A new startup is taking on one of artificial intelligence’s biggest challenges: understanding why large language models behave the way they do. With the launch of Guide Labs Steerling-8B, a newly open-sourced 8-billion-parameter model, the San Francisco-based company claims it has built a fundamentally interpretable LLM where every token can be traced back to its origins in the training data.

At a time when AI systems face scrutiny over hallucinations, bias, and unpredictable behavior, Guide Labs is positioning Steerling-8B as a more transparent alternative to conventional deep learning models.

Guide Labs Steerling-8B Introduces Interpretable Architecture

Unlike traditional large language models, Guide Labs Steerling-8B was trained using a novel architecture that incorporates a “concept layer.” This layer groups data into traceable categories, enabling developers to track how and why specific outputs are generated.

CEO Julius Adebayo and chief science officer Aya Abdelsalam Ismail founded Guide Labs to address long-standing concerns about the opacity of neural networks. Adebayo previously co-authored a widely cited 2018 research paper while earning his PhD at MIT that demonstrated how many existing interpretability methods were unreliable.

Instead of analyzing models after training — often compared to performing neuroscience on a black box — Guide Labs engineered interpretability directly into the model’s foundation.

The result, according to the company, is a system where developers can identify reference materials for factual outputs or analyze how the model understands complex ideas such as humor or gender.

Tackling Bias, Hallucinations and AI Governance

One of the core promises of Guide Labs Steerling-8B is improved control. Current LLMs often require delicate fine-tuning to prevent unwanted behavior, and even then, results can be inconsistent.

Adebayo argues that interpretability allows for more reliable control over model behavior. For consumer-facing applications, this could mean better safeguards against copyrighted content misuse or more precise moderation of sensitive topics such as violence or drug abuse.

In regulated sectors like finance, interpretable models may help ensure decisions are based only on relevant factors, such as financial records, rather than protected attributes like race.

Scientific research is another area where interpretability is crucial. Deep learning has made breakthroughs in protein folding, but researchers often need insight into why specific patterns are identified. Guide Labs believes its architecture could bridge that gap.

Performance and Scalability

Despite its interpretability focus, Guide Labs claims Steerling-8B achieves approximately 90% of the capabilities of existing frontier models, while using less training data. The company attributes this efficiency to its engineered architecture.

The startup, which emerged from Y Combinator and raised a $9 million seed round from Initialized Capital in November 2024, plans to develop a larger model next. It also intends to offer API and agentic access for enterprise users.

One concern with highly structured architectures is whether they limit emergent behaviors — the unexpected capabilities that make LLMs powerful. However, Guide Labs says its system still allows for “discovered concepts,” such as quantum computing, which emerged during training without explicit programming.

Why Interpretable LLMs Matter

As AI systems become more powerful and embedded into everyday decision-making, transparency is increasingly viewed as essential. Industry analysts and researchers have repeatedly warned that black-box AI systems can undermine trust and accountability, particularly in high-stakes environments.

Guide Labs argues that interpretability should not be treated as an optional feature but as a foundational design principle. By embedding traceability into the model architecture itself, the company believes it can scale transparency without sacrificing performance.

With Guide Labs Steerling-8B, the company is making a case that the future of AI may depend not just on larger models, but on models that can clearly explain themselves.