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

Founder Peter Elias stated that Probably's goal is to achieve near-perfect accuracy, comparable to deterministic systems, by developing a rigorous error-catching mechanism. Their initial product is a data science tool that generates answers with citations and audit trails, a growing trend in AI applications.

The system employs a "data science mech suit" approach, where LLM-generated answers are cross-referenced with a deterministic validator that checks against the source dataset. This process, combined with training the LLM against the validator, is designed to ensure accuracy and speed.

Elias highlighted that robust "harness engineering" allows for the use of smaller, less powerful AI models. This is because refining the context and reducing ambiguity means the model doesn't need to be as sophisticated. Consequently, Probably's tool can run on local hardware, significantly reducing the token costs associated with AI usage.

This cost-saving aspect is particularly relevant as AI token costs rise and companies re-evaluate their AI budgets. Elias believes the underlying engine can be extended to other precision-sensitive fields such as accounting and medical services. He also noted that major AI labs have not pursued this approach, suggesting a potential conflict of interest as they profit from increased model usage and corrections.