Codex Was Being Sneakily Used by Non-Engineers Inside OpenAI, and ChatGPT Work Was Born
OpenAI moved Codex's harness over to knowledge workers as-is, changing only UX trade-offs and sandbox defaults; the judgment behind it is that functional boundaries will be flattened by AI, so it refuses to draw product lines by "who you are."
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Non-developers sneakily using Codex was the starting point of this product
Akshay says the most direct trigger for ChatGPT Work was an inflection point in non-developer adoption of Codex inside OpenAI—strategic finance, marketing were all using it. But the signal that really stuck with him wasn't usage, it was emotion: these people were "very proud" that they were using Codex, and Swyx called it on the spot as "I'm not supposed to be using it, but I am," which Akshay admitted was exactly it, and they felt they'd been given superpowers. The conclusion drawn from this: the moment when agent capabilities spill over beyond developers came much earlier than OpenAI itself expected; and ChatGPT's existing distribution is a ready-made channel, so the only question left is how to put that power in front of those people.
— Akshay NathanThe two products share one harness, the difference is only in defaults
Swyx asked directly whether Codex and ChatGPT Work use the same harness, and Akshay's answer was the same one, shared, and improvements made for knowledge work—plugins, computer use, artifacts—are available on both sides. There are only two differences: one is UX trade-offs, where Codex mode defaults to putting Git state and diff front and center, the dynamic island assumes you're in a repo, and the same "make a retirement calculator spreadsheet" task in Work shows no file edits; the other is different security defaults for the sandbox. His principle: users shouldn't be forced to first choose which experience they're in. The reason for merging isn't engineering reuse either, it's that every few months everyone's job content changes, and drawing hard boundaries by identity will eventually be wrong.
— Akshay Nathan32 options is too many, the default is the product
Asked how to min-max between models and reasoning tiers, Akshay first threw out a number: 32 options, and volunteered that "you could say there are too many options right now, we're simplifying." His order is rigid: use the default first, and only change the configuration when you see a clear shortfall in cost efficiency or intelligence quality. Ultra and multi-agent settings suit two kinds of tasks—extremely complex open exploration, or highly parallelizable ones; goal suits tasks that can be continuously and verifiably advanced; and most tasks are neither. The slider's design idea is to forcibly project multiple dimensions onto one: one end is speed and efficiency, the other is quality and thoroughness. After launch they also changed Ultra to require manually turning it on in advanced settings, because it eats through usage quota faster.
— Akshay NathanSites are replacing decks plus spreadsheets
Sites is most discussed as a prototyping tool, but the other side Akshay gives is more noteworthy: the monthly reports the company's finance team makes were historically slide decks plus spreadsheets, and now they're directly a Site, and the Site itself becomes the vehicle for team collaboration. The reason is bandwidth and ceiling—PowerPoint and Excel are nominally infinitely flexible, but you hit two walls: either people don't know how to use a feature, or the product simply doesn't support it; a Site is HTML, so if you want a conclusion to be the hero image it can be, and things that would get buried in a deck can be pushed to the top. Even the model slider in this launch was finalized by design, engineering, and product collaborating on a Site. Akshay's assessment of Markdown: it was never suited for human reading anyway.
— Akshay NathanAI doesn't write performance reviews, it just gathers all the context
Swyx's concern is specific: as a manager, using something an LLM generated to evaluate people makes you look like you don't care. The line Akshay draws is clean—he would never purely rely on AI to finish a review and submit it; AI's value is entirely in collecting context: code, problems this person caught, review records, Slack, all of it can be dug through. He says he tried it six months ago during that review cycle and it was completely unhelpful; this cycle the model did better than he himself could, surfacing contributions he hadn't seen at all. Swyx summed it up as "this is agentic search," Akshay agreed, and added that it's more steerable and targetable than before. As for whether it misses things, his answer: it will, but I miss things too, it's relative.
— Akshay Nathan10 million users isn't an achievement, it's the numerator
Akshay calls 10 million users a culmination, but immediately puts it back into the denominator: ChatGPT has hundreds of millions of users, almost all of whom equate it with AI, so by comparison 10 million is just a start. The key constraint is one OpenAI set itself—ChatGPT Work is currently only open to paying users, and won't default ChatGPT users into it, so it has to iterate round after round through education, trials, and feedback. At the same time he explicitly denies that Codex will be absorbed, in stronger terms than Swyx's question: developers have always been the core market, Codex will continue to go deep on software development, and merging actually raises its utility, because you can seamlessly switch between writing diffs, generating artifacts, and searching across codebases.
