← back
From Arc to Dia: Lessons learned building AI Browsers – Samir Mody, The Browser Company of New York
Takeaway
Build prompt/eval tooling directly into your product so the whole team can iterate AI features with real user context — and use techniques like GEPA over RL.
Summary
- Browser Company shipped Arc (2022) then pivoted to Dia, an AI-native browser; key tooling investments: in-product prompt editors with full personal context, evals, training data collection, and automation for hill-climbing.
- Built prompt-iteration tools directly into Dia so PMs, designers, even the CEO can ship and refine AI features alongside engineers — 10x faster iteration.
- Uses GEPA (reflective prompt mutation paper) to optimize prompts via seed→score→Pareto-select→LLM-reflect→new prompts — sample-efficient vs RL for a small team.
- Treats 'model behavior' as a craft: behavior design, data collection for measurement, model steering (prompts, context, params); iterates with internal + external feedback.
- Tried 'tens of different computer-use strategies' inside the product before committing — prototyping cost matters.
ai-browsergepaprompt-optimization
Original description
What happens when you take a polished, beloved browser and rebuild it from the ground up around AI? In 2024, The Browser Company did exactly that: transforming Arc, a human-designed browser, into Dia, a browser with an AI-native browser experience. The journey forced us to unlearn what we knew about product design, system architecture, and even what "good UX" means when the interface itself can reason, plan, and act. In this talk, I'll share the lessons we learned along the way: how we redesigned for emergent behavior instead of deterministic control, how our engineering and design teams learned to prototype with language models, and how our culture evolved to ship fast in a world where we're still learning what a good system looks like. You'll leave with practical insights for building AI-native products, from technical architecture to team mindset, and a glimpse into how we're thinking about the future of AI browsing. Speaker: Samir Mody | Head of AI Engineering, The Browser Company https://x.com/jedimody https://www.linkedin.com/in/samir-mody/