Skip to content
Y CombinatorY Combinator

Varun Mohan: Why Insights Depreciate and Why Evals Compound

Windsurf rebuilt from GPU virtualization to vibe coding in 48 hours; evals and irrational optimism are the moats when every competitive insight depreciates.

Varun MohanguestGarry Tanhost
May 1, 202552mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Windsurf CEO on rapid pivots, AI agents, and democratized software building

  1. Varun Mohan, co‑founder and CEO of Windsurf, explains how his team pivoted from GPU virtualization (Exafunction) to AI coding tools in a single weekend after realizing transformer models would commoditize their original business.
  2. He details how a lean engineering team rapidly built a Copilot competitor, then an agentic IDE (Windsurf) by leveraging deep GPU/inference expertise, rigorous evals, and relentless experimentation—even when most bets fail.
  3. The discussion covers Windsurf’s architecture, agent design, RAG alternatives, enterprise adoption, and how AI is reshaping software development, hiring, and developer workflows.
  4. Mohan argues that ‘developers’ will evolve into ‘builders’ as AI makes software creation radically more accessible, and that startups must continually generate new insights or risk slow death.

IDEAS WORTH REMEMBERING

5 ideas

Pivot fast when your core thesis breaks, even if current metrics look good.

Exafunction was profitable with millions in revenue and a fresh Series A, but Mohan’s team killed it over a weekend when they realized transformer models would commoditize GPU infra; they treated pivoting as a badge of honor rather than a failure.

Combine irrational optimism with uncompromising realism to build enduring startups.

Mohan emphasizes that founders need enough blind optimism to attempt ambitious things, but must switch to brutal realism when facts change—otherwise they either never start or cling to dead ideas.

Moats in AI are dynamic: compounding insights and execution, not one-time breakthroughs.

Every technical insight depreciates quickly; Windsurf’s strategy is to continually find new edges (e.g., middle-of-line completion, agentic editing, advanced retrieval) and stack them, much like NVIDIA must keep innovating to defend margins against AMD.

Rigorous, domain-specific evaluation unlocks justified complexity and better systems.

Instead of blindly adopting vector DB RAG, Windsurf built detailed evals over real codebases (e.g., commit-intent reconstruction and test-passing) and only added complexity like AST parsing and GPU-based re-ranking when evals proved it materially improved retrieval and edits.

Agents are only useful if they can operate precisely on large, real codebases.

Windsurf invested heavily in understanding huge repositories (100M+ LOC), tracking a unified timeline of human and agent actions, and designing workflows so agents make surgical edits rather than sprawling, unreviewable diffs.

WORDS WORTH SAVING

5 quotes

Every single insight that we have is a depreciating insight.

Varun Mohan

Startups require irrational optimism and uncompromising realism, and they’re in tension with each other.

Varun Mohan

If we don’t continually have insights that we are executing on, we are just slowly dying.

Varun Mohan

The mission of the company is to reduce the time it takes to build technology and apps by 99%.

Varun Mohan

This notion of just a developer is probably going to broaden out to what’s called a builder, and I think everyone is going to be a builder.

Varun Mohan

Exafunction’s pivot from GPU virtualization to AI coding tools (Codium/Windsurf)Competing with incumbents like GitHub Copilot and fast-moving startupsBuilding agentic IDEs, vibe coding, and large-scale codebase understandingArchitecture choices: beyond simple RAG, retrieval systems, and eval frameworksEnterprise go-to-market and supporting massive legacy codebasesHow AI changes software engineering, hiring, and developer workflowsFuture of coding: democratization, builders vs. developers, and just‑in‑time software

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome