OpenAI found software engineers spend 80% of their time on administrative coordination, prompting a pivot to deploy autonomous background agents across general enterprise knowledge work.
## Background
This video was published on YouTube by OpenAI on May 14, 2026. According to the page description, it features an OpenAI Forum conversation discussing how Codex has expanded beyond software engineering into everyday knowledge work, research, planning, file organization, and data analysis.
## TL;DR
In this discussion, Chris Nicholson speaks with Thibault Sottiaux, Head of Codex at OpenAI, about the evolution of Codex into a general-purpose AI agent for non-technical workers. They explore how model upgrades like GPT-5 enabled long-horizon task execution, allowing users to automate routine workflows, delegate administrative responsibilities, and generate custom tools. Sottiaux demonstrates live applications—from creating interactive web apps to managing daily tasks via voice prompts—while addressing enterprise safety controls and prompting best practices. Ultimately, the conversation highlights how autonomous agents are transforming workplace productivity and mitigating burnout by executing tasks that previously required extensive manual effort.
## Atmosphere & style
The session carries an engaging, forward-thinking, and pragmatic tone, striking a balance between technical insight and accessible product demonstration. Nicholson and Sottiaux share a collaborative, enthusiastic dynamic as they frame AI agents not as futuristic abstractions, but as practical tools already reshaping daily work routines. Listening to the conversation feels like sitting in on an inspiring, behind-the-scenes engineering chat filled with practical tips, live demos, and a shared sense of optimism about the future of creative knowledge work.
## Key takeaways
- [00:09](https://www.youtube.com/watch?v=DLP9CagE3dU&t=9s) Codex originally launched as a cloud-based web tool, but high setup friction for developers and model limitations on complex tasks forced a pivot to local integration.
- [00:33](https://www.youtube.com/watch?v=DLP9CagE3dU&t=33s) Software engineers spend only 20% to 30% of their time writing code, while 70% to 80% is spent on non-coding tasks like bug triaging, system planning, and gathering context.
- [00:33](https://www.youtube.com/watch?v=DLP9CagE3dU&t=33s) Major model upgrades like GPT-5 and GPT-5.2 provided the reliability needed for Codex to expand beyond coding into general knowledge work.
- [00:57](https://www.youtube.com/watch?v=DLP9CagE3dU&t=57s) Accelerating product development through AI shifted organizational bottlenecks downstream to marketing and communications teams struggling to keep pace.
- [01:09](https://www.youtube.com/watch?v=DLP9CagE3dU&t=69s) AI agents enable cross-functional independence, significantly reducing employee reliance on centralized teams like data analysts for routine metrics.
- [01:32](https://www.youtube.com/watch?v=DLP9CagE3dU&t=92s) Non-technical users can build personalized 'home-cooked software' using natural language and voice commands in just minutes.
- [15:18](https://www.youtube.com/watch?v=DLP9CagE3dU&t=918s) Codex can execute tasks autonomously in the background while users focus on other work, scaling from simple personal errands to multi-billion-dollar enterprise workflows.
- [17:45](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1065s) Treating Codex as an automated chief of staff allows users to queue up over 100 daily tasks, ranging from local file organization to compute fleet management.
- [19:30](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1170s) By eliminating administrative overhead, AI agents unlock 'infinity time' tasks—projects that would otherwise never be attempted due to lack of bandwidth.
- [24:00](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1440s) The `/goal` command introduces persistent long-horizon autonomy, enabling agents to work on complex objectives like scientific research or codebase translation for hours, days, or weeks.
- [27:03](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1623s) To ensure success with long-horizon tasks, users must provide explicit evaluation criteria and clearly define what 'done' looks like up front.
- [28:40](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1720s) The primary barrier to enterprise AI adoption is not model capability, but human trust regarding security, safety, and data exfiltration risks.
- [31:05](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1865s) Safety architectures like sandboxing and 'Auto Review' secondary agents help mitigate risks by monitoring and halting high-risk primary agent actions in real time.
- [32:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1976s) Successful adoption relies on treating agents like new hires by providing rich context, leveraging tool plugins, and communicating clear instructions.
- [37:00](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2220s) Users should avoid 'over-delegation'—outsourcing total comprehension to the AI—and instead use active recall strategies to retain critical knowledge.
- [38:54](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2334s) True general agency unlocks latent potential, allowing professionals across industries to bring backlog 'stacks of good ideas' to realization.
## Detailed walkthrough
Here is a detailed walkthrough of this portion of the conversation.
---
[00:09](https://www.youtube.com/watch?v=DLP9CagE3dU&t=9s) Introduction and the Evolution of Codex Beyond Engineering
Chris Nicholson from the OpenAI Global Affairs team opens the event by introducing the OpenAI Forum as a space to discuss practical applications of AI. He introduces Thibault Sottiaux, the Head of Codex at OpenAI, to explore how Codex is expanding beyond software engineering into general knowledge work, personal planning, data analysis, and file organization.
When asked about the origin of Codex, Thibault explains that OpenAI has long been focused on creating models that accelerate internal development, with software engineering serving as a early "grand challenge." About two years prior, the goal was to enable AI to code like a prolific software engineer. This led to the launch of "Codex web" roughly a year before this discussion—a cloud-based environment where users specified a task, and the AI examined a code repository and automatically opened a pull request on GitHub.
However, the team realized the cloud-only web setup was too high-friction for everyday workflows. Developers preferred their custom local setups, and forcing them to recreate those environments in the cloud was burdensome. Moreover, the models at the time were not yet reliable enough for long-horizon tasks without easy, continuous local iteration, prompting the team to pivot toward local machine integration.
* **Original Launch Concept:** Codex web operated in an isolated cloud environment to inspect GitHub repositories and generate pull requests automatically.
* **Key Pivot Driver:** High setup friction and model limitations on long-horizon tasks forced the team to shift focus to running locally on users' machines.
---
[00:33](https://www.youtube.com/watch?v=DLP9CagE3dU&t=33s) Expansion to Non-Coding Knowledge Work and GPT-5
The shift toward non-technical use cases accelerated over the previous six months, driven by major model upgrades like GPT-5 and version 5.2. These updates provided a step-change in reliability and generality for complex, long-horizon tasks. Thibault highlights a fundamental realization about technical work: software engineers actually spend only about **20% to 30%** of their day writing code. The remaining **70% to 80%** involves administrative and coordination tasks such as triaging tickets, architecture discussions, investigating bug reports, handling outages, being on call, and gathering context across systems.
As engineers began using Codex to streamline these non-coding parts of their daily jobs, it became clear that the underlying technology could serve broader knowledge workers. By giving the agent access to broader contextual tools like Notion and internal documentation, its problem-solving abilities improved dramatically. Today, the majority of tasks performed within Codex are actually non-coding tasks.
* **Engineering Time Distribution:** Software engineers spend roughly **20–30%** of their time actually writing code and **70–80%** on context gathering, bug triaging, and system planning.
* **Model Milestones:** The releases of **GPT-5** and **GPT-5.2** were crucial in enabling long-horizon task execution and general-purpose reasoning.
---
[00:57](https://www.youtube.com/watch?v=DLP9CagE3dU&t=57s) Internal Adoption: The Breakthrough Moment and Organizational Ripple Effects
Thibault recalls the specific moment he realized Codex was destined for everyone, not just developers. During a product launch, Alexander Americus, the Lead Product Manager on Codex, used the tool to manage the launch state. Instead of manually digging through channels, Alexander deployed multiple Codex agents to track feature updates, gather user feedback, and chase down team members via Slack for status updates—all while maintaining an auto-updated planning document during a meeting with Thibault.
This capability shifted organizational bottlenecks at OpenAI. As engineers and product managers used Codex to build and ship features faster than ever, the core bottleneck shifted downstream to communications and marketing teams, who struggled to keep pace with the sheer volume of output and maintain a coherent public story. Thibault emphasizes that OpenAI could not maintain its rapid pace of development without Codex.
* **Key Internal Example:** Alexander Americus (Lead PM) used multiple Codex agents connected to Slack to automatically chase status updates, summarize user feedback, and maintain live launch documentation.
* **Departmental Bottleneck Shift:** Accelerating engineering and product design shifted the primary operational bottleneck to marketing and communications teams.
---
[01:09](https://www.youtube.com/watch?v=DLP9CagE3dU&t=69s) The Shift in Professional Roles and Self-Service Data Analysis
The widespread adoption of general-purpose AI agents is fundamentally compressing task timelines across science, engineering, and business operations. Tasks that historically required days—such as conducting in-depth market research, aggregating public sentiment on feature releases, or preparing stakeholder presentations—are compressed into a couple of hours or automated entirely.
This shift empowers individuals to perform cross-functional tasks independently without waiting on central teams. For instance, employee reliance on data analyst teams for routine metrics has dropped significantly. Knowledge workers can now ask Codex to locate relevant business dashboards and analyze performance trends across international markets like India or Korea directly, freeing data analysts to focus on higher-value work.
> "I would have expected like doing an in-depth marketing research or analyzing the public sentiment around a new feature... that would have taken a lot of time... Now that's just really being compressed to like a couple of hours of work."
* **Empowerment Example:** Employees independently query business data across regions (e.g., India, Korea) via Codex rather than joining long queues for data analyst teams.
* **Role Blurring:** User Experience (UX) designers on the Codex team now push code changes directly to refine interfaces without needing engineering prioritization.
Chris and Thibault discuss the emerging era of "home-cooked, personalized software," where non-technical users can instantly build custom tools tailored to their hyper-specific personal or professional needs. To demonstrate this concept, Thibault shares a live example on his screen.
Having moved from Europe to San Francisco, Thibault found local bread prices unusually high. To solve this, he prompted Codex to research the city's top bakeries, compile a comparison spreadsheet of sourdough loaves, and evaluate prices. In 5 minutes, Codex searched the web and built an organized spreadsheet featuring local spots like **Jane the Bakery**, **Arsicault Bakery**, **Tartine Bakery**, and **Neighbor Bakehouse** in Dogpatch (which was identified as having a great sourdough option).
To make the data easier to digest, Thibault then asked Codex to convert the spreadsheet into an interactive web application. Within **4 minutes**, Codex generated a functional webpage displaying an interactive map with clickable bakery locations. Thibault notes that he controls the tool primarily via voice commands, using single prompts to dictate what he wants created while he focuses on other tasks.
```
+---------------------+-----------------------+-------------------+
| Bakery Name | Recommended Item | Location / Notes |
+---------------------+-----------------------+-------------------+
| Jane the Bakery | Sourdough Loaf | San Francisco |
| Arsicault Bakery | Artisan Loaves | San Francisco |
| Tartine Bakery | Country Loaf | San Francisco |
| Neighbor Bakehouse | Sourdough Loaf | Dogpatch, SF |
+---------------------+-----------------------+-------------------+
```
* **Spreadsheet Build Time:** 5 minutes of automated background web research and structured data output.
* **Web App Build Time:** 4 minutes to generate an interactive map interface from the raw dataset.
* **Primary Interaction Mode:** Thibault manages his workflows using hands-free voice prompts.
---
What specific aspects of this discussion would you like to explore further?
