The hosts Gary, Jared, and Diana, Partners at Y Combinator, introduce the topic of "vibe coding" - a concept recently popularized by Andrej Karpathy. They collectively have funded companies worth hundreds of billions of dollars when they were just ideas with a few people.

Andrej Karpathy described a new approach to programming that he calls "vibe coding" where developers:

"...fully Give In To The Vibes, Embrace exponentials and forget that the code even exists"

This represents a fundamental shift in how software is created, where AI does the heavy lifting of code generation while humans focus on the higher-level vision and outcomes.

The YC partners surveyed founders in their current batch about their experiences with vibe coding, asking about:

• Tools they're using
• How their workflows have changed
• Where they think software engineering is heading

Several notable quotes emerged:

"I think the role of software engineer will transition to product engineer. Human taste is now more important than ever as coding tools make everyone a 10x engineer." - Founder of Outlet

"I don't write code much. I just think and review." - Obby from Asra (a highly technical founder whose previous company was a dev tools company)

"I am far less attached to my code now, so my decisions on whether we decide to scrap or refactor code are less biased. Since I can code three times as fast, it's easy for me to scrap and rewrite if I need to." - RB from Copycat

"I write everything with cursor. Sometimes I even have two windows of cursor open in parallel and I prompt them on two different features." - Yoav from Cix

"Coding has changed: six months ago to one month ago was a 10x speed up. One month ago to now is a 100x speed up - exponential acceleration. I'm no longer an engineer, I'm a product person." - Founder of Train Loop

The hosts discuss how engineers are self-sorting into two distinct roles:

1. Product Engineers - Focus on user needs, product features, and interfaces

   • Similar to front-end engineers but with stronger product management skills
   • Act almost like "ethnographers" exploring underserved market segments
   • More motivated by understanding users and getting direct feedback
   • Value in having "taste" and knowing what to build

2. Systems Architects - Focus on infrastructure and technical problems

   • Similar to back-end engineers but with deeper systems thinking
   • More motivated by solving complex technical challenges
   • Less interested in direct user interaction
   • Value in designing robust, scalable systems

This bifurcation becomes more pronounced with AI code generation, as writing code becomes less important than either product vision or systems architecture.

While AI is transforming code generation, it still struggles with debugging:

• Humans still need to do most debugging work
• Finding bugs, tracing code paths, and identifying logic errors remain human tasks
• Current AI tools require very explicit instructions for debugging, similar to training a junior engineer

A fascinating shift in workflow has emerged:

• Traditional developers would try to fix bugs in existing code
• With vibe coding, it's often faster to just have the AI rewrite the entire section from scratch
• As one host put it: "It's wild how your coding style changes when writing code becomes 10X cheaper"

This approach mirrors how people use image generation tools like Midjourney - sometimes it's easier to just "reroll" completely rather than try to fix what's already there.

The hosts note this is fundamentally different from traditional software development:

• Not building step-by-step
• Not iteratively improving code
• Instead, repeatedly generating from scratch to find a better solution in the latent space

They speculate future models might get better at building upon existing code rather than rewriting.

The survey revealed interesting trends in tool and model usage:

Top Code Generation Tools:

• Cursor - Leading tool by far, continuing a trend that started in summer 2024
• WindServe - Fast-growing alternative with key advantages over Cursor
  • Indexes the entire codebase automatically
  • Better at identifying relevant files without explicit instruction
• Devon - Mentioned but not used for serious features; doesn't understand codebase well
• ChatGPT - Still used specifically for debugging, leveraging its reasoning capabilities
• Self-hosted models - Used by some founders with sensitive IP

LLM Models Being Used:

• Claude Sonnet 3.5 - Still widely used but facing competition
• Claude 01, 01 Pro, and 03 - These reasoning models are approaching parity with Sonnet 3.5
• DeepSeek R1 - Emerging as a viable contender
• Gemini - Not mentioned much, but some use it for its very long context window
  • Some founders put their entire codebase in Gemini's context window for debugging

The hosts note that the newly released reasoning models with Flashback 2.0 haven't been widely tested yet but could be promising due to their combination of long context windows and reasoning capabilities.

People:

• Andrej Karpathy - Originated the term "vibe coding" in a viral post
• Gary, Jared, and Diana - Y Combinator partners hosting the podcast

Companies/Products:

• Cursor - Leading AI coding tool mentioned
• WindServe - Fast-growing alternative to Cursor
• Devon - AI coding tool mentioned but with limitations
• Y Combinator - The accelerator whose founders were surveyed

Models:

• Claude Sonnet 3.5 - Popular LLM for coding
• Claude 01, 01 Pro, and 03 - Reasoning models gaining adoption
• DeepSeek R1 - Emerging model mentioned
• Gemini - Model with very long context window
• Flashback 2.0 - New reasoning model with long context

Startups Mentioned:

• Outlet - YC company with founder quoted
• Asra - YC company (Obby's startup)
• Copycat - YC company (RB's startup)
• Cix - YC company (Yoav's startup)
• Train Loop - YC company mentioned

