Probably and Andreessen Horowitz: $9 Million for Error-Free AI

In a context where large language models (LLMs) struggle to eliminate errors, Probably, an innovative company, has raised $9 million in seed funding from Andreessen Horowitz. This amount will be used to develop a more reliable method for detecting and correcting errors in LLMs.

Peter Elias, founder of Probably, explains that the goal is to achieve an accuracy of 99.99%, a level common in deterministic systems but still difficult to reach with AI. To accomplish this, the company has designed a data science tool that provides quick answers from complex datasets while ensuring the traceability of results.

The tool operates through a deterministic validation system that checks the initial responses of the LLMs. This system, described by Elias as a "mechanical data science suit," helps reduce errors by refining context, which lightens the workload of AI models. Crucially, the LLM has been trained against the validator to ensure this accuracy.

Probably has managed to make its tool work on smaller AI models, thus reducing operational costs. The current version operates on a model that is four classes weaker than state-of-the-art models, meaning it can run on local hardware, such as a desktop computer, rather than in a data center.

This approach is particularly relevant at a time when token costs are rising and companies are looking to optimize their AI budgets. Elias envisions extending this technology to other sectors sensitive to accuracy, such as accounting and medical services. He emphasizes that large AI labs have not even attempted to do this, as they are incentivized not to, making money every time a model correction is needed.