Control the Ideas, Not the Code

Salvatore Sanfilippo, better known as antirez, is the creator of Redis — one of the most widely used databases in the world. For over 20 years he has written high-quality code by hand and earned a reputation for clean, reliable systems. Now he is making a provocative claim: in the age of advanced AI coding assistants, reviewing the code itself has become largely pointless.

His reasoning is straightforward and hard to dismiss. Modern models can generate thousands of lines of code in a single day. A human developer simply cannot read it all. An 8-hour workday spent line-by-line reviewing AI output leaves almost no time for the higher-value work that actually moves projects forward: defining the direction of the software, inventing features, optimizing architecture, and verifying that the final result works as intended.

From “Writing and Reviewing Code” to “Controlling Ideas”

Control the Ideas, Not the Code Antirez argues that the programmer’s real product is no longer the code — it is the ideas behind the system. AI excels at local implementation details, but it still needs strong human guidance on the big picture. If you truly control the ideas — the data structures, the invariants, the trade-offs, the mental model of how everything fits together — then staring at every generated line becomes suboptimal and often unnecessary.

This is not “vibe coding,” where you vaguely tell an AI to “build me an app” and hope for the best.It requires deep understanding.

Instead of asking “Does this line look correct?”, you ask questions like:

“Explain the overall design of this component.”
“What are the key invariants here?”
“How does this part interact with the rest of the system?”

You then test the outcome — through rigorous QA, automated testing, and real-world usage — rather than trying to audit every token the model produced.

What Antirez Actually Does Today

Even though he believes code review is mostly obsolete, antirez still reviews the code he generates for Redis. He does it partly out of respect for the many people who will later read and modify those files by hand. However, he openly admits that today’s strongest models — he specifically mentions Claude Fable and GPT 5.6 — already find more bugs, race conditions, and subtle issues than he does himself.

Control the Ideas, Not the Code He would much rather spend that time writing clear DESIGN.md documents that describe the ideas in human language: why a particular data structure was chosen, what tricks and optimizations are hidden inside, what the important invariants are. With such documentation, future humans (or AI agents) can understand the why behind the code and modify it intelligently instead of guessing.

“Nobody should anymore look at this code, but only at the ideas the code contains,” he writes.

Software Was Never Perfect — AI Didn’t Invent Bugs

Critics often claim that AI is flooding the world with low-quality code. Antirez pushes back: software has always contained subtle bugs and technical debt. Long before LLMs existed, he personally discovered tricky issues in widely used tools that had gone unnoticed for years. The idea that “everything was high quality before AI” is largely a myth.

What has changed is the volume and speed. AI doesn’t create the problem of imperfect code — it simply makes the old way of catching it (exhaustive human review) impossible to scale.

An Important Caveat: Beginners Still Need to Write Code

There is one group for whom the advice “stop reviewing code” does not apply yet: beginners and junior developers. Antirez is clear that new programmers must still build things from scratch — interpreters, small databases, hash tables, etc. — to develop the mental models and intuition that later allow someone to “control the ideas.”

Without that foundation, you cannot effectively steer AI or evaluate whether the generated design is sound. Once you have that experience, however, you can (and should) stop wasting time on low-leverage tasks like manually reviewing boilerplate JavaScript for client websites.

The Bigger Shift

Control the Ideas, Not the Code This change is not limited to software engineering. The same pattern is emerging across many knowledge professions: lawyers will control legal strategy and arguments rather than drafting every clause; analysts will define the questions and validate insights rather than manually crunching every dataset; marketers will own the narrative and positioning rather than writing every piece of copy.

The core skill is moving from execution to direction and judgment.

The dangerous trap, however, is trying to skip the hard work of building expertise. You can only safely delegate the routine work to AI once you already understand the domain deeply. Everyone else still needs to put in the reps first.

The New Job Description

Control the Ideas, Not the Code In antirez’s view, the programmer of the near future will spend far less time writing or reading code and far more time on:

Clearly articulating the ideas and architecture;
Creating high-quality documentation of intent (not just implementation);
Designing tests and verification processes;
Exploring new directions and optimizations;
Shipping and iterating on real outcomes.

The value is shifting from “I can write code” to “I understand what we are building and why it should work this way.”

It’s a painful transition for many, but according to one of the most respected engineers in the industry, it is also inevitable — and ultimately more joyful. The craft isn’t disappearing. It is evolving into something higher-leverage: controlling the ideas.

---