Probably raises $9M to build a more reliable kind of AI

Probably, an AI startup focused on eliminating factual errors from large language model outputs, has raised $9 million in seed funding from Andreessen Horowitz [1]. The company aims to deliver deterministic-level accuracy in AI-generated answers by rethinking core engineering assumptions [1]. The funding round was led by the venture capital firm Andreessen Horowitz, which manages $90 billion in assets as of January 2026 [4]. The firm was co-founded by Marc Andreessen, a former software engineer who co-authored the Mosaic web browser and co-founded Netscape [5]. Probably's approach arrives as the AI industry grapples with the persistent problem of hallucinations, where models generate plausible but incorrect information [1]. Founder Peter Elias said the company's goal is to prevent hallucinations and simple factual errors from reaching users, targeting 99.99% accuracy — a standard common in deterministic systems but difficult to achieve with AI [1]. "What we learned building this was that the better your harness engineering is, the weaker the model can be," Elias said [1]. "If you can refine the context enough, the model does not have to work very hard to do the right thing. Basically, it's an exercise in reducing ambiguity" [1]. Probably's first product is a data science tool that produces answers from complex datasets, each accompanied by a citation and an audit trail [1]. The system uses a deterministic validator to check the LLM's first-pass answers, rejecting any results that do not match the underlying data [1]. Elias described the supporting infrastructure as a "data science mech suit" [1]. Because the harness engineering reduces the burden on the model, Probably can run on a model that Elias said is "four classes weaker than the frontier models" [1]. This allows operation on local hardware, such as a desktop computer, cutting token costs significantly [1]. The company plans to extend its engine to other precision-sensitive fields, including accounting and medical services [1]. Elias argued that major AI labs have not pursued similar validation architectures. "They're incentivized not to, because they make money the more times you have to correct the model," he said [1]. The launch comes during a period of heightened federal focus on AI. Sriram Krishnan, a former Andreessen Horowitz general partner, has served as Senior White House Policy Advisor on Artificial Intelligence and helped author the American "AI Action Plan" released in July 2025 [6].