Curt Jaimungal · Published 2026-10-05

The Physicist Revolutionizing Physics With AI

privateReview: unreviewedConfidence: high
formal verificationneuro-symbolic aiquantum gravityautomated reasoningtheoretical physicsneuro-symbolicformal-verificationproof-assistantsleandomain-specific-languageauto-formalizationcognitive-debt
Open on YouTube ↗

Summary

Overview

  • Speaker: Jonathan Gorard
  • Channel: Curt Jaimungal
  • Main topic: Applying AI and formal verification to fundamental physics and mathematics
  • Purpose: To explore how AI, formal verification, and automated reasoning will transform mathematics and theoretical physics, and to introduce Lanyon.AI's approach to formalizing scientific computing. Jonathan Gorard, applied mathematician and co-founder of Lanyon.AI and the Wolfram Physics Project, discusses the impact of AI on mathematics and physics. He covers automated theorem proving, formal verification, the timeline of AI capabilities in math, the crisis of physics and math, and why physics requires 'neuro-symbolic' domain-specific languages rather than standard LLMs.

Topic Map

The Impact of AI on Mathematics and Physics

  • Explanation: Discussion on how AI is making mathematics and physics executable, moving beyond code-based progress to algorithmic and scientific proof generation.
  • Key claims:
    • AI is making scientific progress distinct from just mathematical or code-based progress.
    • Physics needs to become executable in the same way software and mathematics have become executable.
  • Examples:
    • Maryna Viazovska's 8- and 24-dimensional sphere packing proof automatically formalized using Lean.
    • OpenAI's formal proof of the Navier-Stokes blowup conjecture.
  • Terminology:
    • Formal Verification
    • Automated Theorem Proving
    • Lean
    • Neuro-symbolic AI
  • Why it matters: It changes the fundamental nature of how mathematicians and physicists work, threatening traditional workflows and requiring new tooling.

The Timeline of AI Capabilities and Grief in Academia

  • Explanation: Gorard reflects on his personal timeline of grief regarding how rapidly AI capabilities advanced, shifting his expectations from decades to months.
  • Key claims:
    • AI capabilities in formal theorem proving advanced much faster than expected in early 2026.
    • Academia is not prepared for neuro-symbolic AI solving foundational mathematical problems at superhuman speeds.
  • Examples:
    • Math Inc.'s automated formalization of Viazovska's sphere packing theorem in March 2026.
  • Terminology:
    • Auto-formalization
    • Cognitive Debt
    • Autonomous Proof Generation
  • Why it matters: It disrupts long-term career planning for researchers and creates an existential crisis for human mathematicians and physicists.

Science as Process vs. Science as Product

  • Explanation: Distinguishing between the end-products of science (the papers and theorems) and the process of discovery and curiosity.
  • Key claims:
    • Society and funding agencies reward the product of science, but scientists are motivated by the process.
    • AI threatens to automate the process, leaving humans disconnected from the intrinsic joy of discovery.
  • Examples:
    • Curie cancer, solve nuclear fusion, generate clean energy as products vs. protein folding and quantum gravity as process.
  • Terminology:
    • Science as Product
    • Science as Process
    • Hedonism
    • Cognitive Friction
  • Why it matters: If AI automates the process of thinking and discovery, the human motivation for doing science collapses.

The Limits of Traditional LLMs in Physics

  • Explanation: Why standard LLMs talking in English are inadequate for physics, and why domain-specific formal languages are needed.
  • Key claims:
    • English is a poorly suited language for efficient physical reasoning.
    • Neuro-symbolic architectures combining LLMs with domain-specific proof assistants are necessary.
  • Examples:
    • Toki Pona as an example of minimal expressive language.
    • Wightman and Haag's axioms in algebraic quantum field theory.
  • Terminology:
    • Neuro-symbolic AI
    • Domain-Specific Language
    • Wightman Axioms
    • Haag's Theorem
  • Why it matters: Explains why Lanyon.AI is building specialized substrate tooling rather than relying on general-purpose chatbots.

