Probably Raises $9M for More Reliable AI: A New Approach to Error Prevention

As language models (LLMs) have grown more powerful, they’ve also become increasingly prone to errors known as 'hallucinations'—mistakes where the AI outputs incorrect or nonsensical information. These errors are particularly troublesome in applications requiring high precision and reliability.

Enter Probably, a startup that just secured $9 million in seed funding from prominent investor Andreessen Horowitz (a16z). Founded by Peter Elias, this company is tackling the challenge of AI errors head-on with innovative solutions designed to prevent hallucinations and factual mistakes before they reach users.

Building a Mech Suit for Data Science

Elias explains that his goal is to achieve an accuracy rate of 99.99%—a level commonly seen in deterministic systems but challenging to attain with AI. To do this, Probably has developed a comprehensive data science tool that incorporates a complex validation mechanism.

The first step involves using a large language model (LLM) to generate initial answers from datasets. These outputs are then rigorously checked against a deterministic system—a kind of 'data science mech suit'—that ensures the results match the input data exactly. If discrepancies arise, they're flagged and corrected before reaching the user.

Engineering for Accuracy

The key to Probably's approach lies in refining the context so that the model doesn't need to work as hard to produce accurate results. By reducing ambiguity, they can use significantly smaller AI models that run on local hardware, such as a desktop computer, rather than requiring data centers.

This not only saves on token costs but also makes the technology more accessible and cost-effective for businesses looking to integrate AI into their operations.

Broader Applications

The same validation engine can be applied beyond data science to fields like accounting, medical services, and any other scenario where precision is critical. Elias believes this approach could revolutionize how businesses handle sensitive information using AI tools.

He notes that while major AI labs haven't attempted similar approaches, Probably's strategy could provide a new pathway for achieving high-accuracy outputs in a cost-effective manner.