Lanyon AI emerged from stealth on Monday with $10.6 million in initial funding to build what it describes as a new kind of scientific AI. The Princeton-based company, founded by mathematicians and physicists from Princeton University and the Princeton Plasma Physics Laboratory, claims its agent can generate simulations, prove theorems, and write algorithms that are provably correct by construction - a standard that goes beyond the "mostly right" output typical of large language models.
The round was led by Dimension with participation from Industrious Ventures. Lanyon AI is targeting physics, engineering, GPU kernel optimization, and frontier AI inference, where correctness is non-negotiable.
A neurosymbolic approach to reliability
Lanyon's agent works in a formal specification language rather than natural language. The LLM proposes ideas and generates specifications, then symbolic methods expand that specification into code and proof simultaneously. If the specification can't be rigorously proven correct, the code doesn't generate, and the system loops back to try again.
This avoids a known failure mode in AI-generated proofs. "Many AI companies have proposed a workflow of 'autoformalization,' wherein an LLM generates a formal proof of correctness for a piece of code using a proof assistant language such as Lean," said Jonathan Gorard, co-founder and CEO. "Such approaches are clever, but they carry the fundamental risk of misformalization, where the code and the proof don't actually match."
Gorard, an applied mathematician previously known for co-founding the Wolfram Physics Project with Stephen Wolfram, said the company's approach separates creative work from verification. "We're pioneering at Lanyon AI is a neurosymbolic approach: having the LLM do what it's good at - being creative and proposing interesting ideas - while exploiting symbolic methods to do what they're good at - being unfailingly reliable and predictable."
The formal specification language is also highly condensed and domain-specific. Lanyon claims it operates at a fraction of the token and compute cost of frontier models like GPT-5.6 and Fable 5, while running far faster.
Selling proofs where near-enough isn't good enough
Simon Barnett, Partner and Head of Research at Dimension, said frontier models fall short for mission-critical engineering. "LLMs have transformed software development and have begun making waves in formal mathematics," Barnett said in a statement. "Despite headlines featuring Olympiad-level proofs and eye-watering benchmark scores, next-token prediction still ends at 'endmostly right', a standard that doesn't clear the bar for flight controls, nuclear systems, or simulating chip tape-outs."
Lanyon AI is initially focused on aerospace engineering, space and atmospheric propulsion, and nuclear energy - industries where simulation errors carry real consequences. "Formally verified, high-accuracy simulations are critical for solving deep scientific and engineering problems," said co-founder and CTO Ammar Hakim. He previously led computational physics work in fluid mechanics, nuclear fusion, and aerospace engineering.
Co-founder James (Jimmy) Juno is a plasma physicist who developed methods for laboratory, space, and astrophysical plasmas. The founding team brings a combined five decades of applied mathematics, computational physics, and scientific AI expertise.
Why this matters for scientists and researchers
For working scientists and engineers, the practical difference here is in what you can trust. Current AI assistants are useful for drafting code or exploring options, but every line needs in-person verification before it goes near a real instrument, test bench, or analysis pipeline. Lanyon's approach promises that the result comes with a mathematical certificate of correctness - meaning you could, in principle, skip part of that manual verification in fields where errors are not an academic fee, but a safety or financial risk.
The more immediate signal is costs. If the company's claims about token efficiency hold up under independent testing, high-precision simulation and formal proof generation will be accessible to research teams that can't reallocate budget for frontier-model inference. That would make verified computational tools a practical option for daily work, not just a showcase exercise.
Whether the approach will be stable at scale remains to be demonstrated. Lanyon is unproven at the level of production engineering workloads. But it's a notch in the practical direction for a field where - so far - "I'm pretty more confident" has been the best available standard of rigor.
Your membership also unlocks:
