A significant hurdle in managing deep learning models lies in deciphering their operational logic. This difficulty is evident in scenarios such as xAI’s persistent efforts to adjust Grok’s unusual political leanings, ChatGPT’s susceptibility to flattery, or commonplace hallucinations, making the task of dissecting a neural network with billions of parameters inherently complex.
Today, San Francisco-based startup Guide Labs, co-founded by CEO Julius Adebayo and chief science officer Aya Abdelsalam Ismail, unveiled a solution to this issue. The company publicly released Steerling-8B on Monday, an 8 billion parameter large language model (LLM) developed with an innovative architecture specifically engineered for clear interpretability: each token generated by the model can be directly linked to its source within the LLM’s training data.
This interpretability can range from straightforward tasks like identifying the source materials for facts the model references, to more intricate analyses such as comprehending the model’s interpretation of humor or gender.
Adebayo shared with TechCrunch, “Should there be a trillion methods to encode a concept like gender, and I’ve embedded it into one billion of those trillion possibilities, one must reliably locate and control all those one billion encoded instances – switching them on or off as needed.” He added, “While current models allow this, the process is extremely delicate… It’s considered a sort of ultimate goal in the field.”
Adebayo initiated this research during his doctoral studies at MIT, co-authoring a highly referenced 2020 paper which highlighted the unreliability of existing deep learning model interpretation techniques. This research culminated in a novel approach to constructing LLMs: developers integrate a concept layer into the model, categorizing data into traceable segments. While this demands increased initial data annotation, the team leveraged other AI models to assist, enabling them to train Steerling-8B as their most substantial proof of concept to date.
“The typical approach to interpretability is akin to performing neuroscience on a model; we, however, reverse that,” Adebayo explained. “Our method involves engineering the model fundamentally so that such ‘neuroscience’ becomes unnecessary.”
A potential drawback of this strategy is the risk of suppressing emergent behaviors, which are a fascinating aspect of LLMs, such as their capacity to generalize about unfamiliar topics. Adebayo assures that this phenomenon persists within his company’s model, with his team monitoring “discovered concepts” that the model independently uncovers, such as quantum computing.
Adebayo contends that this interpretable architecture will become essential across various sectors. For consumer-oriented LLMs, these methods would empower model developers to prevent the use of copyrighted content and more effectively manage outputs concerning topics such as violence or substance abuse. Regulated fields will necessitate more manageable LLMs; for instance, in finance, a model assessing loan applicants must factor in financial history but exclude race. Furthermore, scientific research demands interpretability, another domain where Guide Labs has innovated. While deep learning models have achieved remarkable success in protein folding, scientists require deeper understanding into how their software arrived at effective configurations.
“What this model proves is that developing interpretable models has transitioned from a scientific endeavor to an engineering challenge,” Adebayo stated. “We’ve resolved the scientific aspects and can now scale these models, and there’s no inherent reason why this approach shouldn’t rival the performance of state-of-the-art models,” despite those models typically possessing far more parameters.
According to Guide Labs, Steerling-8B achieves 90% of the capabilities of current models while utilizing less training data, a benefit of its innovative architecture. The company, which originated from Y Combinator and secured a $9 million seed funding round from Initialized Capital in November 2024, plans to next develop a larger model and provide users with API and agentic access.
Adebayo informed TechCrunch, “Our current methods for training models are quite rudimentary, and therefore, making inherent interpretability widely accessible will ultimately benefit humanity in the long run.” He elaborated, “As we pursue the development of highly intelligent models, it’s crucial that the decisions they make on our behalf are not shrouded in mystery.”