Here is a detailed, faithful walkthrough of this portion of the video:
[15:18](https://www.youtube.com/watch?v=DLP9CagE3dU&t=918s) Generating Custom Web Apps for Everyday Decision-Making
Thibault Sottiaux demonstrates how Codex can continuously execute tasks in the background without requiring active human attention. Rather than sitting at the screen waiting, a user can simply issue a prompt—such as finding where to buy a specific loaf of bread in San Francisco along with its pricing—and then switch over to other tasks on their computer. Chris Nicholson notes that this changes the entire paradigm of software: anyone with access to data can now easily build a custom website, analyze the data, visualize it, and share it with friends, family, or colleagues. While this process uses code under the hood, writing code itself feels like an anachronism. The user no longer needs to understand how the agent performs its work; Codex simply uses code as a tool to edit spreadsheet files, generate slide decks, or create full websites.
The two discuss how a project like this used to take a skilled developer an entire weekend during college just to track down information and build a rough prototype. Today, it happens instantaneously, and making adjustments requires zero manual re-coding. Sottiaux illustrates this by modifying his request on the fly: asking Codex to perform the exact same price-versus-quality market analysis for local coffee shops in San Francisco instead of bread. Within minutes, Codex produces a brand-new website reflecting those precise preferences. Nicholson frames this as the universal loop of human decision-making: gathering data, visualizing it, extracting insights about the world, and using those insights to make better choices to reach one's goals. Sottiaux emphasizes that this exact same workflow scales seamlessly from personal errands to complex enterprise problems, noting that OpenAI used this very same process during its latest fundraising round.
- **Key capabilities:** Automatically pulls data, generates visualizations, and builds interactive websites or documents without manual coding.
- **Workflow shift:** Users express natural-language preferences (e.g., optimizing for quality vs. price) and Codex acts on them in the background.
- **Scalability:** Handles small personal tasks (finding local bread/coffee prices) as well as multi-billion-dollar enterprise workflows (OpenAI fundraising analysis).
---
[17:45](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1065s) Codex as an Orchestrator and Personal Chief of Staff
Sottiaux shares his personal workflow as Head of Codex, revealing his sidebar full of tasks queued up before the talk began. He regularly dispatches over 100 individual tasks per day to Codex, treating the agent as a personal coordinator and chief of staff. These automated tasks range from organizing files on his local desktop and managing OpenAI's compute fleet, to analyzing on-call rotations, checking on engineering team health, monitoring upcoming product launch schedules, and proactively flagging at-risk items that require human intervention.
To illustrate how anyone can set up daily routines, Sottiaux walks through a simple automation prompt. By granting Codex access to tools like Gmail, Notion, and Google Calendar, he instructs it to compile a comprehensive morning summary and flag high-risk items every day at 9:00 a.m. automatically delivered to his inbox. Nicholson observes that this fundamentally changes time management: Codex actively helps prioritize focus so human attention is directed only where it matters most, eliminating the low-level administrative tax that used to consume hours every week.
- **Daily task volume:** Sottiaux hands off ~100+ tasks per day to Codex.
- **Admin & Ops duties:** Used for desktop file organization, compute fleet management, monitoring engineering on-call rotations, and tracking launch schedules.
- **Automated daily briefing:** Configured to parse Gmail, Notion, and Calendar at 9:00 a.m. daily, delivering a prioritized summary of risks directly to inbox.
---
[19:30](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1170s) Reclaiming Time and Mitigating Organizational Burnout
Reflecting on what bothered him most before using Codex, Sottiaux explains that he used to abandon valuable small projects or questions simply because he lacked the time to do them himself and felt bad bothering colleagues to pull the data. Now, he can spin up custom internal tools, personal software, and micro-reports without placing demands on anyone else's desk. He shares a personal favorite routine: receiving a tailored, highly curated morning news report built strictly around his specific interests. Additionally, when lower-priority user bug reports pop up on social media (like X/Twitter)—issues that traditionally affect only a handful of users and would otherwise fall through the cracks due to competing priorities—he simply drops them into Codex to investigate and fix.
Nicholson posits that this capability fundamentally shifts our relationship with technology. Rather than saving mere seconds or weeks, Codex unlocks tasks that would have otherwise taken "infinity time" because teams never would have gotten to them. In an era where professionals across all sectors—including doctors and teachers—feel trapped inside administrative software and bogged down by manual data entry, Codex serves as an antidote to information overload and burnout. Sottiaux agrees, describing Codex as a trustworthy partner that shields users from the noise of seven different open apps, allowing them to define their daily goals and let the agent manage the rest.
- **Eliminating "Infinity Time" tasks:** Enables execution of low-priority edge cases, personal software, and niche analyses that previously were never attempted.
- **Bug triage:** Low-impact bug reports found on social media are handed off directly to Codex to investigate and resolve automatically.
- **Burnout reduction:** Acts as a buffer against notification fatigue and administrative overload across fragmented workplace tools.
---
[24:00](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1440s) Long-Horizon Autonomy: The `/goal` Command
Sottiaux highlights a new, advanced capability within Codex designed for long-duration tasks: the `/goal` command. Entering `/goal` mode shifts the agent from short turn-based interactions to executing long-term objectives relentlessly over extended periods. Rather than returning an output in seconds, Codex can work on a complex objective for hours, days, or even weeks until it verifies that the goal has been satisfied. Sottiaux notes that `/goal` is already being deployed for heavy engineering and scientific workloads, such as optimizing software performance, refactoring entire codebases from one programming language to another, and assisting researchers with breakthroughs in mathematics and physics.
Nicholson remarks that "genius is just the ability to think about the same thing longer," noting how rapidly AI capabilities are expanding—moving from 10-minute task horizons just months ago to multi-week autonomous execution today. Sottiaux mentions that while the `/goal` mode initially debuted in the command-line interface (CLI), it will soon roll out across all user surfaces, including the Codex desktop app. Looking ahead, Sottiaux predicts that agents will evolve past turn-based execution entirely, running 24/7 in the background to continuously perform helpful actions, monitor systems, and seek human guidance only when stuck or when higher-level steering is required.
- **`/goal` feature:** A command mode designed for persistent, long-horizon tasks lasting hours, days, or weeks.
- **Use cases:** Whole-codebase language translations, deep code performance optimization, and scientific research in physics and mathematics.
- **Future trajectory:** Shift from turn-based prompt-response interactions to 24/7 background agents that continuously work and self-direct.
---
[27:03](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1623s) How to Prompt for Goal Success
When asked how to ensure complex, long-horizon goals actually succeed, Sottiaux shares practical advice for users transitioning into managing AI agents. He encourages new users to chat openly with Codex to discover its self-aware capabilities, but stresses that the single most critical factor for goal completion is clear evaluation criteria. To succeed, users must explicitly define what "good" or "done" looks like so the agent can independently evaluate its progress and know when it has satisfied the objective.
Sottiaux provides a concrete example of structuring a request for a presentation: rather than vaguely asking for a slide deck, a user should specify exact constraints—such as requesting a 10-slide deck where the first two slides introduce the topic, the middle six provide a deep technical breakdown, and the final two outline open Q&A items. Quantifying goals and being explicit about required outputs mimics how one delegates responsibilities to a human assistant or chief of staff, significantly increasing the likelihood of success.
> "A good trick is to be precise on helping it evaluate its own success. And so, if you're able to describe what good looks like and what solved looks like... then Codex will be able to understand whether it's doing well and whether it has completed the task."
[28:40](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1720s) Community Q&A: Moving Beyond ChatGPT and Enterprise Bottlenecks
Chris Nicholson turns to audience questions submitted by the community. The first question comes from Gal GenKarya, a Principal Engineer in Financial Services, who asks what would motivate non-coders to move from ChatGPT over to Codex. Sottiaux explains that Codex is not meant to replace ChatGPT, but rather to complement it. While ChatGPT remains his go-to tool for quick conversational answers, Codex excels at actively doing work on a computer—manipulating local files, running background automations every few hours, and eliminating the manual "copy and paste" era of working with code and data.
The second question, submitted by Anastasia Chueva, asks what poses the biggest bottleneck to enterprise adoption: model capability, human trust, or organizational process design. Sottiaux firmly rejects the idea that model capability is the limiting factor, asserting that the underlying AI capability is already sufficient. Instead, the primary barrier in corporate environments is human trust—specifically regarding security and safety. Organizational leaders worry about letting autonomous agents run freely across internal networks where they could potentially delete sensitive files, send unauthorized emails, or exfiltrate confidential company data.
- **ChatGPT vs. Codex positioning:** ChatGPT handles rapid q&a and ideation; Codex executes active file operations, continuous background tasks, and direct computer automation.
- **Enterprise adoption bottleneck:** Model capabilities are already adequate; the true barrier is enterprise trust around security, safety, and data exfiltration risks.
[31:05](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1865s) Sandboxing Enterprise Controls and "Auto Review" Agents
Chris Nicholson and Thibault Sottiaux address the critical security concerns around running autonomous AI agents. By default, Codex runs inside a strict sandbox with tight operational controls. Users can constrain the agent's reach to a single folder on a file system or disable network access entirely, giving enterprise teams fine-grained permission over what data is accessed. Thibault draws an analogy to corporate information silos: just as a human employee only gets access to a specific department’s data, a sandboxed agent is restricted to a defined folder and set of allowed actions—such as having read-only access so it can extract and analyze data without the ability to modify or delete original files.
To prevent unintended or high-risk actions, OpenAI introduced an alignment mechanism called "Auto Review," detailed on their research blog. This architecture pairs the primary Codex agent—which is incentivized to complete tasks—with a secondary reviewer agent acting like a "referee or an umpire." This reviewer monitors every action the primary agent takes in real time, stepping in to flag and halt any operation deemed too risky before it executes.
* **Sandboxing Features:** Folder-level scoping, network access toggles, and read-only vs. read-write file system permissions.
* **Auto Review Mechanism:** A secondary agent inspects the primary agent's planned actions and halts high-risk operations, instructing it to seek a safer path.
[32:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1976s) Habits for Success: How Non-Developers Get Results
In response to a community question from Jason DeLuca (Owner and Principal Consultant at Crossing Point IT Solutions) about what separates successful non-technical users from those who get stuck, Thibault highlights three habits. First is actively participating in user communities to share creative use cases. Second is framing prompts with precise instructions rather than vague requests. Thibault suggests treating Codex like a brand-new hire who has zero context about your preferences, organizational norms, or where documents live—they require explicit guidance to perform well.
The third key habit is connecting Codex to a wide array of tools via plugins. With over 100 available plugins covering calendars, documents, and tools like Notion, feeding rich context into the system directly correlates with better results. Chris notes that much of this context exists solely in the user's head, requiring them to "be a good boss" and communicate their goals clearly. Thibault shares his own practice of writing down his daily thoughts, ideas, and goals in local computer files so Codex can continuously read them and align with his mindset.
* **Community & Creativity:** Engaging with other users exposes people to unexpected, high-value workflows.
* **The "New Hire" Mindset:** Providing explicit instructions and background context yields far better output than brief, open-ended prompts.
* **Integration Density:** Connecting tools (calendars, Notion, internal documents) via Codex's 100+ plugins creates a comprehensive workspace context.
