The Browser Company

From Arc to Dia: How to Build AI Tools That Ship 10x Faster

The untold story of Arc's evolution into Dia, the AI browser that lets anyone build products. How The Browser Company mastered model behavior, automated prompt optimization, and transformed their entire organization.

We're not going to win unless we build the tools, the process, the platform, and the mindset to iterate, build, ship, and learn faster than everyone else.

— Samir Mody, Head of AI Engineering, The Browser Company

AI Engineer Conference 2024
Samir Mody, The Browser Company
10x Speed Improvement

Executive Summary

The Transformation

Arc, the innovative browser from The Browser Company, evolved into Dia — an AI-native browser that enables anyone from CEO to intern to build and refine products with full context, achieving a 10x speed improvement in product development.

"From our CEO to our newest hire can ideate and create a new product in DIA and also refine an existing one all with their full context."

— Samir Mody (throughout talk)

The Breakthrough

The secret wasn't better models — it was mastering model behavior as a craft, building automated prompt optimization (Jeba), and embracing non-engineers as AI builders.

"When you recognize that technology shift, you have to embrace it. And you have to embrace it with conviction."

— Samir Mody

What You'll Learn

Why tool optimization matters more than model optimization
The 4 pillars of model behavior engineering
How Jeba automates prompt tuning (tune text, not weights)
Why non-engineers make better AI builders
The "lethal trifecta" of browser AI security
How to transform your organization for AI-native development

The Philosophy: Tools Over Models

We're not going to win unless we build the tools, the process, the platform, and the mindset to iterate, build, ship, and learn faster than everyone else.

— Samir Mody, Head of AI Engineering

Build Better Tools

The competitive edge isn't the model — it's the tooling around it. Speed of iteration determines winners.

Create Better Processes

Systematic prompt optimization, automated testing, and continuous learning loops.

Cultivate the Mindset

Organizational conviction to embrace AI-native development across all teams.

The Result: 10x Speed Improvement

By building the right tools and processes, The Browser Company achieved a 10x improvement in product development speed. Everyone in the company — from CEO to newest hire — can now ideate, create, and refine products in Dia with full context.

"From our CEO to our newest hire can ideate and create a new product in DIA and also refine an existing one all with their full context."

— Samir Mody

Model Behavior: The New Engineering Craft

Traditional software engineering is about code structure and algorithms. Model behavior engineering is an entirely new craft — shaping how AI systems respond, reason, and interact through prompts, context, and feedback loops.

1. Prompt Engineering

Crafting precise instructions that guide model behavior. This is iterative, experimental, and data-driven.

2. Context Design

Determining what information the model needs and how to structure it for optimal reasoning.

3. Feedback Loops

Creating systems where model outputs are continuously evaluated and refined based on real-world results.

4. Evaluation Frameworks

Building automated systems to measure model performance and catch regressions before they reach users.

The Shift from Software to Behavior Engineering

Traditional engineers write code and ship features. Model behavior engineers craft experiences through iterative refinement, automated optimization, and continuous learning. The code you write matters less than the behavior you shape.

Jeba: "Tune Text, Not Weights"

Tune text and not weights. That's the Jeba philosophy. Instead of retraining models, we optimize the prompts themselves through automated iteration.

— Samir Mody, explaining Jeba's core principle

What Jeba Does

Jeba is The Browser Company's automated prompt optimization system. It takes a base prompt, generates variations, tests them against evaluation criteria, and iteratively improves performance — all without retraining the model.

Key insight: Most model improvements don't require fine-tuning. They require better prompts. Jeba automates this.

How It Works

  1. 1Start with a base prompt and evaluation criteria
  2. 2Generate prompt variations automatically
  3. 3Test each variation against eval suite
  4. 4Select best performers and iterate

Why "Tune Text, Not Weights" Matters

Traditional ML optimization requires retraining models — expensive, slow, and resource-intensive. Jeba's approach is fast, cheap, and iterative. You can run thousands of experiments in the time it takes to do one model fine-tune.

100x
Faster iterations
0
Model retraining needed
Continuous
Optimization loop

The Strategy & Ops Story: Non-Engineers as AI Builders

The Surprise: Non-Engineers Built the Best AI Features

One of The Browser Company's biggest revelations: their Strategy & Operations team, not engineers, built some of the most impactful AI features in Dia. Why? Because they understood user problems better than anyone.

