Alexander Embirikos

OpenAI's Codex Lead

Mentions

Appeared in

Discussed in

Key positions and views

The primary bottleneck to achieving AGI is the speed of human validation and typing, not compute power or model architecture.

The most effective way for AI agents to interact with and control computers is by writing code, which is why coding agents have achieved product-market fit while others have lagged.

AI tools like Codex will increase the demand for software engineers over the next five years, shifting their role from writing code to reviewing AI-generated code.

OpenAI's core competitive advantages are its superior compute infrastructure and its ability to produce the best models, which it leverages by providing API access even to direct competitors.

The most significant opportunity for new startups is building vertical AI applications for specific, underserved customer segments rather than competing on foundation models.

Podcast consensus on Embirikos

Points of consensus

▶OpenAI is developing a proprietary web browser named Atlas, specifically to create a secure, controlled environment for agentic tasks in enterprise settings, a point mentioned across both sources.Feb 2026

▶OpenAI's Codex product has experienced explosive growth, with claims from both sources indicating its user base increased 20-fold between August and late 2023.Feb 2026

▶Recent updates to OpenAI's models and infrastructure have yielded significant performance improvements, including faster model serving and more efficient task completion.Feb 2026

▶The development of AI agents for non-coding domains has progressed more slowly than anticipated, with coding agents being the only category to achieve clear product-market fit so far.Feb 2026

Points of debate

▶The exact magnitude of recent API performance improvements is inconsistent across claims, with one stating models are served 40% faster and another citing a 25% speed increase from a specific optimization.Feb 2026

▶There is a potential tension between the view that compute is not the primary bottleneck to AGI (Claim 14) and the assertion that superior compute infrastructure is one of OpenAI's defining competitive advantages (Claim 17).Feb 2026

▶Embirikos presents a future with more software engineers (Claim 13) but also describes their new primary role—reviewing AI code—as a task that is 'often perceived as less enjoyable,' raising questions about the desirability of this future role (Claim 38).

▶The specific versioning of the latest Codex models is presented with slight variations, referring to GPT 5.1 Codex Max, GPT 5.2 Codex, and GPT 5.3 Codex across different claims, suggesting a rapid and perhaps complex iteration cycle.Feb 2026

Key themes

▶The AI-Driven Transformation of Software EngineeringFeb 2026

Embirikos details a fundamental shift in software development, driven by tools like Codex. He claims developers at OpenAI are moving from pair-programming with AI to fully delegating tasks, often bypassing traditional IDEs, and that the vast majority of code is now AI-written and AI-reviewed.

This suggests a future where engineering value shifts from implementation skill to problem decomposition and rigorous validation, potentially creating a new class of 'AI-augmented' senior engineers and altering how talent is evaluated.

▶OpenAI's Vertically Integrated Product EcosystemFeb 2026

The claims outline OpenAI's strategy of building both foundational models (Codex) and end-user applications (Atlas browser, Sora app) that leverage them. This ecosystem is designed for high user stickiness through integrations and is supported by a business model that includes providing API access to competitors, confident in its superior models and compute.

OpenAI is not just a model provider but is building a full stack, from infrastructure to applications, aiming to control key distribution points like the browser to ensure its agents have a secure and effective environment to operate in.

▶The Primacy of Code in Agentic AIFeb 2026

Embirikos posits that writing code is the most effective method for AI agents to interact with computers, which explains why coding agents are the only ones to achieve product-market fit. This belief underpins OpenAI's development of tools like the Atlas browser and OS-level sandboxing for Windows, creating controlled environments where code-writing agents can operate safely.

This 'code-first' philosophy suggests that progress in general-purpose AI assistants may be gated by the ability to translate any user intent into executable code, making the security of these code-executing environments a critical battleground.

▶Redefining the Bottlenecks to AGIFeb 2026

Embirikos challenges the common narrative that compute is the main obstacle to AGI. He argues the true bottleneck is the slow speed of human feedback, typing, and validation, as the models need vast amounts of high-quality, human-verified data to improve.

For investors, this perspective implies that the most valuable long-term investments in the AI space may not be in raw compute, but in companies developing novel human-computer interaction paradigms that can dramatically accelerate the human feedback loop.

Source episodes

Sentiment over time

Not enough data for timeline

Changes over time

August 2023

A period of rapid user adoption for OpenAI's Codex product begins, leading to 20x growth over the following months.

December 2023

OpenAI releases GPT 5.2 Codex, which Embirikos identifies as a key inflection point where developers shifted from pair-programming with the AI to fully delegating tasks to it.

Late 2023

The user base for the Codex product is reported to have grown by a factor of 20 since August.

Recent Period

OpenAI ships multiple model and infrastructure updates, including GPT 5.1 Codex Max and API optimizations that increased model serving speed by 25-40%.

Present

Embirikos states that coding agents are the only AI tool with clear product-market fit, while progress in non-coding and multimodal domains has been slower than expected.

2026 (Projected)

Embirikos anticipates this will be the 'year of agents,' suggesting a significant leap in the capabilities and adoption of agentic AI.

Suggested prompts

How does Embirikos's assertion that human validation speed is the primary AGI bottleneck (Claim 14) square with OpenAI's strategy of leveraging superior compute as a key competitive advantage (Claim 17)? &nearr;What are the strategic implications for Google and Microsoft of OpenAI building its own browser (Atlas) and OS-level sandboxing for Windows (Claim 32) to create a secure environment for its AI agents? &nearr;Embirikos predicts more software engineers in five years (Claim 13), but describes their new primary role of code review as 'less enjoyable' (Claim 38). What are the potential long-term consequences for developer morale and talent retention? &nearr;Given Embirikos's view that only coding agents have achieved product-market fit (Claim 27), what specific technical or product hurdles does he believe are preventing similar success for agents in non-coding domains? &nearr;

Key concepts

Codex 2 ep AI Agents 2 ep OpenAI 2 ep Atlas Browser 2 ep Software Engineering 2 ep GPT Models 2 ep Compute 2 ep API 2 ep Sora App 1 ep AGI (Artificial General Intelligence) 1 ep

Notable quotes

“For the team building the Atlas browser at OpenAI, tasks that previously took 2-3 engineers 2-3 weeks to complete can now be done by one engineer in one week using Codex.”

Alexander Embirikos · Inside OpenAI： 2026 is the year of agents, AI’s biggest bottleneck, and why compute isn’t the issue

“The Sora Android app was developed from scratch and launched to the public in a total of 28 days with the help of OpenAI's Codex.”

Alexander Embirikos · Inside OpenAI： 2026 is the year of agents, AI’s biggest bottleneck, and why compute isn’t the issue

“The models powering OpenAI's Codex are serving many trillions of tokens per week.”

Alexander Embirikos · Inside OpenAI： 2026 is the year of agents, AI’s biggest bottleneck, and why compute isn’t the issue

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report