Theoretical Physics With Generative AI – #101

Introduces the core analogy of working with AI models as consulting a genius colleague who has encyclopedic knowledge and can do lightning calculations, but makes both simple and profound mistakes. Establishes that while these intelligences don't understand physics like humans do, they can be used fruitfully with proper verification.

• •AI models have encyclopedic mastery of literature and can perform rapid calculations
• •Models are capable of deep insights but also simple and profound errors
• •Their understanding of physics differs fundamentally from human reasoning
• •Success requires treating AI outputs with appropriate skepticism and verification

Details the practical methodology of using multiple AI instances in a generate-verify pipeline to suppress errors and hallucinations. Explains testing frontier models (GPT-5, Gemini, Quinn Max) on physics papers and the collaboration with DeepMind's CoScientist tool.

• •Chain different model instances as generators and verifiers to reduce error rates
• •Simple verification prompt: 'You are a world class theoretical physicist. Check the following for errors'
• •Cycling through generate-verify-modify loops suppresses both simple errors and deep conceptual confabulations
• •Frontier models can solve graduate-level physics problems, suggesting research utility
• •DeepMind's CoScientist represents a souped-up version designed for scientific collaboration

Describes how GPT-5 spontaneously proposed applying Tomonaga-Schwinger formalism to nonlinear quantum mechanics while being tested on a 2015 paper. The model generated a correct equation that became the foundation of a peer-reviewed Physics Letters B publication.

• •Testing involved uploading old papers to see if models could understand and suggest follow-up work
• •GPT-5 proposed using Tomonaga-Schwinger formulation for state-dependent quantum mechanics modifications
• •The model generated technically correct equations within seconds that led to publishable research
• •This represents combinatorial innovation - connecting distant areas of physics rather than Einstein-level breakthroughs
• •First published theoretical physics paper where the main idea originated from an AI

Discusses the most time-costly failure mode: when models propose plausible-sounding connections between distant areas that turn out to be wrong. Covers the axiomatic quantum field theory example and strategies for detecting unreliable suggestions.

• •Models can suggest applying technique B to problem A, sounding plausible but wasting researcher time
• •Axiomatic QFT suggestion required learning new area only to discover it was incorrect
• •Lack of convergence across multiple models signals likely incorrect proposals
• •Convergence across diverse models provides positive (but not guaranteed) signal of correctness
• •Generator-verifier cycling helps detect these deep confabulations before significant time investment

Analyzes who can effectively use AI for research, arguing that high-level expertise is still required to avoid generating 'subtly wrong slop.' Discusses the productivity boost for experienced researchers versus risks for PhD students.

• •PhD students 1-2 years from graduation are still novices from frontier research perspective
• •Less experienced researchers risk generating 'subtly wrong slop' without catching errors
• •Thousands of expert physicists/mathematicians could experience major productivity boosts
• •Models excel at computations, literature searches, LaTeX generation, citations, and proofreading
• •Students should learn carefully; experts should experiment but remain vigilant about wild goose chases

Explains using AI models to verify that research results are genuinely novel and not just reproductions of existing work. Demonstrates prompts for careful literature searches and discusses concerns about AI regurgitating training data.

• •Repeatedly asked models: 'Do the TS conditions for state dependent QM exist elsewhere in the literature?'
• •Critical concern: ensuring AI-generated ideas are truly novel, not memorized from training data
• •Models useful for finding papers where results could have been anticipated
• •Post-publication, no physicists have identified prior work containing these specific equations
• •Model likely made combinatorial connection between Tomonaga-Schwinger and Weinberg-style work

Details how GPT-5 performed geometric analysis of past light cones to test the Kaplan-Rajendran model against new integrability conditions. Demonstrates AI's ability to handle complex spatial reasoning and develop elegant notation.

• •Referee requested testing recent Kaplan-Rajendran nonlinear QM model against TS equations
• •Calculation required geometric analysis of overlapping past light cones in spacetime
• •GPT-5 correctly solved the geometry problem despite LLMs typically struggling with spatial reasoning
• •Model developed elegant notation (set J for past light cone, intersection operations)
• •Gemini and Quinn Max initially got it wrong; verification cycling confirmed GPT-5's correctness
• •KR model appears to violate the integrability conditions

Discusses building automated generator-verifier pipelines, collaboration with xAI on browser-based tools, and the path toward better AI research assistants. Emphasizes that models will improve as expert scientists generate training data through actual research use.

• •Built prototype browser-based tool automating generate-verify pipeline with xAI collaboration
• •Tool designed to work with any model, not just Grok
• •Chaining multiple models creates 'supermodel' with better performance than individuals
• •Similar to DeepMind's CoScientist approach; ongoing collaboration continues
• •Models can solve textbook problems and hard competition problems; lack deepest expert-level understanding
• •Training data from expert scientists using models daily will close remaining comprehension gaps

Provides technical overview of the physics problem: whether quantum mechanics must be exactly linear or could have nonlinear corrections. Covers Weinberg's 1980s work, the Schrödinger equation, and implications for quantum superposition and the many-worlds interpretation.

• •Schrödinger equation is completely linear - states only appear to first power
• •Linearity enables quantum superposition and many-worlds interpretation
• •Fundamental question: is linearity exact or an approximation with small corrections?
• •Weinberg studied nonlinear modifications in 1980s; testing these modifications is important
• •Nonlinearity allows distant objects to become instantaneously entangled (problematic)
• •Violates special relativity and enables signaling between Everett branches

Deep dive into the technical physics results: Tomonaga-Schwinger integrability conditions for quantum field theory, how they constrain nonlinear modifications, and implications for microcausality. Discusses whether any nonlinearity in quantum mechanics necessarily destroys locality.

• •Tomonaga-Schwinger equation describes quantum field evolution with 'many-fingered time'
• •Integrability conditions ensure space-like separated regions can't influence each other
• •State-dependent modifications require new functional derivative machinery
• •Weinberg-type nonlinearity satisfies conditions only if microcausality still holds
• •Microcausality assumption questionable once nonlinearity introduced - standard justification breaks down
• •Deep question: any nonlinearity may require abandoning locality entirely in quantum theories
• •Instantaneous entanglement of distant wave packets suggests microcausality violations
• •Fundamental tension: locality requirements may force exact linearity in quantum mechanics