20 takeaways from this talk on building Dia, an AI browser

The Browser Company’s head of AI engineering breaks down how they’re building DIA. Watch the full talk here

On Building Fast

1. Be honest when your product isn’t hitting your vision—even if users love it. Arc was loved by millions but they knew it was incremental.
2. Build AI tools INTO your product, not beside it. They moved prompt editors from dev-only into DIA itself.
3. Context is everything. You can’t judge AI quality without testing in your actual workflow with real data.
4. The best prompt engineers might not be engineers. A strategy/ops guy rewrote all their prompts over a weekend and became the model behavior team.
5. Widen who can build. Their CEO and newest hires can now create AI features. Creativity explodes when you remove gatekeepers.
6. Try many approaches before committing. They tested tens of computer use strategies before building one.
7. Two distinct phases—don’t confuse them. Phase 1: wide funnel, many experiments. Phase 2: evals, hill-climbing, quality gates.

On Optimizing Prompts

1. JEPA = automated prompt optimization. Generate variations → score → reflect → mutate → repeat. Evolution for prompts.
2. Sample-efficient methods matter for small teams. You can’t always do RL or fine-tuning.
3. Model behavior is the new product design. It’s evolving from “prompt in, output out” to shaping personality and autonomous reasoning.
4. The loop is relentless: Build → refine → eval → ship → collect feedback → repeat.
5. New AI capabilities drop weekly. Get as many at-bats as possible.
6. Don’t underestimate the path from prototype to production. Fast ideation AND shipping discipline both matter.

On AI Security

1. Browsers face the “lethal trifecta”: private data + untrusted content + ability to act externally = maximum prompt injection risk.
2. Technical defenses against prompt injection will always leak. Tags? Escapable. Separating data/instructions? Helps but no guarantees.
3. Security through UX, not just code. Confirmation screens showing exactly what the AI will do before it does it.
4. Design every feature assuming prompt injection will happen. Autofill, scheduling, email—all have confirmation steps.

On Company Evolution

1. Dog food relentlessly. Real usage with real stakes reveals what evals can’t.
2. It’s not a product evolution—it’s a company evolution. They changed hiring, training, communication, collaboration.
3. When you recognize a technology shift, embrace it with conviction. Not halfway. All in.​​​​​​​​​​​​​​​​