20VC: Windsurf Founder on Will Model Companies Own the App Layer | Why Moats Do Not Exist in a World of AI | Why the Notion of Single Person $BN Companies is BS | Lovable vs Bolt & Cursor vs Windsurf: How Does it All End with Varun Mohan

Harry Stebbings welcomes Varun Mohan, CEO and Co-Founder of Windsurf, one of the hottest and most talked-about startups from Silicon Valley. Windsurf is the leading AI-native IDE with over a million users, generating over 50% of all committed software across thousands of companies.

The conversation opens with a fundamental insight about startup culture and the nature of persistence versus adaptability in entrepreneurship. Varun emphasizes that the startup world doesn't reward companies for doing the wrong thing consistently over time.

This sets the stage for a deeper discussion about when to pivot, when to persist, and how to maintain the delicate balance between conviction and flexibility that successful startups require.

Timestamp: [0:00-1:29]

Varun explores the fundamental challenge that most founders face: the first thing you believe will rarely be the right thing. This leads to a nuanced discussion about Peter Thiel's philosophy of choosing non-obvious ideas while acknowledging that most non-conventional ideas are simply bad ideas.

The key insight revolves around finding the sweet spot between obvious ideas (where big companies will beat you with their resources) and truly innovative approaches that seem counterintuitive but have merit. Varun explains that startups typically win by selecting non-conventional ideas, but founders must remain humble enough to recognize when their "weird idea" isn't actually viable.

He shares his company's evolution from Exa Function, which focused on GPU virtualization technology, based on the belief that GPU workloads would power the world. While they correctly predicted Nvidia's success, they misjudged the diversity of GPU workloads, expecting hundreds of different model architectures when everything ended up being transformers.

Timestamp: [1:29-3:32]

The conversation shifts to one of the most challenging aspects of entrepreneurship: balancing irrational optimism with uncompromising realism. Varun discusses the dangerous tendency of founders to become too attached to their ideas and the importance of thesis-driven thinking versus emotional attachment.

This segment reveals the psychological complexity of startup leadership, requiring founders to maintain two seemingly contradictory mindsets simultaneously. The discussion touches on how many people confuse persistence and grit with stubborn attachment to failing concepts.

Harry shares his own example of being too in love with disciplined portfolio construction as an investor, while acknowledging that today's market requires breaking most traditional rules. When pressed about his own blind spots, Varun admits they were consistently too attached to their ideas at every stage, wishing they had pivoted from GPU virtualization to code AI three months earlier despite having the signals.

The timing of their eventual pivot coincided with the challenges faced by their autonomous vehicle customers during the peak ZIRP (Zero Interest Rate Policy) era, when many couldn't raise their next funding rounds in mid-2022.

Timestamp: [3:32-6:14]

The discussion explores the critical question of timing and first-mover advantage in an era where "time to clone" - the speed at which competitors can replicate your product - has dramatically decreased. Harry references insights from Fiverr's CEO about this accelerating trend.

Varun presents a two-pronged argument for why being first remains crucial despite shorter clone times. First, it serves as a signal of organizational agility and willingness to disrupt oneself. Second, it enables faster learning cycles that compound over time.

He emphasizes that being first allows companies to learn from the market faster, positioning them to be first to the next idea as well because they can see "where all the dead bodies are" in their category. This creates a compounding advantage where early market presence enables better subsequent product development.

When Harry challenges this with the alternative strategy of learning from others' mistakes without bearing the costs, Varun counters that in software, many critical mistakes happen during internal R&D phases that are never visible to outsiders. His company has developed institutional wisdom about ideas they won't pursue because they've already tested them internally.

Timestamp: [6:14-8:26]

Varun provides concrete examples of internal experimentation and failed attempts that informed their eventual successful products. This segment reveals the iceberg of product development - the vast amount of work that happens below the surface before public launches.

He shares two specific examples of their hidden R&D efforts. First, they shipped a beta code review product last year, experimenting with different approaches including a Chrome extension and a parallel website. Despite multiple attempts, none felt right until they recently shipped a version that provides significantly more value, informed by all their previous failed iterations.

Second, their code agent product launched with Windsurf actually had its first version in early 2023, but it wasn't good enough. This early failure enabled them to identify and fix the fundamental issues: improving codebase understanding while waiting for language models to advance. When the models improved, their internal R&D on codebase understanding had progressed enough to launch successfully.

This approach of continuous internal experimentation and willingness to fail privately before succeeding publicly represents a significant competitive advantage that's invisible to outside observers.

Timestamp: [8:26-9:45]

The conversation shifts to funding philosophy, with Harry asking whether founders should raise as much as possible early to maximize their number of experimental "at-bats." This connects directly to Varun's experience with pivoting and the role of capital in enabling strategic flexibility.

Varun reveals that they raised their Series A from Green Oaks based on their GPU virtualization idea, with the same investor having led their seed round as well. He diplomatically leaves it to the audience to judge whether Green Oaks should have invested in the Series A for that particular concept.

However, he strongly advocates for the confidence that adequate funding provides. Having cash in the bank gave them the freedom to launch Kodium (their extension product) entirely free, leveraging their infrastructure expertise without worrying about immediate monetization pressures.

He emphasizes that the funding strategy only works if you maintain organizational agility. Most companies with 10-20 people become unwilling to completely change direction, but Windsurf demonstrated extreme flexibility when he and his co-founder decided to work on Kodium over a weekend, announced it to the company on Monday, and had everyone working on it immediately - despite generating millions in revenue from their previous business.

