Understanding the inner workings of deep learning models poses a significant challenge, often leaving users puzzled about their decision-making processes. Guide Labs, a start-up based in San Francisco, is addressing this issue with its latest offering. Founded by CEO Julius Adebayo and Chief Science Officer Aya Abdelsalam Ismail, the company has unveiled an open-source large language model (LLM) named Steerling-8B, which boasts 8 billion parameters and a unique architecture designed for enhanced interpretability.
Steerling-8B allows users to trace every token generated back to its original training data. This feature simplifies the process of verifying facts and understanding complex concepts such as humor or gender representation. Adebayo emphasizes the importance of being able to manipulate encoded attributes within the model, ensuring reliability and control over its outputs.
This innovative approach stems from Adebayo's research during his PhD at MIT, where he co-authored a key paper highlighting the unreliability of existing methods for interpreting deep learning models. The breakthrough led to a new model-building strategy that incorporates a concept layer, organizing data into traceable categories. Although this requires extensive data annotation, the team successfully trained Steerling-8B, showcasing it as their most ambitious proof of concept to date.
Adebayo notes that traditional interpretability methods often resemble neuroscience applied to models, whereas their design philosophy prioritizes engineering from the ground up, eliminating the need for complex neuroscience. This shift aims to maintain the emergent behaviors that make LLMs fascinating, allowing the model to discover concepts independently, such as advancements in quantum computing.
According to Adebayo, the demand for interpretable models is growing across various sectors. For consumer-facing applications, developers can utilize these techniques to prevent the use of copyrighted content or regulate discussions on sensitive topics. In regulated industries like finance, where models assess loan applications, interpretability is crucial for ensuring fairness and compliance.
Guide Labs is also committed to enhancing scientific research through its technology. While deep learning has made significant strides in areas such as protein folding, researchers require clearer insights into the decision-making processes of their models. Adebayo asserts that the training of interpretable models has transitioned from a theoretical challenge to an engineering task, with the potential to match the performance of larger, more complex models.
Steerling-8B is reported to achieve 90% of the capabilities of existing models while utilizing less training data, thanks to its innovative design. Looking ahead, Guide Labs plans to develop a larger model and provide API access for users, further democratizing the power of interpretable AI.
Adebayo concludes that the current methods of training models are outdated, advocating for a future where inherent interpretability becomes a standard feature. This evolution is essential as we venture into the realm of super-intelligent models, ensuring that they operate transparently and align with human values.
