Juraj Masar

Some people think venture capital is a sort of Ponzi scheme. This is because VCs report great returns during the first two years of a fund, which is when they’re trying to raise *another* fund. After they raise the new fund, the reported performance of the previous fund collapses.

21

11 Comments

At some point in the past 18 months, it became trendy to say that software engineering was dead. Why learn to code when you can tell an LLM, “make me an app with a login screen and a database” and out pops code? Voilà. Startup in a weekend. Series A by Thursday. Thus was born vibe coding - the art of building software by manifesting it. Mix 1 part natural language, 1 part vague ambition, and 1 part blind confidence as you paste mystery code into production. For a brief, shimmering moment, it almost felt like the future. Until it didn’t. AI code assistants unlocked a wave of creative energy. Non-technical founders spun up MVPs. Engineers offloaded boilerplate. Students built full apps in a weekend. But prototype-grade code isn’t production-grade code. Many teams that began with LLM scaffolds are now spending weeks refactoring. Some have even declared “code bankruptcy.” Because there’s a difference between writing code and building software. The former is syntax. The latter is systems thinking. At some point, every serious technical team has the same realization: you don’t just need code - you need engineering. Vibe coding isn’t a tech failure, it’s a categorization error. It assumes that the problem in software development is generation speed. But for any company past the napkin stage, that’s not the bottleneck. It is: - Understanding and reflecting business logic - Architecting clean, extensible code - Managing state, latency, auth, concurrency, observability - Reasoning through edge cases and failure modes LLMs don’t reason through trade-offs or hold long-term context. They don’t ask, “Why does this route even exist?” So when teams use LLMs to generate full features - or worse, entire codebases - they end up with systems that appear functional but are structurally hollow. Like a house with beautiful wallpaper but no load-bearing walls. There’s a market for vibe-coding. It’s just not software. This is the real distinction: vibe coding and AI in software development are not the same thing. - Vibe coding tries to replace engineering. Hand the keys to the model, hope for the best. - AI in software development amplifies engineering. Accelerate rote work while owning architecture, logic, and trade-offs. The first treats AI as a substitute; the second treats it as a lever. Vibe coding is fantastic 0 → 1. It’s a liability 1 → 100. It’s like bringing a balloon animal to a knife fight. Wonderful at a birthday party. Less helpful in real combat. There’s a real market for fast, disposable, front-of-the-house code. But most tech companies are in the business of building the kitchen, not just plating food. The panic about “engineering being dead” comes from people who don't understand it. Engineering isn’t syntax. It’s constraint management, abstraction design, knowing when to optimize and when to punt. Ironically, as AI makes building easier, the value of engineering judgment goes up. The faster you can go, the more you need someone to steer.

4,000

370 Comments

Are any founders AI/agent maxxing their ops stack so that they can prevent being frame mogged by YC founders working on similar things??? If so comment below. Looking to deploy a lot of capital into any AI infra projects especially that dominate distribu channels. Clavicular is an LP in this fund.

15

4 Comments

The big labs are expansive in their product ambition, especially since foundation models have largely improved in lockstep - in order to compete with them you have to do things they won’t which are: - building a very rich software ecosystem around a primitive - orchestration across multiple models - going insanely deep on product and growth for a narrow vertical domain

58

3 Comments

SaaS monetizes features. Rails monetize mistakes. The main objection to Rails: "Isn't this just network effects?" No. Rails don't create network effects. They create learning effects. Every time work flows through the rail, it doesn't only deliver an outcome, it makes the rail smarter. SaaS scaled revenue. Consulting scaled judgment. Rails scale error correction. While LegacyCos bolt AI agents onto existing workflows, NewCos build differently. They know you can't buy five years of resolved edge cases. That compound learning becomes the moat. Learning Effects > Network Effects Classic SaaS defensibility came from network effects. Slack gains utility with more teammates. But SaaS compounding is bounded. Each instance learns locally. Operating Rails compound differently. Every GitHub Copilot rejection, every Stripe fraud flag, every CrowdStrike attack blocked makes the entire network smarter. Expensify example: When you correct a miscategorized expense, it only improves your workspace, despite 15 million users generating errors daily. Imagine Expensify as a Rail: every correction across thousands of companies teaching the entire network. Your operating costs benchmarked against similar companies. Edge cases from one company preventing errors at another. This isn't about user count. It's about work count. Every pattern resolved becomes reusable logic with provenance. Rails monetize mistakes across the network. SaaS monetizes features in isolation. The Three Laws of Operating Rails Law 1: Learn faster than they copy  Your rail must improve faster than competitors can imitate. A competitor can copy features in weeks. They can't compress five years of edge cases resolved, versioned, and rolled back. The moat isn't what you built; it's how fast you compound. Law 2: Make it tacit, not portable  Perfect documentation is easily copyable. The real moat lives in tacit knowledge, patterns that only work with your specific context, governance, and audit trails. Law 3: Power without transparency kills trust  As rails automate more decisions, they need more governance. One bad auto-execution can destroy years of trust. Constitutions, vetoes, one-click rollbacks aren't nice-to-haves. They're essential. NewCo Playbook: Start Boring, Compound Fast (Exclusive to newsletter subscribers) The winners won't be the ones who automate fastest. They'll be the ones who learn from failure fastest. LegacyCos are adding AI agents to workflows, hoping for magic. But intelligence in local instances doesn't compound. You don't build moats by making each silo smarter. NewCos who grasp Operating Rail laws will target boring shared burdens. They'll turn every error into network intelligence. They'll shoulder bounded liability to earn trust. And they'll compound learning at rates LegacyCo can't match. Strong operators compound errors into moats. Weak operators add features and hope. (Full version sent to subscribers)

55

24 Comments

Cursor hit $500M ARR. Lovable went from $17M to $75M in five months. On paper, AI coding tools look unstoppable. Then you check the traffic data. Lovable is down 40%. Vercel's v0 dropped 64% since May. Bolt.new fell 27% since June. When revenue climbs while usage falls, that's not growth. That's a warning sign. Three forces are working against generic AI coding tools: → Developer sentiment is cooling. Stack Overflow shows usage up from 70% to 76%, but favorability dropped from 77% to 72%. People are using the tools more and liking them less. → Enterprises are hitting compliance walls. FINRA, GDPR, HIPAA. Many tools weren't built for enterprise procurement. → Switching costs are too low. When every tool feels the same, users bounce. There's no moat in being "pretty good at everything." Meanwhile, vertical AI is scaling fast. The market sits at $12.9B in 2024, projected to hit $115.4B by 2034. Big tech is buying specialized applications, not building generic copilots. ServiceNow paid $2.85B for Moveworks. M&A hit 177 deals in Q2 2025, nearly double the quarterly average. We're in Wave 2.0. Infrastructure is maturing. Value creation is moving to the application layer. CB Insights tracked 1,500+ tech markets. The most AI deals in Q2 went to companies solving specific industry problems. LLM developers tied for ninth place. The tools that win won't be the loudest. They'll be the ones that become essential. Full piece in Forbes (link in comments)

396

103 Comments