From Arc to Dia - Lessons Learned in Building AI Browser

Time: 2:45 PM

Speaker Bio: Head of Engineering, AI at The Browser Company. Previously senior engineer at Instagram/Facebook (6 years).

Speaker Profile: Full Speaker Profile

Company: The Browser Company is being acquired by Atlassian ($610M). Developing Arc and Dia browsers with AI integration.

Focus: UX/engineering lessons from rebuilding a beloved product (Arc) with AI. Understanding how AI agents integrate with user interfaces.

Reference: LinkedIn Profile

Notes

• What we learned going from Arc to Dia around AI
• Shipped a few ideas in Arc with AI stuff
  • Put out App 2
• Built a new one
  • With speed and security in mind
  • It gets to know you
  • On the way to achieving the vision
• What did we learn along the way
  • Optimize your tools and process for fast iteration
    • Prototype for AI production features
    • Building and running evals
    • Collection data for trains and evals
    • And automation for hill climbing
• Tools
  • Prompt editor in dev builds
  • Moved all these prompts into the tools itself
    • 10x the speed of ideating and iterating
    • “Ideating”
    • All with their full context
    • Super fun for everything to do it
  • GEPA as a way to refine the prompt
    • Seed them
    • Exec and score
    • Choose top prompts
    • Reflect and generate new prompts

• “Generate a skill based on the user input” -> multipage prompt
• GEPA hill climbing is really exciting
• Treating model behavior as a craft
  • Behavior design
  • Measurement and training
  • Model steering
  • Product design and the craft of the internet moved over time
    • Functional -> agentic behavior
    • What might the future hold?
  • We are in the early days of model behavior
  • The best people it might surprise you
    • Formation of a small behavior team
• Prompt Injections
  • Exfiltrating the data somehow I missed the explanation
  • Lethal trifecta
  • Wrapping untrusted context in tags
  • “While this can help, there are no guarantees”
  • It’s on us to design our product with this in mind
  • Read and confirm the data shared with the form as part of the tool call
    • Human confirmation step

Key Takeaways

1. Tools
2. Treating model behavior and craft and discipline
3. AI security as an emergent property of building products

• Technology shift -> Product Company -> Evolution Company
• When you recognize that it tech shifts, you have to embrace it with conviction

Slides

Slide: 2025-11-20-14-43

Key Point: The slide uses Brian Eno’s quote to make a philosophical point about how AI/automation shifts the challenge from technical skill to strategic judgement and decision-making about what to build.

Literal Content:

• Dark background with colorful gradient at bottom
• Quote: “The great benefit of computer sequencers is that they remove the issue of skill, and replace it with the issue of judgement.” (word “judgement” highlighted in green)
• ”…”
• “So the question becomes not whether you can do it or not, because any drudge can do it if they’re prepared to sit in front of the computer for a few days, the question then is, ‘Of all the things you can now do, which do you choose to do?’”
• Attribution: ”— Brian Eno”
• Source: “Interview in ‘The Wire’ with Paul Schütze (1995)“

Slide: 2025-11-20-14-52

Key Point: The slide illustrates a development methodology that progresses from prototyping and internal testing (dogfooding) through evaluation and human-in-the-loop refinement before shipping features.

Literal Content:

• Pink background
• Title: “For certain features, we can put it all together”
• Diagram showing workflow progression (left to right):
  • Left circle (pink outline) contains two blue circles labeled “Dogfood” and “Prototype” with circular arrows
  • Arrow points to middle section (pink outline) with three orange circles: “Dogfood”, “GEPA / Human hill-climb”, and “Collect Evals” with circular arrows
  • Arrow points to right green circle labeled “Ship”