Why Codex Merged with ChatGPT? OpenAI Leader Explains the Future

Since January, Codex’s weekly active users have risen more than fivefold to five million, with knowledge‑worker adoption three times higher than among developers; the steep growth is attributed to the February launch of a dedicated desktop app that lowered the usage barrier.

Andrew Ambrosino, head of the Codex desktop team, explains that his role sits at the intersection of developer‑centric tooling and a general‑purpose AI workbench, giving him a view that goes beyond raw growth numbers.

He argues that implementation is no longer the expensive part of product development. Instead, the costly element is "taste"—the curation process that decides which of the many parallel attempts (often 90 teams exploring the same feature) should be folded into the product, how to frame them, and how many refinement levels a UI element should have.

Ambrosino defines taste through four dimensions: aesthetics, systems thinking (how a feature fits the overall system), sense of direction (the thematic relevance), and presentation details such as animation speed matching semantic intent.

When asked why AI still struggles with design, he cites two reasons. First, design quality is hard to score objectively because human taste is part of the feedback loop, unlike code which can be compiled and tested. Second, cultural factors shape what counts as "good design," and AI models lack the ability to generate truly novel aesthetics beyond copying existing styles.

He also notes that the abstract‑layer gap makes it difficult for AI to understand deep structural relationships between UI components, which hampers tasks like brand‑wide redesigns that require semantic consistency across many elements.

Regarding product timing, Ambrosino is convinced that releasing Codex in November 2022 would have failed, whereas the February launch succeeded because the underlying model had improved in the intervening months. This illustrates that product usefulness now hinges on the model’s capabilities at a given moment rather than on UI or interaction design alone.

This shift has changed product planning: teams now list all ideas, prototype them with the current model, and defer those that require future model breakthroughs, effectively planning around model milestones instead of fixed feature roadmaps.

He observes that traditional role boundaries between engineers, designers, and product managers are blurring; team members frequently wear multiple hats, yet each discipline still retains its own skill thresholds, preventing a complete collapse of role distinctions.

On the frontier of AI‑assisted development, Ambrosino describes a transition from writing code to "guiding AI"—evaluating whether generated code is supervised or unsupervised. He mentions experiments with autonomous loops, such as having the model run nightly clean‑up tasks, and a persistent weakness where models tend to make code increasingly complex. He wishes for better "code‑deletion" capabilities to avoid this drift.

The decision to merge Codex with ChatGPT stems from the observation that the two products are merely different entry points to the same underlying capability. Codex started as a command‑line developer tool, but internal use spread to non‑technical teams, prompting the team to make the product more universal.

Ambrosino illustrates the "home base" concept: the app should act as a central hub where users can start work, finish work, and automate tasks by invoking external tools (e.g., opening Excel or controlling Premiere Pro via a plugin). He recounts how a photographer used Codex to edit video by having the model generate a Premiere Pro extension and command the editor, demonstrating seamless collaboration with existing professional tools.

The team’s two current objectives are: (1) enable Codex/ChatGPT to work with users’ existing tools through connectors or computer‑use capabilities, and (2) allow users to launch web applications inside Codex, letting the AI perform additional actions within those apps. Both lines are being pursued in parallel.