Guide Labs debuts a new kind of interpretable LLM

Guide Labs, founded by Julius Adebayo and Aya Abdelsalam Ismail, has open-sourced Steerling-8B, an 8-billion parameter Large Language Model (LLM) built with a novel architecture to ensure high interpretability. This design allows every token produced by the model to be traced back to its source in the training data, addressing the difficulty of understanding 'black box' models like Grok or ChatGPT. CEO Adebayo explained that this approach engineers interpretability from the ground up, contrasting it with post-hoc 'neuroscience on a model.' The model requires more upfront data annotation but can still exhibit emergent behaviors like discovering concepts such as quantum computing. Adebayo argues that this interpretable architecture is essential for consumer-facing LLMs (to block copyrighted material or control sensitive outputs) and regulated industries (like finance, ensuring factors like race are excluded from loan decisions). Guide Labs claims Steerling-8B achieves 90% of the capability of frontier models while using less training data, positioning interpretability as an engineering problem rather than a scientific mystery. The company plans to build larger models and offer API access.

(Source：TechCrunch)