Eclipse Founder Neel Somani on Transitioning from Crypto to AI

Neel Somani, the Founder of Eclipse Labs, has been focusing his attention on artificial intelligence. In this article, we discuss his transition from the crypto industry, his reasoning, and his future plans.

A Brief Timeline

Eclipse Labs is a scaling solution for the Ethereum blockchain. Somani founded and led Eclipse through its Series A, raising $65M in capital. From early on, he had planned to step back from the CEO role after the Series A, a transition that was discussed with both team members and investors months in advance.

In May 2024, he transitioned from CEO to Executive Chairman. In that role, he retained ultimate responsibility for company direction and executive leadership, while delegating day-to-day operations to Vijay Chetty as CEO. The company completed and shipped the Eclipse Mainnet product that had been promised to the market.

By mid-2025, Somani reassessed Eclipse Labs’ long-term direction and led a board-level decision to pivot the company toward building an application, appointing Sydney Huang to lead day-to-day execution under the revised mandate.

Throughout these transitions, Somani retained effective control over Eclipse’s strategic direction until his voluntary departure from governance in late 2025. In October 2025, he stepped down as Executive Chairman to focus full-time on machine learning, with particular emphasis on long-term research and intellectual contributions to ML theory. By that point, Somani’s interests had shifted away from crypto and toward building an enduring research-oriented legacy.

Questionnaire with Neel Somani

Q: Would you be able to summarize your career for the readers?

A: Sure. My background is in math and computer science. I initially worked as a software engineer at Airbnb before transitioning to a quant research role at Citadel. I focused on energy markets, where large-scale optimization models are frequently employed to solve for the efficient market price. I had prior experience in formal methods from my work in type systems in undergrad, and it turns out that skillset applies pretty widely across fields. I took that perspective to the crypto industry, where I built Eclipse. While Eclipse was a major chapter of my career, my long-term interests were never limited to crypto infrastructure, and I found myself drawn to research problems that I no longer had bandwidth to pursue. Around mid 2024, I transitioned to an Executive Chairman role at Eclipse, and by the end of 2025, I took that off my plate so I can focus full-time on my new research direction.

Q: Tell us a bit about the new research direction.

A: What I find is that whenever you enter a new field, you have to spend some amount of time just developing taste. By taste, I mean some sense of what problems are worth pursuing. There are tons of research problems out there that you can tackle. But you only have so many hours in the day. So I spent the last year or so getting up to speed in various parts of ML. That includes reinforcement learning, which I wrote a blog series covering recently. It also includes application-layer concepts like agents. But ultimately, from a research perspective, I find myself most drawn to making sense of the LLM itself. That falls under mechanistic interpretability, and I find that formal methods might be able to give us stronger robustness and safety guarantees around LLM behavior.

Q: What are some projects you’ve been working on lately?

A: One project that’s pretty early is called Symbolic Circuit Distillation. The idea is that in mechanistic interpretability, practitioners try to extract “mechanistic circuits” from LLMs, which are small subsets of LLM neurons that encode different types of algorithmic behavior. For example, there might be a circuit for modular addition within the LLM. That’s what the LLM uses when it needs to compute modular arithmetic. The limitation is that extracted circuits are often descriptive but not falsifiable. Symbolic Circuit Distillation is an attempt to search over a program space to automatically match a program to that circuit, and formally prove that the program is equivalent over some domain.

Q: What’s your advice to new grads right now?

A: That’s a tough question. I’ve mentored dozens of undergrads, and there’s never been a more difficult time to advise than now. The world is changing quickly. If you had asked me a year ago, I would have recommended they learn the skill of distribution. Attention is a scarce resource. Table stakes are getting familiar with modern coding tools like Claude Code or Codex. But if I had to really give a piece of advice, maybe I’d just recommend that undergrads consider working at a top AI lab. There aren’t many places where you can have the same impact right now. Of course, there’s no single right path, and my own career is evidence of that.

Disclaimer: This is a paid post and should not be treated as news/advice.