* **Personal Alignment:** Thibault keeps local text files of his personal goals and thoughts so Codex can read his mind "by proxy."
[35:28](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2128s) Real-World Computer Use Beyond Coding
Answering a question from Daniel Green (Kansas City AI Collective), Thibault highlights his favorite non-coding applications for Codex's computer-use capabilities. On his personal computer, he relies on Codex for weekly meal planning and automatically ordering the required grocery ingredients. He also uses it to navigate complex operating system settings on Windows or macOS—asking Codex to click through the sub-menus to toggle specific features—as well as tweaking presentation slides and embedding visuals.
For technical teams, computer use serves as an automated quality assurance (QA) engineer. Codex can physically open applications, click through UI workflows, and verify feature functionality automatically.
* **Personal Automation:** Meal planning paired with automated grocery ordering.
* **System Navigation:** Guiding users through complex OS settings by taking control and clicking the necessary menus.
* **Automated QA:** Opening desktop or web applications to perform UI click-testing.
[37:00](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2220s) Over-Delegation and the "Tutor Skill" Solution
Addressing a query from Sai Sri Hamsini (AI/ML Engineer at AdeptMind) about common prompting mistakes, Thibault points to the danger of "over-delegation"—handing off entire workflows without maintaining a personal understanding of the underlying domain. When users delegate their own comprehension, they become ungrounded and lose productivity. Instead of outsourcing all thought, users should leverage Codex to explain concepts, analyze code bases, or synthesize marketing plans into clear diagrams using image generation models like DALL-E / Images V2.
Chris strongly agrees, noting that "whoever does the work does the learning." To counteract passive delegation, Chris built a custom "tutor skill" for his setup. This skill deliberately flips the dynamic by testing Chris on the concepts it just explained, forcing him to engage in active recall and consolidate his learning.
* **The Pitfall of Over-Delegation:** Offloading comprehension causes users to lose their mental model of the problem space.
* **Visual Explanations:** Codex uses image models with strong text-rendering capabilities to turn complex plans or codebases into explanatory diagrams.
* **Active Recall Design:** Chris's custom "tutor skill" quizzes the user on freshly presented material to guarantee genuine knowledge retention.
[38:54](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2334s) General Agency and the Future of Knowledge Work
Looking toward the future, Thibault describes Codex as a foundation for a broad, powerful general agent. As these systems gain access to broader real-world data and execution capabilities, they will enable individuals to complete ambitious projects that were previously impossible or prohibitively time-consuming. The overarching goal is to democratize these capabilities globally, raising the ceiling on what people can dream of achieving.
Chris echoes this sentiment, sharing that physicists, industry project managers, and executives frequently tell him that Codex has allowed them to finally execute "stacks of good ideas" that would otherwise have remained unrealized. Both agree that knowledge work is entering a "golden era" of creative realization.
> "I had a stack of good ideas and I've been able to start work on them and they never would have been realized in the world before. It feels like we're entering a golden era." — **Chris Nicholson**
[40:31](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2431s) Closing Remarks and Upcoming OpenAI Forum Events
Closing out the session, Chris thanks Thibault and emphasizes the value of the OpenAI Forum as a learning community where skills are modeled and shared. He reflects on the product's name, reminding the audience that while "code" is in the title, *codex* historically means "book"—a general term for structured knowledge that applies to all information workers searching for needles in haystacks, visualizing data, or running background tasks.
Chris challenges every attendee to pick a single manual workflow in their daily routine—whether a research brief, project plan, onboarding sequence, or decision memo—and test Codex on it directly. He wraps up the session by thanking the participants and announcing the next OpenAI Forum event on Tuesday, May 26th, titled "How AI Could Detect and Prevent the Next Pandemic" featuring guest speakers from the Gates Foundation.
* **Etymology Note:** *Codex* stems from the historical word for "book," reinforcing its scope beyond software source code.
* **Call to Action:** Try automating one real workflow (onboarding, reporting, decision memos, or research briefs).
* **Next Event:** Tuesday, May 26th — "How AI Could Detect and Prevent the Next Pandemic" with the Gates Foundation.
## Notable quotes
> [01:09](https://www.youtube.com/watch?v=DLP9CagE3dU&t=69s) I would have expected like doing an in-depth marketing research or analyzing the public sentiment around a new feature... that would have taken a lot of time... Now that's just really being compressed to like a couple of hours of work.
> [27:03](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1623s) A good trick is to be precise on helping it evaluate its own success. And so, if you're able to describe what good looks like and what solved looks like... then Codex will be able to understand whether it's doing well and whether it has completed the task. — Thibault Sottiaux
> [38:54](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2334s) I had a stack of good ideas and I've been able to start work on them and they never would have been realized in the world before. It feels like we're entering a golden era. — Chris Nicholson
## Mentioned
Codex, OpenAI, Chris Nicholson, Thibault Sottiaux, OpenAI Global Affairs, GitHub, GPT-5, GPT-5.2, Notion, Alexander Americus, Slack, India, Korea, San Francisco, Europe, Jane the Bakery, Arsicault Bakery, Tartine Bakery, Neighbor Bakehouse, Dogpatch, Gmail, Google Calendar, X/Twitter, Gal GenKarya, Financial Services, ChatGPT, Anastasia Chueva, Jason DeLuca, Crossing Point IT Solutions, Daniel Green, Kansas City AI Collective, Windows, macOS, Sai Sri Hamsini, AdeptMind, DALL-E, Images V2, Gates Foundation
OpenAI's Codex has shifted from a cloud-based code generator into a locally integrated execution agent that primarily handles administrative, analytical, and operational knowledge work. Head of Codex Thibault Sottiaux argues that while model capabilities now reliably support long-horizon tasks, organizational adoption depends on structured delegation habits and rigorous security sandboxing.
From Code Generation to General Coordination (0:33)
Codex was originally conceived around two years prior as an engineering tool, culminating in the cloud-hosted 'Codex web' that opened automated pull requests on GitHub repositories. However, OpenAI pivoted to local machine environments because cloud setups imposed excessive friction and early models struggled with long-horizon iteration. With the release of GPT-5 and GPT-5.2, reliability for extended tasks increased substantially, prompting an expansion into general knowledge work.
Sottiaux observes that software engineers spend only 20% to 30% of their workday writing code, dedicating the remaining 70% to 80% to system planning, bug triaging, outage response, and cross-system coordination. Integrating Codex with tools like Notion revealed that the system could absorb these non-coding burdens. During an internal product launch, Lead Product Manager Alexander Americus used multiple Codex agents to monitor feature updates, gather user feedback, and prompt colleagues on Slack for status reports. According to Sottiaux, the majority of Codex tasks are now non-coding, which accelerated product cycles to the point where internal bottlenecks shifted downstream to communications and marketing teams.
Personalized Software and Compressed Decision Cycles (1:32)
Codex enables what Sottiaux and OpenAI's Chris Nicholson describe as 'home-cooked software'—bespoke, throwaway applications built to satisfy hyper-specific user requirements. In a live demonstration, Sottiaux prompted Codex by voice to research San Francisco sourdough prices, yielding a structured spreadsheet across local bakeries in five minutes, which the agent converted into an interactive web map in four minutes. Sottiaux notes that identical workflows are applied to enterprise contexts, including analytical modeling for OpenAI's own fundraising rounds.
This continuous background execution compresses research and analysis timelines from days into hours. Cross-functional workers increasingly perform self-service data tasks—such as querying performance metrics in international markets like India or Korea—reducing dependencies on dedicated data analysts. Similarly, UX designers on the team push interface adjustments directly through Codex without waiting for engineering prioritization.
Persistent Autonomy and Disciplined Delegation (24:00)
To transition beyond short conversational exchanges, Codex introduced the '/goal' mode, enabling agents to pursue long-term objectives across hours, days, or weeks. The command is currently deployed for complex engineering and scientific tasks, such as translating complete codebases between languages, optimizing software performance, and assisting in physics and mathematics research. Sottiaux personally dispatches more than 100 tasks daily to manage compute resources, organize desktop files, and compile daily 9:00 a.m. risk summaries from Gmail, Notion, and Google Calendar.
Achieving reliable results over extended durations requires explicit delegation. Sottiaux emphasizes that users must supply testable success criteria—defining precisely what 'done' looks like, such as setting slide-by-slide constraints for a presentation—much like onboarding a new hire. Both speakers caution against 'over-delegation,' where users outsource fundamental reasoning and lose their mental model of the domain. To reinforce retention, Nicholson utilizes a custom 'tutor skill' that actively quizzes the user on newly generated material.
Enterprise Trust and Safety Controls (31:05)
When addressing enterprise bottlenecks, Sottiaux contends that model intelligence is no longer the limiting factor. Instead, institutional adoption stalls over security, safety, and the fear that autonomous agents might alter files, send unauthorized communications, or leak confidential corporate data across internal networks.
To manage these risks, Codex operates within configurable sandboxes featuring folder-level access boundaries, network connection toggles, and read-only permissions that prevent unauthorized modification or deletion. OpenAI also implemented an 'Auto Review' alignment mechanism, which pairs the primary task-executing agent with a secondary reviewer agent. Modeled after an umpire, this secondary agent monitors planned actions in real time and halts high-risk operations before execution.
OpenAI frames autonomous agents as an antidote to administrative burnout, unlocking edge-case tasks that teams historically abandoned due to time constraints. Nonetheless, widespread enterprise viability remains contingent on establishing institutional trust through strict execution guardrails and preventing users from completely abdicating critical domain comprehension.
