Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead)

Today’s guest is Alexander Embiricos, product lead for Codex, OpenAI’s fast‑rising coding agent. We get into what building product at OpenAI really feels like, how Codex helped the Sora team ship an app that quickly hit the top spot in the App Store, the surge in usage, why reviewing code is now the bottleneck, his view on agents, AGI timelines, and a lot more. Alexander, welcome.

Great to be here. The biggest difference at OpenAI is the speed and ambition; it reset my sense of what “moving fast” means and made me far more ruthless about where I spend time.

How is the org set up to move that fast? It sounds more like ready‑fire‑aim than ready‑aim‑fire.

The tech moment is ripe, and even if models paused, we’d still be behind on product. We aim fuzzily on long horizons and learn quickly through bottom‑up, empirical work, which only works because individuals have very high autonomy and drive.

Give us the Codex basics. What is it and how should people think about it?

Codex is a coding agent that lives in your IDE and terminal to help understand, write, run, and ship code. We see it as the start of a true engineering teammate that will ideate, plan, validate, deploy, and maintain code, not just autocomplete.

So the vision is a teammate, not just a smarter editor.

Exactly. You should feel like you have superpowers without constantly thinking, “how do I invoke AI now?” It plugs into how you work and becomes proactive.

How is it going—any numbers you can share?

Usage has exploded about twenty times since August, now serving many trillions of tokens weekly. A tight product‑plus‑research loop let us iterate the model and harness together, and Codex has become the most served coding model in our API.

What unlocked that growth?

Our first cloud‑only agent felt a bit ahead of its time. Moving into the IDE and CLI with a safe local sandbox made it trivial to get value, created a tight feedback loop, and then set people up to delegate more over time; dogfooding had favored the future, but most users needed an easier on‑ramp.

Beyond product, what boosted its raw ability to code?

We shipped GPT 5.1 Codex Max, which is roughly thirty percent faster on tasks and smarter at higher reasoning. It can run for very long periods with a feature we call compaction that spans model, API, and harness, and we’ve optimized for a shell‑first, sandboxed toolchain so the model learns one powerful way to act.

How do you win in this space long term?

Build a default‑helpful teammate and super assistant that works without constant hand‑holding. Chat is great for discovery, and experts can pull up a deeper GUI to go hands‑on, but the magic is that it just knows how to help.

How does that connect to ChatGPT’s broader assistant vision?

Any capable agent needs to act on computers, and the most reliable way to do that is by writing code. Coding becomes a core competency for all agents, even when the end user doesn’t realize code is being written.

Are agents actually useful today?

Coding agents are already strong; outside coding, it’s still early. I expect a big jump once agents can compose with code in more domains.

How will this change engineering work, and what are you fixing next?

Code will show up in more places, which increases demand for people who can reason about systems. The biggest friction now is reviewing AI‑written code, so we’re focused on validation, better evidence, and showing outcomes first so review feels faster and more fun.

Do we move to spec‑driven development, or something else?

Plans help long tasks, but not everyone likes writing specs; I also see chatter‑driven workflows where context across tools triggers changes. Imagine a future where an agent proactively proposes work in a stream and you approve or reject in seconds.

Real talk: the bottleneck seems to be validation and code review.

That’s the constraint. We need teams to configure agents for more autonomy in the later stages and give users strong, quick ways to trust the results.

How has Codex changed your PM workflow?

It compresses the talent stack. I answer questions, explore changes, and prototype faster, and we see a surge of throwaway tools and vibe‑coded prototypes from designers. With Codex, a tiny team built the Sora Android app in eighteen days and it hit number one; the Atlas browser team reports one engineer now does in a week what used to take a few engineers a few weeks, and even Windows support accelerated once the model learned PowerShell.

Any notable non‑engineering uses yet?

There are many creative experiments, but we’re keeping the team focused on coding for now because the surface area is huge.

Best way to try it, and what languages does it handle?

Give it a real, hard task like a nasty bug or a tricky change, not a toy. It supports most mainstream languages; start by letting it learn your repo, co‑write a plan, then have it execute so trust builds naturally.

Career advice as agents do more coding?

Be a doer and master the new tools; systems thinking and team skills still matter. Set agents up to verify their own work—for example, Codex already catches configuration errors in its training stack and is starting to babysit training runs by watching charts and responding.

Your AGI timeline—what unlocks the hockey stick?

Humans are the bottleneck when we must prompt and manually validate; once agents can act and self‑check by default, the curve bends. I expect early adopters to see that next year, with larger companies following as their systems adapt.

Lightning round. A book you recommend?

The Culture series by Iain Banks, because it imagines a hopeful future with AI and helps you think about choices that lead there.

A recent show you loved.

Jujutsu Kaisen. It deals with dark themes, but the main character is kind, and that optimism is refreshing.

A product that impressed you.

Tesla’s driving experience. It is a great example of mixed‑initiative design that keeps you in control while the agent helps.

A motto you live by.

Be kind and candid. Treat candor as an act of kindness and deliver it that way.

One last fun one about your surname.

I identify more with the poet. He loved our family’s island, Andros, and wrote beautifully about it.

This was a joy. Thanks for the optimism, the concrete examples, and the candor. If you enjoyed this, follow the show on your podcast app and consider leaving a rating or review. You can find past episodes at Lenny’s Podcast dot com. See you next time.