What was the inspiration behind launching ChatGPT Work?

The inspiration came from the unexpected adoption of Codex by non-developers at OpenAI, highlighting the broad applicability of advanced AI tools beyond just coding. This led to the idea of merging these capabilities into a more accessible platform for everyone.

Is the underlying harness of ChatGPT Work the same as Codex?

Yes, the underlying harness is shared between ChatGPT Work and Codex. The differences lie primarily in the user experience (UX) and specific opinionated design decisions, such as how much of the Git state is exposed or how agent thinking is displayed.

How does ChatGPT Work handle artifacts like spreadsheets or websites?

ChatGPT Work introduces 'artifacts' as a key feature for knowledge work. Tools like 'Sites' allow users to create and collaborate on web-based documents and prototypes, offering a higher bandwidth alternative to traditional slide decks and spreadsheets.

What is the role of 'productivity' at OpenAI, and how is it measured?

The productivity team's mission is to enable people to do things they couldn't before, focusing on knowledge work. Measurement is challenging due to the diversity of goals, but the aim is to provide leverage and help users achieve their objectives, moving beyond proxies like code commits.

How does AI change the nature of jobs and team structures?

AI is blurring the lines between traditional product development functions (like engineering, design, product management). It enables individuals to become more generalist ('T-shaped') while also specializing, leading to potentially limitless accomplishments.

What are the biggest bottlenecks in AI development and product creation?

The primary bottleneck is shifting towards ideas and taste. With AI enabling rapid creation, the challenge becomes having enough unique ideas and discerning which ones are worth pursuing, highlighting the continued value of human generalists.

What advice is there for users deciding which AI model (e.g., Soul, Terra) to use?

The general advice is to stick with the default configuration, as it's optimized for most use cases. Power users can explore other options if they seek specific efficiency or quality improvements, but the default is intended to be the best starting point.

How has ChatGPT's memory system evolved, and what is Chronicle?

OpenAI has made significant improvements to the memory system (now V3), allowing ChatGPT to 'know' users and retain context. Chronicle is an experimental feature that learns from computer usage to build even deeper memories and provide proactive insights.

What are sub-agents and how do they impact AI tasks?

Sub-agents are smaller, specialized AI models that can be invoked by a main agent to handle specific parts of a complex task. They can improve efficiency and reduce costs by offloading work to cheaper models, and allow for more parallel processing.

What is the difference between 'motion' and 'progress' in the context of AI adoption?

Motion refers to the increased activity and ease of using new AI tools, while progress is about achieving actual goals. The trap is conflating the two; true progress requires being prescriptive and deliberate about what needs to be achieved, not just engaging with the tools.

How does OpenAI view the future of personal productivity with AI?

The vision is to bring useful AI agents to everyone, extending beyond developers and knowledge workers into personal lives. The goal is to provide leverage, creating time for users to pursue their passions, akin to a 'ChatGPT life'.

Key Moments

OpenAI’s Plan to Make ChatGPT the Everything App — Akshay Nathan, OpenAI

Latent Space Podcast

Science & Technology7 min read71 min video

Jul 28, 2026|6,223 views|115|14

Save to Pod

Want to know something specific about what's covered?

We've already dissected every moment. Ask and we will deliver (with timestamps).

Key Moments

TL;DR

OpenAI's ChatGPT Work aims to be an 'everything app' by merging capabilities like Codex and plugins, but its success hinges on user adoption and understanding its broad applicability beyond traditional work tasks.

Key Insights

Codex unexpectedly saw significant adoption among non-developers at OpenAI, signaling a broader demand for powerful AI tools beyond traditional engineering roles.

ChatGPT Work aims to unify various functionalities, deprecating separate experiences like the browser interface to create a singular 'super app' for productivity.

The core 'harness' for capabilities is shared between Codex and ChatGPT Work, with differences primarily in user experience and opinionated product decisions rather than underlying functionality.

OpenAI is focusing on 'productivity' rather than just 'enterprise' to encompass personal productivity, with an aim to provide leverage and time savings for users.

The concept of 'artifacts' is a key push for ChatGPT Work, encompassing outputs like spreadsheets and sites, with ongoing development for collaboration and richer interactivity.

OpenAI views the evolution from developer tools (Codex) to general knowledge work, and then to 'everyone,' as a sequential expansion of AI's utility.