## Background
本视频由 OpenAI 于 2026 年 6 月 3 日发布在 YouTube 平台。视频记录了 OpenAI 全球事务团队的 Chris Nicholson 与 Codex 负责人 Thibault Sottiaux 在 OpenAI Forum 中的深度对谈，展示了 Codex 如何从最初面向开发者的编程辅助工具，演化为涵盖调研、规划、文件整理、自动化分析等广泛日常知识工作的通用 AI 智能体。
## TL;DR
对谈深度复盘了 Codex 从早期隔离的云端网页模式转型为深植本地系统的通用工作流 Agent 的历程，揭示了大模型能力跃迁如何让代码转变为智能体的内部工具。随着工程师、产品经理乃至财务和设计人员全面接入，Codex 已渗透至跨平台日程管理、企业融资数据梳理及长尾缺陷修复等海量非编程场景，并演化出长周期自主执行模式（`/goal`）。同时，讨论详述了默认沙箱隔离与双 Agent 自动审查机制以破解企业信任难题，并重点警示了“过度委托导致思维退化”的陷阱，倡导知识工作者借助人机协同将搁置的构想转化为现实。
## Atmosphere & style
对谈氛围务实、睿智且极具感染力，既有宏大的行业洞察，又穿插着极具生活气息的实机演示（如全语音吐槽旧金山高价面包并在数分钟内生成交互地图）。主持人 Chris 擅长从认知哲学与组织协同的高度提炼反思，而 Codex 负责人 Thibault 则以坦诚亲切的一线工程视角分享团队“吃狗粮”的顿悟时刻与每天调度上百个任务的幕僚长工作流。整场交流在理性剖析技术边界与展望人机协作“黄金时代”之间保持着精妙平衡，给听众带来一种强烈而真实的未来即视感。
## Bulletin
- [09:30](https://www.youtube.com/watch?v=DLP9CagE3dU&t=570s) Codex 战略重心已从云端孤立环境彻底转向深植本地的通用知识工作工具，不仅接管了工程师 70%–80% 的非编码杂务，更深度渗透至 OpenAI 内部跨部门协作与数十亿美元规模的融资统筹中。
- [24:11](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1451s) 随着底层模型长跨度任务可靠性的突破，Codex 引入 `/goal` 等持久执行模式，支持自主运行数小时至数周以攻克高难任务，正推动人机协同从单轮回合制迈向常驻后台的持续自主系统。
- [30:14](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1814s) 企业规模化落地 AI Agent 的核心阻碍在于人类信任与数据安全而非模型能力；OpenAI 采用默认沙箱隔离与 Auto Review 双 Agent 监督制衡机制来规避越权破坏与数据外泄风险。
- [37:29](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2249s) 团队警告切勿陷入“过度委托”导致丧失现实掌控力的陷阱，主张践行“谁负责干活谁获得学习”法则，通过多模态概念可视化与反向提问的辅导模式（Tutor Skill）确保人脑深度参与认知。
## Key takeaways
- [01:21](https://www.youtube.com/watch?v=DLP9CagE3dU&t=81s) Codex 早期完全架设于云端且直接提 PR 的模式因复刻本地复杂开发环境代价高昂、模型纠错摩擦过大而受挫，促使架构全面转向集成于本地工作流与文件系统。
- [03:31](https://www.youtube.com/watch?v=DLP9CagE3dU&t=211s) 真实软件工程中编码仅占工作时间的 20%–30%，其余多为排查与沟通协同；GPT-5 跃迁及集成内部知识库上下文后，Codex 自发演变为处理广泛非编码事务的通用平台。
- [05:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=356s) 产品经理借助调遣小型 Agent 军团实现了跨 Slack、PR 与文档的自主催办与计划同步，彻底打破了工程产出瓶颈，甚至导致组织瓶颈反向转移至沟通与市场叙事团队。
- [10:04](https://www.youtube.com/watch?v=DLP9CagE3dU&t=604s) Codex 消除了跨部门的数据索取壁垒，业务人员可凭自然语言直接调取跨国业务指标，摆脱对共享数据分析团队的需求排期依赖，使“问题持有者”直接转变为“方案构建者”。
- [12:07](https://www.youtube.com/watch?v=DLP9CagE3dU&t=727s) 非编程人员（如 UX 设计师）能够借助 Codex 直接提交和修改功能代码，打破传统工程排期限制，开启了针对个人长尾痛点定制的“家庭自制软件”时代。
- [15:37](https://www.youtube.com/watch?v=DLP9CagE3dU&t=937s) 代码的角色发生了根本性蜕变：它不再是人类直接编写交付的终极产物，而是演化为智能体在后台解决复杂数据呈现、应用构建与逻辑执行的内部中间工具。
- [18:12](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1092s) Codex 可担任团队负责人的“幕僚长”，每天定时穿透 Gmail、Notion 和日程自动化巡检潜在风险与延期事项，将人类的认知注意力重新收束至高价值决策。
- [19:30](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1170s) Agent 带来的颠覆性价值不仅是缩短已有工作的时间，更在于将过去因沟通成本高昂而“耗时无限长”（永远不会动手）的长尾琐事与偶发边缘 Bug 快速转化为现实成果。
- [21:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1316s) Codex 从被动等待点击的冰冷工具转变为可深度委托的“可信赖伙伴”，能够在多应用环境中筑起过滤无用噪音的防护盾，缓解知识工作者的职业倦怠。
- [24:11](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1451s) 进阶指令 `/goal` 赋予 Agent 强大的持久攻坚能力，使其能够脱离即时单次回合应答，在后台连续数小时甚至数周自主拆解并攻坚前沿数学与代码迁移命题。
- [27:39](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1659s) 驾驭自主 Agent 的关键在于像指导业务副手一样清晰定义“完成的形态”（what good looks like），给出量化具体的产物规格（如特定幻灯片结构）以建立评估闭环。
- [28:40](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1720s) ChatGPT 聚焦于快速问答，而 Codex 则通过直接读取并接管本地文件与系统操作，宣告了在本地与终端“复制代码粘贴”旧时代的彻底终结。
- [32:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1976s) 高效使用者的共性习惯在于把 AI 视作毫无背景认知的新入职员工，坚持将个人意图与思考框架完全文字化落盘，并为 Agent 挂载广泛的外部办公插件。
- [37:29](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2249s) 警惕把思考过程全盘外包的“过度委托”陷阱，应善用 Images V2 等能力将抽象方案可视化，并通过反转角色的主动回忆提问来防止思维退化。
- [38:54](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2334s) Codex 的终极愿景是打造具备现实世界行动权限的通用智能体，通过把超人类复杂度任务的边际成本降至极低，释放全球科研与知识工作者积压已久的海量绝妙创意。
## Detailed walkthrough
### [00:09](https://www.youtube.com/watch?v=DLP9CagE3dU&t=9s) 欢迎致辞：Codex 从编程助手迈向通用知识工作工具
OpenAI 全球事务团队的 Chris Nicholson 为本次 OpenAI Forum 论坛开场。他介绍道，该论坛旨在汇聚行业专家，深入探讨人工智能在现实世界中的落地应用。本场对谈的核心议题是 Codex 以及“为何它的意义已远远超出软件工程的范畴”。虽然 Codex 最初是作为面向开发者的辅助工具而诞生，但如今越来越多的知识工作者正将其融入日常工作与个人事务中。用户借助它来消除工作摩擦、处理繁冗杂务，甚至用于剖析复杂问题、梳理庞杂信息以及起草可共享的规划文档。
这使得 Codex 迅速从程序员的专属利器，扩展为研究人员、教育工作者、业务运营人员、中小企业主及各领域管理者广泛依赖的助手。本期对话邀请到了 OpenAI 的 Codex 负责人 Thibault Sottiaux，他将深度复盘 Codex 的演进历程、团队从非技术用户群体中获得的反馈，以及普通用户如何在编程之外场景解锁该工具的潜力。
### [01:21](https://www.youtube.com/watch?v=DLP9CagE3dU&t=81s) Codex 的起源与架构转型：从云端隔离到本地工作流
谈及研发初衷，Thibault Sottiaux 回忆称，团队很早就痴迷于探索如何训练出能加速技术开发自身的高价值模型。在约两年前启动该项目时，核心的宏大挑战在于“让 AI 达到高产软件工程师的编码水平”。大约一年前，团队推出了首个公开发布的 Codex 版本，内部称为“Codex Web”。该版本的运作逻辑完全架设于云端：一个独立的实体在云端运行并提供 Web 交互界面，用户在网页上输入任务描述，Codex 自动检视关联的代码仓库、计算出正确的代码修改方案，并直接在 GitHub 上发起 Pull Request（PR）。
然而，这种完全打包隔离的云端设计在实际使用中遭遇了巨大的阻力。Thibault 坦言，当时的方案使用摩擦过高：其一，环境配置极为繁琐，大多数开发者在自己的本地电脑上都已经搭建了极其精细完善的开发环境，要将这套配置在 OpenAI 的云端环境中逐一复现，代价极其高昂；其二，当时的模型能力尚不足以做到“次次正确”，用户一旦遇到错误很难在云端与模型高效交互和迭代。意识到早期模型还无法在没有干预的情况下稳定完成超长周期（long-horizon）任务后，团队迅速做出关键转向，将 Codex 从远端孤岛搬到了开发者的本地计算机上。
```mermaid
flowchart TD
accTitle: Codex 架构路线演进图
accDescr: 展示 Codex 从最初的云端网页模式演化为本地深植工作流的历程。
subgraph Phase1["初代探索：Codex Web（约一年前）"]
A["Web 界面输入意图"] --> B["云端隔离运行环境"]
B --> C["云端拉取代码并运算"]
C --> D["GitHub 自动提交 PR"]
end
subgraph PainPoints["实践痛点"]
E["本地开发环境难以在云端复刻"]
F["早期模型长跨度任务可靠性不足"]
G["云端调试与协作交互摩擦过大"]
end
subgraph Phase2["架构转向：本地化与通用化"]
H["深植用户本地机器环境"]
I["紧密结合本地工具链与文件系统"]
J["支持快速双向交互与持续迭代"]
end
Phase1 -.-> PainPoints
PainPoints ==> Phase2
```
### [03:31](https://www.youtube.com/watch?v=DLP9CagE3dU&t=211s) 从代码走向日常：软件工程真相与大模型的通用能力跃迁
转变的发生大约始于近六个月内。随着 GPT-5 以及随后的 5.2 版本的发布，模型在通用性和可靠性上实现了质的跃迁，特别是在应对长跨度复杂任务时表现出前所未有的稳健度。与此同时，团队在深度复盘软件工程师的工作本质时发现了一个现实：程序员绝大部分的工作时间其实并不是在写代码。真实的工作切片中，实际敲代码的时间大概只占 20% 到 30%，剩下的 70% 到 80% 全耗费在查看工单、梳理任务优先级、讨论问题解决方案、制定架构决策、排查用户 Bug 报告、分析线上故障（Outage）、参与值班（On-call）以及四处搜集系统信息上。