Key Points

Executable Mathematics and Physics

  • Explanation: Just as software became executable, and mathematics became executable via proof assistants, physics must become formally verifiable and executable.
  • Evidence: Recent automated formalizations of Fields Medal-winning results in Lean.
  • Practical implication: Physicists will need to use formal verification tools to validate their models against physical reality.

The Transmissibility of Understanding

  • Explanation: Solving a problem computationally does not automatically transmit understanding to human brains; understanding requires cognitive friction.
  • Evidence: Reading a math paper vs. reproducing the proofs yourself.
  • Practical implication: We risk relying on AI black boxes without building human intuition or verifiable conceptual frameworks.

Under-constrained Physical Theories

  • Explanation: Most fundamental physics theories are massively under-constrained compared to mathematics, making formalization much harder.
  • Evidence: Einstein's field equations vs. Wightman axioms in quantum field theory.
  • Practical implication: We need new formal languages that capture physical intuition without sacrificing mathematical rigor.

Frameworks, Models & Processes

Input-Process-Output Framework for Intellectual Disciplines

  • How it works: Analyzing fields based on their societal funding inputs, internal reasoning processes, and published outputs.
  • Components:
    • Inputs (Funding, Societal Structure)
    • Process (Curiosity-driven reasoning)
    • Outputs (Papers, Proofs, Theorems)
  • When to use: Evaluating the sustainability of academic disciplines in the age of AI automation.

Neuro-symbolic AI Architecture

  • How it works: Combining neural language models for generative spark with symbolic proof checkers for rigorous verification.
  • Components:
    • Neural Generator (LLM for creativity)
    • Symbolic Verifier (Proof assistant for correctness)
  • When to use: Building reliable AI systems for scientific discovery and formal verification.

Examples & Case Studies

Maryna Viazovska's 8- and 24-dimensional sphere packing proof was fully auto-formalized by Math Inc. in March 2026.

  • Illustrates: The sudden acceleration of automated mathematical reasoning using AI and Lean.
  • Lesson: AI can now handle highly complex, multi-year mathematical achievements end-to-end much faster than anticipated.

Galileo's thought experiment comparing falling bodies of different weights.

  • Illustrates: How physical insight can sometimes be derived through pure thought experiments (thought verification) rather than brute-force physical testing.
  • Lesson: Intuition and internal consistency play a massive role in physical reasoning before empirical data is gathered.

Actionable Takeaways

  • Immediate:
    • Check out Lanyon.AI for tools connecting AI to physical verification.
    • Subscribe to Curt Jaimungal's Substack at curtjaimungal.org for bonus written content and early podcast episodes.
  • Strategic:
    • Prepare for the automation of academic research workflows by learning formal proof assistants like Lean.
    • Recognize that physical theories require more than just raw compute; they require rigorous axiomatic foundations.
  • Questions to investigate:
    • How can we build an analog of the Turing machine for physical and scientific reasoning?
    • Can AI discover new physical laws that are fundamentally uninterpretable to human intuition?

Claims Worth Verifying

  • Math Inc. dropped a multi-hundred-thousand-line autonomous Lean proof of Viazovska's sphere packing in March 2026. (Factual prediction/scenario presented in video)
  • OpenAI generated a formal proof of finite time blowup for Navier-Stokes. (Factual prediction/scenario presented in video)

Notable Quotes

"We're not putting the genie back in the bottle." (at 0:03) "I'm in a continual process of grief." (at 0:44) "The limits of my language mean the limits of my world." (at 100:25)

Compressed Summary

  • AI is rapidly automating formal mathematical theorem proving.
  • Physics is harder to formalize than math due to under-constrained physical principles.
  • Lanyon.AI is building neuro-symbolic compilers and verification substrates for scientific computing.
  • Human scientific motivation relies on the process of discovery, not just the output.
  • Keywords: formal verification, neuro-symbolic ai, quantum gravity, automated reasoning, theoretical physics
  • Core insight: Neuro-symbolic AI and formal verification will revolutionize physics and mathematics, forcing researchers to transition from manual proof-checking to orchestrating executable scientific models.

