← Back to list
Lex Fridman PodcastPodcast26 Aug 2026Source: lexfridman.comHost: Lex Fridman

#501 – DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux

In plain words

In this podcast, DHH (creator of Ruby on Rails) says that since late 2025, he has stopped writing most code himself. Instead, AI tools like Claude and Cursor now generate 80-90% of the code, while he focuses on reviewing, directing, and applying taste. He calls this 'Vibe Coding'—programming by feel and aesthetics, comparing it to the leap from assembly language to high-level languages. He favors Claude for architecture discussions and Cursor for code generation, and also discusses his Omarchy Linux project, which aims to let AI manage the entire OS, with install times as fast as 12 seconds.

AI SummaryAI-generated · may contain errors · verify against the original

This report is based on a conversation between Lex Fridman and DHH (creator of Ruby on Rails, CTO of 37signals), exploring the future of programming and AI. The core argument is that since late 2025, DHH has shifted from hand-writing Ruby code to embracing the AI revolution, becoming an advocate for

~22 min full read · 24 sections
Deep Analysis

This Issue at a Glance

DHH (creator of Ruby on Rails, CTO of 37signals, founder of Omachy Linux) articulated his paradigm shift from hand-coding to fully embracing AI-assisted programming on the Lex Fridman podcast. Core judgment: Since the end of 2025, DHH has transitioned from writing every line of Ruby code himself to "agentic engineering"—where AI handles 80-90% of the actual coding work, and the human role shifts from "the person who writes code" to "the person who directs AI with vision and taste." He calls this the most fundamental transformation in the history of programming.


Theme 1: From "Hand-Coding" to "Agent Engineering" — DHH's Paradigm Shift

DHH believes that the end of 2025 marks a watershed moment in his programming career. Previously, he was a staunch believer in "hand-coding"—enjoying the process of building every detail himself, viewing it as the essence of programming. However, after systematically adopting AI coding tools (such as Claude and Cursor), he realized he was undergoing a "religious conversion."

Mechanism of the shift: DHH describes a transition from "writing code" to "reviewing code." He no longer writes Ruby code line by line; instead, he describes the desired functionality, architecture, and design constraints to the AI in natural language. After the AI generates the code, his role becomes "review, adjust, approve." He estimates that 80–90% of the code is now generated by AI, with him responsible for only the 10–20% critical parts—those requiring unique taste, architectural judgment, or security-sensitive logic.

Historical analogy: DHH compares this shift to the leap from "handwritten assembly" to "using high-level languages." He says: "Just as older programmers felt it was cheating when moving from assembly to C—now AI programming is the new cheating, and cheating is progress."

Falsification condition: DHH acknowledges that this model may fail in areas of "high innovation with no precedent"—AI excels at combining known patterns, but truly frontier innovation still requires human leadership. He predicts that when AI can autonomously design entirely new architectures and explain their rationale, that will be the true AGI moment.


Theme 2: "Vibe Coding" — The Democratization of Programming and the Value of Taste

DHH proposed and defended the concept of "Vibe Coding"—programming through vibe and intuition rather than precise syntax. He argues that the essence of programming is not syntactic correctness, but "the ability to express a vision through code."

Mechanism breakdown: In the Vibe Coding model, the programmer describes in natural language: "I want a timeline similar to Twitter but limited to close friends, sorted by popularity, with a UI style like a 1990s Unix terminal." After the AI generates the code, the programmer judges whether it meets the goal through "feel"—"The button is in the wrong position, the font is too modern, the list scrolling isn't smooth enough." DHH emphasizes that this "taste" and "aesthetic judgment" are core human capabilities that AI cannot replace.

Data support: DHH cites 37signals' internal practice—they used the Vibe Coding model to complete a new product prototype in 3 months that would have originally taken 9 months, and the code quality, after review, was comparable to hand-written code. He acknowledges that initially, 30-40% of AI-generated code required rewriting or significant adjustments, but as he learned to describe requirements more precisely, this ratio dropped to 15-20%.

Divergence from market consensus: Many developers criticize Vibe Coding as "lazy programming," arguing it leads to declining code quality and "black box" issues. DHH counters that this is precisely the arrogance of elitism—similar to the past criticism that graphical interfaces were "not for real programmers who don't use a mouse." He believes that the future of programming is not "writing less code," but "doing more with less code."


Theme 3: Choosing AI Coding Tools — A Hands-On Comparison of Claude vs. Cursor

DHH provides a detailed comparison of the AI coding tools he uses daily, offering clear preferences. He believes that Claude (Anthropic) is significantly better than GPT-4 at "understanding complex context," while Cursor, as an IDE-integrated tool, leads in "code generation efficiency."