这种工作属性促使技术人员自发进行“吃狗粮”（Dogfooding，内部试用）。早期接纳 Codex 的技术人员自然而然地开始用它处理日常杂务。为了让它更好地辅助开发，团队尝试为 Codex 接入了团队在 Notion 知识库和各类说明文档中的上下文权限。这一接入带来了意想不到的连锁反应：当 Agent 掌握了组织文档后，它在各类综合任务中的表现大幅跃升。到如今，Codex 用户所执行的任务中，绝大多数已经完全演化为非编程任务。
- **工程师真实时间分配**：真正用于编码的时间仅占 20%–30%，其余时间均被沟通、文档梳理、排查故障与跨部门协调占据。
- **底层驱动模型**：GPT-5 带来了通用性跃迁，随后的 5.2 版本进一步攻克了长任务跨度下的连续执行可靠性。
- **知识集成放大器**：通过连接团队 Notion 与本地文档库，Agent 的信息整合效率获得了指数级提升。
### [05:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=356s) 顿悟时刻：产品经理的“Agent 军团”与生产力瓶颈的外溢
当 Chris 问及团队何时意识到“Codex 将走向全人类、全行业”时，Thibault 分享了一个关键的转折时刻。在一次 Codex 版本的发布冲刺期，Codex 首席产品经理 Alexander Americus 的工作状态震撼了团队。为了确保各项预定改动顺利落地，Alexander 并没有像传统项目经理那样疲于在不同 Slack 频道、文档和 GitHub PR 之间来回穿梭检索，而是调遣了一大批小型 Codex Agent 作为自己的工作代理。
这些 Agent 自动化地处理了所有令人头疼的催办协调工作：Codex 接入了 Slack，能自主给相关责任人发消息询问某项功能的最新进展，同时汇总所有渠道收集的用户反馈与开发动态，实时整理并更新一份清晰、严密的发布计划。Thibault 惊叹道，当时 Alexander 正在与他面对面交流，而后台的一众 Agent 正在不知疲倦地帮他“追进度、汇信息”。那一刻让团队彻底意识到，他们正在重构的不只是软件工程，而是知识协同的全貌。
```mermaid
flowchart LR
accTitle: Alexander 借助 Codex 实现的自动化项目协同模式
accDescr: 展示产品经理如何通过多 Agent 协作将信息搜集、跟进与文档维护全流程自动化。
subgraph DataSources["分散的数据与沟通渠道"]
A["Slack 团队沟通"]
B["GitHub PR 变更"]
C["用户体验反馈"]
D["内部备忘与文档"]
end
subgraph MultiAgents["Codex Agent 集群"]
E["自主向责任人发信催办进度"]
F["实时提取与归纳关键信息"]
G["自动对齐差异与识别未决问题"]
end
subgraph Output["交付物与人员状态"]
H["动态更新的发布主计划文档"]
I["PM 专注高层决策与战略沟通"]
end
DataSources <--> MultiAgents
MultiAgents --> H
H --> I
```
这种生产力的剧变在 OpenAI 内部引发了强烈的涟漪效应。随着工程师与核心团队的产出效率被极度放大，设计师与产品经理的角色全面转变，组织原有的工程瓶颈被彻底粉碎，但瓶颈随之转移到了沟通与市场营销（Communications & Marketing）部门。团队交付新成果的频率和体量激增，导致如何向外部世界条理清晰地讲述发布故事成为了新的挑战。当 Chris 询问“如果没有 Codex，OpenAI 还能否维持如此惊人的迭代速度”时，Thibault 给出了斩钉截铁的回答：“不可能，如今 Codex 对我们而言是不可或缺的底座。”他认为，这项技术已成熟到足以承担高度通用的桌面复杂工作，这种高效范式完全可以复刻到全球各类行业与团队中。
### [09:30](https://www.youtube.com/watch?v=DLP9CagE3dU&t=570s) 从代码生成到通用任务：Codex 与 OpenAI 的最新融资
Codex 的应用场景早已跨越了最初的软件工程边界。在日常工作中，无论是准备汇报演示以拉齐各方利益相关者（stakeholders）、搜集公众舆论反馈，还是进行深入的市场调研与组织归档，Codex 都扮演着核心角色。在 OpenAI 内部，财务部门也在重度使用它；OpenAI 首席财务官 Sarah Friar 就多次提到，团队在 Codex 的辅助下完成了最新一轮融资的组织协调工作，这在整个融资过程中起到了巨大的加速作用。Thibault Sottiaux 指出，Codex 展现出了一种“绝妙的通用性”（beautiful generality）——它如今的演进方向根本不是单纯的生成代码，而是用于执行广泛的通用任务。
- **多场景渗透**：从市场调研、公众认知情绪分析到高管汇报对齐，日常知识工作的准备流程正被全面重塑。
- **高管背书**：OpenAI 财务负责人 Sarah Friar 将 Codex 作为最新一轮融资的关键加速工具，用于资产归档与数据梳理。
---
### [10:04](https://www.youtube.com/watch?v=DLP9CagE3dU&t=604s) 组织角色的演变与数据分析团队的“解放”
随着工具的进化，职场角色和工作节奏正在发生剧烈转变。最显著的改变在于“一切都在变得快得多”，人们必须以极高的敏捷度去适应这种提速。以往需要耗费数天才能解决的硬核难题，如今往往被压缩到短短几个小时。不仅在科学研究与工程领域如此，在市场分析中也是一样：过去要针对某项新功能做深入的公众情感分析或市场研判，需要逐个排查信息源、归纳要点并整理成各部门可消化的报告；现在这套流程在两小时内就能全自动化完成。
这种变化带来了极为强烈的赋能感，打破了跨部门协作的沟通壁垒。过去，员工遇到业务数据疑问——比如“我们在某个市场表现如何？”“我们在印度的增长势头怎样？”“韩国那边的业务情况如何？”——通常需要向数据分析团队提需求排期。而现在，任何人都能直接唤出 Codex：“嘿，帮我把仪表盘调出来，我不知道它具体存哪了，你去搞定它。”随后员工便能自主挖掘业务细节。Chris Nicholson 回忆道，以往数据分析师作为各部门共享的公共资源，需求队列总是排得老长；如今只要业务人员能接触到原始数据，就能独立提取洞察，从而把数据分析师从繁琐细碎的基础查询中解脱出来，让他们专注于更有深度的工作。
```mermaid
flowchart TD
accTitle: 传统数据查询模式与 Codex 赋能模式的对比
accDescr: 流程图展示了传统模式下业务人员需排队等待数据团队，而 Codex 模式下业务人员可通过代理直接获得洞察。
subgraph Traditional["传统协作流程（耗时数天）"]
A["业务人员产生疑问"] --> B["向数据团队提交工单"]
B --> C["排队等待共享资源"]
C --> D["数据分析师提取并反馈"]
end
subgraph Agentic["Codex 赋能流程（耗时数分钟）"]
E["业务人员自然语言提问"] --> F["Codex 自动定位仪表盘与数据"]
F --> G["自主生成洞察与决策支持"]
end
```
- **问题持有者即方案构建者**：过去“提问题的人”与“建方案的人”彼此割裂，需求沟通周期漫长且成果往往妥协；如今同一名员工能直接完成从发现问题到构建解决方案的敏捷迭代。
- **自主数据探索**：员工无需知晓仪表盘的精确存放路径，仅凭业务意图即可调取印度、韩国等区域市场的深度指标。
---
### [12:07](https://www.youtube.com/watch?v=DLP9CagE3dU&t=727s) 消除沟通隔阂：直接写代码的 UX 设计师与“家庭自制软件”时代
在产品开发一线，这种能力的下沉引发了设计与工程关系的重构。OpenAI 的团队中甚至连 UX 设计师也开始直接提交代码、修改功能，与软件工程师并肩塑造产品。以往，设计师若想推动一些细节改动，必须花精力去说服工程团队将其排入优先级，而工程师往往觉得这些只是“无关痛痒的细枝末节”；如今设计师能够亲自动手实现心中的构想，将产品的工艺细节（craft）与用户体验拔高到足以服务数百万用户的标准。
Chris 将这一现象概括为我们正在步入“家庭自制、个性化软件”（home-cooked, personalized software）的时代。Thibault 对此深表赞同，认为未来每个人都能拥有、维护完全贴合自身独特需求的定制化工具，软件将精准地实现使用者想要的每项微小功能。
- **提升工艺标准**：设计师不再受制于工程排期，可以直接在代码仓库中推行交互细节优化。
- **软件个性化浪潮**：软件生产从大型标准化工业品，逐步转向个人按需定制的“私房菜”工具。
---
### [13:20](https://www.youtube.com/watch?v=DLP9CagE3dU&t=800s) 旧金山天价面包实录：十分钟生成交互式网页与全语音交互
为了直观展现这种个性化软件的构建过程，Thibault 现场展示了一个自己为了解决生活问题而做的小程序。作为刚从欧洲搬到旧金山的新居民，他忍不住吐槽旧金山的面包价格“贵得离谱”（outrageous）。于是他向 Codex 发出指令：“旧金山哪里的面包最好吃？帮我做个电子表格，列出各大烘焙坊的招牌面包、购买地点以及对应的价格。”
Codex 自主运行了大约 5 分钟，便整理出了一张详尽的表格，收录了 Jane The Bakery、Arsicault Bakery、Tartine Bakery 以及位于 Dogpatch 区的 Neighbor Bakehouse 等名店的酸种面包（sourdough loaf）价格与描述。但 Thibault 不满足于死板的表格数据，希望以更直观的方式查看，便追问一句：“你能帮我直接做成一个网页吗？”仅仅又过了 4 分钟，Codex 就生成了一个嵌有地图标记的交互式网页，支持点击查看不同烘焙坊的图文信息。
更为惊艳的是 Thibault 的日常交互习惯——他完全使用语音控制，全程无需碰键盘。他甚至现场示范对着麦克风交代：“嘿，我在旧金山，我很喜欢吃面包，帮我把全城能买到的招牌面包做成一张地图，把购买地点和单价结构化整理出来。”在发出语音指令后，他便直接转头去处理电脑上的其他事务，把注意力完全解放出来，无需在屏幕前焦虑等待。
- **极速交付周期**：从单条自然语言提示词出发，5 分钟抓取清洗数据生成表格，4 分钟构建可视化地图网页，全程不到 10 分钟。
- **全语音异步工作流**：用户以纯语音表达复杂结构化需求后即可转入后台并行工作，无需主动盯盘。
---
### [15:37](https://www.youtube.com/watch?v=DLP9CagE3dU&t=937s) “用嘴写代码”：将代码作为代理工具与从面包到咖啡的极速迭代
针对这种变化，Chris 感慨过去人们总在讨论“编写代码”（writing code），现在这种说法似乎已经成了一种过时的历史名词，大家实际上是在“口述代码”（speaking code）。Thibault 纠正并深化了这一观点：底层确实依然在运行代码，但代码现在变成了**智能体手中的一项工具**。用户甚至根本不需要懂代码，也不必关心 Codex 底层是如何调动技术栈的，它会自动创建电子表格、演示文档、幻灯片或网页应用；正是由于把代码视作工具，系统才获得了这种普适的通用性。
系统的另一大威力在于极低的修改与试错成本。如果用户改变主意，只需补上一句：“嘿，帮我把这套分析逻辑完全复制一份，换成旧金山的咖啡店，不要只做面包。”系统就会在几分钟内生成一个专门针对旧金山精品咖啡的全新交互网站。用户可以任意输入自己权衡“性价比与风味品质”的主观偏好，Codex 会将这些偏好全部吸纳进决策逻辑中。
Chris 总结称，这形成了一个普世的生活决策闭环：获取底层数据 → 数据可视化呈现 → 提炼世界运行的洞察 → 做出更符合自身目标的决定。Thibault 强调，这套闭环的精妙之处在于它兼顾了微观与宏观：无论是寻找一根高性价比的酸种法棍，还是前文提到的 OpenAI 几十亿美元融资的筹备归集，底层依赖的都是完全一致的通用认知循环。
- **代码工具化**：代码由人类直接撰写的“最终产物”转变为 AI 代理用于解决问题的“内部中间介质”。
- **多维度权衡优化**：系统能够轻松处理“品质与价格平衡”等主观偏好，而非机械地返回最便宜的选项。
