Guide Labs: Interpretable foundation models FAQs

Question 1

What are interpretable foundation models?

Accepted Answer

Interpretable foundation models are AI models that can explain their reasoning and outputs, unlike traditional 'black box' models. Guide Labs has developed interpretable versions of large language models (LLMs), diffusion models, and large-scale classifiers that can provide explanations for their decisions while maintaining high performance.

Question 2

What capabilities do Guide Labs' interpretable models have?

Accepted Answer

Guide Labs' interpretable models can: 1) Explain and steer outputs using human-understandable features, 2) Provide reliable prompt attribution to indicate important parts of the prompt, 3) Identify influential tokens in pre-training and fine-tuning data, 4) Customize and explain the model using high-level concepts from domain experts.

Question 3

How do these models compare to standard foundation models in terms of performance?

Accepted Answer

According to Guide Labs, their interpretable models are as accurate and high-performing as standard foundation models, but with the added benefit of providing explanations for their outputs.

Question 4

Can users fine-tune these interpretable models?

Accepted Answer

Yes, Guide Labs offers a fine-tuning API that allows users to customize the models with their own data, inserting high-level concepts to steer and control the model's output.

Question 5

Who developed these interpretable foundation models?

Accepted Answer

The models were developed by the team at Guide Labs, led by Julius Adebayo (PhD in ML interpretability from MIT) and Fulton Wang (PhD in ML interpretability from MIT). Both have extensive experience in machine learning interpretability from companies like Google, Meta, and Apple.

Question 6

How can I access Guide Labs' interpretable models?

Accepted Answer

Guide Labs is currently working with selected users to test these models. Interested parties can sign up for a waitlist on their website to get early access.

Guide Labs: Interpretable foundation models Howto

More Information

How to Use Guide Labs: Interpretable foundation models