Theorem, a startup, raises $6M to develop tools ensuring AI-generated code is reliable, addressing the growing trust gap in automated software development.
The Challenge of Trusting AI Code
In San Francisco, Theorem is making a bet. The bet is on trust. They raised $6 million to verify AI-written code. Jason Gross, their co-founder, knows the stakes. He says AI writes code faster than humans can check it. The gap is growing. Theorem wants to close it. They use formal verification to prove code works. It’s a method that once took years. Now, with AI, it takes days. The stakes are high. Financial systems and power grids depend on it.
Gross worked on verified cryptography at MIT. It took fifteen person-years to finish. He knows the cost of mistakes. Theorem’s system uses fractional proof decomposition. It allocates resources based on code importance. Recently, it found a bug at Anthropic. It was missed by traditional tests. Theorem’s approach is different. It’s about catching bugs without wasting resources. It’s about trust in a world where AI writes code faster than we can read.
The Power of Formal Verification
Formal verification isn’t new. It’s been around for decades. But it was too costly for most. Eight lines of proof for one line of code. Theorem changes that. They use AI to write proofs. It’s faster and cheaper. They turned a 1,500-page specification into 16,000 lines of code. No human reviewed it. Yet, it worked. Theorem’s system verified it. A few hundred lines of executable specification ensured it matched the original. The customer got a parser operating at 1 Gbps. Confidence without manual checks.
Theorem’s system runs agents in parallel and sequentially. It handles interdependent code. This is where others fail. It’s about more than speed. It’s about correctness. It’s about ensuring AI-generated software doesn’t fail when it matters most. Theorem is already working with AI research labs and other industries. They prove their system’s worth with real-world applications. They show that AI can write code. But someone, or something, must verify it.
Facing the Risks and Opportunities
AI is changing software fast. It’s in financial markets, medical devices, and grids. The risks are real. Subtle bugs can slip in. Gross talks about asymmetric defense. AI makes hacking cheaper. Defense must scale without more resources. Theorem offers a solution. Verification that lasts through AI improvements. It’s about security. It’s about ensuring systems don’t fail. Gross says not using formal verification might be negligence. The stakes are high. The risks are real.
Theorem is different from other startups. They focus on scaling oversight, not just verification. Their team has deep expertise. Gross knows programming language theory. Rajashree Agrawal trains the AI models. They aim to harness AI’s full potential. Not just average AI, but Linus Torvalds-level AI. Theorem plans to grow. More team members, more compute resources. They aim for new industries: robotics, energy, cryptocurrency. The race is on. To verify AI code before it controls everything. To ensure speed doesn’t compromise safety.
A New Era of Software Verification
Theorem’s emergence signals a shift. Enterprise leaders must evaluate AI tools differently. The first wave promised speed. The next demands proof. Proof that speed doesn’t cost safety. AI systems are improving fast. Superhuman software engineering is likely. But without oversight, we lose control. Theorem’s mission is clear. To verify AI code. To ensure it works. To build trust in a world where machines write the code.
I see the world changing. Machines write the code now. They do it fast. Faster than we can check. But someone has to check. That’s where Theorem comes in. They use AI to verify AI. It’s a new era. An era where trust is built on proof. I’ve seen the cost of mistakes. I know the stakes. Theorem is betting on trust. It’s a bet we can’t afford to lose.
