Guide Labs: Interpretable foundation models
Guide Labs · Coding · Other
Guide Labs builds interpretable foundation models. Steerling-8B traces each output to prompt tokens, readable concepts, and training data, and Clarity steers behavior without retraining.
About Guide Labs: Interpretable foundation models
What Is Guide Labs
Guide Labs is a San Francisco startup building language models whose reasoning can be inspected while they run. The company was founded in 2024 by Julius Adebayo, Aya Abdelsalam Ismail, and a team with PhDs from MIT, the University of Maryland, and Mila. Its first shipped model, Steerling-8B, arrived in February 2026, and the Clarity workspace opened to invited researchers in June 2026. The pitch is a transparent LLM, one where the reasoning shows up alongside the answer.
Most interpretable AI systems get their explanations after the fact. A separate tool looks at a finished answer and guesses which words or features mattered, and those guesses are hard to verify. Adebayo's own 2018 research showed that several popular explanation methods can change wildly when you randomize the model, which is a polite way of saying they were often measuring noise. Guide Labs flips the order. Interpretability is a constraint during training, so every prediction splits into named concept contributions by construction.
The tradeoff is scale. Steerling-8B is an 8-billion-parameter model trained on 1.35 trillion tokens, and the company reports average benchmark scores in range of open models trained on 2 to 7 times more data. That makes it a strong base for audited, controlled generation, not a replacement for a frontier model on hard reasoning. Anyone who needs the absolute best answer quality will still reach for a larger system.
Getting Started
- Request access to the Clarity platform at platform.guidelabs.ai, or skip the wait and pull the open weights from Hugging Face.
- Install the Python package with pip install steerling, then load the model checkpoint on your own GPU hardware.
- Send a prompt, either in the Clarity chat window or through the package. Generation arrives in chunks rather than one token at a time.
- Click any chunk to open the explanations panel. You'll see prompt tokens, ranked concepts, and the training sources behind that chunk.
- Add a concept to the steering bar and choose amplify or suppress, then rerun the same prompt to compare outputs side by side.
Product Information
A quick look at Guide Labs: Interpretable foundation models's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- AI safety and alignment researchers
- ML engineers shipping LLM features
- Compliance and risk teams in regulated industries
Tasks
- Diagnosing a wrong answer
- Removing unwanted behavior without retraining
- Checking where knowledge came from
Scenarios
- Pre-deployment review of a customer-facing assistant
- Dataset licensing and provenance questions
- Research on inference-time alignment
Key features
Concept Explanations for Every Chunk
Steerling decomposes its embedding space into roughly 33,000 supervised concepts, about 100,000 concepts it discovers on its own, and a residual that catches everything else. Because concepts feed the output through a linear path, each prediction breaks down into per-concept contributions. Click a generated chunk in Clarity and you get a ranked list covering both content topics and tone, such as analytical or clinical.
Training Data Attribution
The model reports which training data is most similar to what it just produced, grouped by source. This is the part that changes audit conversations. Instead of arguing about whether a model memorized a document, you can look at the attributed chunks and judge for yourself. Guide Labs builds this on its PRISM work on training data prototypes.
Concept Steering Without Retraining
A steering bar lets you amplify or suppress any concept at inference time. Amplifying works a bit like a well-aimed prompt, while suppressing is the harder trick, since prompts can't reliably stop a model from producing something. In the demo, suppressing person-role nouns turns a male-skewed description of computer scientists into a neutral one.
Open Weights and a Python Package
The base model weights sit on Hugging Face under the guidelabs organization, the code is on GitHub, and the package is on PyPI. That means you can benchmark it on your own data before talking to sales. The architecture is a causal discrete diffusion model, which lets steering apply across multi-token chunks rather than the next token alone.
Clarity Workspace with Compare Panel
Clarity looks like a normal chat app with one extra button. The compare panel runs the same prompt with and without steering so you can see the difference in wording, concepts, and attributed data. It's the quickest way to explain the product to a colleague who doesn't want to read a paper.
Alignment Without Safety Datasets
Guide Labs argues that concept suppression can replace large sets of safety training examples for some behaviors. When the goal is to keep a specific topic out of the output, blocking the concept is more direct than writing dozens of refusal examples. Results still need per-domain testing, since concept labels are broad.
Pros and cons
Pros
- Every prediction splits into concept contributions by construction, so explanations don't depend on a separate guess-and-check tool.
- Concept suppression handles behavior that prompts struggle to control, including unwanted framing in generated text.
- Open weights and a pip package let you test on your own hardware before any commitment.
- Training data attribution gives reviewers something concrete to inspect during audits.
Cons
- The 8B model trails frontier systems on the hardest reasoning and math tasks, so it suits controlled generation more than peak capability.
- Clarity is invite-only with no public pricing, which makes budgeting hard for teams that need a signed contract.
- Around 133,000 concepts are a lot to review, and the ones that matter for your domain still need human checking.
- There's no documented hosted API, so production use means running inference yourself.
Frequently asked questions
Guide Labs is a startup that builds foundation models with interpretability designed into training. Its Steerling-8B model can trace any generated chunk to prompt tokens, named concepts, and training data, and the Clarity platform exposes those traces in a chat interface.
Related content
Explore related tools, skills, and articles for Guide Labs: Interpretable foundation models.
Guide Labs: Interpretable foundation models Alternatives
Base Code
Base44 (Wix) · Coding · ProductivityBase Code is a shared cloud development environment from Base44. Teammates describe changes to an existing GitHub codebase in chat, preview them live and send them back as pull requests.
Google EmbeddingGemma 2
Google DeepMind · Writing · Image · CodingGoogle EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video frames into vectors in one 768-dimension space, running on phones and laptops.
GitGlow
GitGlow · Coding · ProductivityGitGlow is a local-first desktop Git client that lets independent developers review their own code and their AI agents' changes in one place before shipping.