5 guests 6 episodes 2,484 words

Coding Is Dead, Long Live Coding: What Builders Actually Need to Learn Now

Do product managers and designers still need to know how to code in the AI era?

The head of Claude Code at Anthropic has not edited a single line of code by hand since November 2025. One hundred percent of his code is written by AI. Meanwhile, the CEO of Devin -- the AI software engineer -- says learning computer science is more important than ever. The CEO of Vercel says learn to code AND learn to prompt. OpenAI's CPO says the real skill is writing evals, not writing code. And a PM leader at Meta says learning to code is a direct competitive advantage for anyone working in AI products.

Everyone agrees the world is changing. Nobody agrees on what that means for you. The answer depends on which of these voices you trust -- and as you will see, the disagreement is not superficial. It reflects fundamentally different beliefs about how quickly AI will replace human technical skill, and what "technical skill" even means anymore.

If AI can write code, should product builders, aspiring PMs, and career-changers still invest hundreds of hours learning to program? Or has that time become better spent elsewhere?

The 5 Positions

1 / 5

Learn systems thinking, not syntax -- computer science still matters

Bret Taylor

Bret Taylor, co-founder of Sierra, Board Chair of OpenAI -- who co-created Google Maps, invented the Like button, served as CTO of Facebook, and co-CEO of Salesforce -- draws a distinction that most people miss: coding and computer science are not the same thing. The syntax of Python or JavaScript is becoming a commodity. The ability to understand how systems interconnect, how data flows, and how architecture decisions create constraints -- that is becoming more valuable, not less.

Taylor's argument goes deeper than the standard "learn fundamentals" advice. He points out that as AI produces code you might not completely understand, constraining and directing that AI requires more sophistication than writing the code yourself. This is counterintuitive but critical: the person reviewing AI output needs to understand systems at a deeper level than the person who would have written the code from scratch.

"Studying maths is quite important. It teaches you how to think. It teaches you how the world works, physics, math, and I think computer science especially, at least the foundations of it, will continue to be the foundations of how we build software and understanding that when you're interacting particularly with something that's smarter than you, producing code you might not completely understand how you constrain it and how you get it to produce these outcomes. I think it will require a lot of sophistication actually."
Bret Taylor
▶ 00:36:01

He illustrates the point with history. When Salesforce was founded in 1998, putting a database in the cloud was genuinely hard -- a technical moat. That moat is now "comically narrow" with AWS, but the product moat remains enormous. As the marginal cost of writing code approaches zero, the same pattern will repeat everywhere: technical moats erode, product moats persist.

"When Salesforce was created in 1998, just putting a database in the cloud was hard and that alone was a technical moat that is now trivial with Amazon Web Services and that technical moat is comically narrow, but the product moat is quite large."
Bret Taylor
▶ 00:38:54

Taylor's position is the most durable because it separates the medium (code) from the discipline (computer science). The medium is changing rapidly. The discipline is not.

When This Applies

(Rauch): You are a working PM or designer who wants to prototype, you want to evaluate AI capabilities firsthand, or you want to close the gap between spec and shipping. Vercel's product team creating live PRDs is the model here.

Absolutely yes -- coding teaches structured thinking that transcends syntax

Scott Wu

Scott Wu, CEO of Cognition, the company behind Devin -- the AI software engineer used by companies from two-person startups to Fortune 100 firms -- offers the most surprising voice in this debate. The person building an AI that does engineering work still insists humans should learn to code.

His reasoning is twofold. First, computer science teaches decomposition, abstraction, and systematic problem-solving in ways other disciplines rarely match. Second, we have already been through multiple rounds of this exact debate. Someone from 50 years ago might already call Python "explaining in English what you want and the computer does it." Each round of abstraction made practitioners feel obsolete, but those who understood the layers below consistently outperformed.

"My answer would be absolutely yes. I think, to a large extent, when you take computer science classes and when you learn these fundamentals, sure you're learning a little bit about how a particular language is, syntax works or something like that. But honestly, most of what you're learning really is about the ability to logically break down problems."
Scott Wu
▶ 00:25:41

Wu frames Devin's mission not as replacing engineers but as moving them "from bricklayer to architect." He estimates the creative, architectural work of software engineering -- defining problems, designing solutions, mapping what you want to build -- accounts for only about 10% of the average engineer's time. The other 90% is Kubernetes errors, bug reports, code migrations, and version upgrades.

"That's probably in the neighborhood of 10% of the average software engineer's time, right? Because 90% of the time is you've got this Kubernetes error, then you've got to debug, and you have to see what went wrong and the system crash."
Scott Wu
▶ 00:24:21

Devin automates the 90% so humans can focus on the 10% that matters. But that 10% requires deep understanding of systems, not just the ability to prompt.