Timestamp: [9:45-11:08]

• The startup world doesn't reward persistence with the wrong idea - failing fast is better than failing slowly
• Most non-conventional ideas are simply bad ideas, but obvious ideas lack competitive advantage
• Successful entrepreneurship requires balancing irrational optimism with uncompromising daily realism
• Being first matters because it signals organizational agility and enables faster learning cycles
• Hidden R&D and internal experimentation create invisible competitive advantages
• Adequate funding enables more experimental attempts, but only if paired with rapid pivot capability
• Organizational culture that embraces dramatic direction changes is more valuable than the specific idea
• Internal wisdom about what doesn't work is as valuable as knowing what does work

Timestamp: [0:00-11:08]

The conversation shifts to organizational focus and the critical decision-making required when companies pivot. Varun articulates a fundamental principle that challenges conventional startup wisdom about diversification and hedging bets.

He explains that companies don't succeed by doing many things well, but by doing one thing exceptionally well. Using the concept of exponential growth curves and R-values attached to products, Varun argues that startups should always focus on the product with the higher exponential growth potential rather than dividing resources across multiple initiatives.

When executing a pivot, Varun advocates for going "cold turkey" - completely abandoning the previous direction rather than trying to maintain both. This approach requires difficult conversations with investors, customers, and team members, but he believes it's essential for success. The revenue multiples and business dynamics of different products are typically so different that maintaining both dilutes focus and reduces the likelihood of success.

Harry acknowledges the difficulty of this approach, especially when abandoning millions in existing revenue, but agrees with the fundamental principle.

Timestamp: [11:14-12:31]

Varun reveals a fascinating trend in Windsurf's user base: many users are non-developers who are building apps for productive use cases. This segment provides a concrete example of how AI-powered development tools are democratizing software creation.

He shares a specific example from within their own company, where a non-developer who leads partnerships has used Windsurf to build apps that replaced over $500,000 worth of annual sales tool spending. These applications include a partner portal and quoting tools - traditionally expensive, bespoke solutions that companies would purchase from enterprise software vendors.

The transformation occurs because these tools, while highly specialized for individual company needs, were previously expensive due to their bespoke nature and the high cost of custom software development. With AI making software creation more accessible, non-technical people can now build these specialized applications themselves.

However, Varun identifies a strategic tension: making non-developers more productive might require trade-offs in making professional developers working on large codebases more productive. This represents a classic focus dilemma for the company.

Timestamp: [12:38-14:18]

When Harry poses the fundamental question about whether Windsurf is about making engineers 10x better or enabling normal people to become engineers, Varun provides clear strategic direction.

Windsurf's primary focus is making professional engineers significantly more productive, with the democratization of coding for non-developers being a natural consequence rather than the primary goal. This represents a deliberate strategic choice about market positioning and product development priorities.

This decision reflects a sophisticated understanding of technology development and market strategy. By building deep capabilities for professional use cases first, they create technology that naturally scales down to simpler applications, rather than building for simple cases and struggling to scale up to complex professional needs.

Timestamp: [14:18-14:42]

Harry raises a strategic question about market convergence between different categories of development tools - whether Windsurf and Cursor will eventually compete with tools like Lovable and Bolt, and whether those consumer-focused tools will try to move upmarket.

Varun predicts that tools focused on non-developers will need to build more configurability and customization options over time to handle more complex use cases. Conversely, he believes that tools with deep codebase understanding will naturally be able to support simpler use cases with minimal additional effort.

He envisions convergence where companies that deeply understand codebases will be able to build entire applications using natural language that remain consistent with existing company code standards and architecture. This suggests a future where the technical depth becomes the key differentiator.

Timestamp: [14:42-15:47]

When Harry asks what Varun would do differently with infinite resources, the answer reveals a counterintuitive approach to innovation and risk management.

Rather than focusing on scaling existing successful products, Varun would make many more experimental bets within the company. He reveals that approximately 50% or more of their internal initiatives fail, but they're currently constrained in how many experiments they can run simultaneously.

This philosophy reflects a sophisticated understanding of innovation economics - that in high-uncertainty environments, the key to success isn't avoiding failure but increasing the number of attempts while managing the cost of each failure. The exponential returns from successful innovations can subsidize numerous failed experiments.

Timestamp: [15:47-16:23]

Harry, referencing insights from Neil Mehta, asks about Windsurf's unconventional approach to product development cycles. Varun challenges the conventional wisdom that more engineers automatically lead to better outcomes.

He argues that treating software development like a factory process misses the fundamental nature of innovation. When ideas are unproven, having too many people creates alignment problems because everyone has opinions and no one can definitively prove what's right or wrong.

Instead, Windsurf uses small, opinionated teams (3-4 people, typically engineers and possibly a designer) to prove out concepts. The critical insight is recognizing when an idea has "legs" - and Varun provides a key heuristic from his hard tech background:

He illustrates this with Windsurf's agent development, where even the early, imperfect version could accomplish tasks that were impossible 8-9 months earlier. Once this breakthrough moment is clear, they then resource the project more heavily.

Timestamp: [16:29-18:23]

Harry digs deeper into the operational details of how Windsurf structures their experimental teams and makes resource allocation decisions. Varun provides specific insights into their organizational approach.

