Understanding the internal workings of deep learning models, particularly large language models (LLMs), has long been a significant challenge for developers and researchers. Issues like inexplicable behaviors, political biases, sycophancy, and hallucinations in models like xAI's Grok and ChatGPT highlight the difficulty of plumbing through neural networks with billions of parameters. Guide Labs, a San Francisco-based startup co-founded by CEO Julius Adebayo and Chief Science Officer Aya Abdelsalam Ismail, is addressing this interpretability problem head-on with a novel architectural approach.

Guide Labs recently open-sourced Steerling-8B, an 8 billion parameter LLM. What distinguishes Steerling-8B is its new architecture, specifically designed for inherent interpretability. Every token generated by the model can be traced back to its precise origins within the LLM's training data. This capability allows for both straightforward and complex insights, from identifying the reference materials for factual statements to comprehending the model's understanding of abstract concepts like humor or gender. Adebayo emphasized the fragility of current methods for controlling such aspects in existing models, referring to robust interpretability as a 'holy grail question'.

The foundation of this work stems from Adebayo's PhD research at MIT, where his 2018 paper demonstrated the unreliability of then-current deep learning model understanding methods. This research paved the way for a new LLM construction paradigm: developers embed a 'concept layer' into the model. This layer organizes data into traceable categories, allowing for clear accountability of the model's decisions. While this requires more upfront data annotation, Guide Labs has streamlined the process by employing other AI models to assist, enabling them to scale this approach and train Steerling-8B as their largest proof of concept to date.

Adebayo described their method as an engineering solution rather than a 'neuroscience on a model' approach, which is typical for many interpretability efforts. They engineer the model from the ground up to eliminate the need for post-hoc analysis. A common concern with such structured interpretability is the potential loss of 'emergent behaviors'—the model's ability to generalize to new, untrained concepts. However, Guide Labs asserts that Steerling-8B retains this capacity, with the team tracking 'discovered concepts' that the model identifies independently, such as quantum computing.

The demand for interpretable LLMs spans multiple sectors. For consumer-facing applications, this technology can enable model builders to prevent the use of copyrighted material or better regulate outputs related to sensitive topics like violence and drug abuse. In regulated industries such as finance, controllable LLMs are crucial; a model evaluating loan applicants, for instance, must consider financial records but strictly exclude factors like race. Scientific research also benefits significantly; while deep learning excels in tasks like protein folding, scientists require insight into the model's reasoning for proposing promising combinations. Adebayo stated that this model proves training interpretable models is now an engineering problem, not just a scientific one, asserting their ability to scale these models without sacrificing performance compared to frontier-level LLMs, even with fewer parameters.

Guide Labs claims that Steerling-8B achieves approximately 90% of the capability of existing models while utilizing less training data, a testament to its innovative architecture. Following its emergence from Y Combinator and a $9 million seed round in November 2024, the company's next steps include developing an even larger model and offering API and agentic access to users. Adebayo envisions that democratizing inherent interpretability will be a long-term benefit for humanity, especially as AI models become super-intelligent, ensuring that decisions made on our behalf are not shrouded in mystery.