The Future of Work: AI Generalists, Ideas, and Taste — Akshay Nathan, OpenAI

Summary

• ChatGPT Work converted a developer-agent wedge into a 10 million-user product, but the larger distribution unlock remains ahead. Nick Turley called the launch a “culmination” of Codex’s internal adoption beyond engineering, while stressing that Work is paid-only and not ChatGPT’s default; against ChatGPT’s hundreds of millions of users, “we need to get this to everyone.”
• OpenAI is consolidating around one agent harness while retaining distinct interfaces for coding and general work. Codex and Work share capabilities, including plugins, computer use, and artifacts; Codex foregrounds Git state, diffs, and developer-oriented sandboxing, while Work abstracts those details. The product thesis is explicit: “We should enable users to choose, but we shouldn’t box them in.”
• Enterprise demand is no longer the constraint; converting broad enthusiasm into concrete workflows is. Akshay saw companies establish AI deployment teams with “enormous budgets,” yet their use cases exploded in every direction because a box that accepts anything also leaves users unsure what to do. Even with hundreds of millions already familiar with AI, he sees a “ten X or a hundred X bigger market” that does not yet understand agents.
• Artifacts and Sites are positioning ChatGPT Work above the application layer, where models generate the interface appropriate to each task. Logan Kilpatrick said OpenAI’s model slider was developed almost entirely inside a Site, while an internal corporate-finance team moved recurring reports from decks and spreadsheets into Sites. His premise is that traditional tools eventually hit a boundary, whereas “with a Site you can kinda do anything.”
• Persistent context could become Work’s compounding advantage, but permissions and trust are the corresponding liabilities. Nick said Memory V3 carries personal context between ChatGPT and Work; Adam Fry described Chronicle as another source based on computer activity; and Logan called the context from plugins and local files “deeply personal.” Shawn Wang noted that the human currently acts as a permissions layer when an AI-derived answer contains information the requester may not be authorized to see.
• The rollout sequence is developers, general knowledge workers, and eventually everyone’s personal life. Logan described Codex as starting with friction-tolerant early adopters, with Work adding artifacts, computer use, and guided discovery; persistent files, scheduled tasks, finance, meal planning, and household coordination point toward a personal operating layer. “You see how easily work becomes personal and personal becomes work.”
• AI increases organizational throughput enough to make traditional productivity proxies actively misleading. Adam expects T-shaped workers who can generalize across functions while retaining a specialty, with ideas and taste becoming scarcer than implementation. His management test is the quality and frequency of complete “at-bats”—idea, build, feedback, validation—not tokens, pull requests, or story points: “The trap is conflating motion and progress.”

Deep dive

Not yet available upstream; scheduled sync will retry.