Themis AI: Teaching AI Models What They Don’t Know

In an era where artificial intelligence systems like ChatGPT can generate seemingly plausible answers, the challenge of recognizing their own knowledge gaps and areas of uncertainty becomes increasingly critical. MIT’s spinout, Themis AI, aims to tackle this issue by quantifying model uncertainty and addressing potential errors before they escalate into larger consequences. This is particularly important as AI applications extend into high-stakes environments, such as drug development and autonomous driving.

Introduction to Themis AI

Themis AI has developed the Capsa platform, which integrates with any machine-learning model to quickly identify and correct unreliable outputs. The primary function of Capsa is to modify AI models, enabling them to recognize patterns that indicate uncertainty, incompleteness, or bias in their processing.

The Vision Behind Themis AI

Co-founded in 2021 by MIT Professor Daniela Rus, along with Alexander Amini and Elaheh Ahmadi, Themis AI seeks to improve the reliability of AI systems in a variety of industries. Rus emphasized the importance of offering solutions that not only enhance model performance but also provide guarantees regarding their correctness. Themis AI’s technology has already helped companies ranging from telecommunications to oil and gas in optimizing their operations.

Addressing Knowledge Gaps in AI

Themis AI’s research is rooted in understanding model uncertainty, a focus that Rus and her team have engaged with since 2018, particularly in the context of developing reliable AI for autonomous vehicles. Their exploration revealed that improving AI’s awareness of its limitations could mitigate risks inherent in decision-making processes where incorrect outputs could have severe repercussions.

Applications of Capsa

Through the Capsa platform, Themis AI allows models to self-assess and quantify their uncertainty, enhancing their ability to provide reliable answers. This capability is vital for companies interested in deploying large language models (LLMs) based on their own data but concerned about the implications of unreliable outputs. By self-reporting their confidence levels, AI outputs can be thoroughly vetted for accuracy.

Future Prospects and Impact

Moreover, Themis AI is exploring how its technology can enhance accuracy in AI reasoning processes, particularly in the context of chain-of-thought reasoning used by LLMs. This advancement has the potential to streamline operations and reduce computational demands while improving overall user experience.

As Themis AI continues to make strides in AI reliability and transparency, Rus emphasizes the importance of the company’s mission: to ensure that AI not only transforms industries positively but also addresses the societal challenges and concerns raised by its rapid adoption.