Small experimental teams consist of 3-4 people, typically engineers and possibly a designer. For purely systems technology, it might be just engineers. An advantage of building developer tools is that their own developers can evaluate early versions and provide informed feedback about utility and potential.

Regarding budgets and timelines, Varun reveals they don't set specific constraints for experimental projects. This approach is enabled by being in an "unconstrained market" where great technology for accelerating software development is extremely valuable to customers.

Instead of budget constraints, they focus on tracking progress over time and making top-down decisions about whether to continue, pause, or pivot projects. This is explicitly not a democratic process - leadership makes these strategic calls about resource allocation.

When Harry asks about projects they should have continued longer, Varun admits they probably tabled their autocomplete experience too early. They delayed building a better product experience because they were constrained by VS Code's UI limitations, but once they had Windsurf as their own platform, they could invest more heavily in the capability.

Timestamp: [18:35-20:21]

• Startups succeed by doing one thing exceptionally well rather than many things adequately
• When pivoting, going "cold turkey" and abandoning the previous direction completely is more effective than trying to maintain multiple focuses
• Non-developers using AI tools can replace hundreds of thousands of dollars in enterprise software spending
• Strategic focus should prioritize professional use cases first, with consumer applications following naturally
• Development tool markets will likely converge as companies build deeper technical capabilities
• Innovation requires increasing the number of experimental attempts rather than avoiding failure
• Small, opinionated teams (3-4 people) are more effective for proving unvalidated ideas than large groups
• The key signal for resourcing an idea is when even the "crappy version" demonstrates breakthrough capability
• In unconstrained markets, tracking progress matters more than setting budget limits for experiments
• Platform constraints (like VS Code limitations) can delay optimal product development

Timestamp: [11:14-20:21]

Harry challenges Varun with a provocative statement about moats being dead in the AI world, using the easy switching between Cursor and Windsurf as evidence. This leads to a fundamental discussion about defensibility in AI-powered software.

Varun argues that the concept of startup moats is generally premature, referencing Hamilton Helmer's "7 Powers" as a great framework that's simply too early for most startups to apply. Even with exceptional talent - noting that most of their engineers come from MIT and represent the top few percent of graduates - having 50-100 engineers still represents only hundreds of engineering years invested in the product.

He argues that the only real moat in their category is speed - learning where the failures are, understanding how to compound advantages, and maintaining velocity of innovation. This challenges conventional wisdom about technical defensibility in favor of execution-based advantages.

Timestamp: [20:28-21:42]

Varun uses Nvidia as a compelling example to illustrate his point about moats and competitive dynamics. While many believe CUDA represents Nvidia's true moat, he argues this perspective misses the fundamental driver of their success.

He points out that large companies like OpenAI, Anthropic, and Google - who spend tens of billions on chips and even build their own TPUs - would find alternative solutions if CUDA disappeared. They would write assembly code or find other ways to utilize Nvidia hardware because the economic value is so substantial.

Instead, Nvidia's success comes from continuous execution: making hardware faster, improving interconnects, enhancing memory bandwidth, and maintaining this pace year after year. They face a "ticking time bomb" where failure to improve performance annually would erode profit margins and invite competition from AMD.

This example reinforces his thesis that even the most successful tech companies succeed through relentless execution rather than insurmountable technical barriers.

Timestamp: [21:42-22:54]

The conversation explores why hyperscalers with massive resources and distribution advantages struggle to compete effectively in fast-moving categories like AI development tools.

Harry challenges Varun's emphasis on speed over distribution, noting Google's early success through partnerships. However, Varun provides a nuanced response acknowledging that while Google had great partnerships, they also had the best product, which made others want to partner with them.

He observes that despite having significant distribution advantages, many hyperscalers have created products in the AI development category that "are not amazing products." The fundamental issue is the difficulty of marshaling talent to execute quickly in rapidly evolving spaces.

Varun attributes this velocity gap to "existential dread" - startups face survival pressure that drives intense focus and speed, while employees at large companies lack the same urgency. An Amazon employee working on competing products doesn't worry about losing their job if they don't ship great products quickly, whereas Windsurf would be "out of the market a year ago" if they had shipped the same quality as some hyperscalers.

Timestamp: [22:54-25:13]

Harry probes whether companies can build as efficiently remotely compared to the intense, seven-days-a-week Silicon Valley culture that's returning. Varun, whose company is 100% in-person, provides insights into why physical proximity creates competitive advantages.

While acknowledging that remote companies might succeed if they're more principled from the start, he argues that in-person collaboration creates "unfair advantages" in speed and flexibility. The ability to gather everyone in a room within five minutes, rather than coordinating across time zones, significantly enhances organizational agility.

This advantage becomes particularly crucial during the early stages when ideas frequently prove incorrect and companies need to pivot extremely quickly. The ability to marshal the entire organization rapidly for direction changes represents a significant competitive edge in fast-moving markets.

Timestamp: [25:19-26:29]

When Harry asks about brand importance and whether people identify as "Windsurf users" or "Cursor users," Varun provides a realistic assessment of brand value in rapidly changing markets.

While acknowledging that strong brands help with launching new products and building user bases, Varun emphasizes that their category requires constant innovation to maintain relevance. He provides historical context showing how quickly leadership can change in their space.

He stresses that companies in their category must "prove themselves almost every day" - if they don't ship something amazing every three months, they risk becoming irrelevant regardless of their brand strength. While brand can be helpful, it doesn't provide the right to move slower or reduce innovation pace.

Timestamp: [26:29-27:46]

Harry poses a critical question about valuation multiples for companies whose value can deteriorate rapidly, using Devin and GitHub Copilot as examples of how quickly market leadership can shift.

Varun acknowledges the inherent tension but points to aspects of their business that provide some stability. They have a substantial enterprise customer base with wall-to-wall implementations at Fortune 500 companies, which creates some switching costs, though not as substantial as products like Salesforce.

However, he identifies a fundamental trade-off: focusing too heavily on making products difficult to switch away from can slow innovation, and in fast-moving markets, customers will still switch to better products despite switching costs.

He provides a concrete example of this dynamic, referencing a tweet from Tom Blomfield (founder of Monzo) who switched from Cursor to Windsurf for a project because it was "better or just as good." This illustrates how easily users can move between products in their category.

Timestamp: [27:46-29:05]

Varun addresses the challenge of maintaining competitive advantages through concrete product innovation. He explains that while they ship releases every couple of weeks, complex technological capabilities require months of development.

He provides a specific example of their model suite 1, shipped a couple weeks prior to the interview, which performs comparably to frontier models for agentic workloads but is much faster and cheaper to run. This model can operate over large codebases and was developed by learning from internal user behavior.

The model now processes hundreds of billions of tokens of code daily and runs on their own GPUs, representing the kind of substantial technical investment that takes months to develop but creates meaningful differentiation. This illustrates how companies can build lasting advantages through deep technical work, even in rapidly changing markets.

He also notes the rapid pace of change in their category - many people hadn't heard of Windsurf 10 months ago but are now familiar with the company, demonstrating both the opportunity and risk in fast-moving markets.

Timestamp: [29:05-30:33]

Varun reveals that over 50% of their revenue comes from enterprise customers and shares insights about enterprise needs that aren't obvious to outside observers.

A key insight is that many large enterprises have substantial Java developer populations who use IntelliJ from the JetBrains family rather than VS Code. While Windsurf created their own editor by forking VS Code to have flexibility for building new agentic experiences, they also maintain full plugin support for all JetBrains IDEs with the same functionality.

He provides JP Morgan Chase as an example, where over 50% of developers are JetBrains users. By supporting these environments, Windsurf can serve all developers in an enterprise rather than limiting their product to only VS Code users. This comprehensive approach allows them to tell enterprises that 100% of their developers can use the product, not just 40%.

This strategy demonstrates how enterprise success often requires accommodating existing toolchains and developer preferences rather than forcing adoption of new environments.

Timestamp: [30:33-31:56]

• Traditional startup moats are largely irrelevant in AI-era software - speed of execution is the only sustainable advantage
• Even companies with exceptional technical talent (top MIT engineers) can be replicated given enough time and resources
• Large companies with massive distribution advantages struggle to compete due to lack of existential urgency
• Nvidia's success comes from continuous execution and improvement rather than insurmountable technical barriers
• In-person collaboration provides significant speed advantages over remote work in fast-moving markets
• Brand power exists but doesn't provide permission to slow innovation in rapidly evolving categories
• Enterprise success requires supporting existing developer toolchains (like JetBrains) rather than forcing new adoptions
• Complex technical differentiation (like custom models) still takes months to develop but can create meaningful advantages
• Market leadership can shift dramatically within 6-12 months, requiring constant innovation to maintain relevance
• Over 50% of enterprise revenue demonstrates the importance of wall-to-wall implementations despite switching risks

Timestamp: [20:28-31:56]

Harry asks a fundamental question about the future workforce composition and what defines an "engineer" as AI tools become more powerful. Varun provides a nuanced perspective on how the engineering profession will evolve.

He draws parallels to the historical evolution of programming languages, from assembly to C, C++, Java, Python, and JavaScript. Each progression made programming more accessible, but also created a new tier of developers who couldn't work at lower abstraction levels. Today's JavaScript developers often can't write assembly code, yet they're still valuable engineers.

Varun predicts a spectrum will emerge: people who operate purely on natural language abstractions with AI tools explaining how software components work, and those who can "go down to the weeds" for production-critical applications. He uses JP Morgan Chase's transaction processing system as an example - handling millions or billions of daily transactions requires deep technical validation that goes beyond natural language interfaces.

The key insight is that while more people will become "engineers" through natural language interfaces, there will always be a need for engineers who can work at the lowest levels for mission-critical systems.

Timestamp: [32:01-33:54]

The conversation shifts to the future of product management, with Harry referencing the common belief that PMs are "CEOs of the product" and citing ElevenLabs' approach of not having traditional PMs.

Varun argues that PM expectations will fundamentally change from telling others what to build to actually building it themselves. The role will require much more agency and hands-on capability, but AI tools will provide the technical capabilities that many PMs currently lack.

He sees technical PMs who understand and can write code becoming "even more deadly" than before, no longer needing teams of 10 people to prove out their ideas. This addresses one of his frustrations with current PM culture: spending excessive time writing documents to convince organizations rather than building prototypes to demonstrate concepts.

The future PM will be more of a builder-PM hybrid, using AI tools to rapidly prototype and validate ideas rather than relying on persuasion through documentation.

Timestamp: [33:54-34:58]

Harry probes whether the traditional design stage will be bypassed in favor of direct prototyping. Varun provides insights into how AI tools are already changing their internal development process.

