AI summary by TwinMind
Vibe Coding Is The Future
Y Combinator
31m 33s
Introduction to Vibe Coding
- Hosts: Gary, Jared, and Diana - Y Combinator partners who funded companies worth hundreds of billions
- Topic: "Vibe coding" from viral Andrej Karpathy post - "fully give in to the vibes, embrace exponentials, forget code exists"
- Surveyed current YC batch founders on AI coding tools usage and workflow changes
- Phenomenon described as "giant beanstalk seeds dropped at night" - sudden transformation
Founder Survey Response Highlights
- Outlet founder: "Role of software engineer will transition to product engineer - human taste more important than ever as coding tools make everyone 10x engineer"
- Obby from Asra: "Don't write code much, just think and review" (highly technical founder from dev tools background)
- RB from Copycat: "Far less attached to my code now - decisions less biased since I can code 3x faster, easy to scrap and rewrite"
- Yoav from Cix: "Write everything with cursor, sometimes two windows open in parallel prompting different features"
- Train Loop founder: "6-12 months ago 10x speedup, 1 month ago to now 100x speedup - exponential acceleration. No longer engineer, I'm product person"
Role Evolution: Product vs Systems Engineers
- Two emerging paths for engineers:
- Product engineers: Focus on user needs, taste, ethnographic work in GDP segments
- Systems engineers: Infrastructure, hard technical problems, don't want user interaction
- Parallel to existing front-end vs back-end divide
- Evals becoming most important part of process
- From TripleByte experience: Beyond technical threshold, key differentiator was willingness to talk to users
Current Limitations and Debugging Challenges
- AI tools terrible at debugging - humans still required
- Must give explicit instructions like teaching first-time software engineer
- Alternative "vibe" approach: Ignore bugs, rewrite from scratch
- Coding style changes when writing becomes 1000x cheaper
- Humans would never scrap long work for bugs, but AI can rewrite 1000 lines in 6 seconds
- Similar to image generation - just "roll" multiple times instead of iterative fixes
Tools and Models Usage Trends
- IDEs: Cursor dominant leader, Windsurf fast follower
- Cursor requires telling which files to examine
- Windsurf indexes entire codebase, figures out relevant files automatically
- Devon: Mentioned but limited to small features, doesn't understand codebase well
- Models:
- Claude Sonnet 3.5 still big contender
- O1, O1 Pro, O3 reasoning models gaining ground, neck-to-neck with Sonnet
- GPT-4 virtually no use for coding
- DeepSeek R1 viable contender
- Gemini used for longest context window - entire codebase input for one-shot bug fixes
- ChatGPT still used for reasoning models in debugging
AI-Generated Code Statistics
- 25% of founders report >95% of codebase AI-generated
- All founders highly technical, capable of building from scratch year ago
- Some founders so young they learned to code in last 2 years - don't know world without Cursor
- Example: Highly technical founders with math/physics degrees, not classically trained CS, extremely productive with AI tools
Historical Context and Hiring Evolution
- Comparison to Rails/Active Record debates about abstraction vs understanding internals
- Successful companies (Stripe, Gusto) shifted hiring from whiteboard algorithms to practical productivity
- Interview evolution: "Teach us how you think" → "Build todo app in 3 hours"
- Current hiring hasn't caught up to vibe coding reality
- Two valuable skills emerging:
- Great taste and product understanding
- Systems thinking and architecture
- Speed of coding may become less valuable differentiator
Zero-to-One vs Scale Challenges
- Vibe coding excellent for zero-to-one development speed
- Historical parallel: Rails/PHP enabled fast development but hit scaling walls
- Twitter fail whale example - architecture couldn't scale
- Facebook hired hardcore systems people to build custom compiler (HipHop) for PHP
- Current AI tools not good at low-level systems engineering
- Different skill sets needed for different company stages
Technical Assessment Evolution (TripleByte Insights)
- TripleByte: Technical assessment platform, founders did thousands of engineering interviews
- Key insight: Companies want different things, must design assessments accordingly
- Modern assessment would need to account for AI tool proficiency
- Questions need to be 100x harder or different format since ChatGPT can solve traditional problems
- Debate: Force coding without AI vs allow AI with new question types
- Code reading/debugging skills remain critical
- Need taste to distinguish good vs bad AI output
Developing Taste and Expertise
- AI coding natives question: How develop taste without classical training?
- Market forces will determine - startups with bad code will fail at scale
- Two classes emerging:
- "Good enough" engineers (low barrier to entry)
- Exceptional top 1% requiring deliberate practice
- Picasso analogy: Master realistic drawing before abstract art
- Deliberate practice (Anders Ericsson research) vs just putting in hours
- Technical founders advantage: Can't be bullshitted by employees or AI agents
Technical Founder Advantages
- Examples: Max Levchin, Toby Lütke (Shopify) - deep technical knowledge as CEOs
- Story: Interaction designer catching dev team lying about implementation feasibility
- Technical knowledge prevents being misled by employees or AI agents
- Workplace cultures too polite to call out lies
- AI agents will lie just like human employees if not held accountable
Future Outlook
- Not a fad - dominant way to code going forward
- "If you're not doing it, you might be left behind"
- Exponential acceleration continuing
- Vibe coding here to stay, time to accelerate adoption