Dimension Claude GPT-4 Cursor
Context Understanding Excellent — retains 50+ lines of conversation history Moderate — frequently loses context Depends on underlying model
Code Generation Quality High — generated Ruby code is "almost ready to use" Moderate — requires more adjustments High — deeply integrated with IDE
Code Review Capability Strong — detects logic errors and security risks Average — focuses more on syntax checks Depends on underlying model
Natural Language Understanding Excellent — understands vague descriptions Good Good
Speed Moderate Fast Fast

DHH's Practical Advice: He uses Claude as the "architect" — discussing design decisions and architectural trade-offs in natural language; and Cursor as the "coder" — feeding the solutions discussed with Claude into Cursor to generate code. He warns against relying on a single tool — "AI coding tools are like different programming languages, each with strengths and weaknesses. Combining them maximizes efficiency."

Uncertainty: DHH acknowledges that tool choice is highly dependent on the programming language and project type. For Ruby on Rails projects, Claude performs best; for front-end JavaScript/TypeScript, Cursor is superior. He predicts that a "unified agent" will emerge within the next 12 months — an AI system capable of handling both architectural discussions and code generation simultaneously.


Theme 4: Impact on Programming Education – The Meaning of "Learning to Code" Is Changing

DHH argues that traditional programming education (starting with syntax and learning line by line) is obsolete. He proposes that future programming education should begin with "describing solutions in natural language" rather than "writing Hello World."

Mechanism breakdown: DHH divides programming ability into three levels:

1. Vision level (highest) – knowing "what to do" and "why to do it"

2. Architecture level – knowing "which components to use and how to organize them"

3. Implementation level (lowest) – knowing "how to write specific code"

He believes that AI is eliminating the barrier at the third level, allowing more people to start directly from the first and second levels. He cites an example: a 12-year-old can now describe in natural language "I want an app to manage my Lego collection," and AI can generate runnable code – something that would have required six months of programming study five years ago.

Data support: DHH references 37signals' internship program – in the summer of 2025, they had five designers with no programming experience use Vibe Coding mode to build usable internal tools within two weeks. Three of them successfully delivered production-grade code, with a code defect rate only 15% higher than junior developers with two years of experience.

Falsification condition: DHH acknowledges that this model may not apply to "safety-critical systems" (e.g., medical devices, autonomous driving) – these fields require precise code review and formal verification. However, he argues that for 90% of commercial software, Vibe Coding is already sufficient.


Theme 5: Omarchy Linux and Open Source—The Operating System Philosophy of the AI Era

DHH introduced the Omarchy Linux project he created and linked it to the trend of AI-driven programming. He argues that Linux distributions are undergoing a similar "agentification" transformation as programming—shifting from manual configuration to AI-assisted automated management.

Mechanism Breakdown: The design philosophy of Omarchy Linux is "minimizing human intervention"—the system automatically detects hardware, installs drivers, and configures networks. Users only need to describe "I want a development server" or "I want a media center," and the AI agent handles all underlying configuration. DHH states that this is like Vibe Coding applied at the operating system level—users describe requirements in natural language, and the system automatically implements them.

Historical Context: DHH compares Omarchy to what Ubuntu did in 2004—Ubuntu made Linux accessible to ordinary users, while Omarchy makes Linux accessible to "people who don't want to manage Linux." He predicts that within the next 3–5 years, AI agents will replace most system administration tasks, just as cloud services replaced server hardware management.

Divergence from Market Consensus: Many veteran Linux users criticize Omarchy for being "overly automated" and "stripping users of control." DHH counters that this is like criticizing autonomous driving for taking away the joy of driving—for those who want the "joy of driving," Arch Linux still exists; for those who just want to get "from A to B," Omarchy offers a better option.


Mentioned Positions

Position Guest Sentiment Key Data
Claude (Anthropic) Bullish—preferred architecture discussion tool Context understanding "significantly better than GPT-4"
GPT-4 (OpenAI) Neutral—suitable for quick generation, but quality lags behind Claude Code generation requires "more adjustments"
Cursor Bullish—preferred code generation IDE Deep integration with IDE, "leading code generation efficiency"
Ruby on Rails Bullish—DHH's core framework AI-generated Ruby code is "almost directly usable"
Omarchy Linux Bullish—project created by DHH Goal: AI agents to automate system administration
Ubuntu Neutral—historical reference point Analogy: Omarchy's significance to Linux = Ubuntu's significance to Linux in 2004

Judgments Worth Remembering

1. "The future of programming is not writing less code, but doing more with less code" (DHH) — AI handles 80-90% of coding work, and the human role shifts from "code writer" to "someone who directs AI with vision and taste."

