Themis AI: Teaching AI Models What They Don’t Know to Enhance Reliability
In the rapidly evolving landscape of artificial intelligence, systems like ChatGPT often present answers with unwavering confidence, even when their knowledge is incomplete. This poses a significant challenge, especially as AI is increasingly deployed in critical sectors such as drug development and autonomous driving. Themis AI, an MIT spinout, is tackling this issue head-on with its Capsa platform, designed to quantify model uncertainty and rectify unreliable outputs.
The Capsa platform is versatile, capable of integrating with any machine-learning model to swiftly detect and correct inaccuracies. By modifying AI models, Capsa enables them to identify patterns indicative of ambiguity, incompleteness, or bias in their data processing.
“The idea is to take a model, wrap it in Capsa, identify the uncertainties and failure modes of the model, and then enhance the model,” explains Daniela Rus, co-founder of Themis AI and Director of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). “We’re excited about offering a solution that can improve models and offer guarantees that the model is working correctly.”
Founded in 2021 by Rus, Alexander Amini, and Elaheh Ahmadi, Themis AI has already made strides in various industries. The company has assisted telecom companies with network planning and automation, aided oil and gas companies in interpreting seismic imagery using AI, and contributed to research on developing more reliable chatbots.
“We want to enable AI in the highest-stakes applications of every industry,” states Amini. “We’ve all seen examples of AI hallucinating or making mistakes. As AI is deployed more broadly, those mistakes could lead to devastating consequences. Themis makes it possible that any AI can forecast and predict its own failures, before they happen.”
Rus’s research into model uncertainty dates back several years, including a 2018 project funded by Toyota to assess the reliability of machine learning-based autonomous driving solutions. Her lab also developed an algorithm capable of detecting and mitigating racial and gender bias in facial recognition systems, showcasing the potential for AI to be more equitable.
In 2021, Rus and her colleagues demonstrated a similar approach to aid pharmaceutical companies in predicting drug candidate properties using AI models, leading to the formation of Themis AI later that year.
Today, Themis AI collaborates with various enterprises, particularly those developing large language models (LLMs). Capsa enables these models to assess their own uncertainty for each output, enhancing the reliability of question answering and flagging potentially unreliable outputs.
“Many companies are interested in using LLMs that are based on their data, but they’re concerned about reliability,” observes Stewart Jamieson, Themis AI’s Head of Technology. “We help LLMs self-report their confidence and uncertainty, which enables more reliable question answering and flagging unreliable outputs.”
Themis AI is also in talks with semiconductor companies to integrate AI solutions into their chips for use outside of cloud environments, promising efficient edge computing without compromising quality.
Pharmaceutical companies are leveraging Capsa to refine AI models for identifying drug candidates and predicting clinical trial performance. According to Amini, Capsa offers immediate insights into whether predictions are grounded in training data or are merely speculative, thereby accelerating the identification of the most promising candidates.
Looking ahead, Themis AI is exploring Capsa’s capacity to improve accuracy in chain-of-thought reasoning, potentially guiding LLMs to identify the most reliable reasoning chains. Rus emphasizes her commitment to ensuring that her MIT research translates into real-world impact, fostering trust and understanding between people and AI technologies.