The ability to quickly mock up designs that are consistent with company design standards and rapidly prototype them is becoming much faster. Windsurf's internal approach exemplifies this shift - they don't immediately jump to Figma when building apps if the application can be built quickly internally.

This suggests a tiered approach to design: rapid prototyping for initial validation and internal tools, with formal design processes reserved for core product experiences that require higher standards. The "laborious design stage" can be skipped for many use cases where speed of iteration matters more than pixel-perfect initial designs.

Timestamp: [35:04-35:36]

Harry explores the fascinating convergence happening in the development tools space, where different companies are meeting in the middle. Lovable and Bolt are moving upmarket, while design tools like Figma are adding "design to code" capabilities, creating a complex competitive landscape.

Varun provides important context about the scope of software development that goes beyond what these converging tools address. He argues that companies like Lovable and Figma are tackling "maybe a very small fraction of what software is" - primarily websites with backends attached.

He uses Google as an example of software complexity, pointing to their re-ranker systems and Spanner (their geographically consistent database with tens of millions of lines of code). These systems require deep engineering work that goes far beyond design-to-code conversion or simple web application building.

Rather than focusing on competitor movements, Windsurf concentrates on understanding all the ways software engineers spend their time and making them more efficient across the entire spectrum of software development, not just the "tip of the iceberg" that design tools address.

Timestamp: [35:36-37:05]

• The definition of "engineer" will become more flexible as AI tools enable natural language programming
• A spectrum will emerge: natural language operators and deep technical specialists for critical systems
• Fewer engineers will need to "go down to the weeds" but they'll still be essential for production-critical applications
• Product managers will evolve from coordinators to builder-PM hybrids who prototype rather than just document
• Traditional design stages can be skipped for rapid prototyping, with formal design reserved for core product experiences
• The convergence of development tools primarily addresses simple web applications, not complex software infrastructure
• Most software development involves complex systems (databases, infrastructure) that go far beyond front-end interfaces
• AI tools will give non-technical people agency to build, reducing the need for large teams to validate ideas
• The future favors those who can build and demonstrate rather than persuade through documentation

Timestamp: [32:01-37:05]

Harry explores the future of asynchronous remote agents, asking Varun to expand on how he envisions this technology developing and which tasks will be automated first.

Varun explains that Windsurf currently operates as a local agent capable of very long-running tasks, like migrating entire codebases from one version to another without requiring keyboard interaction once instructed. The concept of doing this asynchronously - such as firing off a request from your phone to generate a pull request - represents a powerful evolution.

However, he identifies three critical factors that make building products in this category complex: latency (response time), quality (correctness), and correctability (how quickly changes can be made). These factors become exponentially more important in asynchronous workflows.

The fundamental challenge is that when people wait hours for an asynchronous result, their quality expectations skyrocket. If the output requires significant correction, users lose faith in the product because the feedback loop becomes too slow for effective iteration.

Timestamp: [37:12-39:11]

Varun dives deeper into why quality standards become so critical for asynchronous agents, using specific examples to illustrate the user experience challenges.

When an asynchronous workflow takes hours to complete and produces a pull request on GitHub, even a 10-20% error rate becomes problematic. Users expect at least 90% correctness, but many believe it needs to be 99% correct because the stakes are higher when people have waited hours for results.

The core issue is that debugging partially correct code requires understanding what went wrong, which consumes valuable human time. This creates a paradox where async agents need to be nearly perfect to be useful, but complex tasks requiring high accuracy are exactly what's hardest for current AI to achieve.

Varun's prediction is that async remote agents will initially succeed only on tasks simple enough to be completed with high accuracy, while complex tasks requiring rapid iteration will remain in local environments like Windsurf where the feedback loop is immediate.

Timestamp: [39:11-39:43]

When Harry asks about user sensitivity to latency, Varun provides a striking example that demonstrates how even tiny delays significantly impact user behavior and product adoption.

For their tab completion product - which provides autocomplete and automatic refactors as users type - every 10 milliseconds of latency affects acceptance rates by percentage points. This granular sensitivity to response time highlights how critical speed is for real-time developer tools.

This insight reveals why the transition to asynchronous agents faces such challenges. If users are sensitive to 10-millisecond delays in real-time interactions, the psychological shift to waiting minutes or hours for results requires a fundamental change in expectations and workflow patterns.

The data suggests that latency sensitivity varies dramatically based on context - users will tolerate longer waits for more complex tasks, but the quality expectations scale proportionally with wait time.

Timestamp: [39:43-40:02]

The conversation explores the practical challenges of designing user interfaces for async remote agents, particularly around mobile interaction patterns and review workflows.

Varun identifies uncertainty around the optimal form factor for these interactions. While firing requests from an IDE makes sense since that's where developers are in their work state, the appeal of mobile interactions raises complex usability questions.

The fundamental challenge is code review on mobile devices. When an async agent generates thousands of lines of code changes, reviewing that output effectively on a phone becomes nearly impossible. This creates a constraint on interaction patterns.

Varun suggests that mobile interactions might work best for "one and done" scenarios - simple requests that produce outputs requiring minimal review. For these cases, the workflow would involve:

1. Send request from phone
2. Agent completes task with near-100% accuracy
3. Developer returns to laptop to quickly approve and merge

This highlights how form factor limitations influence both the types of tasks suitable for async agents and the quality standards required for success.