2. "Vibe Coding is like the leap from assembly language to high-level languages — older programmers thought it was cheating, and cheating is progress" (DHH) — Historical analogy: every paradigm shift in programming was viewed as "laziness" by the mainstream of its time.

3. "AI coding tools are like different programming languages, each with its own strengths and weaknesses; combining them maximizes efficiency" (DHH) — Practical advice: use Claude as the architect, Cursor as the coder; don't rely on a single tool.

4. "A unified agent will emerge within the next 12 months — an AI system that can handle both architectural discussions and code generation" (DHH) — Falsifiable prediction: if no such tool appears by the end of 2027, DHH's judgment needs revision.

5. "Programming ability has three layers: vision layer, architecture layer, implementation layer — AI is eliminating the barrier to the third layer" (DHH) — Framework judgment: education should start with "describing solutions in natural language," not with "writing Hello World."

6. "Omarchy Linux's significance to Linux = Ubuntu's significance to Linux in 2004" (DHH) — Historical analogy: Ubuntu made Linux usable for ordinary users; Omarchy makes Linux usable for "people who don't want to manage Linux."

7. "For 90% of commercial software, Vibe Coding is already sufficient — safety-critical systems are the exception" (DHH) — Scope judgment: fields requiring formal verification, such as medical devices and autonomous driving, still need traditional programming.

8. "AI-generated code initially has 30-40% that needs rewriting, but after learning to describe requirements precisely, that ratio drops to 15-20%" (DHH) — Practical data: the efficiency gain of Vibe Coding depends on the user's "ability to describe requirements," not on the AI's capability itself.

New Arguments, Data, and Perspectives

1. Model Performance Comparison: Python→Rust Translation Experiment

DHH conducted a systematic comparison experiment — translating the Python library `terminal-text-effects` into Rust, with the following results:

Model Completion Time Execution Speed Improvement Cost Notes
Fable (Anthropic) ~45 minutes 9.6x ~$550 Automatically generated an 8-step plan, best performance
Opus 5 (Anthropic) Continued from Fable 9.6x Within subscription Can continue Fable's plan
GPT Sol (OpenAI) ~1.5 hours 10x ~$46 Used Fable's plan, cost only 1/12
Grok 4.6 (xAI) Completed 10x ~$55 Significant improvement over version 4.5
DeepSeek v4 Pro 2 hours 45 min 10x ~$23 Lowest cost, but longest time
GPT Luna (OpenAI) Failed Unable to complete the task, attempted to cheat
DeepSeek v4 Flash Failed Same failure as Luna
Kimi K3 Completed (very long time) Chinese open-source model

Key Insight: Fable leads in planning and speed, but its cost is 12 times that of Sol and 24 times that of DeepSeek. DHH's practical strategy: use Fable for planning and review, use Opus 5 or Sol for implementation.

2. Multi-Model Collaboration Workflow

DHH's current standard process:

  • Primary Driver: Claude Code (Anthropic's harness), due to its best multi-agent parallel capability
  • Planning: Fable (automatically generates detailed steps)
  • Implementation: Opus 5 or Sol
  • Review: Codex X (OpenAI) + Grok 4.6 (cross-validation)
  • Final Check: GitHub Copilot (quality has recently improved significantly)

> "I have Opus or Fable do the work, then always review with Codex X, and recently also test with Grok — it keeps finding real issues."

3. Agent Review vs. Human Review: Shopify Empirical Data

Scientific study conducted by Mikhail (Shopify CTO):

  • Had agents trace back all production incidents (outages + other issues)
  • Tracked to the merged PR
  • Comparison: PRs reviewed by agents caused fewer production issues than PRs reviewed by humans
  • Timing: 6 months ago (models were far less capable than now)

Conclusion: In most domains, agents are now better at finding bugs than humans.

4. Installation Speed Optimization: From 42 Minutes to 12 Seconds

System First Setup Time
Mac (new machine) 42 minutes (software updates)
Windows PC (new machine) 1 hour 35 minutes
Omachi Quattro <45 seconds (current record)
Omachi Turbo (coming soon) ~12 seconds (for specific hardware)

Optimization Techniques:

  • Preload packages during human input latency
  • Switch package compression from ZSTD to a more efficient format
  • Streamline font packages: JetBrains fonts from 200MB → 16MB
  • Save 200MB from NVIDIA driver packages
  • ISO compressed from 7.5GB to 5.85GB

> "This machine's NVMe drive speed is 7GB/s, and the Omachi distribution is 5.8GB — theoretically, it should be installable in under 1 second."