---
### [17:46](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1066s) 任务指挥官：每天下发上百个代理任务的全新工作流
对话最后，Chris 半开玩笑地询问 Thibault，如今的工作方式是不是已经变成了像亚历山大大帝（Alexander the Great）统领军团那样，只需坐镇后方进行调度与编排。
Thibault 展示了自己的操作界面侧边栏，坦言在开场对谈前的片刻时间里，他就已经派发了多个自动化任务。他目前的常规工作状态是每天向 Codex 派发超过 100 个独立任务，将其作为全能的数字协作者。这些任务的跨度极大，小到整理散乱的桌面文件，大到动态管理复杂的算力资源配给，全部交由 Codex 在后台异步处理完成。
- **超高吞吐调度**：个人日均下发并托管超过 100 项独立任务。
- **操作跨度**：覆盖本地文件自动整理归档、系统级任务处置以及后台算力（compute）调配。
### [18:12](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1092s) “幕僚长”式的日常自动化与注意力解放
Thibault Sottiaux 详细描绘了他每天如何将 Codex 当作自己的“幕僚长”（Chief of Staff）来使用。作为一个管理工程团队的负责人，他需要统揽整个集群舰队（Fleet）、轮值排班（On-call Rotation）状态、工程师的工作负荷以及即将上线的发布计划。现在，他只需让 Codex 每天执行一项跨平台自动化流程：自动扫描检索 Gmail 邮件、Notion 文档和日程日历，提取出整日工作摘要，并重点标出潜在的延期风险或阻碍事项。这项任务被设定在每天上午 9 点定时触发，报告准时递送到收件箱中。
Chris Nicholson 感叹道，这种机制不仅是在列待办清单，更是在主动帮助人类确立优先级。Thibault 极为赞同，指出这彻底改变了他分配注意力的方式。在没有 Codex 之前，许多琐碎的事情需要耗费数小时的繁琐电脑操作，或者因为抽不出整块时间而最终彻底被抛诸脑后。现在，通过委托给 Codex，日常运维的机械劳动被全部接管，注意力得以完全收束在真正关键的核心决策上。
```mermaid
flowchart LR
accTitle: Codex 作为幕僚长的工作流程
accDescr: 展示 Codex 聚合多数据源、自主评估风险并在每日固定时间交付简报的自动化闭环
subgraph DataSources ["日常数据源输入"]
G["Gmail 邮件"]
N["Notion 知识库"]
C["Calendar 日程"]
end
subgraph AgentAction ["Codex 幕僚长代理"]
direction TB
Agg["交叉聚合 & 分析"]
Risk["识别阻塞与潜在风险"]
Sched["按计划定时触发 (每日 9:00 AM)"]
Agg --> Risk
Sched --> Agg
end
subgraph Output ["交付输出"]
Brief["每日风险摘要简报\n直达收件箱"]
end
G --> Agg
N --> Agg
C --> Agg
Risk --> Brief
```
- **每日定时巡检**：每天固定在早晨 9:00 自动执行聚合任务，生成包含当日日程、待办与风险提示的晨报。
- **跨平台打通**：一站式穿透 Gmail、Notion 和日历，无需人工在不同应用间来回切换整理。
- **重新锚定工作重心**：免去繁琐的手工数据整合，帮助管理者把注意力精准投放在真正重要的问题上。
---
### [19:30](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1170s) 消除拖延与打扰他人之苦：从“无限时间”到微型软件与长尾修复
当 Chris 问及在 Codex 出现前最令人深恶痛绝的痛点时，Thibault 坦言，最折磨人的是那些“本来想做但不得不放弃”的事情。过去许多需求自己没空写脚本搞定，但去麻烦同事又显得小题大做，于是事情只能一拖再拖。Chris 精辟地总结道，人们常常宣称 AI 把“几周的工作缩短到几秒”，但实际上更多的场景是“把耗时无限长（永远不会动手）的事情变成了现实”。
如今，不仅构建专属的个人小工具（Personal Software）变得随手可得，面对社交媒体上的用户反馈也完全变了样。以往 Thibault 在 Twitter（X）上偶遇用户反馈的边缘 Bug 时，往往会因为“全球可能总共就三个人会遇到”而将其归入最低优先级，最终任其淹没在积压列表中；现在他可以直接把问题丢给 Codex，让它自主定位并修复。这种“不让任何事情从缝隙中滑落”的安全感，极大地降低了日常工作中的认知负荷（Cognitive Load），甚至连每天早晨依据个人独特偏好定制的新闻速递，都带来了一种纯粹的探索乐趣。
- **告别社交心理负担**：不再需要为了零碎的信息搜集或临时报表而去打扰同事，自己即可随时在 Codex 中按需构建。
- **拯救长尾偶发缺陷**：过去影响极小、通常直接被弃置的边缘 Bug，如今可以无负担地交给 Codex 闭环解决。
- **大幅减轻认知负荷**：杂事有条不紊地自动完成，不必在后台大脑进程中时刻担心有事项遗漏。
---
### [21:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1316s) 重新定义人与工具的关系：从信息过载与职业倦怠中突围
Chris 将话题引向了更广泛的职场现实：如今教师、医生等各行各业的专业人员，往往都被深陷在手工录入数据等工具牢笼中，本应提供帮助的软件反而加剧了倦怠感（Burnout）。Thibault 认为，Codex 正在引发人与工具关系的根本性转变——它不再是一个被动等待点击的冰冷软件，而是一个高度值得托付的“可信赖伙伴”（Trustworthy Partner）。
这种信任体现在深层委派能力上：你将一项工作交给它，确信它能够以不输甚至超越人类亲自操刀的质量标准交付；如果遇到阻碍，它会及时向上汇报并请求输入，同时它还能充当抵御信息洪流的防护盾（Shield），将七八个不同应用的噪音拦截在外。在 Thibault 的愿景中，未来的知识工作者甚至完全不需要亲自阅读繁杂的邮件收件箱，私人智能体会负责全量翻阅、提取要点、协助决策，把原本要在十几个干草堆里找针的折磨，变成一份条理清晰的专属晨间资讯。
> “在七八个不同的应用中反复翻找、被铺天盖地根本不关心的噪音狂轰滥炸的日子将成为过去……未来我甚至不必亲自去读邮件，私人智能体会替我阅读收件箱，只有在真正重要、需要决策时才提示我。”
---
### [24:11](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1451s) `/goal` 与时间维度的跃迁：从短周期响应到连续自主演进
谈及近几个月代理能力的飞速演进，Thibault 揭秘了 Codex 最新推出的进阶指令模式——`/goal`。Codex 内部原生支持各类 Slash 命令以切换工作范式，而 `/goal` 能够让代理进入一个具备持久韧性（Relentless）的高级长周期模式。过去人们还在为 AI 代理能够持续自主工作 10 分钟而雀跃，如今给出一个极其艰深的数学猜想、复杂的代码重构或大型跨语言重写任务，Codex 能够自主拆解、排错并持续运算数小时、数天甚至数周，直至达成既定目标。
目前 `/goal` 已在 CLI 命令行界面中先行应用，正逐步拓展到图形化应用程序中。两人展望道，当前的智能体大多还是基于回合制（Turn-based）的被动响应，而真正的技术飞跃将是持续不断、7×24 小时运行且随时接受人类掌舵的自主系统。即便在人类没有输入具体指令的间隙，它也能在后台主动巡检、处理有益事务，完成任务后自主休眠待命。针对如何精准定义目标，Thibault 建议初次接触的用户不必拘谨，完全可以像与真人交谈一样直接向 Codex 询问其自身的能力边界，探索的关键正在于对最终期望的精准界定。
- **Slash 命令体系**：Codex 内部集成了多种模式切换指令，`/goal` 专为长周期、高难度的复杂目标而设计。
- **任务耐力跨越**：从原先单次几分钟的问答交互，进化到支持连续运行数小时、数天甚至数周攻克高阶命题。
- **落地应用前沿**：已被前沿研究者用于硬核数学与物理命题推导、跨编程语言完整工程代码迁移以及底层性能深度调优。
- **走向常驻自主系统**：从单轮的“找某地某种面包的价格”类任务，迈向全天候在后台自主发现价值、主动推进事务的持续常驻模式。
### [27:39](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1659s) 明确定义“完成”标准：像管理下属一样管理 AI
对话接续了如何帮助 Codex 自主评估任务成功与否的讨论。Thibault Sottiaux 强调，关键在于用户能否清晰描述出“好的结果是什么样”（what good looks like）、“问题解决时的形态”（what solved looks like）以及最终交付时你期望看到的具体产物。如果能给出清晰的标准，Codex 就能自主判断当前进展是否良好、任务是否真正完成。
主持人 Chris Nicholson 追问是否可以通过量化指标来实现精确沟通。Thibault 给出了一个制作演示文稿的经典案例：例如用户可以具体要求制作一个包含 10 张幻灯片的演示文稿（slide deck），前两张幻灯片放置特定背景信息，接下来的六张幻灯片深入核心问题并进行技术拆解，最后两张幻灯片则留给未决问题与问答互动（Q&A）。只要你对交付成果的要求足够具体明确，成功的几率就会大幅提升。Chris 赞同道，这与现实工作中指导业务助理或副手（lieutenant）的逻辑如出一辙。
---
### [28:40](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1720s) 从 ChatGPT 到 Codex：告别“复制代码粘贴”的时代
现场进入社区提问环节。金融服务领域的首席工程师 Gal Gen Karya 提问：对于非编程人员而言，究竟是什么动力会促使他们从 ChatGPT 转向使用 Codex？
Thibault 认为这是一个渐进演变的生态过程。他并不认为所有人都会彻底离开 ChatGPT，两者更像互补关系：
- **ChatGPT** 依然是快速获取解答的日常首选工具。
- **Codex** 的真正价值在于“替你动手做具体的事情”（do things for you）。只要涉及本地电脑的文件操作、后台自动运行任务或每隔数小时自动批处理工作，Codex 都是更适用的选择。
Chris 感慨在过去的时代，人们使用 ChatGPT 时不得不把生成的代码复制、粘贴到本地文件或终端中运行；如今“复制粘贴的时代已经彻底终结”（The copy and paste era is over）。Codex 可以直接在本地操作代码与数据。Thibault 补充道，无论电脑里存的是文件还是图片，直接让 Codex 去读取处理即可，用户再也不需要手动到处点击搬运。
---
### [30:14](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1814s) 企业落地的瓶颈与安全体系：沙箱隔离与双 Agent 自动审查机制
Anastasia Chueva 提问：阻碍企业规模化采纳 AI Agent 的最大瓶颈究竟是什么？选项包括：模型能力（model capability）、人类信任（human trust）还是组织流程设计（organizational process design）？
Thibault 直言瓶颈绝非模型能力，现有的能力储备完全足够。核心阻碍在于**人类信任**，而信任的前提是系统的**安全性与保密性**（safe and secure）。如果放任一个 Agent 在企业内网肆意乱窜，误删核心敏感文件，或者擅自把机密数据写进邮件外发泄密（exfiltrate data），任何企业都不可能敢用。
针对这一问题，OpenAI 采取了多层防御机制：
```mermaid
flowchart TD
accTitle: Codex Enterprise Security Architecture
accDescr: Diagram showing how Codex operates within sandboxes and is supervised by an Auto Review agent.
