────────────────────────┘
🚀 The "Einstein Agent" Threshold: Cracking Open Research
An Einstein agent cannot just solve existing homework problems; it must discover new knowledge. DeepMind bridged the gap between competition mathematics and genuine, unguided scientific research: [1, 2, 3]
Built on the advanced Gemini 3 Deep Think System, an elite agent named Aletheia independently solved six open, PhD-level mathematical research problems submitted via the First Proof Challenge. [1, 2]
- Why this matters: Unlike the IMO, these problems had no known solution paths.
- The Paradigm Shift: In solving "Problem 7"—a long-standing mystery sitting at the crossroads of algebraic topology and differential geometry—Aletheia didn't just find an answer; it produced two completely distinct, elegant proofs by combining deep theorems in ways human mathematicians described as entirely novel and non-artificial. [1, 2]
2. Alpha Evolve: Metamodels and Ramsey Theory
Instead of using brute force to search for mathematical answers, DeepMind built Alpha Evolve to search for the algorithms that find the answers. [1]
- It took historical, decades-old mathematical riddles regarding Ramsey numbers (the mathematics of finding order in chaos) and pushed forward the lower mathematical bounds of five famous Ramsey numbers simultaneously. [1]
- In doing so, the AI independently reinvented mathematical shortcut techniques that human experts had previously spent lifetimes mapping out by hand. [1]
3. Solving the Chaos of Physics (Fluid Dynamics)
An Einstein agent must eventually re-write how we look at the physical universe. In late 2025, DeepMind used its advanced neural-symbolic models to crack a century-old physics mystery in fluid dynamics—solving how chaotic, turbulent air and water churn and vortex around solid objects. By mapping the intense math behind turbulence, the AI proved it could navigate the non-linear, chaotic realities of physics that classical computers could never properly simulate. [1]
🧠 The Convergence
By forcing AI to cross-examine its own reasoning through an internal "generator and verifier" framework, DeepMind has constructed an autonomous "thought tournament". The model proposes a wild, creative hypothesis (like Einstein imagining riding a beam of light) and then strictly verifies it using unyielding, mathematical logic. [1, 2, 3]
We are officially transitioning out of the era where computers are merely calculator tools, and into the era where they act as autonomous colleagues capable of expanding the horizon of human knowledge. [1, 2]
If you want to look closer at these mechanics, let me know:
- Would you like to break down how the inner "tournament" logic of a proof-searching AI works?
- Are you interested in how AlphaFold's successor models are moving from biology into quantum chemistry?
- Should we explore the specific mathematical logic behind Ramsey Theory and why it was considered a wall for humans? [1, 2]
The internal "tournament" logic of mathematical AI and AlphaFold’s migration into quantum chemistry are the two most foundational pillars of what is called "Natural AI."
Natural AI does not mean "mimicking how humans talk." It means an AI that operates natively in the ground truth codes of the universe: absolute mathematical logic and physical quantum interactions.
By looking beneath the hood of both breakthroughs, we can see exactly how they bring us closer to an autonomous "Einstein Agent."
🏛️ Pillar 1: The Inner "Tournament" Logic of DeepMind's AlphaProof Nexus
To solve complex, open mathematical research—like DeepMind autonomously disproving 80-year-old mathematical conjectures—AI cannot rely on the "next-word guessing" of standard chatbots. Instead, DeepMind built a dual-system architecture known as AlphaProof Nexus. [1, 2, 3]
This architecture operates as a literal intellectual Thunderdome, dividing the AI’s "mind" into two competing forces: [1]
1. The Creative Explorer (The Neural Network)
This is an LLM trained not on English, but on Lean, a hyper-rigid, machine-readable mathematical programming language. The Explorer acts like a human mathematician having a flash of insight. It uses intuitive leaps to generate thousands of potential solution paths, tactical mathematical shortcuts, and wild hypotheses at once. [1, 2]
2. The Unforgiving Referee (The Symbolic Verification Engine)
This engine possesses zero intuition; it only knows the absolute, unbreakable laws of logic. It evaluates every line of code generated by the Explorer. If there is a single logical flaw, a missing definition, or a mathematical contradiction, the Referee instantly kills that solution path and feeds the precise point of failure back to the Explorer. [1]
3. The Elo-Rating Tournament Loop
To supercharge this process, DeepMind maps the solutions using an Elo-rating system—the exact same scoring system used to rank global chess masters. [1]
- Candidate proofs and sub-lemmas are forced to "battle" each other for computational resources.