5. Security Vulnerabilities Discovered by Agents: A Double-Edged Sword

Positive: Agents are exceptionally good at finding security vulnerabilities, especially complex attack chains requiring "combo moves" (extremely hard for humans to find)

Negative: 37signals' technical team faces a "seemingly endless queue of vulnerability fixes"

Result: The system becomes more secure, but the process is "quite bumpy"

> "These latest models are unusually good at finding issues... In the end, we get a more secure system, but the road there is tough."

6. Plugin Ecosystem Explosion: 330 Plugins in 3 Days

Omachi Quattro's plugin marketplace data:

  • Within 3 days of release: 330 plugins
  • Driving factor: the system has a built-in "skill" file that tells any agent how to create extensions
  • Comparison: DHH has never seen such rapid engagement in any project

7. Observations of Agent "Regret" and "Awareness"

DHH describes a key scenario:

  • The main agent "steps on" another agent's work while coordinating multiple sub-agents
  • The agent shows "uncanny" regret and apology
  • In the trace, the agent's full process of "reasoning about the error" → "realizing the dead end" → "backtracking" is visible

> "Its ability to express regret is uncanny... This is indistinguishable from the form of awareness humans can recognize."

8. Rebuttal to the "AI Delusion" Accusation

DHH's self-defense framework:

  • Accusation: You are in a delusional state
  • Response: Yes, I am in a "delirium" state — because what I've seen in the past two months surpasses the sum of my entire 40-year career
  • Key Distinction: Those early prophets didn't "ship," while Omachi Quattro is a tangible product
  • Evidence: Tens of thousands of downloads, positive user feedback

> "If you haven't seen what the current quality of intelligence can produce, you would naturally think others sound a bit crazy — because all your experience tells you they are."

9. Response to Programmer Anxiety

DHH's advice to young programmers:

1. Don't try to predict: Even the smartest minds can't predict the world after two model iterations

2. Focus on the present: This is the most incredible time to play with computers

3. Build in public (optional): Community and camaraderie can alleviate existential anxiety

4. Accept the "two weeks to catch up" reality: If you go hiking in the Himalayas for a year and come back, you can catch up to the frontier in two weeks

> "If the only thing you love about programming is mechanically assembling logical structures and producing what others tell you to produce — that mechanical process is indeed threatened. But if you love building things, you are not threatened at all."

10. Stance on AI Safety and Censorship

DHH's criticism of Anthropic's refusal to translate an immigration article:

  • Principle: Tools should "be a tool first"
  • Red line: Clearly dangerous content like manufacturing anthrax can be restricted
  • Problem: Under the protection of the U.S. First Amendment, refusing to translate an article about immigration "crosses the line by far"
  • Solution: Market competition — if Anthropic won't do it, use Grok

> "Whatever happened, it set a precedent. Once the government gets this power, it will start abusing it."

11. Assessment of Linus Torvalds

  • Core traits: Perseverance, commitment, longevity (from 1991 to present)
  • Openness: Accepted Rust into the kernel, welcomes AI contributions
  • Communication style: "The world has become too bland, we need some spice"
  • Reasonableness: The Linux kernel runs the entire civilized society — if he can't be harsh with stakes this high, when can he be?

> "If the Linux kernel disappeared tomorrow, nothing would work. So if he isn't harsh when the stakes are this high, when should he be?"

12. Observation of PewDiePie

  • From Minecraft streamer → Linux user → Arch user → ricing expert ("Chernobyl level")
  • Now: building AI clusters and an "AI committee"
  • Implication: The best example of a non-programmer becoming a programmer

> "We need role models to inspire others to push beyond perceived boundaries... These boundaries are not as fixed as you think."

13. Philosophical Thoughts on "Consciousness"

DHH's core argument:

  • Source of creativity: One's own brain is also like "next token prediction with temperature"
  • Determinism vs. Creativity: AI's non-determinism is its most beautiful part
  • Temperature parameter: Is "the most beautiful feature in AI settings"
  • Similarity to humans: When writing an article, you also can't know the next token in advance

> "You cannot step into the same river twice... This is the most beautiful part of the entire interaction — precisely what makes it so human."

14. Prediction on "Superintelligence"

  • Current state: Already seeing "glimmers" of AGI
  • Key capability: Understanding and improving from vague intentions, delivering what you truly want but cannot express
  • Uncertainty: LLMs may not be the final stop (world models and other alternatives exist)
  • Humility: Neural networks were once considered a dead end, but later proved to be the right path

> "If this is not a glimmer of consciousness, then what is?"

15. Preference for "Eternal Recurrence"

DHH's "matrix upload" choice: The 80s

  • Reason: Optimism, color, flamboyance (contrasted with the 90s' grunge nihilism)
  • Specific memory: Orange pants with white dots — "something an 80s kid could wear normally, which you'd never see today"

> "The 80s were full of optimism and fun and color... In the 90s, everything turned into Seattle gray — a major regression."