Theo: What we lose and gain when AI writes the code

Theo Browne mourns the hands-on craft while arguing that AI rewards curiosity and lets more people build.

TBTheo BrowneSpeaker · Ping Labs
Channel
Theo - t3․gg
Published
Nov 19, 2026
Video
28:16
Read
6 min
Original
English
Open on YouTube
Scroll to read and the video shrinks to the bottom-right corner. Captions follow your reading language.
Reading languageAvailable in 2 languages
You're reading the original English.
SUMMARY

In 30 seconds

Theo Browne says Claude writes more than 90% of his code, and he expects hand-written production code to become a niche craft. Reacting to an essay by engineer Lauren Tan, he acknowledges the loss of solving a difficult bug by hand while celebrating how AI lets curious people learn and build without nearby experts. He also argues that the label “vibe coder” and gender can cause people to discount an engineer’s capabilities, pointing to Tan’s React Compiler work.

  1. 1Claude writes most of Browne’s codeHe says the share is above 90% and believes very little code still needs to be typed entirely by a human.
  2. 2Web developers are accustomed to rapid changeBrowne points to the succession of web frameworks as experience that may have prepared some engineers for AI tools.
  3. 3Solving a hard bug by hand gave a feeling he missesHe recalls working late until a different perspective made it possible to remove half the code and solve the problem.
  4. 4AI can help a curious learner go deeper without a nearby expertBrowne compares AI to search and describes a student exploring the Linux kernel despite limited access to mentors.
  5. 5Automating parts of building leaves room to try more ideasHe says projects he had contemplated for a decade are now built and that he is exploring more ambitious ideas.
  6. 6A “vibe coder” label can obscure other abilitiesBrowne says social-media audiences can reduce people to one role, and argues that women engineers face additional skepticism.

Takeaways

  1. Browne says AI writes more than 90% of his own code; he expects manual production coding to become uncommon.
  2. He urges experienced developers to recognize the real pleasure and effort behind hand-written code while adapting to change.
  3. Use AI to follow questions and experiment with projects that previously demanded skills or people beyond reach, Browne suggests.
  4. Browne argues that labels such as “vibe coder” can obscure an engineer’s existing skills.
FULL TEXT

Read the edited text by chapter

Transcript only
01

AI writes most of Browne’s code

Browne says the work of typing production code has changed dramatically. He introduces Lauren Tan’s essay as both a tribute to manual coding and an acceptance of its changing role.

A year ago, I was still writing substantial amounts of code myself. Eighteen months ago, I did large migrations by hand with occasional AI autocomplete. Now Claude writes more than 90% of my code, and I think the familiar act of typing production code has largely been automated.

There are still occasions when an AI gets something wrong and you can correct it by hand or ask it to fix the problem. But in my own work, very little needs to be written entirely by a human.

Engineer Lauren Tan, whose work on React has long inspired me, has written an essay that is both a tribute to the coding that brought her here and an acceptance of what has changed. I want to respond to her reflections.

02

React Hooks and the excitement of changing tools

Browne recalls learning from software-design debates and discovering React Hooks. He sees rapid change as a familiar condition for web developers.

Tan remembers programming as a hobby: the pleasure of discovering a feature, making things, and considering every line for elegance and performance. I relate, even though I have tended to care more about how the result works than the form of the code.

A Brian Will video criticizing object-oriented programming helped change how I thought about software and showed me the kind of discussion I wanted to see on YouTube. I returned to it for years and eventually made my own response.

I also remember the React Hooks presentation by Dan Abramov and Ryan Florence. At the time, sharing and synchronizing state between components frustrated me. Hooks gave me a way to express reusable state and made the code easier for me to understand.

That interest in frontend composition and application architecture eventually led into the T3 stack and my channel. I was so excited by new patterns that I needed more places to talk about them than conversations with friends.

Looking back at old videos comparing two versions of a pattern feels nostalgic. I understand why people feel a loss. I happen to be excited more by where things are going, and I encourage people to find enjoyment in the process of change.

Many people working on AI development tools have backgrounds in React and TypeScript. I think web developers’ experience with frequent changes in frameworks helped some of them adapt. The pace could be frustrating, but changing tools was familiar.

Think of the succession of frontend tools: jQuery, Knockout, Ember, Angular, React and many newer options. The list and the debates over which tool belongs on it illustrate how much the ecosystem has shifted.

03

Mourning the hands-on craft

A comparison with handmade shoes captures the shift toward automation. Browne recalls the satisfaction of solving a hard bug late at night and argues that those grieving deserve empathy.

Tan cites a comparison between hand-written code and handmade shoes. Artisans still make shoes, but most people wear manufactured ones. I think hand-coding will similarly remain a niche craft rather than the ordinary way production software is made.

Tan says our tools depend on the discoveries and work of people who came before us. Those who miss the direct, tactile experience of solving problems deserve empathy rather than shame.

I remember struggling with a bug or architectural problem late at night for hours, then suddenly seeing a different approach: perhaps half the code could be deleted. You make a burst of changes, look up at the clock and realize it is four in the morning, but the problem is solved.

I miss that feeling too, though I have found other satisfying activities. People who built kernels, languages and frameworks created the foundations on which today’s tools run. Their work deserves recognition even as AI generates more new code.

