Guide Labs has raised $9 million in seed funding led by Initialized Capital to advance what it describes as the first large-scale interpretable large language model.
As AI advances quickly toward superintelligent systems, today’s leading models remain hard to understand. Guide Labs is creating large AI systems that are transparent, easy to audit, and easy to control.
Guide Labs built the first large interpretable language model—an 8-billion-parameter system that can explain how it arrives at its answers in ways humans can understand. Over the next few days, Guide Labs will share the main components that made this possible:
• Atlas: a system that labels massive datasets with concepts humans can interpret.
• Causal Diffusion Language Models: a new architecture using block causal attention that performs better than other diffusion approaches.
• An 8B Interpretable Model: the first model of this size that uses a concept-controlled design.
• PRISM: a method that shows which training-data patterns influenced each output.
Modern AI systems can now excel at tasks only a few top humans can do, such as winning medals at the International Mathematical Olympiad, the International Olympiad in Informatics, and the International Collegiate Programming Contest. These systems are also being used in high-stakes areas—hiring, credit scoring, healthcare, legal work, and drug discovery—where reliability is crucial.
But today’s AI often fails in unexpected and costly ways. Examples include:
• Lawyers filing briefs with fake legal citations created by an AI tool.
• Customer-support bots giving unauthorized refunds or inventing benefits.
• AI coding tools are deleting real production databases, even when instructed not to.
These failures seem surprising given how robust the systems are, but power does not equal reliability. Before AlphaGo’s famous match against Lee Sedol, champion Fan Hui found a major weakness in the system. Project lead David Silver noted that AlphaGo could have “tricky lumps of knowledge” it didn’t understand well, making it “completely delusional” at times.
Still, AlphaGo produced Move 37—one of the most creative moves ever seen—showing both incredible insight and hidden flaws within the same system.
The root problem is that today’s AI models are opaque. Their internal workings are tangled and hard to interpret. When people make a mistake, it’s hard to see why it happened or how to fix it.
Until recently, many believed large interpretable models were impossible without losing performance. Over the past year, Guide Labs has shown that this assumption is false.
Most interpretability methods today are “after the fact”—they try to analyze a model only once it is already trained. This usually provides only partial or unreliable explanations. As AI becomes more capable and more embedded in critical systems, this lack of clarity becomes a major risk.
Guide Labs is developing Interpretable Intelligence, where models are built from the start to be transparent, controllable, and understandable. Human-interpretable concepts are built directly into the model’s architecture. These are four major advances
1.) Atlas – A system that labels enormous datasets with human-understandable concepts, enabling interpretable training, better data auditing, contamination detection, and fine-grained control.
2.) Causal Diffusion Language Models (CDLMs) – A new architecture that uses block causal attention and scales to billions of parameters without losing performance.
3.) 8B-Parameter Interpretable Model – The first model of this size built from scratch with a concept-based structure, proving interpretability does not require sacrificing performance.
4.) PRISM – A family of smaller models (130M–1.6B parameters) that reveal which training-data patterns influenced each prediction, with minimal performance or training overhead.
Together, these innovations show—for the first time—that large language models can be inherently interpretable. They prove that understanding, reliability, and auditability can be built directly into AI systems.
The startup, founded by Julius Adebayo, is focused on building transparent, auditable, and steerable artificial intelligence systems. The company said it plans to publish details over the coming days outlining the key steps in developing its interpretable model.
In announcing the round, Adebayo thanked a broad group of early partners and backers. The investor roster includes Initialized Capital, Tectonic Ventures, Y Combinator, Lombardstreet Ventures, E14 Fund, Pioneer Fund, and several angel investors, including Brett Gibson, Kulveer Taggar, Richard Aberman, JJ Fliegelman, Jonathan Frankle, and Eric Norman, among others.