Bridging the gap between code and accessibility

Akshay Nathan, leading Core Product Engineering at OpenAI, discusses his career's consistent theme: making software power accessible to non-coders. His journey, from no-code tools at Airtable to developing automated testing with AI, culminated in the realization that Large Language Models (LLMs) were the missing piece for truly democratizing code's capabilities. The launch of ChatGPT Work is presented as the manifestation of this long-held vision, aiming to bring the magic of code to everyone without requiring technical expertise. This echoes his earlier work on platforms like Walrus and Airtable, which sought to lower the barrier to entry for complex software functionalities. The transition to OpenAI allowed him to leverage advanced LLM technology to fulfill this mission on a grander scale.

The unexpected power of Codex among non-developers

A significant impetus for the development of ChatGPT Work stemmed from the surprising internal adoption of Codex by non-developers at OpenAI. Despite being initially conceived for developers, teams in finance, marketing, and other non-technical departments began using Codex for their specific use cases. Participants expressed pride in using the tool, feeling like they possessed a 'superpower.' This realization highlighted that the power of AI agents was not confined to engineers and that a massive distribution base of existing ChatGPT users could benefit from these advanced capabilities. The challenge then became how to effectively bring these powerful agents to the familiar ChatGPT interface, leading to the 'merge' strategy and the eventual launch of ChatGPT Work.

ChatGPT Work: The 'super app' for productivity

ChatGPT Work is positioned as the ultimate no-code platform, evolving from earlier iterations like Codex and the separate browser experience. The deprecation of the browser interface signifies a move towards a unified 'super app' model. The decision to merge functionalities from Codex into ChatGPT Work aims to provide a seamless experience, ensuring users don't get 'stuck in a tab or an experience where they don't get the power of the product.' While Codex remains a durable brand, the principle is that users should access the full power of the product regardless of the specific UI affordance. The core 'harness' of capabilities is shared, with differences largely residing in opinionated UX choices and safety defaults, such as how agent thinking is displayed or how sandboxing is implemented. This unified approach is intended to simplify user interaction and leverage.

Defining 'productivity' beyond the enterprise

The team responsible for ChatGPT Work is focused on 'productivity,' a term intentionally chosen over 'enterprise' or 'work.' This broader definition acknowledges personal productivity as well, recognizing that many AI-driven tasks can blur the lines between professional and personal life. An example cited is a user who utilized ChatGPT Work to track down a misplaced package by analyzing images and neighborhood listings. This highlights the agent's tenacity and capability in solving real-world problems. The goal is to provide leverage across all 'worky' or productivity-related activities, empowering individuals to achieve more and create time for their personal pursuits, ultimately aiming for users to leverage AI in every aspect of their lives.

The evolution of the AI harness and user experience

The conversation delves into the evolution of OpenAI's AI 'harness,' comparing the existing ChatGPT harness with the newer Codex harness. The older ChatGPT harness was optimized for latency and personality, ideal for tasks like search and personal communication. In contrast, the Codex harness, designed for knowledge work, leverages an 'infinitely flexible environment as a computer' to perform powerful tasks. While the core capabilities are shared, the UX and the way users interact with the system differ. The strategy is to bring the power of the computer environment to knowledge workers, abstracting away complexities while retaining utility. This is part of a broader vision to ensure AI power is accessible everywhere, meeting users where they are, and continuously working towards making all scenarios equally capable.

Artifacts and the future of collaboration

A significant focus for ChatGPT Work is the concept of 'artifacts,' which refer to generated outputs like spreadsheets, sites, and documents. Recent model improvements (e.g., GPT-4.5 and 5.6) have led to dramatic enhancements in the quality of these artifacts. The product side is also evolving, with features like hosted sites that eliminate the need for users to manage their own webpages. An example provided is the creation of a playable board game interface directly within ChatGPT, showcasing its ability to generate complex, interactive outputs. Internally, 'sites' are becoming the new canonical artifact for team collaboration, replacing traditional slide decks and spreadsheets. This format offers higher bandwidth and flexibility, allowing users to emphasize specific elements and iterate more effectively, pushing the boundaries of what can be created and shared.

Navigating the complexity of AI models and user defaults

With a multitude of AI models and configuration options (like 'Terra,' 'Sol,' and various reasoning levels), users can feel overwhelmed. OpenAI's strategy is to provide a strong, opinionated default that is best for most use cases, aiming for simplicity and optimal performance out-of-the-box. For power users, advanced settings allow for customization of reasoning levels and model classes, though the company acknowledges there might be 'too many toggles' and is working on simplification. The advice for most users is to stick with the default, only exploring other configurations if specific efficiency or quality issues arise. The default aims to balance speed and thoroughness, projecting these complex dimensions onto a user-friendly slider for easier selection.

