Interview, Podcast
Vibe Coding Is The Future
Y CombinatorAndrej Karpathy, Gary, Jared Harge, Diana, Abhi von Copycat, Mark Mandelmann, Yoav, Leslie Kendricks, Francesc Campoy Flores, Mark Mirchandani, Melanie Warrick, Trevor, Mark Blyth, Anders Ericsson, Malcolm Gladwell, Picasso, Max Levchin, Toby Lutke, Mark Zuckerberg
- Dominant Trend Confirmation: Participants characterize "vibe coding" (fully embracing AI tools while minimizing direct code writing) not as a fad, but as the emerging dominant standard for software development; failure to adopt it risks being left behind.
- Role Evolution to Product Engineering: Founders report a transition from "software engineer" to "product engineer," where human taste and product judgment become more critical than raw coding ability as LLMs enable 10x to 100x speedups.
- Exponential Acceleration Metrics: One founder from TrainLoop cited a specific timeline of velocity increases: a 10x speedup six months prior, a 10x increase one month later, resulting in a cumulative 100x speedup "now," described as exponential acceleration.
- Workflow Parallelization: Tools like Cursor allow for parallelized coding workflows; founders report running multiple instances of AI agents simultaneously to develop distinct features concurrently, with one instance stating they could easily run three or more.
- AI-Generated Code Volume: In the current YC batch, approximately 25% of founders estimate that more than 95% of their codebase characters are AI-generated, even among highly technical founders with non-classical CS backgrounds.
- Shift in Hiring Assessments: Traditional whiteboard algorithms and classical CS knowledge are being replaced by assessments focused on "taste," ability to work with users, and proficiency with AI tools, though the broader market has not yet fully adapted to these new hiring standards.
- Tool Landscape Shifts:
- Cursor remains the market leader.
- Windsurf is identified as a fast follower with a key advantage: full codebase indexing, allowing the AI to locate relevant files without explicit user direction.
- Devon is mentioned but limited to small features due to a lack of deep codebase understanding.
- ChatGPT (Reasoning Models) is still utilized for complex debugging tasks where pre-reasoning models (like Cursor's current iteration) fail.
- DeepSeek R1 and Gemini (via long context windows) are emerging as viable contenders for specific use cases.
- Debugging Limitations: Current LLM tools struggle significantly with debugging; users must either provide "first-time engineer" level explicit instructions or adopt a "reroll from scratch" strategy similar to image generation tools like Midjourney.
- The "Zero to One" vs. "One to N" Split:
- Zero to One: Vibe coding is ideal for rapid prototyping and initial product-market fit.
- One to N: Scaling to massive user bases requires "hardcore systems engineering" and classically trained architects to optimize infrastructure, a capability currently beyond the scope of AI-generated code and "good enough" engineers.
- Historical Precedent: The discussion cites Facebook's shift from PHP to a custom compiler (HipHop) and Twitter's struggles with scalability as examples where rapid initial development eventually necessitated deep systems expertise.
- The "Good Enough" vs. "Exceptional" Engineer Dichotomy:
- AI tools lower the barrier to entry, creating a "fat class" of engineers who are "good enough" through rapid iteration.
- Exceptional engineers (top 1%) will still require "deliberate practice" and deep classical knowledge to achieve world-class status, analogous to how Picasso mastered realism before mastering abstraction.
- Critical Evaluation Skills: The ability to "read and review" code remains the most constant and valuable skill; engineers must possess enough technical knowledge to detect AI hallucinations or "bullshit" from both AI agents and human subordinates.
- Future Outlook: Participants predict that as coding becomes exponentially cheaper, the primary differentiator will be the ability to judge code quality, make high-level architectural decisions, and solve complex system-level problems rather than syntax generation.