I draw a line when people attack other developers for using AI without recognizing their prior work. Tan helped build the React Compiler; dismissing her as merely a “vibe coder” misses what she has done. At the same time, it is reasonable to mourn the old craft.

04

AI gives curiosity a new way forward

Browne reads Tan’s optimism about a new generation of builders and adds that AI can help curious people find answers without nearby experts. He compares the reaction to earlier arguments over search and Wikipedia.

Tan writes that coding lives on in model weights, frameworks and programming languages. As a self-taught programmer, she is excited that more people may discover the pleasure of building and believes natural-language tools could help a new generation produce software.

I would broaden that point to curiosity. A teenager fascinated by the Linux kernel used to have documentation, source code and talks, but might lack a knowledgeable person nearby to encourage questions. I think AI can make it easier to keep investigating instead of giving up.

I compare this to the arrival of Google: children could pursue questions their parents might not know how to answer. AI takes that further in my view, although it also provokes a familiar kind of resistance to new ways of finding information.

In school, a teacher suspected I had copied from Wikipedia when I had used it to find the sources behind an assignment. I see a similar objection today when people dismiss a tool rather than examine how it was used.

05

Automating work makes more experiments possible

Browne considers the idea of improving prompts rather than editing every output by hand. He welcomes new ideas from people who lack conventional programming experience.

Tan believes AI-assisted builders may eventually produce software beyond what individuals could write by hand. She quotes the idea that the right sequence of inputs could produce the desired code without editing, so a builder might keep improving the prompt instead of immediately changing the output.

I like the experimentation in that framing. If AI automates more of the implementation, I can try more ideas. Projects I had imagined for years are now built, and I am thinking about more ambitious projects such as an operating system or browser.

Poorly prepared pull requests can be frustrating, but I am also seeing young people rethink video editing, animation, games and deployment. Someone who does not know the conventional approach may try something an experienced developer would never consider.

06

How labels shape perceptions of engineers

Browne discusses Tan’s engineering work and his concern that gender and public labels can obscure a person’s abilities. He ends by encouraging viewers to apply their curiosity to building with AI.

Tan’s technical depth and willingness to replace an approach when appropriate impressed me long before this discussion. I also think women engineers face skepticism that men with equivalent ability may not, and that public labels affect how people are perceived.

On social media, people often have a limited picture of someone. I have been told that viewers are surprised I held a real engineering job, skateboard or know about music because they see me primarily as a YouTuber.

I think being labeled a “vibe coder” can similarly lead people to discount engineering ability. When that label is combined with being a woman, I worry that audiences will question expertise in ways they might not otherwise.

I felt my own public identity change around the time my channel reached 100,000 subscribers: instead of a developer who made videos, some people seemed to see a YouTuber pretending to code. I have seen similar reductions of people’s identities, disproportionately affecting women in my view.

I worried that as Tan became more visible for AI coding and social media, people would overlook her engineering depth and humor. Instead, many came to recognize her technical work and her AI experiments together.

An online dispute also brought more attention to the humor I had long noticed. I see her public reception as a challenge to my assumption that someone could not be recognized for all of those qualities at once.

Coding may be changing, but the skills and curiosity developed while learning it still matter. If we apply the excitement of building and learning to AI-assisted work, I think we can explore projects that previously felt out of reach.

More from this channel

2 slides45:45EN
Theo - t3․gg· Nov 18, 2026Theo Browne: Astra 6.2 excels at review but frustrates in real workTheo Browne · Ping LabsAstra 6.2 scored highly in Theo Browne's tests, but opaque answers, stalled tasks and resource use kept it from becoming his default coding agent.Coding agentsCode review & verificationDeveloper workflows & testing+2
4 slides27:57EN
Theo - t3․gg· Nov 16, 2026Theo Browne: Use Different AI Models to Review the Same ChangeTheo Browne · Ping LabsBrowne’s council workflow asks models from different families to review and test code, but repeated rounds can burn tokens and expand scope.Code review & verificationCoding agentsAgent skills+2
2 slides42:19EN
Theo - t3․gg· Nov 13, 2026Opus or Fable? Match the coding model to the workTheo Browne · Ping LabsTheo Browne compares two Claude models and explains how a thorough tester and a more exploratory second opinion helped solve a performance problem.Coding agentsDeveloper productivityAI cost & performance+2
3 slides19:39EN
Theo - t3․gg· Nov 13, 2026Theo revises his Gemini 4 Carbon assessment after measuring coding delaysTheo Browne · T3His own agent runs showed long response-start waits and incomplete reviews; he asks Google to investigate hosting and cache behavior.Coding agentsAI cost & performanceAgent engineering
2 slides29:37EN
Theo - t3․gg· Nov 11, 2026Theo finds a security use for Gemini 4 Carbon despite coding frustrationsTheoHis tests show fewer false code-analysis claims than Argon, but latency and reliability still push him toward other models.Coding agentsModels & trainingAI cost & performance+1
4 slides48:20EN
Theo - t3․gg· Nov 10, 2026Fable 5.2 is powerful, but Theo Browne still prefers Opus for everyday codingTheo Browne · T3After a day of tests, Browne compares benchmarks, usage limits and the mistakes that still require a human reviewer.Coding agentsBenchmarksModels & training