New Startup 'Probably' Aims to Eliminate AI Hallucinations with Innovative Validation System

As advances in Large Language Models (LLMs) continue, the issue of hallucinations—or errors in AI outputs—remains persistent. Probably, a new startup, has raised $9 million in seed funding from Andreessen Horowitz to tackle this challenge.

• Founded by Peter Elias, the company is developing a robust system to prevent these errors, striving for 99.99% accuracy akin to deterministic systems.
• Probably's initial product is a data science tool that generates quick answers from complex datasets, including a citation and an audit trail for each result.
• The system employs an innovative 'data science mech suit,' where an LLM's output is validated against a deterministic system, ensuring accuracy and reducing ambiguity.
• This allows the use of smaller AI models, which can run on local hardware and decrease token costs significantly.
• The model aims for use in various precision-sensitive fields beyond data science, including accounting and medical services.
• Elias criticizes large AI labs for not pursuing such accuracy-focused approaches, suggesting their financial incentives are misaligned.