Session brief
The Hidden Life of Embeddings: Linus Lee
Overview
This talk explores the concept of latent spaces within AI models, particularly focusing on embeddings, as a means to gain more direct control and understanding of model behavior. Instead of relying solely on indirect prompting, the speaker advocates for interacting with the internal representations of models to achieve more precise manipulation and insight into their decision-making processes.
Who should watch
- AI Engineers
- Product Managers
- Builders experimenting with language and image models
- Those interested in understanding model internals
- Researchers exploring new methods for model control and interpretability
Key takeaways
- Prompting language models offers indirect control, akin to steering a car from the back seat with a pool noodle.
- Latent spaces, particularly those found in embeddings, offer a more direct layer of control by allowing interaction with a model's internal representations.
- By encoding text or images into vectors within these latent spaces, meaningful attributes and features can be disentangled and manipulated.
- Manipulating embeddings can lead to controlled changes in generated text, such as altering length or sentiment while preserving the core topic.
- Embeddings from different models can be adapted, allowing a decoder trained on one model's embeddings to interpret embeddings from another, like OpenAI's Ada.
- Similar techniques can be applied to image models, enabling interpolation between different visual representations and manipulation of attributes like emotion or style.
- The speaker has open-sourced models and a Colab notebook to facilitate experimentation with these concepts.
- Discovering interpretable and controllable representations within latent spaces is a significant area of ongoing research.
Notable quotes
*Prompting language models feels a lot like you're steering a car from the back seat with a pool noodle.*
*The purpose of a thinking medium is to bring thought outside the head to represent these Concepts in a form that can be seen with the senses and manipulate with the body.*
Unofficial community note. Prefer the recording for nuance.