/ Princeton mathematicians raise $10.6M for verified AI
Mike Alvarez Joined TheTweaks as an emerging tech , startup and launches journalist. He comes up with those things or products which are not there yet and are hardly possible to come close to. He covers the latest technology, startups and new product launches and releases. He worked 4 years in a startup company in Austin and later he joined TheTweaks and he is proving himself here with his commendable writing performance. He is also the publisher of a small consumer robotics newsletter.
A new AI lab is betting on the future of scientific computing will be more about guarantees than bigger language models. Princeton-based Lanyon AI, a research firm, has emerged from stealth mode with a $10.6 million seed round co-led by Dimension and Industrious Ventures.
The startup is not targeting the Hot Take of general-purpose chatbots. It’s the much more narrow, much more difficult realm of hard-science calculations: physics simulations, optimizing GPUs, engineering workflow, and frontier AI inference, where a “mostly right” answer isn’t sufficient (PR Newswire).
What Lanyon Actually Does
At the heart of the launch is Lanyon’s AI agent, also named Lanyon, which, according to the company, can simulate complicated physical systems, solve sophisticated math problems, and create new algorithms but at a fraction of the compute and token cost of today’s frontier models. But it’s not speed, that’s the thing. It’s correct.
According to the company, any code or output that Lanyon produces is provably correct, that is, it is guaranteed to not be different from its formal specification mathematically. This is a much more solid guarantee than is currently provided by most AI coding tools, which may only be validated after use when they are tested, checked by humans or compared to benchmarks rather than guaranteed beforehand.
The technique of autoformalization, which involves a large language model generating the code and then a separate large language model producing a formal proof (typically in a proof assistant like Lean) that the code was correct, has been tried in the AI industry already.
Lanyon’s co-founder and CEO Jonathan Gorard says this technique has a structural weakness; there is no assurance that the proof and the code were talking about the same thing. He calls this “misformalization”: that is, the code passes, the proof passes, and they don’t actually match one another.
Lanyon’s solution is a neurosymbolic pipeline, however. As LLMs excel at generating creative concepts and suggesting a specification for the problem at hand, the LLM can be used to do what it is good at doing.
The specification is then expanded, from the same source, to both the executable code and correctness proof simultaneously, using symbolic, rule-based methods. If the specification is not hard to prove then the system does not generate incorrect code, it just goes around again, until it finds one that does hold up.
The practical side benefit is that the language Lanyon specifies programs is compact and built for scientific and technical applications, meaning that the system can perform considerably faster and more cost-efficiently than general purpose frontier models applied to the same types of applications, according to the company.
Simon Barnett, Partner and Head of Research at Dimension, sees an opportunity that exists now in the field of AI: in the midst of competitive maths scores and media star achievements, they only have approximate correctness and that is not sufficient for flight control systems, nuclear infrastructure or chip design simulations.
That’s the space where Dimension is placing its bets that Lanyon’s neurosymbolic domain-specific languages can fill in the gap between an engineer’s initial intent, the formal spec and the code and math that emerges.
Lanyon’s initial focus is on industries where the penalty of an unrecognized mistake is very large: aerospace engineering, space and atmospheric propulsion systems, and nuclear energy.
As much as a safety measure, the technology is designed to be an extension tool, said co-founder CTO Ammar Hakim, and high-accuracy simulations that are formally verified don’t just limit the risks; in addition, they increase the scope of solutions that scientists and engineers can practically explore when addressing complex technical challenges.
Lanyon AI started with three researchers who have close ties to the mathematics and physics community at Princeton. Gorard is an applied mathematician who has also been previously involved in the Wolfram Physics Project (with Stephen Wolfram).
Ammar Hakim, co-founder, has a background in computational physics, including fluid mechanics, nuclear fusion and aerospace engineering. James “Jimmy” Juno, the third co-founder, is a plasma physicist with expertise in tackling problems in laboratory, space and astrophysical plasmas.
And together, the founding team has over five decades of experience in applied math, scientific computing and applied physics, which reads more like a list of national labs than an AI startup team. With a mission statement as bold as “formal verification for a computable universe,” the company is based in Princeton, N.J.
Lanyon’s idea is a counterpoint to the prevailing narrative of AI that focuses on record leaderboards and larger models. Lanyon is saying that there are types of science and engineering tasks where a mistake is not an inconvenience, but a disaster that require a fundamentally different sort of AI system, one that doesn’t just have to be “good” in the training set; it needs to be “right,” and that has to be demonstrated mathematically before the output ever leaves the system.
It will be interesting to see if that’s the case in a real seed-stage pitch deck. Formal verification is notoriously hard to scale to messy real world engineering problems and “provably correct” is a tall bar to clear as scope of work expands.
Lanyon AI has at least chosen an arena that few other AI firms are engaging in, with a founding team that has formal-methods and physics chops, and a clear path into industries that do not tolerate approximations.
Lanyon AI doesn’t aim to be a standard AI helper, making its appeal all the more fascinating. The majority of the AI industry is on a scale game at the moment larger models, larger benchmarks, larger promises. Lanyon, however, is not playing a trust game, rather, in industries such as aerospace, nuclear power and chip design, trust is the only measure that truly counts.
It is a very clever architectural solution to have the LLM do the creative work and have the symbolic work do the correcting and the “correct by construction” promise sounds great and if it does on stress testing in the real world it would be a step-up from the currently available generate then verify AI coding tools.
The founding team’s credentials don’t hurt either: This is not a bunch of hype-chasers; it’s a team of researchers who’ve been working in just the same fields they’re now attempting to automate for the past 30 years.
But a $10.6 million seed round and a stealth-exit press release is just the start. The key challenge will be to see if Lanyon’s formal verification pipeline scales well as problems become more complicated and less ideal for textbook solutions.
The real question will be whether Lanyon’s formal verification pipeline scales well as problems become more complicated and less textbook like in their simplicity. A thumbs-up so far for Lanyon: technically an ambitious, well-credentialed bet on a problem that really does need to be solved, but one that still needs to be shown to be more than a mere launch announcement.
For more AI and tech news, visit TheTweaks AI News or check out our latest tech launches, tech trends and big tech coverage.
Tim Cook’s 15 years as Apple’s chief executive officially ended on September 1 and now the torch has been passed on to John Ternus, who is the company’s long-time hardware. It’s Apple’s first leadership change…
Be respectful and constructive. Have a question or feedback? We’d love to hear from you. Contact us at contact@thetweaks.com