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

Despite advancements in large language models (LLMs), errors and "hallucinations" remain a significant challenge. These inaccuracies appear even in the most sophisticated models, and the industry is actively seeking effective solutions to mitigate them.

Probably, a new company backed by $9 million in seed funding from Andreessen Horowitz, is tackling this problem head-on. Their goal is to prevent factual errors and hallucinations from ever reaching users, striving for an accuracy level of 99.99%—a standard common in deterministic systems but notoriously difficult to achieve with AI.

Probably's initial product is a data science tool designed for rapid analysis of complex datasets. Each result includes a citation and a detailed audit trail, reflecting a growing trend in AI transparency. To maintain accuracy, the system employs an "elaborate harness" where an LLM's preliminary answers are validated against a deterministic system that flags inconsistencies.

This innovative approach allows Probably's data science tool to operate effectively on significantly smaller AI models. Founder Peter Elias notes that their current model is "four classes weaker than the frontier models," enabling local hardware operation and substantially reducing the token costs associated with AI processing. The company plans to extend this technology to other precision-sensitive applications like accounting and medical services.