Breaking AI & Tech News Analyzed

The Quest for Reliable AI The rapid growth of Large Language Models (LLMs) has brought significant advancements in AI capabilities. However, hallucinations and factual errors have proven challenging to eliminate. Probably, a startup founded by Peter Elias, aims to address this issue by developing a more rigorous approach to catching errors. The Funding and Vision Probably has secured $9 million in seed funding from Andreessen Horowitz. The company's primary goal is to prevent hallucinations and simple factual errors from reaching users, achieving the high accuracy levels common in deterministic systems but difficult to attain with AI. The Data Science Tool Probably's first product is a data science tool designed to produce quick answers from complex datasets. Each result comes with a citation and an audit trail for its development. This approach is becoming increasingly common among AI tools. The Innovative Approach The tool uses an elaborate harness system, described as a "data science mech suit," to keep errors from creeping into summaries. The LLM's first-pass answers are checked against a deterministic validator system, which rejects any results that don't match the dataset. The LLM has been trained against the validator, and the entire system is optimized for fast and accurate answers. The Impact on AI Engineering The approach requires rethinking basic assumptions of AI engineering. As Elias notes, "the better your harness engineering is, the weaker the model can be." By refining the context, the model does not have to work hard to do the right thing, essentially reducing ambiguity. The Future Outlook This innovation allows Probably's data science tool to run on significantly smaller AI models, reducing token costs associated with AI use. The company plans to extend its engine to cover use cases like accounting or medical services, essentially any precision-sensitive use case. Elias remarks, "I think it's really interesting that the big AI labs have not even attempted to do this. They're incentivized not to, because they make money the more times you have to correct the model."