Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI

Episode Summary

Executive Summary: The conversation centers on OpenAI’s shift from ChatGPT and Codex into ChatGPT Work, a unified “productivity” layer for knowledge work and personal tasks. Akshay argues that LLMs make code-like leverage accessible to everyone, that enterprise adoption is fragmented by use case, and that the best product strategy is to meet users where they are while hiding complexity. The discussion also covers artifacts, sites, memory, sub-agents, model selection, and how AI is changing measurement of productivity from output proxies to goal attainment.

Main Topics: ChatGPT Work as a unified productivity product (Priority: 5/5): Akshay explains why OpenAI merged Codex and ChatGPT experiences into ChatGPT Work: users shouldn’t have to choose between apps when the underlying capability is the same, especially as work increasingly spans code, documents, spreadsheets, search, and personal tasks. Enterprise adoption and meeting users where they are (Priority: 5/5): The enterprise lesson is that AI use cases vary widely across teams and companies, so success depends on surfacing the right workflow and teaching users how to apply AI in context rather than expecting them to infer it from a blank box. Harness engineering, modes, and product abstraction (Priority: 5/5): The transcript compares the classic ChatGPT harness and the Codex harness, explaining that the underlying harness is shared but the UX differs by use case, with opinionated defaults around sandboxing, diffs, and visibility into agent behavior. Artifacts, sites, and knowledge-work outputs (Priority: 4/5): A major theme is that AI outputs are evolving from plain text to high-fidelity artifacts such as spreadsheets, slides, sites, and research panels that can replace decks and docs as the canonical team output. Memory, Chronicle, and persistent context (Priority: 4/5): Akshay describes memory as a key differentiator that makes ChatGPT feel personal over time, and notes that work mode extends that via persistent context, file storage, and experimental tools like Chronicle that observe computer usage. Model choice, sub-agents, and defaults (Priority: 4/5): The discussion covers the new model lineup, reasoning levels, sub-agents, and why OpenAI prefers strong defaults with optional advanced controls for power users who want to optimize for quality, cost, or speed. Productivity measurement and the motion vs progress trap (Priority: 5/5): Akshay argues that AI makes it easier to create motion than real progress, so teams need clearer goal-based evaluation and more focus on iteration quality, learning loops, and at-bats rather than legacy proxies like commits or story points.

Key Arguments: LLMs are the missing technology that lets many more people do code-like work without knowing what’s underneath. Enterprise AI fails if it assumes a one-size-fits-all workflow; product must map to discrete user goals. Codex and ChatGPT Work share the same underlying harness, but the UX should adapt to the task and the safety model. The power of agents is not just for developers; non-developers quickly adopted Codex internally once they saw it as a superpower. Artifacts and sites are becoming the new collaborative medium for knowledge work, replacing static docs and decks. Memory and persistent context are essential to making AI feel personal and useful over time. The best default is an opinionated one; power users can adjust reasoning, model class, and sub-agents when needed. Progress should be measured by whether teams hit goals, not by superficial activity metrics that AI can inflate. AI is shifting teams toward T-shaped generalists: broad capability plus deep specialty. Show, don’t tell, is now a core product challenge because users discover value faster when the product demonstrates it in context.

Data Points: OpenAI headcount when Akshay joined: 500 people - He recalls joining in 2023 and feeling the company was even more startup-like than expected. ChatGPT user base: hundreds of millions - Referenced as the distribution base that ChatGPT Work can build on. Launch adoption: 10 million users - Mentioned as the early launch milestone for ChatGPT Work. Model/version comparison: 5.4, 5.5, 5.6 - Used to describe improvements in artifacts and launch timing around the new product experience. Time horizon for internal review cycle improvement: 6 months - Akshay said the model’s help in reviewing people’s work was dramatically better than six months earlier. User prompt exploration scale: 1.7 billion tokens - The guest described a personal project that generated a playable site for a board game after heavy iteration. Project workflow example: 18 minutes 53 seconds - He contrasted a Codex-built version of a game prototype completed in this time frame. Personal/enterprise usage pattern: 4 threads on one project - Used to illustrate how knowledge work becomes multi-session and benefits from persistent memory/context.

Pivotal Quotes: "the missing technology required to bring the magic of code to everyone without them having to know what's going on underneath the hood" — Akshay: Explaining why LLMs completed the no-code/low-code thesis from his earlier work. "we want the user to not need to choose which experience they're in" — Akshay: Describing the rationale for merging Codex and ChatGPT Work into a unified experience. "Motion is much easier now than ever before because of the tooling that we have. But progress requires you to be very prescriptive and deliberate about what you're actually trying to achieve" — Akshay: Discussing productivity measurement and the risk of conflating activity with outcomes.

Implications: AI products are moving from chat interfaces to task-native systems that blend text, files, tools, memory, and agents. For users and teams, the competitive edge will come from better defaults, better context, and clearer success metrics—not from exposing more knobs.