— Akshay NathanThe bottleneck has shifted from execution to not being able to come up with new ideas
When everyone can build things, Akshay thinks the bottleneck becomes ideas and taste. He says the one automation he most wants but still can't build is "give me new ideas," and LLMs just can't do it. His explanation isn't insufficient model capability, it's the structural source of ideas: ideas aren't produced in a vacuum, and in product development they come from conversations with users, from friction and feedback you observe, from the foundation you've already laid. So he judges that generalists who can close this loop will always have value. On role evolution, his view is T-shaped: AI makes everyone a generalist—he says he could never do design before, and still has no visual taste, but can iterate with AI—and the specialty is the vertical stroke.
— Akshay NathanDon't mistake adding dashboards for the team making progress
The old proxy metrics for productivity—commits, lines of code, story points—are failing after AI: token usage and PR counts are no longer highly correlated with "is the team achieving its goals." Akshay's replacement for managers isn't a new dashboard, it's at-bats: not just the count, but the quality—from generating an idea, building it, getting feedback, reacting to feedback, validating or falsifying the hypothesis, then moving to the next idea, whether this whole loop can run efficiently; and whether the team has a culture that keeps its motivation and humility through repeated runs. He explicitly names the biggest trap: adding a bunch of LLMs and building all kinds of dashboards, but nothing changes, and that is mistaking motion for progress.
— Akshay NathanIn their own words · checked verbatim
And, like, we should enable users to choose, but we shouldn’t box them in.
Akshay Nathan11:43
at these tools like PowerPoint and Excel are like infinitely flexible, but at some point you reach the boundary of like either as a human you may not know how to use some feature or something, or the product itself doesn’t support it. But with a site you can do anything.
Akshay Nathan23:52
Oh, I think the etiquette is that, like, I would never write something via, like, well, solely via AI and, like, present it as, like, a review for someone. What I was talking about is more, like, gathering context. That’s the place where it’s incredibly helpful.
Akshay Nathan38:27
I would say that the one automation that I would love to work and it doesn’t work is bring me new ideas, right? somehow LLMs are just not it. One interesting part about ideas is, like, they’re not, like, in a vacuum.
Akshay Nathan1:03:07
I think maybe the trap is like conflating motion and progress. I think motion is much easier now than ever before because of the tooling that we have. But progress requires you to be like very prescriptive and deliberate about like what you’re trying to achieve
Akshay Nathan1:08:06
Figures
| Headcount at OpenAI when Akshay joined | About 500 (2023) | 1:33 |
| Number of model and reasoning tier combinations | 32 options | 16:24 |
| ChatGPT Work + Codex users | 10 million | 41:43 |
| ChatGPT overall user base | Hundreds of millions | 42:46 |
| ChatGPT Work availability | Paying users only, and not on by default for ChatGPT users | 42:46 |
| Time comparison for the same task (retirement calculator) | ChatGPT Work 8 minutes, Codex still hadn't finished at the time | 19:29 |
| Time for Vibhu to generate the same tabletop game with goal | 18 minutes 53 seconds | 31:19 |
Glossary
- harness
- The tools, loop, and environment layer wrapping the model, determining what the agent can call and how it executes.
- artifacts
- Structured outputs the model generates and can iterate on repeatedly, such as tables, slides, and web pages.
- Sites
- HTML pages ChatGPT generates and hosts directly, now replacing decks and spreadsheets as deliverables.
- Chronicle
- An experimental feature that learns from your computer usage and writes to memory, off by default.
- at-bats
- A baseball metaphor for the number and quality of times a team completes the full "idea—build—feedback—validate" loop.
- OpenClaw
- A separate open-source personal agent project, used in this episode as a reference point for the personal OS form.
How to listen
PMs and designers building agent products, leads currently rolling out AI tools for non-engineering teams, engineering managers who need to redefine productivity metrics.
26:02–31:19, the two hosts take turns demoing the tabletop game Site they made; low information density.