Timestamp: [40:02-41:02]

• Async remote agents will initially succeed only on simple tasks that can be completed with near-perfect accuracy
• Complex tasks requiring rapid iteration will remain in local environments due to feedback loop constraints
• User quality expectations scale dramatically with wait time - async results need 90-99% correctness vs lower standards for immediate feedback
• Even 10-millisecond latency differences affect user acceptance rates by percentage points in real-time tools
• Mobile form factors create fundamental limitations for code review, constraining async agent interaction patterns
• The "correctability" of AI output becomes critical when users can't easily iterate and fix errors
• Debugging partially correct code requires understanding what went wrong, consuming valuable human time
• Form factor choice influences both suitable task types and required quality standards
• One-and-done mobile interactions may work for simple requests with near-perfect output quality
• The transition from real-time to async workflows requires fundamental shifts in user expectations and behavior

Timestamp: [37:12-41:02]

Harry brings up a provocative statement from Satya Nadella about apps collapsing into agents and becoming just databases with business logic, suggesting companies like Salesforce and HubSpot would be reduced to databases that agents crawl over.

Varun provides a nuanced disagreement with this vision, acknowledging that while any company's state can be distilled to a database, the reality is more complex. He points to the significant organizational inertia and workflow complexity that has been built around existing systems.

The key insight is that complex workflows aren't just technical constructs—they're deeply embedded in how humans operate within organizations. The switching costs aren't purely technical but organizational and behavioral.

Additionally, Varun argues that current agents aren't capable enough to operate autonomously on databases without human supervision. While agents excel at reading from data sources, they're not yet trusted to write to databases at scale for arbitrary workflows, representing a significant capability gap that needs to be bridged.

Timestamp: [41:07-43:03]

Harry, speaking from his perspective as a venture investor where "the only thing we give a shit about today is agents," asks whether the industry is overly optimistic about agent capabilities and progression.

Varun delivers what he calls "the most obvious statement" - that people, especially investors, believe these systems are more capable than they actually are today. He uses the Devon launch from March as a perfect example of this disconnect.

However, he identifies a critical blindspot in the opposite direction: people fail to grasp how quickly the exponentials are improving. What models can accomplish in six months will be dramatically different from today's capabilities.

This creates a paradox where people simultaneously overestimate current capabilities while underestimating the rate of improvement. The example he gives is telling: eight or nine months ago, people weren't using agents much at all, yet here we are discussing their widespread adoption.

Timestamp: [43:03-44:30]

When Harry asks what will seem crazy in 12 months that we don't have today, Varun focuses on the broader software development lifecycle rather than just code writing.

Currently, AI tools primarily help with the code writing experience but largely ignore other crucial aspects of software development: designing software, deploying it, reviewing it, building, navigating, debugging, and testing. The limitation exists because software engineers work with many distinct data sources - logging systems, databases, browser data, and more.

Varun predicts that agents will soon gain access to all this disparate data, enabling them to provide much more leverage for debugging complex tasks and designing systems. This represents a fundamental expansion from current narrow applications to comprehensive development support.

This prediction suggests a shift from AI as a coding assistant to AI as a comprehensive development partner that understands and can operate across the entire software development ecosystem.

Timestamp: [44:30-45:56]

Harry asks a challenging question about what Windsurf has learned from Cursor's successes and potential mistakes. Varun's response reveals an important strategic insight about product development priorities.

He credits Cursor with taking "a really good approach on building high quality UI/UX" and notes that this proved to be what people really loved about the product. This was impressive because they executed it "right off the bat."

However, this wasn't initially Windsurf's intuition or approach. Varun reveals that they felt they needed a "massive technical breakthrough" and a genuinely new product experience before building their own IDE. Their philosophy was technology-first rather than UX-first.

This represents a fundamental philosophical difference: Cursor prioritized exceptional UX even with existing technology, while Windsurf waited until they had technical breakthroughs to justify building new interfaces. Both approaches have merit, but Varun acknowledges learning from Cursor's success with the UX-first approach.

Timestamp: [45:56-47:15]

Harry probes the balance between UI/UX quality and underlying model progression from providers like OpenAI and Anthropic in defining customer experience quality.

Varun acknowledges that frontier model capabilities are crucial - each new capability unlocks new types of product experiences that weren't previously possible. However, he emphasizes that successful products require both excellent models and thoughtful product development.

He provides Windsurf's launch as an example: they were the first agentic IDE experience, and while Sonnet (the underlying model) had existed for some time, it required significant work to combine these models with their own internal models and create a usable experience with appropriate UI/UX for quickly reviewing software.

This reinforces that while powerful models are necessary, they're not sufficient - successful products require thoughtful integration and user experience design to unlock the value of underlying capabilities.

Timestamp: [47:15-48:26]

Harry draws an interesting comparison between the model landscape and cloud infrastructure, noting that in cloud computing, companies typically stick with one provider (AWS, Google, etc.) rather than switching between them. He asks whether the model landscape will develop similar preferential relationships or remain multi-provider.

Varun argues that the model space is fundamentally different from cloud infrastructure due to low switching costs and lack of state. Currently, using different model providers is more like using Twilio for text messaging - a simple API call without persistent state or significant switching friction.

However, he envisions this could change as models become more sophisticated. Future models might maintain internal state about users, company data, and local usage patterns. Large codebases are already reaching billions of tokens of code, and if this becomes stateful context that can be injected and ejected from specific model providers, switching costs could increase significantly.

The key insight is that in rapidly evolving categories, it would be counterproductive for model providers to overinvest in switching cost mechanisms if they might miss the next major improvement (like test time compute). The analogy to cloud providers isn't accurate today, but could become relevant as the technology matures.

