The phrase "reduce hallucinations" has been overused, but what actually gets customers to pay is turning it into an engineering tool.
TechCrunch reported on June 16 that Probably closed a $9M seed round led by Andreessen Horowitz. A funding entry from The SaaS News also confirmed the amount, round, and date, positioning the company as a tool for improving AI accuracy.
Probably's direction isn't to train another larger model, but to build a more rigorous validation layer. According to CryptoBriefing, it aims to check outputs using deterministic validators and smaller models, targeting nearly 99.99% accuracy. This goal is ambitious, so take it with a grain of salt for now, but the product focus is spot on.
The AI reliability market will only grow. Because enterprises don't just care whether a model gets the answer right; they also care whether mistakes can be caught before they happen, where the errors lie, and whether they can be audited. Especially in scenarios like finance, healthcare, legal, and data analysis, hallucinations aren't just a UX issue—they're a risk issue.
A seed round also means it's still very early. How many tasks the validators can cover, whether they will introduce new false positives, and whether the cost can be lower than direct multi-model cross-checking, all remain to be seen in real-world testing.
But I'm willing to keep watching. The stronger frontier models become, the more they need an external validation layer; otherwise, everyone is just using more expensive methods to produce more confident mistakes.
Primary Sources: