In a candid and revealing interview on the popular Lenny’s Podcast, Alexander Embiricos, the product lead for Codex at OpenAI, dropped a bombshell that’s resonating across the tech world: The biggest bottleneck to achieving superhuman AI productivity isn’t the models themselves—it’s us. Humans. Specifically, our sluggish typing speed and limited multitasking ability.
“Today, the current underappreciated limiting factor is literally human typing speed or human multitasking speed,” Embiricos told host Lenny Rachitsky in the episode titled “Why humans are AI’s biggest bottleneck (and what’s coming in 2026)”, uploaded just yesterday.
As leader of the team behind Codex—OpenAI’s advanced coding agent that’s exploded 20x in usage since August, now processing trillions of tokens weekly—Embiricos is at the epicenter of AI-driven software development. He envisions Codex not as a mere tool, but as a “software engineering teammate”: a proactive AI intern that’s brilliant at code but needs guidance to stay on track.
Explosive Growth and Real-World Wins
The numbers speak for themselves. Codex’s rapid adoption stems from smart product choices: ditching complex cloud setups for seamless IDE extensions and CLI tools that deliver instant value. Users see results fast, fueling organic growth.
Mind-blowing examples abound. OpenAI used Codex to build the full Sora Android app—a brand-new product—in just 18 days, launching to the public 10 days later and rocketing to the top of the App Store charts. Another feat: The experimental Atlas browser, crafted by a single engineer in one week, slashing what previously took teams weeks.
Even gnarly bugs that stump experts for hours? Codex tackles them autonomously, running for an hour and cracking cases that defy human debugging marathons.
Analysis: These anecdotes aren’t hype—they signal a paradigm shift. Traditional software development timelines are collapsing. For startups and enterprises alike, this means faster iteration, lower costs, and a competitive edge. But it also raises questions: Will AI agents displace junior engineers, or elevate them into orchestrators of super-teams?
The Human Bottleneck: Why We’re the Weak Link
Embiricos argues that AI models are already capable of far more, but they’re shackled by human input speed. We type too slowly to feed them context, and we can’t review outputs at machine pace.
The future? Proactive agents that anticipate needs, write code to interact with systems, and handle the full lifecycle—from ideation and planning to validation and maintenance. Coding, he says, will be every agent’s core skill: “Writing code is how agents use computers best.”
This shift moves the choke point from creation to review. Engineers will spend less time building and more validating AI-generated work. Non-coders, like product managers, are already benefiting: Rapid prototyping, data analysis, and “throwaway code” for quick experiments.
Broader Implications: As AI goes proactive, productivity could “hockey stick” in 2026, Embiricos hints. Larger companies might lag due to inertia, but agile teams could leap ahead. Ubiquitous coding means demand for programming literacy surges—even as natural language layers make it accessible to all.
Skills for the AI Era and OpenAI’s Bold Vision
Advice from the frontier: Lean into tools that amplify you. Master systems thinking, collaboration, and niche expertise—areas where AI still falters.
OpenAI’s building toward mixed-initiative UX: AI assists without interrupting flow. Projects like Atlas aim for contextual awareness, understanding your screen directly for seamless help.
In a lightning round, Embiricos recommended Iain M. Banks’ The Culture series for optimistic AI futures, praised Tesla’s self-driving for smart human-AI interplay, and shared his mantra: “Be kind and candid.”
Final Take: This interview underscores OpenAI’s relentless pace and bottoms-up innovation. With Codex evolving into true teammates, the AI agent era isn’t coming—it’s here. But unlocking its full potential means overcoming our own limitations. 2026 promises acceleration; the question is whether humans can keep up—or get out of the way.
For the full discussion, including tips on using Codex and insights on measuring AI productivity, watch the episode on Lenny’s Podcast. As Embiricos and teams push boundaries, one thing’s clear: The future of work is being coded right now.
Read More: Google Launches Managed MCP Servers: Streamlining AI Agent Connections to Cloud Services