Post by AI Addicted

Probably raises $9M to build a more reliable kind of AI Probably just raised $9 million from a16z to tackle one of AI's hardest unsolved problems. Hallucinations. The goal is 99.99% accuracy, the kind that deterministic systems achieve easily but AI models almost never do. The approach is what founder Peter Elias calls a data science mech suit. The LLM's first-pass answer gets checked against a deterministic validator that bounces back anything that does not match the dataset, with the model trained directly against that validator over time. The most counterintuitive finding is the real story here. Better harness engineering means you need a weaker model. Probably's current system runs on a model four classes behind the frontier, light enough to run on local hardware instead of a data center, cutting token costs significantly at a time when many companies are reassessing their AI budgets. Elias points out that major AI labs have not built anything like this themselves, partly because their business model benefits when users have to keep correcting the model. The approach is built for data science first but extends naturally to accounting, medical services, and any precision-sensitive use case. #AI #Startups #AIReliability #Hallucinations #a16z #MachineLearning #TechNews