The blurring lines of work and the 'T-shaped' professional

The rapid advancement of AI is blurring traditional job roles, leading to a more generalized skillset. Akshay Nathan predicts a future where professionals are 'T-shaped'—possessing a broad set of AI-enabled generalist capabilities, complemented by a deep specialty. AI will allow individuals to iterate and explore in areas outside their core expertise, while also enabling them to go deeper into their chosen field. The primary bottleneck in this new era is identified as 'ideas and taste,' as the ability to build is becoming universally accessible. The challenge for individuals and teams will be generating novel ideas grounded in user feedback and market friction, a process that still requires human intuition and strategic thinking, even with AI as a co-pilot.

The future of productivity measurement and team building

Measuring productivity in the age of AI is shifting from traditional proxies like code commits or story points, which are becoming less correlated with actual goal achievement. The focus is moving towards 'at-bats'—the ability of a team to efficiently cycle through the entire product development process, from idea generation and validation to iteration based on feedback. This requires both quantity and quality of attempts, supported by a culture of humility and motivation. A key trap for teams is conflating 'motion' (ease of doing things with new tools) with 'progress' (actual advancement towards goals). Teams need a prescriptive and deliberate view of what progress looks like to avoid this, ensuring that AI tools enhance, rather than merely facilitate, meaningful work.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

ChatGPT Work Best Practices

Practical takeaways from this episode

Do This

Stick to the default model configuration unless you're experiencing issues with quality or efficiency.

Broaden your imagination to explore what's possible with AI, as capabilities are rapidly advancing.

Provide the model with more context and save artifacts over time to increase its value and utility.

Embrace the 'show, don't tell' approach by using AI to tailor demonstrations to user needs.

Focus on 'at-bats' – the full cycle of generating ideas, building, getting feedback, and iterating.

Be prescriptive and deliberate about what progress looks like for your team and how to measure it.

Avoid This

Don't get stuck in a single tab or experience; leverage the integrated capabilities of ChatGPT Work.

Avoid drawing hard boundaries based on user roles; AI blurs the lines between different functions.

Don't expect LLMs to magically generate new ideas; they are best at building on existing foundations and feedback.

Don't conflate motion with progress; focus on achieving defined goals rather than just being busy with new tools.

Avoid presenting AI-generated reviews of people as solely your own work; use AI for context gathering.

Common Questions

ChatGPT Work is designed for productivity and aims to be a 'super app' for work-related tasks. It integrates capabilities from tools like Codex and offers a more persistent environment for complex tasks and collaboration, unlike the more general conversational interface of standard ChatGPT.

Topics

Mindset & Self-Improvement AI & Machine Learning Technology & Innovation AI & Productivity AI Adoption Large Language Models Product Development Future Of Work Agentic AI ChatGPT Features

Mentioned in this video

Software & Apps

Walrus

Mentioned as a no-code/low-code tool Akshay Nathan worked on.

Airtable

Mentioned as a no-code/low-code tool Akshay Nathan worked on, sharing the thesis of making database parameters accessible.

ChatGPT Enterprise

Akshay Nathan learned from working on ChatGPT Enterprise that there's no one-size-fits-all solution in the enterprise space.

ChatGPT

The product Akshay Nathan leads core product engineering for, aiming to be a super app for productivity.

Codex

An earlier OpenAI product that saw significant adoption among non-developers, inspiring the development of ChatGPT Work.

Sites

A feature within ChatGPT Work that serves as a prototyping tool and an artifact for knowledge work, allowing for higher bandwidth collaboration.

Open-Claw

An open-source personal OS that inspired the persistent computer environment in ChatGPT Work and is seen as a potential predecessor.

Wolfram Alpha

Mentioned in the context of financial planning and budgeting, with ChatGPT Work now being able to handle the component of retirement planning.

Companies

OpenAI

The company where Akshay Nathan leads core product engineering for ChatGPT.

Amazon

Mentioned as the courier in an example where ChatGPT Work helped a user locate a missed package.

Media

Strata

A board game that the speaker used as a case study to test ChatGPT Work, 5.6, Sites, and auto research capabilities.

Organizations

San Francisco Chronicle

An experimental feature that learns from computer usage to build deeper memories for ChatGPT, acting as an input source for memory.

Ask anything from this episode.

Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.

Get Started Free