Guide Labs

Guide Labs was founded in 2023 by Julius Adebayo and Aya Abdelsalam Ismail, both seasoned ML researchers with PhDs and industry experience at Google, Meta, and Genentech. Adebayo originated much of the underlying research during his doctoral work at MIT, where he was supported by Open Philanthropy; his 2018 paper demonstrated that many existing deep learning explanation methods were unreliable or actively misleading. Guide Labs graduated from Y Combinator's Winter 2024 batch and closed a $9.3 million seed round in December 2025, led by Initialized Capital with participation from Tectonic Ventures, Y Combinator, Lombardstreet Ventures, E14 Fund, and Pioneer Fund. The company's core thesis is that interpretability must be built into models from the ground up. They developed Atlas, an automated system for annotating trillion-token pretraining corpora with human-understandable concepts; Causal Diffusion Language Models (CDLMs), a new architecture using block causal attention; and PRISM, a family of smaller models (130M to 1.6B parameters) that reveal which training patterns influenced predictions. In February 2026, Guide Labs open-sourced Steerling-8B, an 8-billion-parameter model trained on 1.35 trillion tokens that routes all predictions through approximately 33,000 supervised concepts and 100,000 discovered concepts, enabling inference-time alignment via concept control, training data provenance for any generated token, and suppression or amplification of specific concepts without retraining. Guide Labs targets regulated industries — finance, life sciences, and other domains where model auditability is a legal or practical requirement — as well as AI safety researchers who need reliable mechanistic understanding of model internals. The team of nine has published more than 24 papers at top ML venues including ICLR and ICML. The company is headquartered in San Francisco and is actively hiring machine learning engineers, full-stack developers, and researchers.