Timestamp: [48:26-50:27]

• Agent-only workflows face significant organizational inertia beyond just technical capabilities
• Current agents excel at reading data but aren't trusted for autonomous database writes at scale
• People overestimate current AI capabilities while underestimating the exponential rate of improvement
• The next major AI breakthrough will expand beyond code writing to the entire software development lifecycle
• UI/UX excellence can be as important as technical breakthroughs for product success
• Successful AI products require both powerful models and thoughtful integration/user experience design
• Model switching costs are currently low (like Twilio) but could increase as models become more stateful
• In rapidly evolving categories, overinvesting in lock-in mechanisms can be counterproductive
• Complex organizational workflows represent hidden switching costs beyond technical considerations
• Every aspect of software engineering will likely become 10x more effective through AI assistance

Timestamp: [41:07-50:27]

Harry explores the commoditization of language model APIs, noting that there are 12-13 different providers in China racing to reach the same utility level, with the US likely following suit. He asks whether we'll reach complete commoditization or if specific providers will dominate particular use cases.

Varun argues that it's unlikely any single model provider will "run away" with a specific category in the short term, especially if that category is valuable. While he believes in the scaling hypothesis - that more capital enables better models - he doesn't think we're near a point where one company will have an untouchable lead for years.

His reasoning centers on the rapid pace of innovation: if a new technique emerges in six months and someone else discovers it, they can implement it more efficiently. Missing a breakthrough means falling behind, regardless of previous advantages.

The key insight is that unless one company monopolizes both global capital and all great ideas simultaneously, sustained dominance is unlikely in such a fast-moving field.

Timestamp: [50:33-52:25]

Given the commoditization of the model layer, Harry asks whether model providers must move into the application layer to achieve meaningful differentiation.

Varun sees opportunities for API-level differentiation through specialization. Model providers can optimize for specific applications and workloads, creating better experiences for particular use cases.

This specialization approach allows model companies to become better vendors for specific customers without necessarily building full applications. However, he acknowledges that model companies are indeed moving up to the app layer in areas they see as valuable.

His philosophy is pragmatic: companies should be rational about the best way to win rather than adhering to rigid opinions about the "right" approach. The market will ultimately determine which strategies succeed.

Timestamp: [52:25-53:35]

Harry references Dario Amodei's prediction that 2026 will see solo billion-dollar companies and asks for Varun's perspective on this widely discussed concept.

Varun delivers a blunt disagreement: "No, I don't believe in that." His reasoning is rooted in basic market dynamics and competitive pressure.

He outlines the inevitable consequences of competition: compressed margins, reduced brand pervasiveness, and increased customer acquisition costs. The fundamental issue is that if one person can build something valuable, others will notice and compete.

Varun questions the logical foundation of the concept: "What is a company? A company is the sum of the discounted cash flows over time." If AI becomes so capable that humans aren't needed, then everyone would have access to the same AI, increasing rather than decreasing competitive pressure.

Timestamp: [53:35-55:20]

Harry asks a deeply personal question about when Varun questioned his love for building companies, acknowledging that inspiration isn't constant and some days are harder than others.

Varun identifies the period before strategic pivots as the most anxiety-inducing time in his founder journey. The pain comes from believing something could be much better while not yet acting on that belief.

He shares a powerful insight from their pivot from Exa Function to Codium: the post-pivot period was actually the freest he'd ever felt, despite expecting failure. The relief came from finally pursuing something they believed in, even with low probability of success.

The key insight is that the most anxiety-inducing periods aren't about failure itself, but about the gap between knowing what should be done and actually doing it. Taking action, even with uncertain outcomes, provides psychological relief and renewed energy.

Timestamp: [55:20-56:52]

Harry challenges Varun's contrarian approach to hiring - waiting until "drowning" in a role before hiring for it - against traditional board advice to hire six months ahead to avoid bottlenecks.

Varun provides a concrete example with their VP of Sales hire. Before bringing on Graham, they successfully closed large enterprise deals without any dedicated salespeople. This approach served two purposes: it demonstrated what enterprise sales should look like internally and proved to the potential hire that Windsurf was an attractive company to join.

His philosophy centers on a fundamental belief about startup failure: domain experts are valuable, but if the company can't achieve even 1-10% of the desired outcome internally, it's probably not the right time to hire externally.

The worst outcome, in his view, is hiring someone who isn't immediately important and ends up manufacturing work to justify their existence. This dilutes focus and creates artificial priorities that don't drive real business value.

Timestamp: [56:52-59:00]

• No single model provider will dominate categories long-term due to rapid innovation cycles and competitive dynamics
• Model companies can differentiate through API specialization and workload optimization rather than just moving to app layers
• The solo billion-dollar company concept ignores basic market competition and the accessibility of AI tools to all players
• The most anxiety-inducing periods for founders occur before pivots, when action is needed but not yet taken
• Taking action on uncertain outcomes provides psychological relief even when expecting failure
• Hiring should be driven by demonstrated internal capability rather than anticipatory planning
• Successful companies often look chaotic internally but excel at executing core business functions
• Domain experts are valuable only when the company has already proven basic competency in that domain
• Premature hiring creates artificial work and dilutes focus from essential business priorities
• Competitive pressure increases when powerful tools become widely accessible rather than decreases

Timestamp: [50:33-59:00]

Harry shares his own management evolution, crediting Julian (who led Workplace in Europe for 10 years) for teaching him that management can't be learned from books alone. He asks Varun about his biggest change of mind on management.