subgraph UserBoundary["用户与权限边界"]
Admin["管理员 / IT 策略"]
Constraints["沙箱限制\n(文件夹隔离 / 禁用网络 / 只读权限)"]
end
subgraph Execution["Agent 执行层"]
PrimaryAgent["主 Agent (Codex)\n目标：全力完成交付任务"]
RiskyAction{"操作风险评估"}
end
subgraph Oversight["监督与审计层"]
ReviewAgent["Auto Review Agent\n(充当裁判员 / 审计员)"]
SafeExecution["安全执行并落盘"]
BlockedAction["拦截高危操作并纠偏"]
end
Admin --> Constraints
Constraints --> PrimaryAgent
PrimaryAgent --> RiskyAction
RiskyAction --> ReviewAgent
ReviewAgent -- "动作合规" --> SafeExecution
ReviewAgent -- "存在泄密/破坏风险" --> BlockedAction
```
1. **默认沙箱机制（Default Sandbox）**：Agent 默认在受限沙箱中运行，权限受到严格管控。企业可以将其限制在特定文件夹内，或者彻底禁用其外网访问权限。对于非技术人员，Thibault 解释沙箱就像企业内部的部门信息壁垒（silos），IT 部门可以设定 Agent 只能读取某些数据以进行分析提取，但绝不允许覆写或删除原数据。
2. **Auto Review 机制（双 Agent 监督）**：OpenAI 在对齐（Alignment）技术博客中披露了该功能。主 Agent 的优化目标是全力帮用户搞定任务，有时可能会采取激进高危的操作；因此系统引入了独立的监督 Agent（Auto Review Agent），它像球场上的“裁判员”（referee / umpire），实时盯梢主 Agent 的每一步动作。一旦发现高风险行为，裁判员会当场喊停并发出警告：“立即住手，这个操作风险过高，请换一种低风险方案。”
---
### [32:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1976s) 高效使用者的共性习惯：把想法写下来，提供充分上下文
Crossing Point IT Solutions 的主理人兼首席顾问 Jason DeLuca 提问：那些让 Codex 发挥出卓越效能的非开发者平时做对了什么？拉开“拿到亮眼成果”与“卡住受挫”之间差距的习惯是什么？
Thibault 总结了三条核心实践经验：
1. **融入社区与保持创意互动**：积极交流并参考他人的成功案例，社区中涌现出的许多全新玩法连开发团队自身都感到意想不到。
2. **给出精确指令，视 AI 为新入职员工**：放弃模糊笼统的提问，明确写下期望结果。应将 Agent 视为一个刚刚入职公司的新同事——他对你的喜好、工作习惯没有任何预设背景。管理新员工时，你必须告诉他去哪里查找资料、如何看待核心逻辑以及该阅读哪些文档；对待 Agent 同样不能偷懒。
3. **连接多元工具与外部数据源**：Codex 目前已支持超过 100 款插件（Plugins），涵盖日历、文档、Notion 以及各类日常办公软件。接入的信息与工具越多，Agent 展现出的能力就越强大。
针对 Chris 提到的“很多目标和经验其实深藏在人脑中，需要像一个优秀管理者那样表达出来”，Thibault 分享了自己的私人工作流：他现在养成了**把一切都记录成文（writing everything down）**的习惯。他将个人的思考框架、规划与目标全部保存在本地文件中并授权 Codex 读取，从而让 AI 深度理解自身意图。“它毕竟无法直接读取你的大脑”，想要 Agent 完美协同，就必须将其用文字显式表达出来。
---
### [35:28](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2128s) 现实世界中的非编程妙用与 Prompting 常见陷阱
Kansas City AI Collective 的 Daniel Green 提问：除了传统编程之外，你最喜欢的电脑端真实应用场景（computer use）有哪些？
Thibault 分享了几个生动的生活与办公案例：
- **日常买菜与餐饮规划**：在个人电脑上让 Codex 制定一周的膳食计划（meal planning），随后由它直接去电商平台下单采购食材。
- **系统设置导航助手**：在 Windows 或 macOS 极其繁琐的设置菜单中寻找特定开关往往令人抓狂。直接对 Codex 说“我想改某个设置，教我怎么过去”，Codex 就会接管界面并逐步点击示范，顺便帮用户熟悉系统。
- **界面微调与幻灯片排版**：利用视觉控制能力进行 PPT 幻灯片排版调整或图片嵌入。
- **软件质量测试（QA）**：这是技术团队中极受欢迎的场景，Codex 可以直接启动客户端应用，模拟真实用户进行点击测试，验证软件各项功能是否正常。
随后，AdeptMind 的 AI/ML 工程师 Sai Sri Hamsini 提出了大家普遍关心的问题：人们在给 Codex 编写 Prompt 时，最容易犯的大忌是什么？Thibault 在篇末指出，随着用户将任务委托给 Codex 的程度越来越深，很多人会滑入一个极其诱人的陷阱——企图将所有事情全部甩手甩出去，甚至包括把原本属于自己的思考过程也一并委托掉……
### [37:29](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2249s) 过度委托的陷阱与“辅导技能”：用 AI 辅助理解而非放弃思考
Thibault Sottiaux 明确指出，当前使用者最容易走入的一个误区就是“过度委托（too much delegation）”，而没有充分利用 Codex 来提升自己对问题的深层认知。如果用户完全抱着甩手掌柜的心态把一切工作都委派出去，很快就会陷入危险境地——你可能根本不知道系统底层究竟发生了什么，失去对业务脉络的现实掌控（grounding），长远来看不仅无法保持高效，反而会损耗个人的实际生产力。相反，正确的做法是花费更多精力让 Codex 充当解释者与知识梳理者。比如，利用多模态生成能力（如 Images V2 对文本渲染与结构图的高效支持），将复杂的新品发布方案、市场营销材料乃至庞大的工程代码库直接转化为清晰易懂的概念图表，帮助自己在研究过程中迅速吃透逻辑。
Chris Nicholson 对此深表赞同，并总结出一句发人深省的学习法则：“谁负责干活，谁才真正获得了学习（Whoever does the work does the learning）。”为了避免自己在 AI 时代退化为被动的信息接收者，Chris 分享了自己构建的一个实用小工具——“辅导技能（tutor skill）”。该技能直接颠倒了人机互动的主客体关系（flips the tables）：不再是人类问 AI 答，而是由 AI 根据刚刚讲授过的知识点反向向 Chris 提问，强迫他运用主动回忆（active recall）来作答，以此确保核心逻辑真正内化到人类大脑中。
- **过度委托的代价**：脱离问题底层脉络，丧失自主判断力与敏锐度，最终反噬工作生产力。
- **Images V2 的可视化实践**：凭借出色的文字排版与图表渲染能力，帮助用户迅速将抽象计划（如发版规划、代码库结构）转化为概念示意图。
- **反转提问的辅导模式**：利用自定义技能构建主动回忆机制，迫使人类在人机协同中保持思维深度。
### [38:54](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2334s) 超越软件工程的终极愿景：释放全球创造力的“黄金时代”
随后，对谈探讨了 Codex 超越传统编程边界后的终极走向与未来愿景。Thibault 强调，OpenAI 的本质目标是打造一个极其通用、极其强大的“行动智能体（general and powerful agent）”。当这个智能体接入人类世界源源不断的信息源，并获得直接在真实世界采取行动（take actions upon the world）的权限后，它将有能力完成用户托付的几乎任何任务。这背后的核心承诺是“难以估量的价值创造”——那些以往由于耗时极度漫长、人力成本过于高昂甚至人类单凭自身极限无法企及的宏大构想，都将因此具备落地的可行性。OpenAI 的长期目标是将这种能力以尽可能广泛的方式分发给全球每一个人，从根本上拔高全人类敢于梦想和行动的上限。
Chris 对此深有感触地指出，无论是顶尖的诺贝尔奖得主级别物理学家，还是各行各业的一线项目经理，都曾向他倾诉过类似的震撼体会：“我手里曾经压了一大堆绝妙的想法，但放在过去，它们根本不可能有机会被真正实现；而现在，我已经把它们一个个动手落地了。”这种创造力的大爆发让 Chris 感叹道：“我们仿佛正在迈入一个黄金时代（It feels like we're entering a golden era）。”Thibault 也对此给出了十分坚定的认同。
```mermaid
flowchart TD
accTitle: Codex 通用智能体价值闭环
accDescr: 展示 Codex 从连接世界信息到执行现实世界操作，进而释放人类海量搁置创意的完整价值路径。
subgraph S1["能力底座"]
A["接入全球权威信息源"] --> B["通用强力 Agent"]
C["环境交互与现实行动权限"] --> B
end
subgraph S2["落地价值"]
B --> D["破解极度耗时与超人类复杂度难题"]
D --> E["唤醒搁置创意<br/>（科学家/项目经理等）"]
E --> F["面向全人类普惠分发"]
end
```
### [40:36](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2436s) 总结与行动倡议：从“代码”到“书籍”，拥抱知识工作的全新超能力
在对谈的总结阶段，Chris 特别提到了 OpenAI Forum（OpenAI 论坛）这一社区的核心价值：用户可以在这里彼此传授技巧、互相示范并迁移经验，从而摸索出工具融入日常生活的全新方式。针对“不写代码的人为何要关注 Codex”这一核心议题，Chris 从词源学上给出了一个极富洞察力的阐释：虽然 Codex 名字中包含“Code”，但在拉丁语中，“Codex”的原意其实是“书籍（book）”——它是承载人类普遍经验与知识的系统化媒介，其外延远比狭义的计算机代码更为宽广。无论是知识工作者在海量数据中大海捞针、进行即时可视化分析，还是在后台默默承担复杂的排序与任务执行，Codex 都在为各行各业赋予前所未有的超能力。
在圆桌闭幕之际，Chris 向所有听众发起了一个具体的行动呼吁：在自己的生活或组织中挑出一个具体的实际工作流来亲自试用 Codex，无论是一份研究简报（research brief）、规划方案、团队入职培训流程，还是一份数据汇报或决策备忘录（decision memo）。他强调，听取再多的官方介绍，也比不上自己上手体验带来的真实感触。在向现场提问观众表达谢意后，Chris 还预告了下一场即将于 5 月 26 日（周二）与比尔及梅琳达·盖茨基金会（Gates Foundation）联合举办的特别活动，探讨“AI 如何检测并预防下一次大流行病（How AI Could Detect and Prevent the Next Pandemic）”。
- **“Codex”的词源意涵**：源自拉丁语的“书籍”，象征着超越计算机代码的通用知识处理系统。
- **知识工作者的核心赋能场景**：海量资料检索（大海捞针）、复杂数据分析、即时图表渲染、后台优先级排序与多步骤任务执行。
- **推荐试点的初始工作流**：研究简报、行动规划、员工入职引导（onboarding）、业务报表与决策备忘录（decision memo）。
- **下期活动预告**：5 月 26 日（周二），OpenAI 将与盖茨基金会联合探讨利用 AI 检测与预防全球下一场流行病的前沿议题。
## Notable quotes
> [01:21](https://www.youtube.com/watch?v=DLP9CagE3dU&t=81s) 让 AI 达到高产软件工程师的编码水平 — Thibault Sottiaux
> [05:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=356s) 不可能，如今 Codex 对我们而言是不可或缺的底座。 — Thibault Sottiaux
> [12:07](https://www.youtube.com/watch?v=DLP9CagE3dU&t=727s) 家庭自制、个性化软件 — Chris Nicholson
> [19:30](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1170s) 把耗时无限长（永远不会动手）的事情变成了现实 — Chris Nicholson
> [21:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1316s) 在七八个不同的应用中反复翻找、被铺天盖地根本不关心的噪音狂轰滥炸的日子将成为过去……未来我甚至不必亲自去读邮件，私人智能体会替我阅读收件箱，只有在真正重要、需要决策时才提示我。 — Thibault Sottiaux
> [28:40](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1720s) 复制粘贴的时代已经彻底终结 — Chris Nicholson
> [30:14](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1814s) 立即住手，这个操作风险过高，请换一种低风险方案。 — Thibault Sottiaux
> [32:56](https://www.youtube.com/watch?v=DLP9CagE3dU&t=1976s) 它毕竟无法直接读取你的大脑 — Thibault Sottiaux
> [37:29](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2249s) 谁负责干活，谁才真正获得了学习 — Chris Nicholson
> [38:54](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2334s) 我手里曾经压了一大堆绝妙的想法，但放在过去，它们根本不可能有机会被真正实现；而现在，我已经把它们一个个动手落地了。 — Chris Nicholson
> [38:54](https://www.youtube.com/watch?v=DLP9CagE3dU&t=2334s) 我们仿佛正在迈入一个黄金时代 — Chris Nicholson
## Mentioned
Codex, OpenAI, Chris Nicholson, Thibault Sottiaux, Codex Web, GitHub, GPT-5, GPT-5.2, Notion, Alexander Americus, Slack, Sarah Friar, Jane The Bakery, Arsicault Bakery, Tartine Bakery, Neighbor Bakehouse, Gmail, Twitter (X), /goal, Gal Gen Karya, ChatGPT, Anastasia Chueva, Jason DeLuca, Crossing Point IT Solutions, Daniel Green, Kansas City AI Collective, Sai Sri Hamsini, AdeptMind, Images V2, OpenAI Forum, 比尔及梅琳达·盖茨基金会 (Gates Foundation)
## Video description
Codex began as a tool for developers. Today, people are using it for much more: research, planning, file organization, automation, data analysis, presentations, and other everyday knowledge work.