Core insights

6
Architecturemedium noveltymoderate evidence

Reliable AI-for-science systems are architected as a neuro-symbolic split: an LLM neural generator produces candidate reasoning, while a separate symbolic proof checker performs rigorous verification. The generative spark and the correctness guarantee live in different components.

Why it matters

Separating probabilistic generation from deterministic symbolic verification is the mechanism that makes AI scientific reasoning trustworthy; it places the trust boundary at the verifier rather than the model, which changes where evaluation, guardrails, and failure handling belong.

Generalization

For any high-stakes generative pipeline, pair a neural proposal step with an independent symbolic or rule-based verifier rather than trying to make the generator itself reliable.

Combining neural language models for generative spark with symbolic proof checkers for rigorous verification.
Open source video
Neuro-symbolic architectures combining LLMs with domain-specific proof assistants are necessary.
Open source video
Mental Modelhigh noveltymoderate evidence

Physics is harder to formalize than mathematics because its theories are massively under-constrained: unlike math (e.g. Wightman axioms) there is no tight axiomatic scaffold, so the same formal machinery that works for theorems does not transfer directly and new formal languages are needed.

Why it matters

It signals that 'just point a prover at the domain' fails when the domain lacks well-defined constraints, implying that modeling the constraints themselves is the hard part of the tooling.

Generalization

Automated reasoning works best over tightly constrained, well-specified formalisms; domains with loose or ambiguous specifications need an intermediate formalization layer before any solver can help.

Most fundamental physics theories are massively under-constrained compared to mathematics, making formalization much harder.
Open source video
Einstein's field equations vs. Wightman axioms in quantum field theory.
Open source video
We need new formal languages that capture physical intuition without sacrificing mathematical rigor.
Open source video
Mechanismhigh noveltymoderate evidence

Natural language (English) is a poorly suited substrate for physical reasoning, so effective systems should target domain-specific formal languages rather than general-purpose chat interfaces.

Why it matters

Argues against the default 'English-in, English-out' agent interface for technical domains and for a DSL/substrate layer, which shifts where the interface and reasoning burden live in the architecture.

Generalization

For technically precise domains, investing in a formal/declarative intermediate representation can outperform expanding natural-language prompting.

English is a poorly suited language for efficient physical reasoning.
Open source video
why physics requires 'neuro-symbolic' domain-specific languages rather than standard LLMs
Open source video
Toki Pona as an example of minimal expressive language.
Open source video
Failure Modehigh noveltymoderate evidence

Computational problem-solving does not transmit understanding to humans; understanding requires cognitive friction (reproducing proofs yourself vs. reading a paper). Automating the output can leave humans with black boxes and no intuition.

Why it matters

Identifies a systemic risk of AI-accelerated science: the artifacts of knowledge can be produced without the comprehension, creating dependence on unverifiable-by-humans machinery.

Generalization

Systems that generate correct outputs without forcing the human through the reasoning risk creating unmaintainable, unauditable knowledge debt ('cognitive debt').

Solving a problem computationally does not automatically transmit understanding to human brains; understanding requires cognitive friction.
Open source video
Reading a math paper vs. reproducing the proofs yourself.
Open source video
We risk relying on AI black boxes without building human intuition or verifiable conceptual frameworks.
Open source video
Empirical Resultmedium noveltymoderate evidence

Auto-formalization of extremely hard results (e.g. Viazovska sphere-packing proof into Lean, Math Inc., March 2026) shows formal-reasoning capability advancing faster than even the speaker expected—years compressed into months.

Why it matters

Concrete evidence that automated formalization is progressing on the fastest timeline, which should drive earlier investment in proof-assistant tooling and benchmark design.

Generalization

Capability timelines in formal verification are compressing; planning assumptions measured in decades should be re-baselined in months.

