Exploring Unbounded AI Products: Lessons from Dawn & Vision Pro Design

Welcome to my talk on making good AI products, where I'll share three key lessons from designing unbounded products at Dawn, including highlighting what matters, establishing hierarchy, and leveraging familiarity.

  • 1. Ben Hidalgo is the founder of Dawn, a company that helps businesses build better AI products.
  • 2. He has a passion for building "unbounded products," or products that go beyond traditional input modalities like typing, talking, and showing images.
  • 3. Unbounded products can be confusing for users, as they often assume the product can do things it cannot.
  • 4. Many people learn how to use chatbots through word of mouth, rather than trial and error.
  • 5. In this talk, Hidalgo will discuss the past, present, and future of unbounded products.
  • 6. Most software lives on a screen and is used through typing, swiping, clicking, and tapping.
  • 7. Multi-touch technology was a significant change in recent years, as it allowed for relative distance and rotation.
  • 8. Unbounded products can be found everywhere, from software roaming the streets to virtual reality.
  • 9. When designing unbounded products, it's important to consider "what if" scenarios, such as what should happen when a user moves from one room to another.
  • 10. Hidalgo worked on the Vision Pro and learned three key lessons in designing unbounded products: highlighting what matters, establishing hierarchy, and leveraging familiarity.
  • 11. In the present, AI products are incorporating structure into their features both good and bad.
  • 12. The right structure is unique to each app and helps users understand its purpose.
  • 13. Dot uses sliders within a conversation to adjust settings, but this can become confusing when trying to follow up on the conversation.
  • 14. Instead of putting structured elements into unstructured conversations, it's better to pull them out and update them separately.
  • 15. Claude has successfully done this with their "artifacts" feature, which allows users to iterate on UI without disrupting the conversation.
  • 16. Version control is another effective tool for AI apps, as it allows users to go back and iterate on previous versions.
  • 17. ChatGPT has introduced memory across all chats, which can be both helpful and unnerving.
  • 18. Agents are unfamiliar to most people, but using familiar structures like spreadsheets can help make them more accessible.
  • 19. Examples and presets can also help users understand what an app is for and skip the "prompt hacking" problem.
  • 20. In the future, interfaces will likely have less prompt engineering and more intuitive ways of adjusting settings.
  • 21. Sparse autoencoders may be a promising path towards offering millions or billions of presets to avoid being reductive.
  • 22. Ranked presets can be personalized, searchable, and even invoked through natural language.
  • 23. Developer-defined personalization will become more important as apps become increasingly different per user.
  • 24. Shifting from evals to analytics will help understand if the needs of users are being met.

Source: AI Engineer via YouTube

❓ What do you think? What is the most effective way to add structure and clarity to unbounded AI products, allowing users to easily navigate and understand their capabilities? Feel free to share your thoughts in the comments!