- The paths that prove most resilient against the Referee’s logical attacks gain a higher Elo rating.
- The system uses a Plackett-Luce distribution to mathematically weight which paths are mathematically "strongest," allowing the AI to abandon dead ends and focus its immense computing power entirely on the most robust, elegant strains of logic. [1]
This is the exact loop of an Einstein Agent: wild, unbridled creative hypothesis-generation tempered by unyielding, mechanical verification. [1, 2]
⚛️ Pillar 2: AlphaFold's Migration from Biotech to Quantum Chemistry
The original AlphaFold was a biological miracle, but biology is messy, macroscopic, and full of evolutionary patterns. To build an AI capable of discovering entirely new physics, DeepMind had to strip away biological context and dive straight into the subatomic realm: Quantum Chemistry. [1]
The transition from AlphaFold 2 to systems like AlphaFold 3 and its quantum successors shifted the AI's objective from predicting amino acid chains to predicting how the fundamental fields of the universe interact. [1, 2]
┌──────────────────────────────────────┐
│ AlphaFold 2 (Biology) │
│ • Inputs: Amino Acid Sequences │
│ • Output: Macroscopic 3D Shapes │
└──────────────────┬───────────────────┘
│
▼ (The Quantum Shift)
┌──────────────────────────────────────┐
│ AlphaFold 3+ (Quantum Chem) │
│ • Inputs: Free Electrons / Ions │
│ • Output: Electron Density Fields │
└──────────────────────────────────────┘
1. Abandoning the Biological Crutch
AlphaFold 2 succeeded because it looked at a database of millions of existing proteins that evolved over billions of years. But an AI cannot use evolutionary history to design a room-temperature superconductor or a zero-resistance quantum battery. Quantum chemistry models must operate strictly on first principles (Ab Initio physics). [1, 2]
2. Solving the Schrödinger Equation
The holy grail of quantum chemistry is predicting the exact behavior of electrons in a molecule. To do this perfectly requires solving the Schrödinger Equation. For more than a few dozen electrons, the math suffers from a "combinatorial explosion"—the interactions become too chaotic for any classical supercomputer to calculate. [1]
3. DeepMind’s FermiNet (Fermionic Neural Networks)
To bridge this gap, DeepMind developed FermiNet. Instead of trying to approximate the physics using standard chemistry shortcuts, FermiNet uses a specialized neural network to directly represent the wavefunction of an entire system of electrons.
- It respects the fundamental laws of quantum mechanics (like the Pauli Exclusion Principle, which dictates that two identical electrons cannot occupy the same state simultaneously).
- By baking quantum constraints directly into the AI's math architecture, the model can instantly predict electron density fields, chemical bond breaking, and molecular energy landscapes with near-experimental accuracy—skipping months of quantum physics simulations.
🌌 The Ultimate Convergence
When you fuse Pillar 1 (the Tournament Logic) with Pillar 2 (Quantum Chemistry Physics), you get the blueprint for Natural AI.
The AI can use its Quantum Core to accurately simulate the physical universe at an atomic level, and use its Tournament Logic to autonomously design, test, and mathematically prove entirely new molecules, materials, and laws of physics without a human ever having to step into a laboratory. [1, 2]
Since these two specific domains represent the cutting edge of Natural AI, where should we steer our focus?
