Guide Labs releases interpretable 8-billion-parameter LLM

On Monday, San Francisco-based AI startup Guide Labs released Steerling-8B, an 8-billion-parameter large language model (LLM) designed to be interpretable. Unlike traditional models where decision-making is a “black box,” every token produced by Steerling-8B can be traced back to its origins in the training data. The company, co-founded by CEO Julius Adebayo and Chief Science Officer Aya Abdelsalam Ismail, aims to address the struggles of existing models like ChatGPT and xAI’s Grok.

To achieve this traceability, Guide Labs builds its models with a concept layer that buckets data into traceable categories. Adebayo, who co-authored a widely cited 2018 paper during his PhD at MIT, explained that this approach shifts the paradigm of model analysis. “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,” Adebayo said. This architecture allows developers to pinpoint specific data encodings. For instance, if there are 1 trillion ways to encode gender and a model encodes it in 1 billion of those ways, developers must find and reliably control all 1 billion instances. Despite this architecture, Adebayo asserted that the model retains emergent behaviors. He also noted that the model demonstrates that training interpretable models has transitioned from a scientific challenge into an engineering problem.

According to Guide Labs, Steerling-8B can achieve 90% of the capability of existing models while using less training data. Adebayo argues that this interpretable architecture will be necessary for consumer-facing LLMs to block copyrighted materials or control outputs regarding violence or drugs. It also holds potential for regulated industries, where models must be safer and more controllable. He added that democratizing inherent interpretability will be beneficial for humanity in the long term, particularly as developers build increasingly intelligent models.

Guide Labs emerged from Y Combinator, a startup accelerator, and raised a $9 million seed round in November 2024. The funding round was led by Initialized Capital, a venture capital firm.

Why it matters

Guide Labs’ Steerling-8B attempts to solve the “black box” problem in deep learning by making model actions traceable, potentially enabling safer and more controllable AI for regulated industries.