Math Inc.'s automated formalization of Viazovska's sphere packing theorem in March 2026.
Open source video
AI capabilities in formal theorem proving advanced much faster than expected in early 2026.
Open source video
Maryna Viazovska's 8- and 24-dimensional sphere packing proof automatically formalized using Lean.
Open source video
Predictionmedium noveltyweak evidence

The paradigm being proposed is making physics 'executable'—formal verification tooling to validate physical models against reality, mirroring how software and then mathematics became executable via proof assistants.

Why it matters

Frames a new tool category (executable/verifiable physics) that changes the expected workflow from informal modeling to formal, checkable model validation.

Generalization

As domains become 'executable' via verification tooling, the tooling substrate becomes the bottleneck and the primary place to build.

Physics needs to become executable in the same way software and mathematics have become executable.
Open source video
Just as software became executable, and mathematics became executable via proof assistants, physics must become formally verifiable and executable.
Open source video

Deep dives

5

Verification boundary and auto-formalization error attribution

Research question

In neuro-symbolic scientific pipelines, where do errors actually originate: in the LLM's auto-formalization step or in proof generation against the formal statement, and can stage-isolated verification reduce end-to-end failures?

Why

Determines whether symbolic proof checking truly eliminates hallucination risk or merely moves it upstream; dictates where to instrument, evaluate, and gate scientific AI systems.

Combining neural language models for generative spark with symbolic proof checkers for rigorous verification.
Open source video
Neuro-symbolic architectures combining LLMs with domain-specific proof assistants are necessary.
Open source video
Math Inc.'s automated formalization of Viazovska's sphere packing theorem in March 2026.
Open source video
Source video

Formal languages for under-constrained physical theories

Research question

What formal language or constraint-annotation layer can make under-constrained physical theories machine-checkable without losing physical intuition, and how should its fidelity to reality be validated?

Why

Physics lacks the tight axiomatic scaffold of math; without a formalization layer, automated provers cannot meaningfully operate. The design of that layer is the bottleneck.

Most fundamental physics theories are massively under-constrained compared to mathematics, making formalization much harder.
Open source video
We need new formal languages that capture physical intuition without sacrificing mathematical rigor.
Open source video
Einstein's field equations vs. Wightman axioms in quantum field theory.
Open source video
Source video

Cognitive friction and explanation reconstruction from formal proofs

Research question

Can machine-checked proofs be automatically transformed into human-facing derivations that build understanding without reintroducing the full cognitive friction of manual proof reproduction?

Why

If verified outputs become black boxes, science gains correctness but loses human comprehension, auditability, and intrinsic motivation; explanation generation is the mitigation.

Solving a problem computationally does not automatically transmit understanding to human brains; understanding requires cognitive friction.
Open source video
Reading a math paper vs. reproducing the proofs yourself.
Open source video
We risk relying on AI black boxes without building human intuition or verifiable conceptual frameworks.
Open source video
Source video

Executable physics: architecture and validation oracle

Research question

What system architecture makes physics executable in the way software and mathematics are, and what serves as the validation oracle for a formal physical model?

Why

Frames a new tool category; if physics becomes executable, verification tooling becomes primary; but physics under-constraint means validation against reality is unresolved.

Physics needs to become executable in the same way software and mathematics have become executable.
Open source video
Just as software became executable, and mathematics became executable via proof assistants, physics must be formally verifiable and executable.
Open source video
Source video

Forecasting formal-reasoning capability compression

Research question

What benchmarks and milestone records substantiate faster-than-expected progress in formal theorem proving, and how should research investment timelines be re-baselined?

Why

Capability timelines shifting from decades to months affect career planning, funding, and tooling investment; need falsifiable tracking.

AI capabilities in formal theorem proving advanced much faster than expected in early 2026.
Open source video
Math Inc.'s automated formalization of Viazovska's sphere packing theorem in March 2026.
Open source video
Maryna Viazovska's 8- and 24-dimensional sphere packing proof automatically formalized using Lean.
Open source video
Source video

Article ideas

4

The Verifier Is the Product, Not the Model

Trustworthy AI-for-science should be architected around an independent symbolic verifier; model capability is secondary, and the verifier becomes the critical dependency, audit point, and failure boundary.

