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
- Vibe coding is projected to remain the dominant development methodology, with software engineer roles expected to transition into product engineer positions where human taste and user-centric judgment supersede raw coding speed.
- Development velocity is predicted to increase exponentially from a 10x speedup six months ago to a current 100x speedup, facilitating workflows where developers easily scrap and rewrite code rather than maintaining attachment to existing versions.
- The industry is expected to bifurcate into two distinct categories: product engineers focusing on user needs and taste, and specialized backend engineers tackling infrastructure and complex systems problems that current AI tools cannot solve.
- Current LLM limitations include poor debugging capabilities requiring explicit human instruction and an inability to build effectively upon previous code, necessitating a "reroll" approach similar to image generation tools.
- Future improvements are anticipated within a six-month timeframe, with next-generation reasoning models ("0.5" and "Flashback 2.0") expected to significantly enhance debugging and enable code bases to be iterated upon rather than rewritten from scratch.
- Specific tools are positioned with varying capabilities: Windsurf is expected to become a fast follower to Cursor by indexing entire codebases automatically, while Devon remains limited for serious features due to a lack of codebase understanding.
- Coding composition is shifting rapidly, with one quarter of founders estimating that over 95% of their codebase is AI-generated, signaling the emergence of "AI coding natives" who may lack traditional degrees but possess high productivity.
- Hiring practices are expected to lag behind technological shifts, though future assessments will likely evolve to screen for the ability to use tools "100 times harder," requiring candidates to prove competence by coding without LLM assistance or under strict supervision.
- Core skills such as reading code, debugging, and exercising "taste" to distinguish between high-quality and flawed AI output are expected to remain constant and critical, particularly for technical founders who must verify agent accuracy.
- A crisis is predicted for startups where rapid "zero to one" development via vibe coding succeeds, but "one to n" scaling phases will encounter bottlenecks if founders cannot descend into systems engineering to build custom compilers or infrastructure as open-source tools fail at scale.
- Technical founders are expected to require a "superpower" in detecting AI hallucinations and lies, while the most elite engineers will distinguish themselves through deliberate practice and classical training against a broader class of "good enough" developers relying on low-barrier tools.