Will AI Automate Coding? OpenAI Codex Lead Explains

TL;DR

Coding is one of the first domains where large language models excel, but automation won't shrink the profession. Just as moving from assembly to higher-level languages let engineers write more code and increased demand, AI will produce more builders, not fewer. The main bottleneck to AGI is human typing speed and validation work, not compute or architecture.

Transcript

You still need software engineers today. You still need designers. I'm a PM. Do you need PMs? You know, you can have some fun jokes about that. I don't think you need them. Today, joining us in the hot seat, we have Alexander and Bericos, product lead for codeex at OpenAI. This is an incredible discussion. Time to get the notebook out. For me, the ... Read More

Key Insights

• Coding is one of the first domains where LLMs became genuinely capable, according to OpenAI Codex product lead Alexander Embiricos, though he argues 'coding is automated' is a heavy, imprecise claim worth unpacking rather than accepting at face value.
• Automation historically expands demand rather than eliminating jobs. When engineers stopped writing assembly and moved to higher-level languages, they wrote far more code, which created greater demand for code and required many more software engineers overall.
• The word 'computer' originally referred to humans doing manual tabulation, reportedly at Bletchley Park decoding the Enigma, and early spreadsheet software mimicked offices of desks arranged in a grid, showing specific tasks get automated while output demand explodes.
• The talent stack is compressing: engineers are becoming more full-stack rather than split into separate backend and front-end roles, so teams still need builders but fewer narrowly specialized positions than a few years ago.
• Product manager is an explicitly undefined role whose goal is adapting to whatever the team or business needs, so its functions can be absorbed by a strong engineering lead or product-minded designer until the team grows very large.
• Human typing speed and validation work, not model compute or architecture, is described as the key bottleneck to AGI, because people cannot prompt AI often enough to capture its full potential value.
• AI should assist tens of thousands of times per day rather than the tens of times most users currently reach, but people are too lazy to type that many prompts and too uncreative to imagine every way AI could help.
• The ideal future removes prompting effort entirely: AI should connect to your context and chime in helpfully without requiring users to learn special prompting techniques, similar to packaged tools like Claude for legal or Claude for Excel.

Questions & Answers

Q: Will AI automate software engineering jobs?

Alexander Embiricos, Codex product lead at OpenAI, agrees coding is one of the first domains where LLMs excel but resists the claim that coding is fully automated. He compares it to when engineers stopped writing assembly and adopted higher-level languages. That shift did not eliminate coding; it let people write much more code, which increased demand and required many more software engineers, suggesting automation of specific tasks expands rather than destroys the profession.

Q: Will there be more or fewer software engineers in five years?

Embiricos believes there will be many more builders, not fewer. He notes that every time a specific manual task has been automated historically, there has been an explosion in demand for the output, requiring even more people to do that kind of work even though the task itself changed. He also observes that terminology shifts over time, just as 'computer' once meant a person and now the term 'software engineer' may come to describe more full-stack roles.

Q: What is the compression of the talent stack?

The compression of the talent stack describes how specialized roles are merging as AI capabilities grow. Embiricos notes that teams still need software engineers and designers, but the traditional split between backend and front-end engineers is fading. On the Codex team, work is much more full-stack than a few years ago. He expects the stack to keep compressing so fewer narrowly specialized people are needed, while builders remain essential.

Q: Why does the Codex lead say you may not need product managers?

Embiricos jokes that PMs may not be strictly necessary because the product manager role is incredibly hard to define, which he considers explicitly undefined. A PM's job is to adapt to whatever the team or business needs, such as looking around corners, collaborating with go-to-market, and being the team's cheerleader and quality raiser. But he argues those tasks could be handled by a strong engineering lead or a product-minded designer until a team becomes really large.

Q: What is the key bottleneck to AGI according to OpenAI's Codex lead?

Embiricos argues that human typing speed and validation work, not model compute or architecture, is the most clickbait but real bottleneck to AGI. Although he uses AI around 30 times a day and knows he should use it far more, he admits he is too lazy to type out that many prompts and too uncreative to find every way AI could help. The limit is human capacity to direct and validate AI, not the models themselves.

Q: How often should people be using AI each day?

Embiricos says AI should be helping people tens of thousands of times per day, compute budget permitting, compared with the tens of times most Codex users currently reach. He cites OpenAI engineers who constantly keep Codex running, never close their laptops, and feel they are wasting time if it is not working during meetings. He envisions inference effectively running around the clock across nearly every task a person does.

Q: How should AI tools remove the prompting burden for users?

Embiricos wants a world where using AI requires no effort to figure out the right way to prompt. Instead of forcing users to learn techniques, AI should connect to a person's context and chime in helpfully, or be added to a group chat where it simply starts assisting. He says most people should not need to work hard to benefit from AGI. He praises Claude's packaging, like Claude for legal and Claude for Excel building a DCF model, as examples of this direction.

Q: Should companies adopt AI through tools for individuals or top-down enterprise automation?

Embiricos favors building tools for people rather than relying solely on top-down automation. Responding to Matt Fitzpatrick of Invisible AI, who argued on the podcast that enterprises cannot adopt AI without FTEs automating workflows, Embiricos disagrees entirely. He notes that a top-down approach is limited by what leadership can envision and what staff can be assigned, whereas empowering individuals with strong tools scales beyond those constraints and lets everyone feel superhuman.

Summary & Key Takeaways

• Alexander Embiricos, product lead for Codex at OpenAI, agrees coding is one of the first domains where LLMs are strong but pushes back on 'coding is automated' as an oversimplification, comparing it to abandoning assembly language for higher-level languages that expanded rather than eliminated engineering work.

• He predicts more builders in five years, not fewer, citing how automation of manual tasks historically explodes demand for output. The talent stack is compressing toward full-stack engineers, and he jokingly argues product managers, whose role is inherently undefined, may not be strictly necessary on smaller teams.

• The key bottleneck to AGI is human typing speed and validation, not compute. Embiricos wants AI helping people effortlessly tens of thousands of times daily, connected to their context, so benefiting from AGI requires no special prompting skill, and he favors building tools for individuals over top-down enterprise automation.