Haize Labs has developed Sphynx, a tool for detecting and addressing hallucination issues in language AI models. The hallucinations referred to here are inaccurate and nonsensical outputs produced by these models.

The Sphynx tool employs the fuzz-testing technique, which simplifies the evaluation process of language AI models by involving a basic beam search algorithm for iterative testing. It generates variations of a given question, tests the model against them, ranks these variations based on their likelihood of causing an error, and detects hallucinations within the AI system.

Haize Labs claims that the key advantage of using Sphynx is its potential to improve the robustness and reliability of hallucination detection in AI models. Through its simple testing process, Sphynx can reveal significant weaknesses in the models, helping developers ensure better-prepared models for real-world deployment.

Haize Labs was launched in 2023 to commercialize AI model jailbreaking for the benefit of AI companies. Haize Labs helps identify security and alignment weaknesses in models. The startup's "Haize Suite," or "haizing suite," is a collection of algorithms designed to probe LLMs like ChatGPT and Claude for vulnerabilities.