Angle

Architecture argument with engineering consequences for eval, guardrails, and release gates.

Source video

English Is the Wrong Substrate for Physics

General-purpose natural-language LLM interfaces are inadequate for precise physical reasoning; progress will come from domain-specific formal languages and translators, not from larger chat models.

Angle

Contrarian engineering thesis on where to invest: DSLs and formal intermediate representations over prompt engineering.

Source video

Cognitive Debt Is the Hidden Cost of Automated Discovery

Automating scientific discovery can produce correct artifacts without producing human understanding; without explanation reconstruction, verified outputs become unauditable black boxes and erode the motivation for science.

Angle

Risk and process essay connecting verification to human comprehension.

Source video

Why Physics Is Harder Than Math for AI

Under-constrained physical theories cannot be fed directly to automated provers; the hard engineering work is building a constraint-modeling formalization layer before any solver can help.

Angle

Comparative analysis of domain formalizability and pipeline design.

Source video

Project ideas

4

AutoFormalization Error Attribution Benchmark

beyond-evals

In neuro-symbolic scientific pipelines on under-constrained physics problems, auto-formalization errors cause more end-to-end failures than proof-generation errors against the formalized statement.

Proof of concept

Build a small suite of physics derivations; have an LLM auto-formalize them into Lean; separately generate proofs; inject known errors at each stage; measure which stage propagates to accepted false results.

Measurement

Percentage of end-to-end failures attributed to formalization versus proof generation; false-accept rate when only proof checking is applied.

Source video

Constraint-Annotated Physics DSL Prototype

new

A minimal domain-specific language with explicit constraint annotations improves machine-checkability and human review of under-constrained physical models compared with English-only LLM prompting.

Proof of concept

Implement a toy DSL for classical field equations with under-constraint flags; translate 10 physics problems; attempt formal consistency checks; compare against natural-language prompting baseline.

Measurement

Proof-check success rate, ambiguity/error count, and expert-rated fidelity to physical intent for DSL versus English.

Source video

Proof-to-Explanation Reconstruction Tool

movement-lab

Automatically reconstructed human-facing derivations from machine-checked proofs improve comprehension and reduce cognitive debt compared with raw formal proof artifacts, without requiring manual proof reproduction.

Proof of concept

Prototype a tool that takes Lean proof traces and generates stepwise explanations; run a small user study comparing comprehension of raw proof versus reconstructed explanation.

Measurement

Comprehension quiz scores, time-to-understanding, and perceived auditability; compare raw proof versus reconstructed derivation.

Source video

Formal Reasoning Capability Compression Tracker

beyond-evals

A benchmark suite of formal theorem-proving tasks and milestone dates can quantify capability compression and forecast near-term progress better than expert intuition.

Proof of concept

Collect public formalization milestones such as sphere-packing auto-formalization and define recurring benchmark tasks; track solve rates and time-to-formalize over model generations.

Measurement

Slope of capability improvement per month; forecast error against future formalization milestones.

Source video

Architectural implications

3

Trust is anchored at a symbolic verifier rather than the generative model.

Before

Reliability is pursued by improving the LLM or by human review of natural-language output.

After

A separate deterministic proof/checker component validates every generated artifact before it is accepted.

Consequence

The verifier becomes the critical dependency and creates a clear trust boundary, failure-handling path, and audit point in the system.

Source video

The reasoning interface is a domain-specific formal language, not English.

Before

Agents interact through natural-language prompts and replies.

After

A DSL/formal intermediate representation mediates between the user/agent and the solver.

Consequence

Shifts engineering investment from prompt engineering toward building and maintaining formal substrates and translators.

Source video

Under-constrained domains block direct application of solvers.

Before

Assume any well-known domain can be fed to an automated prover.

After

Introduce an explicit formalization/constraint-modeling layer before the solver stage.

Consequence

Adds a required pipeline stage whose quality determines whether downstream verification is even meaningful.