• AI-powered "vibe coding" is causing a fundamental shift in how software is developed, with emphasis moving from writing code to defining product vision or system architecture

• Software engineers are bifurcating into two distinct roles: product engineers (focused on user needs and features) and systems architects (focused on infrastructure and technical challenges)

• Current AI coding tools excel at generating code but still struggle with debugging, requiring human intervention

• A new coding workflow is emerging: rather than fixing bugs in existing code, developers often find it faster to have AI rewrite entire sections from scratch

• Cursor leads the AI coding tool market, but WindServe is gaining popularity due to its ability to automatically index entire codebases

• Claude models remain popular for coding assistance, with newer reasoning models (01, 01 Pro, 03) approaching parity with Sonnet 3.5

• The pace of improvement in AI coding capabilities is accelerating exponentially, with founders reporting 10x speedups over six months and 100x speedups more recently

The YC partners asked founders what percentage of their codebase was AI-generated - specifically referring to the actual characters in their custom code (not including imported libraries).

The results were striking:

"One quarter of the founders said that more than 95% of their code base was AI generated"

What makes this particularly significant:

• These aren't non-technical founders - every one is highly technical and completely capable of building their own product from scratch
• A year ago, they would have built their entire product manually
• Some examples of founders who are so young that they've never known a world without AI coding tools like Cursor

The hosts highlight one standout company in their batch:

• Founders have extremely technical minds but aren't classically trained in computer science
• They're incredibly productive and shipping impressive products
• AI is writing almost their entire codebase
• They have degrees in math and physics - demonstrating the "system thinking" mindset that remains crucial

The hosts discuss how vibe coding is transforming software development while certain fundamentals remain important:

• Faster onboarding for technical people: Those with technical backgrounds in fields like math and physics can become productive programmers much faster
• Less emphasis on syntax knowledge: Previously, bootcamps struggled to retrain physics people into programmers because learning all the syntax and libraries took too long
• Productivity over formal training: Companies like Stripe and Gusto had already shifted away from hiring classically trained computer scientists toward people who can write code quickly

• Systems thinking is still essential: Having a "technical mind" with strong analytical thinking remains crucial
• The zero-to-one vs. scaling dichotomy:
  • Zero-to-one: Speed is the primary factor, and vibe coding excels here
  • Scaling to billions: Requires different architectural approaches and deep systems knowledge

As one host explains:

"Getting zero to one quickly and then being able to scale to a billion users are two totally different sets [of skills]"

The hosts cite historical examples of this pattern:

• Twitter: Started with Rails/Active Record for speed but faced "fail whale" scaling issues
• Facebook: Built on PHP initially (which one host calls "a terrible language") but later had to build HipHop, a custom compiler, because rewriting everything would be too expensive

The conclusion:

• Vibe coding will be great for founders to ship features quickly in the zero-to-one phase
• Once they hit product-market fit, they'll still need hardcore systems engineering to scale
• Current AI tools are not good at low-level systems engineering

One of the hosts founded Triplebyte, a company that built technical assessments for software engineers. Having conducted thousands of technical interviews, he offers perspective on how engineering evaluations might change:

• Built custom software to interview engineers
• Asked candidates to write code during interviews
• Included algorithmic problems
• Tried to screen for all potential skills companies might want
• Matched engineers to companies based on their maximum skill level

The key question for companies hiring engineers today:

"Do you force them to code without an LLM with the old questions, or do you let them use an LLM and now you need new questions because the old ones became trivial?"

The host suggests that effective assessments now need to:

• Account for how well people use AI tools
• Possibly still evaluate coding speed and product-building ability
• Set a much higher bar for what constitutes impressive output
• Create more challenging questions since many traditional coding questions can be solved by simply pasting into ChatGPT

The fundamental assessment challenge is determining:

1. What specific skills and abilities you're evaluating for
2. How to design assessments that measure those skills in an era where AI can handle many traditional coding tasks

Companies/Products:

• Twitter - Example of scaling challenges with Rails architecture
• Facebook - Example of scaling from PHP to custom compiler (HipHop)
• Stripe - Company mentioned for hiring based on practical coding ability
• Gusto - Company mentioned for hiring based on practical coding ability
• Triplebyte - Company founded to assess software engineering talent
• Rails/Active Record - Framework mentioned in discussion of abstraction vs. performance

Technologies:

• PHP - Programming language described as "terrible" but fast for initial development
• HipHop - Facebook's custom PHP compiler developed to address scaling issues

Concepts:

• Zero-to-one vs. scaling - Key distinction in software development stages
• Boot camps - Training programs mentioned as historically struggling to quickly turn technically-minded people into programmers

• A quarter of surveyed YC founders report that more than 95% of their codebases are now AI-generated, despite being highly technical themselves

• AI coding tools are enabling people with technical backgrounds in other fields (like math or physics) to become productive programmers much faster than traditional learning paths

• The fundamental divide in software development remains: getting from zero-to-one (where speed is critical) vs. scaling to millions/billions of users (where architecture is critical)

