Guide Labs debuts a new kind of interpretable LLM

Understanding the inner workings of deep learning models poses a significant challenge, as many developers grapple with unpredictable behaviors and outputs. In light of this, Guide Labs, a San Francisco-based startup led by CEO Julius Adebayo and Chief Science Officer Aya Abdelsalam Ismail, has introduced an innovative solution. This week, the company has made its latest creation, the Steerling-8B—a language model with 8 billion parameters—available as open-source. What sets Steerling-8B apart is its unique architecture designed to enhance interpretability. Each token generated by the model can be traced back to its training data origins, allowing for a clearer understanding of the facts it presents, as well as more nuanced aspects such as humor and gender interpretations. Adebayo explained, "If I have a trillion ways to encode gender, and I encode it in 1 billion of the 1 trillion things that I have, you have to make sure you find all those 1 billion things that I've encoded, and then you have to be able to reliably turn that on, turn them off." This interpretability stands as a pivotal advancement in the field of AI. Adebayo’s journey began during his PhD studies at MIT, where he co-authored a critical 2020 paper highlighting the unreliability of current methods for interpreting deep learning models. This research laid the groundwork for a new approach to building language models, which features a concept layer that organizes data into easily traceable categories. Although this method requires more initial data annotation, the integration of other AI models facilitated the training of Steerling-8B, making it the most substantial proof of concept to date. Adebayo emphasized the difference in their methodology: "The kind of interpretability people do is…neuroscience on a model, and we flip that. What we do is actually engineer the model from the ground up so that you don’t need to do neuroscience." While there are concerns that this approach may suppress some of the emergent behaviors that enhance the intrigue of language models, Adebayo assured that Steerling-8B still exhibits these capabilities, as his team tracks "discovered concepts" identified by the model independently. The implications of this interpretable architecture extend beyond theoretical applications. For consumer-focused language models, this technology can help developers eliminate the use of copyrighted material and exert greater control over sensitive topics like violence and substance abuse. Additionally, industries with stringent regulatory requirements, such as finance, will benefit from models that can evaluate loan applicants based on relevant criteria without bias. Guide Labs has also ventured into scientific applications, where the need for interpretability is crucial. For instance, while deep learning models have excelled in predicting protein folding, scientists require insights into the reasoning behind the successful outcomes generated by these models. Adebayo stated, "This model demonstrates that training interpretable models is no longer a sort of science; it’s now an engineering problem. We figured out the science and we can scale them, and there is no reason why this kind wouldn’t match the performance of the frontier-level models." According to the company, Steerling-8B can achieve 90% of the performance of existing models while utilizing less training data due to its innovative structure. Looking ahead, Guide Labs, which originated from Y Combinator and secured a $9 million seed investment from Initialized Capital in November 2024, aims to develop a more advanced model and provide API access to users. Adebayo concluded, "Democratizing inherent interpretability is actually going to be a long-term good thing for humanity. As we pursue models that are super intelligent, we want to ensure that the decisions they make on our behalf are transparent."