Back
AI Certs

2 months ago

Can Math Solve AI’s Hallucination Problem? Exploring New Paths to More Accurate Chatbots

Chatbots like ChatGPT have made significant strides in answering questions, writing essays, and generating images. However, they often produce errors or false information—a phenomenon called hallucination. Researchers are now exploring how mathematics could help eliminate these inaccuracies and make AI systems more reliable. Harmonic, a Silicon Valley start-up founded by Tudor Achim and Vlad Tenev, aims to build an AI that never hallucinates, focusing first on mathematical accuracy.

Math is a field with strict rules for proving right or wrong answers, making it a good starting point for reliable AI. For instance, Harmonic's bot, Aristotle, solves complex math problems and generates computer programs to verify its answers. Google’s DeepMind lab is working on a similar AI system called AlphaProof, which achieved high performance in the International Mathematical Olympiad, solving problems on par with top students.

The use of math to verify AI-generated answers could extend beyond math into areas like computer programming and other logic-based disciplines. Systems like Aristotle can generate what’s called “synthetic data,” trustworthy digital information produced by AI and used to train other AI systems. Over time, this process could help build AI systems that are smarter and more accurate than human experts.

Although promising, these systems still have limitations. While math offers definitive answers, many real-world situations don't. Read more.