Lucas Dixon and Michael Terry, co-leads of PAIR, Google Research, have been working to make AI more understandable, fun, and usable for everyone. They believe that AI can be even more powerful and beneficial if it is designed with people in mind. PAIR focuses on human-AI interaction and machine learning, and they have been actively involved in generative AI research.
Generative AI, which includes technologies like large language models (LLMs) used in chatbots and generative media models, has been a major area of interest for PAIR. They have explored the use of language models to create generative agents and studied the adoption of generative image models by artists. These models allow users to input a text-based description of an image, and the model generates it accordingly. PAIR found that users not only aim to create beautiful images but also strive for unique and innovative styles. Some users even seek unique vocabulary from architectural blogs to develop their visual style.
One of the challenges faced by prompt creators in generative AI is programming without using a programming language. PAIR has been researching solutions to this problem and has developed methods for extracting meaningful structure from natural language prompts. These structures can be used in prompt editors to provide features similar to those found in programming environments. This enables prompt creators to have better control and understanding of their generative models.
Another advancement in the field of generative AI is the development of agile classifiers. PAIR has been leveraging the strengths of LLMs to solve classification problems related to online discourse. Instead of collecting millions of examples to create universal safety classifiers, agile classifiers can be created by individuals or small organizations for their specific use cases. These classifiers can be iterated and adapted quickly, leading to better moderation of online content.
PAIR also focuses on visualization and education to make ML-related work more understandable. They publish visual, interactive online essays called AI Explorables, which provide hands-on ways to learn about ML concepts. They recently published Explorables on model confidence and unintended biases, which explain these concepts and provide interactive examples.
Transparency is another important aspect of AI research for PAIR. They have developed model cards and Data Cards to promote transparent documentation. The recently released Data Cards Playbook is a toolkit that helps teams and organizations overcome obstacles when implementing transparency efforts. It contains various resources and participatory activities to customize Data Cards according to their needs.
Finally, PAIR develops software tools to improve the understanding of ML models. Know Your Data allows researchers to test a model’s performance and identify unintended biases in datasets. The Learning Interpretability Tool (LIT) helps with model debugging and understanding, and the latest version includes support for image and tabular data and improved performance.
Overall, PAIR’s work in AI research is focused on making AI more understandable, interpretable, and usable for all. They continue to explore generative AI, develop new methods and tools, and promote transparency and education in the field.