Varun identifies his most significant shift: understanding that team size doesn't correlate with output, and that ruthless focus is the only thing that truly matters. He learned that having many people working on multiple priorities is far more difficult than having the same revenue with everyone focused on one thing.

The challenge for startups like Windsurf is the constant temptation of new opportunities. Social media amplifies this with daily announcements about solved problems, creating pressure to chase multiple initiatives simultaneously.

Timestamp: [59:04-1:00:45]

Harry asks what makes Neil Mehta special as an advisor, noting that as fellow ambassadors they see different sides of him than founders do.

Varun highlights Neil's remarkable ability to analyze companies with intellectual flexibility, praising both his breadth of knowledge across different businesses and his insight into what makes them work and how they could improve.

Beyond analytical capabilities, Varun emphasizes Neil's generosity with time - something unexpected given his extensive portfolio of companies. This combination of deep insight and accessibility makes him particularly valuable as an advisor.

Timestamp: [1:00:50-1:01:31]

When Harry asks who Varun would most like to add to his board, he immediately names Scott Cook, highlighting the value of having a CEO and founder of an extremely successful software company.

Varun describes walking away from conversations with Scott with deep understanding of competitive dynamics and what it takes to run an enduring, long-term successful business. This reflects his appreciation for operators who have built lasting companies rather than just investors or advisors.

The choice reveals Varun's focus on learning from practitioners who have navigated the full journey of building and scaling software companies over decades.

Timestamp: [1:01:31-1:02:00]

Harry poses a hypothetical about making Windsurf open source, asking how Varun would feel and what he'd do differently to compete.

Varun's response reveals his confidence in where Windsurf's true value lies. He draws an analogy to Google, noting that if Google open-sourced their front-end search box, it wouldn't fundamentally change anything because the real value is in the backend systems.

While acknowledging that GitHub Copilot has taken an open-source approach, Varun emphasizes that most of Windsurf's logic and differentiation exists in the backend. The frontend interface, while important for user experience, isn't where their competitive advantage lies.

Timestamp: [1:02:00-1:02:48]

When asked what advice he'd give an 18-year-old sibling about university and positioning for the new world of work, Varun focuses on fundamental skills rather than specific technical knowledge.

His core advice centers on problem-solving capabilities and being a "high agency individual." He views computer science as essentially the study of problem-solving, making it valuable regardless of specific programming languages learned.

He emphasizes that the specific programming languages matter less than the ability to break apart problems into distinct pieces and execute on solutions. This reflects a broader philosophy about adaptability and fundamental thinking skills being more valuable than technical specifics in a rapidly changing field.

Timestamp: [1:02:48-1:03:26]

When Harry asks what OpenAI did wrong that Varun learned from, there's a pause before he lands on "naming" as his answer.

While Varun doesn't elaborate extensively on this point, the choice is intriguing given OpenAI's branding challenges around the "Open" part of their name as they've become less open over time. This suggests Varun recognizes the importance of choosing company and product names that can scale with the business without creating contradictions or limitations.

The brevity of his response and slight hesitation suggest this might be a diplomatic answer, but it points to the often-underestimated importance of branding decisions in tech companies.

Timestamp: [1:03:26-1:03:37]

Harry asks what aspect of Varun's CEO performance he'd most want to change, prompting honest self-reflection about leadership weaknesses.

Varun identifies his tendency to get deeply involved in details that don't fundamentally matter to the company's long-term success. He describes himself as "neurotic" and wanting to understand everything happening internally, including spending patterns, at a granular level that's probably unnecessary.

He recognizes this creates mental waste - using cognitive cycles that could be better applied elsewhere. This self-awareness about the tension between wanting control/understanding and focusing on what truly matters reflects mature leadership thinking.

Timestamp: [1:03:37-1:04:06]

For the final question, Harry asks what impact Varun wants future generations to remember him for - what his grandfather did for humanity.

Varun frames his answer around Windsurf's mission rather than personal achievement, emphasizing the team's collective goal to "reduce the time it takes to build technology by 99%." This ambitious vision goes far beyond just improving developer productivity to fundamentally transforming how quickly human ideas can become technological reality.

The 99% reduction target represents a moonshot goal that would fundamentally change innovation cycles, democratize technology creation, and potentially accelerate human progress across all fields that depend on software. Rather than claiming personal credit, he positions himself as part of a team effort toward this transformational outcome.

Timestamp: [1:04:06-1:04:34]

• Team size doesn't correlate with output - ruthless focus on one priority is more valuable than distributed effort across multiple initiatives
• Management wisdom comes from experience and nuanced understanding that can't be learned from books alone
• Intellectual flexibility and generosity with time are rare and valuable qualities in advisors
• The best board members are operators who have built enduring businesses, not just investors or advisors
• Backend systems and logic provide more defensible competitive advantages than frontend interfaces
• Problem-solving skills and high agency matter more than specific technical knowledge for career preparation
• Naming and branding decisions can create long-term constraints as companies evolve
• CEO neuroses about unimportant details waste mental cycles that could be better applied to strategic priorities
• Ambitious mission statements should focus on team achievement rather than individual legacy
• The ultimate goal is reducing technology building time by 99%, fundamentally transforming innovation cycles

Timestamp: [59:04-1:04:45]