AI summary by TwinMind
My Thoughts On "Vibe Coding" (And Prime)
Theo - t3․gg
2h 55m 15s
Opening Discussion on YC's "Vibe Coding" Claims
- YC claims "Vibe coding is the future" - developers fully embracing AI without looking at code
- Theo disagrees with Gary's "you'll be left behind" argument
- References "it's good to be late" philosophy
- Got into React 5-6 years after release, still successful
- Being late often better than being early (avoids learning wrong patterns)
- Prime agrees - doesn't believe in FOMO around new tools
- Mobile development analogy: early adopters learned wrong platforms (Java applets) vs later adopters who learned iOS/Android
Survey Results from YC Current Batch
- 25% of founders claim 95%+ of codebase is AI-generated
- Most popular tools: Cursor, Claude 3.5 Sonnet, some O1/O3 usage
- Shift from classical engineering to "product engineering" with human taste as differentiator
- Founders describe 10x to 100x speed improvements
Problems with AI-Generated Code Quality
- Game development particularly problematic due to lack of good open-source training data
- Very few quality open-source games vs web applications
- AI generates "mirrors" of game dev patterns without proper implementation
- Debugging remains major weakness of current AI tools
- Models can't effectively identify and fix their own bugs
- Often easier to regenerate code than debug existing issues
The "Vibe Coding" Phenomenon
- Andre Karpathy's definition: "fully give in to the vibes, embrace exponentials, forget code exists"
- Creates non-deterministic development process (like "RNG programming")
- Theo distinguishes between "vibe of code" (code quality intuition) vs "vibing out while coding"
- Prime concerned about loss of deterministic software development principles
Model Comparisons and Tool Evolution
- Claude 3.5 Sonnet: Best for CSS/frontend, good "vibes" but poor debugging
- O3 Mini: Better reasoning/debugging but slower, more clinical tone
- Cursor vs Windsurf debate:
- Theo sees Windsurf users as mostly beginners
- Cursor dominates among experienced developers (35% in poll vs vanilla VS Code)
Impact on Developer Skills and Hiring
- Concern about "AI coding natives" who never learned fundamentals
- Debate over whether classical training still matters
- Hiring advice: Don't hire for skills you don't understand/appreciate
- Theo's rule: hire for work you'll "miss doing" not work you want to avoid
- Need to match excitement level of employees to retain talent
Scaling and Architecture Concerns
- Serverless revolution changes scaling dynamics vs old Rails/Twitter fail whale problems
- Simplicity scales better than complex solutions
- Companies often over-hire and over-engineer too early after funding
- "Yacht problems" - scaling issues are luxury problems most don't reach
The Future of Engineering Roles
- Split emerging between product engineers (user-focused) and systems engineers (infrastructure-focused)
- Taste and debugging skills become more critical as AI handles syntax
- Physics/math backgrounds may translate better to programming via AI tools
- Need deliberate practice to reach top 1% performance levels
Concerns About Non-Technical Founders
- 95% AI-generated codebases create maintenance nightmares for future engineering hires
- Founders who can't evaluate code quality will struggle to hire/retain good engineers
- "Bozo explosion" risk when non-technical leaders can't identify competence
- Luck vs skill in startup success - resilience more important than timing
Workplace Culture and Trust Issues
- Discussion of "fake nice" people in tech who are toxic behind scenes
- Kent Dodds cited as genuinely positive example
- Importance of consistency between public and private behavior
- Prime and Theo prefer working with people who are authentic even if occasionally abrasive
Technical Depth Still Matters
- Examples of technical founders (Max Levchin, Toby Lutke) who became successful CEOs
- Being technical prevents getting "bullshitted" by engineering teams
- Database indexing anecdote - simple technical knowledge saving projects
- Need to understand failure points and scaling bottlenecks
Closing Thoughts
- Interest in hearing from YC companies using <20% AI-generated code
- What problems are so complex that AI tools actually hurt productivity?
- Balance between embracing new tools and maintaining engineering fundamentals
- Video ran nearly 3 hours vs planned 1.5 hours
