Guide Labs open-sources Steerling-8B, a new interpretable LLM that traces every token to source

Guide Labs Tackles AI's Black Box Problem with New Architecture

Guide Labs, a San Francisco startup founded by CEO Julius Adebayo and Chief Science Officer Aya Abdelsalam Ismail, has open-sourced Steerling-8B, an interpretable LLM that fundamentally changes how developers understand what their AI systems are doing1

1

. The 8-billion parameter model was released on Monday with a novel architecture designed to address one of AI's most persistent challenges: understanding why deep learning models make the decisions they do. From xAI's struggles to fine-tune Grok's political leanings to ChatGPT's issues with hallucinations, the black box problem has plagued AI developers for years1

1

.

How Traceability of Generated Tokens Works

The breakthrough lies in Steerling-8B's ability to trace every token produced by the model back to its origins in the training data1

1

. This capability ranges from simple tasks like verifying reference materials for cited facts to complex analyses of how the model encodes abstract concepts like humor or gender1

1

2

2

.

Adebayo explained the complexity: "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"1

1

. While this is technically possible with current models, the process remains fragile and unreliable.

Engineering Interpretability Through a Concept Layer

The architecture fundamentally alters the standard transformer structure by inserting a concept layer that categorizes data into traceable buckets during training2

2

. "The kind of interpretability people do is...neuroscience on a model, and we flip that," Adebayo told TechCrunch. "What we do is actually engineer the model from the ground up so that you don't need to do neuroscience"1

1

. This approach requires more upfront data annotation, but Guide Labs used other AI systems to assist in the labeling process1

1

2

2

.

Research Roots at MIT and Emergent Behaviors

Adebayo began this work during his doctoral studies at MIT, co-authoring a widely cited 2018 paper that demonstrated the unreliability of existing methods for understanding deep learning models1

1

2

2

. One concern with engineering interpretability is whether it might eliminate emergent behaviors that make language models valuable. Adebayo says Steerling-8B still exhibits these capabilities, with his team tracking "discovered concepts" that the model generates autonomously without explicit training, such as quantum computing1

1

2

2

.

High-Stakes Applications and Control Over LLM Outputs

Guide Labs positions the technology as essential for high-stakes applications requiring strict control over LLM outputs2

2

. In consumer-facing AI systems, the architecture enables developers to block copyright infringement and better manage outputs around sensitive subjects like violence or drug abuse1

1

2

2

. For regulated industries, particularly finance, the model can evaluate loan applicants based on financial records while explicitly excluding protected attributes like race, addressing regulatory requirements1

1

2

2

. In scientific research, including protein folding, the model provides insight into why specific predictions succeed, addressing a critical gap in computational biology1

1

2

2

.

Performance Metrics and Future Plans

Steerling-8B achieves approximately 90% of the capability of existing frontier models while using less training data, thanks to its novel architecture1

1

2

2

. Adebayo argues that "training interpretable models is no longer a sort of science; it's now an engineering problem," suggesting the approach can scale to match frontier models with significantly higher parameter counts1

1

2

2

. The company emerged from Y Combinator and raised $9 million in seed funding led by Initialized Capital in November 20241

1

2

2

. Next steps include building a larger model and offering API access and agentic capabilities to users1

1

2

2

. "As we're going after these models that are going to be super intelligent, you don't want something to be making decisions on your behalf that's sort of mysterious to you," Adebayo said1

1

.