Best of the Pod: Would You Shut Down Your Most Successful Product? The Arc to Dia Story

Josh and Hursh explain the difficult decision to shut down Arc despite having millions of monthly users. They discuss how AI fundamentally changed their vision from fixing Arc's 'novelty tax' problem to realizing they needed to build something entirely new from scratch. The realization that browsers would become personal intelligence layers, not just tab managers, drove the pivot.

• •Arc had millions of users but faced a 'novelty tax'—too much new stuff attracted early adopters but repelled mainstream users
• •The team spent months trying to evolve Arc before realizing they needed to start completely fresh
• •Key insight: The browser isn't the product—it's the enabling technology for a personal intelligence layer that sits across all devices
• •Decision wasn't about Arc failing, but about seeing a fundamental platform shift that required ground-up rethinking
• •Internal challenge: convincing a 70-person team to abandon years of work for an uncertain AI future

The team built Arc Search as a simple mobile experiment—instead of returning links, it would generate custom web pages answering your query. This 'skunkworks' project unexpectedly exploded in popularity and taught them critical lessons about simplicity, focus, and how AI could transform the browser's role as the internet's 'choke point.'

• •Arc Search was built by a contractor as a low-stakes experiment while Josh was in Paris
• •Core idea: Generate the perfect webpage for your query instead of showing links—change one thing, do it extremely well
• •Josh tweeted it casually on a Sunday; it became their most successful launch ever
• •Revealed that search (the command-T box) is the most-used text box on desktop—a leverage point they'd previously ignored
• •Taught them the 'novelty budget' principle: keep it simple, change one thing, make that the story

Josh and Hursh candidly discuss the hardest year of their professional lives—from February realization to June announcement to September final commitment. They faced intense public criticism, employee attrition, hiring challenges, and personal doubt while betting on AI scaling laws that hadn't yet materialized. The conversation reveals the emotional toll of leading through uncertainty.

• •Josh describes waking up every morning with 'a pit in my stomach' for months
• •The team had built trust through Arc's success, but drew down that trust to 'maybe a negative balance'
• •External world (family, friends, press) was skeptical or hostile to AI pivot while founders had deep internal conviction
• •Key challenge: employees joined for Arc, not 'the browser company'—had to rebuild motivation around uncertain vision
• •Hirsh had a kid, Josh had two kids during this period—compounding the stress with no sleep
• •The gradual evolution (Feb realization → June off-site → September final decision) made internal communication especially difficult

The team made a critical decision to rebuild Dia's architecture from scratch using Swift's structured concurrency, fixing Arc's performance issues while maintaining rapid prototyping capabilities. They anticipated that launching would kick off an industry-wide race to build AI browsers and needed infrastructure to move fast.

• •Arc used Redux-style state architecture optimized for fast prototyping but had performance costs
• •Dia rebuilt in Swift with structured concurrency to move work off main thread—dramatically faster
• •Strategy: give team members 'blank page' opportunities to fix long-standing regrets and dream big
• •Anticipated competitive response: 'when we launch this, if we are the first ones, we're gonna kick off a race'
• •Built architecture specifically to enable rapid iteration once competitors emerged

Josh explains how Arc's original 'internet computer' vision—a layer sitting across all devices because apps and files live on the web—evolved into Dia's personal intelligence layer. Hursh's prescient insight was that the browser itself isn't the product; it's the enabling technology for an AI assistant that knows everything you do.

• •Original Arc vision: your 'computer' in 5-10 years is actually a layer across all devices, not a single machine
• •Hursh's key insight: the value isn't the browser features—it's the personal assistant the browser enables
• •Browser provides unique advantages: intercepts queries, sees your docs, remembers your actions, has your authentication
• •Analogy: iPhone isn't valuable because it's a cell phone—that's just the enabling technology
• •This realization helped team understand why they couldn't just 'duct-tape' AI onto Arc

Dia's launch exceeded all expectations. Despite releasing what the team considered a 'basic' product, users immediately understood the vision. Most surprisingly, the hastily-added 'skills' feature—allowing users to create custom AI apps with simple prompts—exploded in popularity, revealing a desire for 'handmade software' and validating the platform strategy.