The insight: When you give non-technical teams the tools to build AI features, they solve real problems because they're closest to the users and the workflows.

Why Non-Engineers Excel

  • Domain expertise: They understand the actual problems
  • No constraints: They don't know what's "impossible"
  • User empathy: They feel the pain points directly
  • Fast iteration: No code review bottlenecks

What This Means for Your Org

  • Democratize AI building: Give everyone the tools
  • Remove engineering bottlenecks: Let teams ship directly
  • Invest in low-code AI tools: Dia is the blueprint
  • Celebrate non-tech wins: Feature parity for all
"It wasn't just a product evolution. It was a company one."

— Samir Mody, on how AI transformed their entire organization

AI Security: The "Lethal Trifecta" of AI Browsers

Browsers sit at the middle of what we can call a lethal trifecta. It has access to your private data. It has exposure to untrusted content and it has the ability to externally communicate.

— Samir Mody, Head of AI Engineering

1. Access to Private Data

Browsers see everything: passwords, emails, banking info, personal messages. AI agents in browsers have access to your most sensitive data.

2. Exposure to Untrusted Content

Browsers load arbitrary websites, malicious scripts, and adversarial content designed to trick or exploit AI systems.

3. External Communication

Browsers can make API calls, send messages, transfer funds, and execute actions in the real world. A compromised AI agent can do real damage.

The Browser Company's Security Approach

AI browsers require fundamentally new security paradigms. Traditional sandboxing isn't enough when an AI agent is reasoning about what to do with your data.

Prompt Injection Defense

Detect and mitigate malicious instructions from web content

Data Access Controls

Fine-grained permissions for what AI can see and do

Action Validation

Human confirmation for high-risk operations

Continuous Monitoring

Audit logs and anomaly detection for AI behavior

Organizational Transformation: Embrace with Conviction

"It wasn't just a product evolution. It was a company one. When you recognize that technology shift, you have to embrace it. And you have to embrace it with conviction."

— Samir Mody

1. Recognize the Shift

The Browser Company realized AI wasn't just a feature — it was a fundamental shift in how software gets built. This required acknowledging that traditional engineering practices wouldn't suffice.

2. Embrace with Conviction

Half-measures don't work. They committed fully to AI-native development, building tools like Jeba, empowering non-engineers, and rethinking their entire product development process.

3. Democratize Building

Dia isn't just for engineers. They democratized AI development across the entire organization, enabling everyone from CEO to intern to ship AI-powered features.

4. Build Iteration Infrastructure

Speed is the competitive advantage. They invested in tools, processes, and platforms that enable rapid iteration, automated testing, and continuous learning.

7 Key Takeaways

1

Tools Beat Models

"We're not going to win unless we build the tools, the process, the platform, and the mindset to iterate, build, ship, and learn faster than everyone else." The competitive edge isn't model access — it's the tooling and processes around it.

2

Model Behavior is a Craft

Master the 4 pillars: prompt engineering, context design, feedback loops, and evaluation frameworks. This is a new engineering discipline, not just "prompt engineering."

3

Tune Text, Not Weights

Build automated prompt optimization systems like Jeba. Most improvements don't require model retraining — they require better prompts. Iterate on text, not weights.

4

Empower Non-Engineers

The best AI features often come from non-technical teams who understand user problems. Give everyone the tools to build AI features, not just engineers.

5

Security is Non-Negotiable

AI browsers face a "lethal trifecta": private data access, untrusted content, and external communication. Build prompt injection defense, access controls, and action validation from day one.

6

Embrace with Conviction

"It wasn't just a product evolution. It was a company one." AI transformation requires organizational conviction, not half-measures.

7

Speed is Everything

The Browser Company achieved 10x speed improvements by building the right infrastructure. Invest in iteration speed — it's your only sustainable competitive advantage.

Video Reference

From Arc to Dia: Lessons Learned Building AI Browsers

Samir Mody, Head of AI Engineering at The Browser Company

AI Browsers
Model Behavior
Prompt Optimization
Jeba
AI Security
Organizational Transformation
Watch Full Video

Speaker: Samir Mody
Role: Head of AI Engineering
Company: The Browser Company
Event: AI Engineer Conference 2024
Video ID: o4scJaQgnFA

Analysis based on Samir Mody's talk at AI Engineer Conference 2024. All insights extracted from the full transcript with contextual accuracy.