Source video

Tradeoffs and failure modes

3

Verifier-backed automation vs. human understanding

Benefit

Correct outputs can be generated and machine-checked at superhuman speed.

Cost or risk

Humans may rely on black boxes without building intuition, producing 'cognitive debt' and unauditable conceptual frameworks.

We risk relying on AI black boxes without building human intuition or verifiable conceptual frameworks.
Open source video
Source video

Specialized DSL substrate vs. general-purpose LLM

Benefit

Domain-specific formal languages are better suited for precise physical reasoning.

Cost or risk

Requires building and maintaining specialized tooling instead of leveraging general chatbots, limiting reuse.

why physics requires 'neuro-symbolic' domain-specific languages rather than standard LLMs
Open source video
Source video

Automating discovery vs. human motivation

Benefit

The process of scientific discovery can be automated.

Cost or risk

If AI automates the process of thinking and discovery, the human motivation for doing science collapses.

If AI automates the process of thinking and discovery, the human motivation for doing science collapses.
Open source video
Source video

Open questions

4

How should understanding be transmitted from a verified artifact back to humans without reintroducing the cognitive friction that slowed the process?

Why unresolved

The summary asserts cognitive friction is required for understanding but does not propose a mechanism to preserve it under automation.

Research direction

Design tooling that reconstructs human-facing derivations from machine-checked proofs (explanation generation from formal proofs).

Source video

What does a formal language for under-constrained physical theories actually look like, and how is its fidelity to physical reality validated?

Why unresolved

The summary states new formal languages are needed but offers no concrete design or validation approach.

Research direction

Prototype candidate physical DSLs and benchmark whether they capture intuition while remaining checkable.

Source video

Does symbolic verification eliminate hallucination in scientific reasoning, or merely move the failure point to the formalization step?

Why unresolved

The verifier guarantees consistency with the formalized statement, but the auto-formalization step itself is model-driven and not addressed.

Research direction

End-to-end benchmarks measuring errors introduced during auto-formalization vs. during proof generation.

Source video

How does the under-constrained nature of physics change the reliability targets for neuro-symbolic pipelines compared to mathematics?

Why unresolved

The summary contrasts math and physics but does not quantify how much harder verification becomes.

Research direction

Comparative evaluation of verification success rates between axiomatized math and under-constrained physics theories.

Source video

Key claims

6
comparativeVerification needed

Standard English-based LLMs are inadequate for physics; domain-specific formal languages are required.

Evidence

English is a poorly suited language for efficient physical reasoning.

Question

Do DSL-based/formal pipelines measurably outperform natural-language LLM prompting on physics reasoning tasks?

Source video
causalVerification needed

Combining LLM generation with symbolic proof checkers is necessary for reliable scientific discovery systems.

Evidence

Neuro-symbolic architectures combining LLMs with domain-specific proof assistants are necessary.

Question

Does adding a symbolic verifier measurably reduce errors versus a generator-only pipeline in scientific reasoning?

Source video
comparativeVerification needed

Fundamental physics theories are far more under-constrained than mathematics, making formalization harder.

Evidence

Most fundamental physics theories are massively under-constrained compared to mathematics, making formalization much harder.

Question

Can the degree of under-constraint be quantified to predict formalization difficulty across domains?

Source video
causalVerification needed

Computational solving does not transmit understanding; understanding requires cognitive friction.

Evidence

Solving a problem computationally does not automatically transmit understanding to human brains; understanding requires cognitive friction.

Question

Can any automated process produce equivalent human understanding of a result compared with reproducing it manually?

Source video
factualVerification needed

AI formal theorem-proving capability advanced much faster than expected in early 2026.

Evidence

AI capabilities in formal theorem proving advanced much faster than expected in early 2026.

Question

What concrete benchmarks or results substantiate the claim of faster-than-expected progress in early 2026?

Source video
predictionVerification needed

Physics must become formally verifiable and executable, as software and mathematics did.

Evidence

physics must become formally verifiable and executable.

Question

Is a formal, executable physics substrate achievable, and what would it validate against?

Source video

Connections

4