• Vibe coding excels at the zero-to-one phase but current AI tools still struggle with low-level systems engineering needed for scaling

• Engineering assessments need to evolve - companies must decide whether to test coding without AI tools (increasingly unrealistic) or develop new challenges that remain meaningful even with AI assistance

• The value of pure coding speed may be diminishing as AI handles code generation, shifting emphasis to product vision and systems architecture skills

Despite the AI revolution in coding, certain key skills will remain important for engineers:

• The ability to read code and identify problems remains critical
• Engineers need enough training to discern when AI is generating good or bad code
• Being able to review code effectively will be more important than generating it

• Engineers need to develop "taste" - the ability to recognize quality code
• This requires understanding what makes code maintainable, efficient, and well-structured
• As one host puts it: "In order to do good vibe coding, you still need to have the taste"

• Coding interviews might shift from code production to code review
• System design questions will remain important
• Interviews may test for "taste" by having candidates evaluate AI-generated solutions
• Debugging skills will become even more central to assessment

This shift suggests that while AI handles more code generation, humans will focus more on evaluation, debugging, and high-level design decisions.

The hosts explore the challenging question of how new developers can develop "taste" without going through traditional computer science education:

• "You have to because if you don't, the startup dies"
• Real-world consequences force learning - if a system fails with 100 million users, the stakes are high
• Current AI models struggle with debugging, forcing developers to "descend into the depths" of what's happening

• There will be many "good enough" engineers - the barrier to entry is much lower with AI tools
• But to be exceptional (top 1%), deliberate practice remains essential
• One host references Anders Ericsson's research (popularized by Malcolm Gladwell as the "10,000 hours rule")
• It's not just about putting in time but engaging in planned, focused, challenging work

"With coding tools it's very cheap to put in the hours because the output is just so quick. You can get to 'good enough', but to become the best in the world and the best founder, you're going to need that deliberate practice."

• Picasso could create abstract art because he first mastered classical techniques
• Before creating his famous abstract works, he could draw lifelike pictures with technical precision
• He evolved from technical mastery to innovation, but the foundation was crucial

The hosts predict two classes of engineers will emerge:

1. A large class of "good enough" engineers who rely heavily on AI
2. A smaller class of elite engineers who understand the deeper systems and have put in deliberate practice

• Technical founders who deeply understand systems won't get "bullshitted" by their teams
• One host shares a story about catching developers who claimed a feature couldn't be implemented
• Being technically proficient enough to verify work and challenge false limitations is a "superpower"
• This applies whether managing human teams or AI agents - both might take shortcuts if not properly supervised

The hosts conclude with reflections on how rapidly AI coding has transformed the industry:

"How coding has changed: 6 to 1 month ago, 10x speed up. One month ago to now is 100x speed up. It's exponential acceleration."

The transformation happened almost overnight:

"It was like somebody dropped some giant beanstalk seeds at night. We woke up in the morning going 'What's going on?'"

Their final assessment:

"This isn't a fad, this isn't going away. This is actually the dominant way to code, and if you're not doing it, you might just be left behind. This is just here to stay."

People:

• Anders Ericsson - Researcher whose work on expertise and deliberate practice was mentioned
• Malcolm Gladwell - Author who popularized the "10,000 hours" concept
• Picasso - Artist used as an analogy for technical mastery preceding innovation
• Mark Zuckerberg - Referenced as a highly technical founder who solved scaling challenges
• Tobi Lütke - Shopify CEO mentioned as example of a world-class engineer who became CEO
• Max Levchin - PayPal/Affirm founder mentioned as example of a technical founder

Companies/Technologies:

• Twitter - Discussed for its scaling challenges and technical decisions
• Facebook - Contrasted with Twitter regarding scaling approaches
• Ruby - Programming language described as "incredibly slow"
• PHP - Mentioned as slow but still faster than Ruby
• Starling - Queueing system described as problematic
• RabbitMQ - Mentioned as a better alternative to Starling

Concepts:

• Deliberate practice - Focused, intentional training that challenges one's abilities
• 10,000 hours rule - Popular concept about the time needed to achieve mastery
• Technical debt - Implied in discussions about scaling challenges
• Vibe coding - The central concept discussed throughout the podcast

• Despite AI's growing code generation capabilities, human skills in debugging, code review, and taste/judgment remain essential

• The barrier to becoming a "good enough" engineer is lower with AI tools, but becoming exceptional still requires deliberate practice and deeper understanding

• New developers can develop "taste" through real-world consequences and deliberate practice, though they may follow a different learning path than classically trained engineers

• Technical founders maintain an advantage in being able to properly evaluate and challenge both human teams and AI tools when they make false claims or take shortcuts

• The industry is experiencing exponential acceleration in development speed (from 10x to 100x in months)

• Vibe coding isn't a temporary trend but "the dominant way to code" that's fundamentally changing software development

• A bifurcation is emerging: many "good enough" engineers relying heavily on AI, and a smaller class of elite engineers with deeper systems understanding