When This Applies

(Nika): You are transitioning into AI product management and need to communicate with researchers, scope ML projects, and make launch-quality decisions about model accuracy.

Coding is solved -- become a generalist instead

Boris Cherny

Boris Cherny, head of Claude Code at Anthropic -- the AI coding tool that has transformed software engineering in its first year -- holds the most radical position, and his credibility is impossible to dismiss. He personally ships 10 to 30 pull requests daily with zero hand-edited code. During his recording with Lenny, he had five AI agents running simultaneously in the background.

"100% of my code is written by Claude Code. I have not edited a single line by hand since November. Every day, I ship 10, 20, 30 pull requests."
Boris Cherny
▶ 00:00:00

His timeline is the most aggressive in the debate:

"In a year or two, it's not going to matter. Coding is virtually solved. I imagine a world where everyone is able to program, anyone can just build software any time."
Boris Cherny
▶ 00:00:22

But Cherny's advice is not "stop learning." It is "stop specializing." On the Claude Code team, everyone codes -- the PM, the engineering manager, the designer, the finance person, the data scientist. The strongest people are not the deepest specialists but the broadest hybrids: engineers with great design sense, product thinkers who understand the business, builders who love talking to users.

"Some of the most effective engineers that I work with every day, and some of the most effective product managers and so on, they cross over disciplines. So, on the Claude Code team everyone codes. Our product manager codes, our engineering manager codes, our designer codes, our finance guy codes."
Boris Cherny
▶ 00:41:25

If everyone can code, coding is no longer a differentiator. The differentiator becomes taste, judgment, communication, and the ability to combine multiple disciplines. The specialist coder loses to the generalist builder every time.

When This Applies

(Taylor, Wu): You are a student or in your first three years of your career, you want to build AI products, or you are entering a technical PM role at a company like Perplexity or OpenAI where understanding systems is table stakes.

Code and prompt are complementary skills -- learn both

Guillermo Rauch

Guillermo Rauch, founder and CEO of Vercel -- who created Next.js (which powers Claude, Grok, and Midjourney) and launched v0 to over 1.3 million users -- sees no contradiction. v0 deliberately shows users the generated code because understanding what the code does improves your prompting, and prompting improves your understanding. They form a feedback loop.

"A lot of the programming jobs to be done that used to be specializations, I think, are going away, in a way. They're translation tasks, but knowing how things work under the hood is going to be very important for you because you're going to be able to influence the model and make it follow your intention a lot better."
Guillermo Rauch
▶ 00:00:37

The practical evidence is striking. At Vercel, they went from 150 engineers and 600 total headcount to effectively 600 builders. The best things created with v0 came not from engineering but from marketing, sales, and product management. The product team now builds "live PRDs" -- interactive prototypes so detailed that engineering's response is often "just ship it."

"We have 150 engineers that can write code and 600 total headcount. Now we have 600 engineers. Some of the best things that I've seen created with v0 have not come from our engineering team."
Guillermo Rauch
▶ 00:33:41

When This Applies

(Weil): You are building AI-native products where accuracy thresholds, fine-tuning, and model behavior define the product experience. OpenAI's deep research product was built this way.

The must-have skill is writing evals, not writing code

Kevin Weil

Kevin Weil, CPO of OpenAI -- previously VP of Product at Instagram and head of data at Twitter -- reframes the question entirely. The technical skill that matters most for product builders is not coding or even prompting. It is writing evals: structured assessments that measure how well AI performs on specific tasks.

"Writing evals is going to become a core skill for product managers."
Kevin Weil
▶ 00:18:58

Weil's logic is structural. Traditional software is deterministic: if your database works once, it works every time. AI is probabilistic: a model might be 99.5% accurate on one task and 60% on another, and you build fundamentally different products for each accuracy level. The product decisions that matter -- what to ship, what accuracy to require, when something is good enough -- all depend on eval quality.

"If the model's 60% right on something, you're going to need to build your product totally differently. And by the way, these things aren't static either."
Kevin Weil
▶ 00:20:20

He adds a meta-observation that connects to all other positions: at every previous company, the underlying technology was stable -- databases improve 5% per year. At OpenAI, "every two months computers can do something they've never been able to do before." The definition of "technical skill" is itself a moving target.

When This Applies

(Cherny): You already have strong technical skills and need to differentiate on taste, communication, design sense, or business judgment. The Claude Code team at Anthropic already operates this way.