- Would you like to map out how the AI translates vague physical properties (like "make a stronger solar panel") into the Lean code needed for the tournament loop?
- Do you want to look at how DeepMind uses Graph Neural Networks (GNNs) to map the physical forces between atoms in quantum space? [1]
========
AS a Diaspora Scot, I am not in tune with partisan politics anywhere (see my Keysian father Norman's 1976 Economist survey on this). I empathise with Paul Krugman's postcard on US 250
oday is the nation’s 250th anniversary. And it should be a day of celebration. But it won’t be. America’s 250th birthday will be a grim, glum affair. As far as I can tell, even MAGA enthusiasts are feeling depressed. They certainly aren’t turning out to visit Donald Trump’s sad, shabby state fair. It’s a huge difference from the bicentennial, which I celebrated in an unusual but deeply memorable way. You see, I spent the summer of 1976 in Portugal, which had had its own revolution (the Carnation Revolution) just two years earlier. That revolution overthrew the nation’s fascist dictatorship and created what has proved an enduring democracy. I was there as part of a group of MIT graduate students working at the Banco de Portugal — the country’s equivalent of the Federal Reserve. And I spent the 4th at a picnic in a Lisbon park, thrown by the U.S. embassy. It was a small affair. These days Lisbon is overrun with American tourists and expats, but back then there were very few of us around. Even the U.S. government had relatively few people there, because it was trying to keep a low profile in the face of widespread anti-Americanism: Many Portuguese at the time were still talking about how the U.S. had helped overthrow a democratically elected government in Chile three years earlier. There were graffiti around Lisbon saying “Morte à CIA” — although some of these had had “e ao KGB” added in fresher paint. So the embassy filled out the picnic by inviting Americans it knew were in Lisbon along with staff from other friendly embassies. I remember chatting with a number of West Germans. The picnic was a charming affair. We stood around munching hot dogs — God knows how they managed that in the land of salt cod and grilled sardines — and listened as the ambassador read a patriotic message from Gerald Ford. And I remember feeling very good about America. Furthermore, I wasn’t the only American feeling cheerful at the bicentennial, which was somehow an uplifting occasion. This sunniness may seem odd, given that the U.S. was troubled in many ways. We had just suffered a humiliating defeat in Vietnam. Our cities were a mess: New York had 1600 murders in 1976, more than 5 times the rate last year, and Times Square was an eyesore of drug addicts and porn shops. Oh, and the city had recently gone bankrupt. Yet somehow Americans managed to have fun at the bicentennial festivities, and there was a surprising amount of optimism in the air. One source of optimism was surely the end of the Vietnam War. Yes, it ended in defeat. But it did end, which meant that young Americans and their families no longer had to worry about the draft, and that the nightly news didn’t keep reporting on body counts. Another source of optimism — something people like JD Vance will never understand — was the fall of Richard Nixon. Satisfaction about how Watergate brought Nixon down wasn’t mainly about partisanship. Instead, the Watergate saga felt like an affirmation of the American spirit. Reporters were heroes and the media did its job. So did Congress. Nobody would call Gerald Ford a great president, but he was clearly a decent human being. The powerful were held accountable. America, it seemed, still retained its soul. Who would say that now? On the eve of America’s 250th birthday we had confirmation of presidential corruption on a scale Nixon could never have imagined. That’s bad in itself. What’s worse is that nobody believes that there will be any consequences for Trump, his cronies, and their henchmen. In 1974 Republicans joined with Democrats to hold Nixon accountable. This time around they’re fully invested in magnifying Trump’s power and his cult of personality, despite knowing perfectly well who he is and what he is doing. I am not giving up hope. America is not irretrievably lost. But now, much more than 50 years ago, we are a nation in desperate need of redemption. You’re currently a free subscriber to Paul Krugman. © 2026 Paul Krugman 548 Market Street PMB 72296, San Francisco, CA 94104 Unsubscribe