• •Team expected brutal feedback for shipping 'basic' product; instead got immediate resonance and understanding
• •Dan Shipper switched from Safari to Dia as default browser—significant given his investor relationship and personal standards
• •Skills feature added one week before launch as afterthought; became the most exciting part of the product
• •Users creating highly personal, niche skills (e.g., extracting guitar tab chords from YouTube videos)
• •Insight: AI enables 'just-in-time' software creation—millions of tiny use cases vs. one big TAM
• •Head of design Dustin took a week alone to prototype, came back saying 'you're inventing a new programming language'
• •Entire company now pivoting to build out skills platform—a direction they thought was 6-12 months away

The team worked on memory and personalization for nine months before killing the project entirely. Six weeks before launch, they tried one more approach enabled by new model capabilities (GPT-4o) and it suddenly worked. This illustrates how rapidly the AI landscape evolved and the importance of betting on scaling laws.

• •Core thesis: Dia should 'get better with every tab you open' like TikTok gets better with every swipe
• •Memory/context would differentiate when model intelligence commoditizes—it's what makes AI personal
• •Worked for 9 months, killed the project, then revived it 6 weeks before launch with new approach
• •GPT-4o made Dia possible—Josh messaged OpenAI contact saying 'this model made our thing happen'
• •Demonstrates importance of betting on scaling laws even when current capabilities don't support vision

With Perplexity, OpenAI, and others entering the browser space, Josh argues traditional moats don't apply. Their competitive advantage is maintaining conviction, moving fast without franticness, and building with 'character' rather than top-down business strategy. The parts of Dia that resonate most are where they said 'fuck it, let's have fun.'

• •Traditional moats are questionable in AI era—need to rethink competitive strategy assumptions
• •Best strategy: stay locked into energetic conviction rather than frantically beating competitors
• •Products built from genuine conviction (like ChatGPT) feel different from derivative competitors
• •Company culture of 'assume you don't know' and prototype-driven development is structural advantage
• •Taste and design sensibility matter more than feature parity
• •Regret: spending too much time thinking about external competition vs. just building with joy

Josh observes that non-tech people are using AI for emotional support and subjective reasoning—asking for advice on health issues, relationship problems, vacation planning. He argues the next frontier isn't just agents clicking buttons, but AI that handles the qualitative, subjective reasoning that's actually the hard part of knowledge work.

• •Non-tech people at parties admit using AI for personal advice, health questions—feeling 'weird but better' afterward
• •Current AI industry too focused on benchmarks, raw IQ, computer use capabilities—missing emotional intelligence
• •Real value isn't clicking buttons (minor nuisance)—it's subjective reasoning like 'which town for family vacation?'
• •Prediction: In 24 months, the AI interface people value most will be the one they have deepest emotional connection to
• •Browser companies building 'agents that book things' are missing the forest for the trees
• •Value will move up Maslow's hierarchy as basic intelligence commoditizes

Josh is studying Romanticism—the 19th century movement responding to Enlightenment rationalism—to understand how society adapts when relationships with technology fundamentally change. He argues language models are 'romantic coded' technology because they're irreducibly squishy, requiring us to grapple with what we actually want from life rather than just what's mechanically possible.

• •Reading Romanticism to understand historical parallels for how culture responds to technological relationship shifts
• •Interested in objects/spaces that create comfort and safety despite not being perfect—like worn-in drum sets
• •AI industry is 'extremely rational, mechanical' focused on benchmarks, but users experience it differently
• •Language models are first software that can't be reduced to rules—fundamentally vibes-based technology
• •Society will need to grapple with philosophical questions: what do we want from this life with AI?
• •Studying how to build emotional connection and sense of ownership in AI interfaces