Evidence from the Archive

Anthropic

Boris personally ships 10-30 pull requests daily with zero hand-edited code since November 2025

Five AI agents running simultaneously during the podcast recording, producing code in parallel

Boris Cherny · Head of Claude Code

Vercel (v0)

Vercel going from 150 engineers to 600 builders using v0 across marketing, sales, and product management

Product management team creating 'live PRDs' so detailed the engineering team says 'just ship it'

Guillermo Rauch · Founder and CEO ▶ 00:00:37

Sierra / OpenAI

The Google Maps rewrite -- once cutting-edge R&D, now achievable by many React developers -- as evidence that...

The Google Maps rewrite -- once cutting-edge R&D, now achievable by many React developers -- as evidence that technical moats erode but product moats persist

Bret Taylor · Co-founder of Sierra, Board Chair of OpenAI ▶ 00:36:59

Cognition (Devin)

Python as an existing example of 'explaining in English what you want and the computer does it' -- from the...

Python as an existing example of 'explaining in English what you want and the computer does it' -- from the perspective of a 1970s programmer

Scott Wu · CEO ▶ 00:25:41

OpenAI

The Lenny & Friends Summit panel where the eval-writing comment resonated most with the audience

Building OpenAI's deep research product: designing evals and product simultaneously, hill-climbing on eval performance

Kevin Weil · Chief Product Officer ▶ 00:18:58

Meta / Google

Deciding whether to launch a model at 70% or 80% accuracy -- a product decision that requires technical understanding

The advice to 'fake the AI with a Figma prototype' for MVPs instead of investing in model training prematurely

Marily Nika · AI PM Leader ▶ 00:08:39

The Synthesis

Everyone in this debate is right -- they are answering different questions for different people at different stages. The question "should I learn to code?" has been replaced by a more useful question: "what layer of the stack should I understand?"

Stack Layer Value

What layer of the technical stack should you understand?

Frontier Disagreement Signal

What does genuine expert disagreement tell you about career bets?

Coding Redefined

Is prompt engineering actually a form of programming?

The hierarchy of value: syntax (least valuable, rapidly commoditizing), frameworks and tools (still useful, also commoditizing), systems architecture (highly valuable, slow to commoditize), product judgment about what to build (most valuable, not commoditizing), and eval/quality judgment (newly critical, not yet widely distributed).

The people closest to the AI frontier disagree most sharply. One says coding is solved in a year, another says learning CS is more important than ever, another says the real skill is evals. This genuine uncertainty tells you to hedge: learn enough to understand systems, but do not make any single technical skill your only investment.

Effective prompting involves role-playing, chain-of-thought reasoning, few-shot learning, RAG pipelines, and LLM-as-judge evaluation. These are programming skills involving logic, debugging, and systematic iteration -- but they look nothing like writing functions in Python. The definition of 'coding' itself is changing.

Which Approach Fits You?

Answer 3 questions about your situation. We'll match you to the right approach.

Question 1

What is your current role?

Question 2

What layer of the tech stack matters most for your work?

Question 3

How do you think about AI's impact on coding?

Notable Absences

The Bottom Line

There is one more dimension that none of these voices explicitly names but all of them imply: the definition of "coding" itself is changing. Lenny's newsletter on prompt engineering techniques reveals that effective prompting involves role-playing, chain-of-thought reasoning, few-shot learning, RAG pipelines, and LLM-as-judge evaluation. These are programming skills -- they involve logic, debugging, and systematic iteration -- but they look nothing like writing functions in Python. The person who learns to orchestrate AI agents effectively is doing a form of programming that is arguably more valuable than traditional coding, because the output scales differently.

Lenny's own newsletter reinforces this convergence. His guide to AI prototyping for product managers shows that PMs can now build functional prototypes with zero coding ability using v0, Bolt, Replit, and Cursor. His survey of vibe coding produced over 1,000 examples of non-engineers building real, useful software -- from nicotine pouch trackers to custom flight radar apps. And Perplexity's co-founder predicts that "technical PMs or engineers with product taste will become the most valuable people at a company over time." The thread connecting all of this: the definition of "technical" is broadening, not disappearing.

Sources

Bret Taylor — "He saved OpenAI, invented the “Like” button, and built Google Maps: Bret Taylor on the future of careers, coding, agents, and more" — Lenny's Podcast, July 31, 2025
Scott Wu — "How Devin replaces your junior engineers with infinite AI interns that never sleep | Scott Wu (Cognition CEO)" — Lenny's Podcast, September 8, 2025
Guillermo Rauch — "Everyone’s an engineer now: Inside v0’s mission to create a hundred million builders | Guillermo Rauch (founder and CEO of Vercel, creators of v0 and Next.js)" — Lenny's Podcast, April 13, 2025
Marily Nika — "AI and product management | Marily Nika (Meta, Google)" — Lenny's Podcast, February 5, 2023