The real bottleneck in AI physics is descriptive language, not compute
Math has formal verification tools that let AI iterate unsupervised; physics lacks them. The same puzzle becomes obvious when described differently. Language bottlenecks progress, not intelligence.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
A major math proof came two years earlier than expected
In March 2026, Math Inc. used AI to automatically prove the high-dimensional sphere-packing problem that Maryna Vyazovska won the Fields Medal for. Gorard had expected this to happen no sooner than late 2027 or 2028, so it arrived roughly two years ahead of schedule. His reaction wasn't ‘AI is so powerful’ but ‘I got the problem wrong—my entire timeline is completely wrong.’ It wasn't just that his research plan needed adjusting; it fundamentally altered how he thinks about his life's trajectory. Worse, academia had no time to adapt.
— Jonathan GorardAcademic funding thrived on blurring science-as-product and science-as-process
Academia has long deliberately blurred the distinction between ‘science as product’ (solving real problems) and ‘science as process’ (pursuing curiosity). It funds the latter by claiming it serves the former—suggesting that curiosity naturally yields useful applications. Gorard gives an example: number theorists routinely claim their work connects to cryptography, but in reality, cryptography relies on mathematics developed thousands of years ago. This justification system took over a century to establish, but once AI can independently accomplish the ‘product’ side, no one will pay for ‘process’ anymore.
Math has verification tools that physics still lacks
This gap explains why AI progresses unevenly across fields. AI breakthroughs in coding and math depend on formal verification tools—compilers, Lean, Agda and similar proof assistants—that let AI improve unsupervised because it knows immediately whether it's right. Physics, chemistry, biology, and engineering have no equivalent verification pipeline. Hilbert attempted to formalize physics the way mathematicians formalized their field in the early 20th century, but he failed. Without this tooling, AI's progress in natural science lags far behind code and math.
Observation never exists independent of theoretical framework
The fundamental difference between physics and math: math doesn't require experimental validation, but physics must. Kuhn and Hanson both showed that ‘pure observation’ is impossible—every observation occurs within a theoretical framework, whether semantically (‘detecting a particle’ assumes dozens of layers of theoretical assumptions) or perceptually (optical illusions prove this). To automate scientific reasoning, you can't simply match observations to predictions; you must parameterize the entire theoretical-framework space. Without that, AI can't disambiguate what it's looking at.
AI's capability jumps come from the tool stack, not models
Gorard's claim: current AI ability isn't improving because transformers are getting better, but because of post-training, reinforcement learning, harnesses, tools, and skills scaffolding around them. He thinks transformers were essential to starting this revolution, but won't be the crucial part in the end. To test this hypothesis: swap the underlying LLM for a diffusion model while keeping all post-training and reinforcement-learning pipelines intact. If results are the same, the model architecture isn't the limiting factor.
— Jonathan GorardPhysics' mysteries dissolve when you change the descriptive language
What appears to be a ‘fundamental mystery’ is mysterious only within a specific model. The black-hole area law looks mysterious in general relativity, but switch to the holographic dual description in boundary field theory and it becomes thermodynamics—‘totally obvious’. Wave-particle duality seems incomprehensible in the particle framework but makes sense in the wave framework. From this, Gorard infers: many apparent physics dead-ends are actually dead-ends in descriptive language, not constraints reality itself has imposed.
— Jonathan GorardHe fears AI most because it automates the thinking itself
Gorard is puzzled by how older scholars readily embrace Claude for research, saying things like ‘Great, now I can read papers, solve equations, work through code’. But his response is ‘Wait—that's the part I actually love.’ This reveals a generational fault line: it's not about whether AI is capable enough, but about who gets to own ‘the joy of thinking’. His resistance to this working style isn't new. As a young scientist, he was already disappointed by the gap between the romantic natural philosophy of the 17th century and today's debugging-heavy computational physics.
— Jonathan GorardAI's real bottleneck is descriptive language, not computing power
Gorard reduces the whole episode to one sentence: AI capability is currently bottlenecked by descriptive language. This doesn't mean AI isn't smart enough; it means the formal language we use to express physics still isn't good enough. Without a verification pipeline, AI instead risks ‘slopification’—generating hypotheses and writing papers without the ability to check theoretical consistency or experimental validity. The result is hypothesis explosion with no real growth in understanding. The problem is in the tools, not in the model.
— Jonathan GorardIn their own words · checked verbatim
The output happens more or less independently of the input or the process. And so now the justification mechanisms that exist societally don't exist anymore.
Jonathan Gorard13:31
anytime you make an observation, anytime you perform a measurement, you are doing so relative to an existing theoretical framework.
Jonathan Gorard46:10
the curve in AI capabilities at the moment is not coming from improvements in the transformer model. It's coming from improvements in post-training, reinforcement learning, harnesses, tools, skills, all these kinds of things.
Jonathan Gorard54:16
we're getting smarter not because our brains are getting better, but because our tools are getting better.
Jonathan Gorard56:17
the pleasure for me was in you know was in writing out the equations it was in figuring out the algorithms it was in thinking about these questions it was in reading you know other people's thoughts about these things the idea that uh reducing frictions in those areas was somehow a good was completely alien to me
Jonathan Gorard1:46:44
Figures
| How early the sphere-packing proof came | Approximately 2 years | 3:04 |
| Theory layers behind a single observation | 20–50 | 47:10 |
| Safe timeline for experimental physics | Within 2 years | 1:42:37 |
| Stages of descriptive language evolution | 4: natural language, geometry, analysis, computation | 1:57:17 |
Glossary
- description language
- Formal language systems for expressing and modeling physical or mathematical reality.
- proof assistant
- Tools such as Lean or Agda that computationally verify mathematical proofs step by step.
- Sapir-Whorf hypothesis
- The hypothesis that a language's structure shapes how its speakers think.
- slopification
- Uncontrolled proliferation of unverified AI-generated hypotheses and papers.
- Flynn effect
- The observed phenomenon that average IQ test scores have risen over time across populations.
- harness
- The infrastructure of tools, skills, and process controls built around an AI model—distinct from the model itself.
How to listen
Researchers studying AI's impact on academia and research; AI product builders and AI investors; anyone wanting to understand where AI's actual bottlenecks lie, rather than harboring vague anxieties.
The mid-section discussion of transferability and educational philosophy gets academic; you can skip ahead.