A Look Inside The Thinking Of The Consultant And Founder Working To Bring Proof And Accountability To Machine Intelligence
When Neel Somani talks about artificial intelligence, he does not speak in abstractions or predictions about the future. He speaks in first principles.
Before Somani’s work in artificial intelligence, he was drawn to the deeper questions underneath computation itself. As an undergraduate studying computer science, he encountered a problem that every student eventually learns about, the halting problem, a proof that reveals fundamental limits to what algorithms can determine.
“That tight relationship between logic and computer science is what led me to formal methods,” Somani said.
Formal methods are a branch of computer science grounded in mathematical proof. Instead of testing software repeatedly and hoping failures appear, they attempt to prove with certainty that a system behaves exactly as intended. These techniques have long been essential in security and cryptography. In machine learning, however, they remain largely absent.
That absence has shaped Somani’s evolving career.
Today, Neel Somani works with companies as a consultant and entrepreneur, helping teams navigate the reliability, interpretability, and safety challenges emerging alongside advanced AI systems. He is the founder of Eclipse, Ethereum’s fastest Layer 2 platform powered by the Solana Virtual Machine, which raised $50 million in Series A funding. Earlier in his career, he served as a quantitative researcher in Citadel’s commodities group, applying mathematical optimization to global energy markets. He holds a triple major in computer science, mathematics, and business administration from the University of California, Berkeley, where his research focused on type systems, differential privacy, and scalable machine learning frameworks.
Across every domain he has worked in, from financial markets to artificial intelligence, Somani has returned to the same belief. Power without understanding is not innovation. It is risk.
From Theory To Real-World Systems
Neel Somani’s introduction to formal methods was not philosophical. It was practical.
One of the first research labs he joined was led by Professor Dawn Song at UC Berkeley, a group focused on computer security and privacy. There, Somani worked on proving that a machine learning algorithm satisfied a strict mathematical definition of privacy known as differential privacy.
Instead of assuming data was private, the system could demonstrate that privacy formally.
“That was the first serious project I worked on in formal methods,” he said.
After graduating, Somani encountered an even clearer example of how abstract computer science governs real-world infrastructure. The U.S. power market, he learned, is solved using extremely complex optimization problems classified as NP-hard, problems so computationally difficult that even modern systems cannot brute-force solutions.
To work on those systems, he joined Citadel’s commodities group as a quantitative researcher.
For several years, he applied mathematical reasoning to markets where precision mattered. Optimization models influenced decisions shaping energy distribution and financial outcomes. The experience reinforced a lesson that would follow him long after finance.
When systems operate at scale, small errors rarely stay small.
Eventually, Neel Somani’s curiosity pulled him back toward machine learning. He began experimenting with applying formal verification techniques to GPU kernels, the low-level programs responsible for much of modern AI computation.
That work revealed something unexpected. While machine learning models were advancing rapidly, experts trained in formal methods were almost entirely absent from the field.
“I played around with different subfields,” he said. “It was a bit of a hammer in search of a nail.”
Over time, he found that interpretability offered the most promising ground.
Why AI Safety Remains Unsolved
Despite AI’s widespread adoption, the industry still lacks a reliable way to certify safety.
“Safety and interpretability in machine learning are preparadigmatic,” Somani said. “There isn’t any established way to certify that a system is safe, or that we fully understand it.”
Most current approaches rely on empirical testing. Engineers run models through thousands or millions of examples and monitor failure cases. But testing can never account for every scenario.
Formal methods offer a fundamentally different approach.
“They’re the gold standard,” Somani said. “They’re the only way to establish strong, principled guarantees about programs.”
One of the clearest examples is robustness. In a reliable system, small changes in input should not cause dramatic changes in output. In machine learning, proving that property is extraordinarily difficult.
Inputs are continuous. There are infinitely many variations that cannot all be tested.
“The only way to establish a guarantee like that is via formal methods,” Somani said.
Another vulnerability lies deeper in the technical stack. GPU programming often relies on domain-specific languages or low-level C and C++ environments that lack many safety protections.
“A tiny bug in GPU code can lead to hidden errors that are extremely difficult to uncover,” he said.
To address that problem, Somani developed Cuq, a project that applies formal verification to GPU kernels written in Rust. The goal is to identify correctness issues before they propagate through large-scale models.
These problems may sound technical, but their implications are broad. Reliability failures at the infrastructure level directly affect AI systems used across industries.
Neel Somani: ‘We Might Expect More Out of Them’
Interpretability remains one of the most contested challenges in artificial intelligence.
Large language models can produce fluent, confident responses, yet their internal decision-making remains opaque even to researchers. Analysts often develop theories about what specific components might be doing.
The problem is falsifiability.
“There’s no way to really prove or disprove those hypotheses,” Somani said.
To close that gap, he developed Symbolic Circuit Distillation.
Interpretability researchers often isolate small parts of neural networks, known as circuits, and attempt to infer their function. Somani’s approach goes further. His system extracts the program encoded by a circuit and then mathematically proves whether that program is equivalent to the model’s actual behavior.
“This way,” he said, “we know for a fact whether the stories we tell ourselves are metaphors, or whether they precisely represent behavior.”
While the technique currently works only in simplified settings, it establishes an important precedent. Understanding no longer depends on intuition alone. It becomes verifiable.
For high-stakes applications in finance, health care, and infrastructure, that distinction matters.
Somani believes the path forward will require both patience and ambition. As machine learning systems grow larger and more autonomous, formal specification will become increasingly necessary. Individual components will need clearly defined behavior before verification is possible at scale.
He points to historical parallels such as CompCert, a multi-year effort that formally verified compiler correctness once those systems matured.
At the same time, he believes machine learning itself may evolve in response.
“Rather than formal methods adapting to machine learning,” he said, “we may see machine learning architectures adapt to formal methods.”
Future models may prioritize verifiability alongside performance.
Looking ahead, Somani is especially interested in rethinking core components such as LayerNorm and attention, as well as deeper questions around reasoning itself. Some approaches frame reasoning as a form of search, but whether models truly perform search internally remains an open question.
“One project on my backlog,” he said, “is investigating whether models actually have the circuits necessary to do things like backtracking.”
It is a question that sits at the boundary between mathematics and machine intelligence.
For Somani, that boundary is precisely where the most important work remains.
As artificial intelligence grows more powerful, his focus is not on accelerating its capabilities, but on making its behavior understandable.
Because in a world increasingly shaped by algorithms, understanding may be the most valuable form of intelligence we have.
“As models write a larger percentage of our code,” he said, “we might expect more out of them.”