In this OpenAI Forum conversation, Chris Nicholson of OpenAI Global Affairs speaks with Thibault Sottiaux, Head of Codex, about how Codex is evolving beyond software engineering and what that shift means for workers, teams, and organizations.
[Source](https://www.youtube.com/watch?v=DLP9CagE3dU)
Source quotes
[1:21] Thibault Sottiaux: 让 AI 达到高产软件工程师的编码水平
[5:56] Thibault Sottiaux: 不可能，如今 Codex 对我们而言是不可或缺的底座。
[12:07] Chris Nicholson: 家庭自制、个性化软件
[19:30] Chris Nicholson: 把耗时无限长（永远不会动手）的事情变成了现实
[21:56] Thibault Sottiaux: 在七八个不同的应用中反复翻找、被铺天盖地根本不关心的噪音狂轰滥炸的日子将成为过去……未来我甚至不必亲自去读邮件，私人智能体会替我阅读收件箱，只有在真正重要、需要决策时才提示我。
[28:40] Chris Nicholson: 复制粘贴的时代已经彻底终结
[30:14] Thibault Sottiaux: 立即住手，这个操作风险过高，请换一种低风险方案。
[32:56] Thibault Sottiaux: 它毕竟无法直接读取你的大脑
[37:29] Chris Nicholson: 谁负责干活，谁才真正获得了学习
[38:54] Chris Nicholson: 我手里曾经压了一大堆绝妙的想法，但放在过去，它们根本不可能有机会被真正实现；而现在，我已经把它们一个个动手落地了。
[38:54] Chris Nicholson: 我们仿佛正在迈入一个黄金时代
Source quotes
[1:09] I would have expected like doing an in-depth marketing research or analyzing the public sentiment around a new feature... that would have taken a lot of time... Now that's just really being compressed to like a couple of hours of work.
[27:03] Thibault Sottiaux: A good trick is to be precise on helping it evaluate its own success. And so, if you're able to describe what good looks like and what solved looks like... then Codex will be able to understand whether it's doing well and whether it has completed the task.
[38:54] Chris Nicholson: I had a stack of good ideas and I've been able to start work on them and they never would have been realized in the world before. It feels like we're entering a golden era.
Original transcript
[0:09] Good afternoon.
[0:11] Welcome to OpenAI forum. Uh my name's
[0:13] Chris Nicholson. I'm with the Global
[0:14] Affairs team and I'm glad to be here
[0:17] with all of you.
[0:18] So, the forum, as some of you know, is a
[0:20] place where we talk with experts about
[0:22] how AI is being [clears throat] used in
[0:24] the world.
[0:25] Today's conversation is about Codex and
[0:28] why it matters beyond software
[0:30] engineering.
[0:31] So, more and more people are using Codex
[0:34] to help with knowledge work and personal
[0:36] work. They're using it to reduce
[0:38] friction, to perform tedious tasks, or
[0:41] just to understand a problem, organize
[0:43] information, and plan a document that
[0:46] they can share.
[0:48] So, that makes Codex helpful for
[0:50] researchers, educators, operators in
[0:52] industry, small business owners,
[0:55] leaders, knowledge workers of of all
[0:56] kinds. Um and today I'm joined by
[0:59] Thibault.
[1:01] Thank you for coming. Of course. Uh
[1:03] Thibault is the head of Codex at OpenAI
[1:05] and he's going to help us understand
[1:07] where Codex came from,
[1:09] and what the team is learning from
[1:11] people using this tool, especially
[1:13] non-technical users, and how folks can
[1:15] begin to use it outside of coding. So,
[1:17] Thibault, great to have you here. Great
[1:19] to be here. Yeah. Um well, let's get
[1:21] started. So, to ground this, Codex was
[1:23] created as a tool for developers. Um
[1:26] maybe you could tell us when did Codex
[1:27] first launch and what problems were you
[1:29] and the team trying to solve at the
[1:31] beginning? Yeah, so the journey started
[1:34] a long time ago. At OpenAI, we've always
[1:36] been fascinated with uh helping
[1:39] just figuring out like how we can create
[1:42] very useful models that would then in
[1:45] turn help us, you know, also like
[1:47] accelerate the development. And, you
[1:49] know, one of the grand challenges for a
[1:51] while has always been uh coding.
[1:53] And being able to code at the level of a
[1:57] prolific software engineer. That was
[1:59] like a grand challenge um like roughly
[2:01] like 2 years ago when we started on this
[2:03] journey. The first public version of
[2:06] Codex was something that we now refer to
[2:08] as Codex web.
[2:09] It was this idea that we would have
[2:12] uh
[2:13] an entity running in the cloud
[2:15] accessible over a web interface. You
[2:18] would put your task in there. It would
[2:20] go and look at the repository like a
[2:22] code repository that you have. It would
[2:24] figure out what the right set of changes
[2:26] were in order to fulfill the task and it
[2:28] would open a pull request on uh GitHub.
[2:31] So it was like fully isolated fully
[2:34] packaged. You would just state your
[2:35] intent and then you would receive like a
[2:36] code change at the end. And that was
[2:38] like the entire thing that we shipped.
[2:39] That was like roughly a year ago. What
[2:41] we found was um it was a little bit too
[2:44] high friction. Um it was too hard to set
[2:46] up. And then also a lot of developers
[2:49] just have you know a beautiful setup on
[2:51] their own computer. And you know you
[2:53] need to do go and like reproduce all of
[2:55] that on our own cloud environment setup.
[2:57] So the friction was too high. Um the
[2:59] models were also not at the level where
[3:01] they would you know get it right all the
[3:02] time. And so it was kind of like hard to
[3:04] iterate there with the model. And we
[3:06] shifted our approach after that which
[3:08] I'm glad to go into. Yeah great. So so
[3:10] you were building a tool for yourselves
[3:12] and it sounds like there was a pivot to
[3:13] where you realized it it should work on
[3:15] everybody's machine. Locally. That's
[3:17] right. We realized that doing this on
[3:19] our own cloud machines and asking
[3:21] everyone to just configure it there was
[3:24] too difficult because the models were
[3:26] not yet at the level where it was
[3:27] reliable for long-term horizon tasks. It
[3:30] was just not working. When did you first
[3:32] see people using it for tasks outside of
[3:35] coding?
[3:36] Oh that's that's been um more of a
[3:39] recent shift in the last last uh six six
[3:42] uh months I would say ever since we
[3:44] released GPT-5 we've had that step
[3:47] change in like generality and
[3:48] reliability for general-purpose tasks I
[3:51] would say. Around 5.2 as well, where it
[3:53] became more reliable for long horizon
[3:55] tasks. And then, one thing to realize as
[3:59] well is like even for software engineers
[4:01] and technical people, like the majority
[4:03] of the day is not actually coding. The
[4:06] majority of the day is like looking at,
[4:08] you know, tickets, prioritizing, having
[4:10] discussions around, you know, what how
[4:12] how we should actually solve problems,
[4:14] >> Mhm.
[4:15] deciding on like the architecture,
[4:17] solving, um, you know, like bug reports,
[4:19] investigating bug reports, and figuring
[4:22] out, you know, whether it's actually
[4:23] true, whether it's a user problem, where
[4:25] the problem actually is, dealing with
[4:27] outages, and you know, being on call,
[4:29] and then, you know, just like general
[4:30] like having an understanding of the
[4:32] systems, and doing a lot of information
[4:33] gathering. I would say like maybe, you
[4:36] know, software engineers spend like 20,
[4:38] 30% of their time actually coding.
[4:40] [clears throat]
[4:41] Uh, and so, what we saw is like a lot of
[4:43] the people who had adopted Codex were
[4:45] already, you know, the technical ones
[4:46] were just using it for their day-to-day
[4:49] work.
[4:49] >> Mhm. Uh, and then, you know, it was
[4:51] obvious that this we were holding sort
[4:53] of like a technology that was like much
[4:54] more powerful, and that we wanted to,
[4:56] you know, was yearning to be useful for,
[4:59] you know, Mhm. at scale, you know, for
[5:01] for the entirety of the society.
[5:02] >> Mhm.
[5:03] >> [clears throat]
[5:03] >> So, it sounds like
[5:04] to be a software engineer, you have to
[5:06] do a lot of things that aren't coding.
[5:07] >> That's right. And and people were
[5:08] dogfooding. They were using their own
[5:10] tool to do these to do these non-coding
[5:12] tasks. And in order to be useful at
[5:14] coding, we were, okay, you need access
[5:16] to more context. Maybe if you have
[5:18] access to, you know, all of the
[5:19] information that we have in Notion, you
[5:21] know, information that you have in
[5:23] documents, um,
[5:25] then actually we saw that the agent
[5:28] would become more useful at solving the
[5:30] task.
[5:30] >> Yeah. And then, you know, we were trying
[5:33] to just make that more and more reliable
[5:35] up until the point where now like the
[5:37] majority of the tasks that are being
[5:38] performed in Codex are actually
[5:39] non-coding tasks.
[5:40] >> That's really interesting. So,
[5:42] it First of all, Codex was searching
[5:44] code, and then you realized
[5:46] it's also super useful to search a bunch
[5:48] of documents and come back with
[5:49] information. And obviously that's
[5:51] something that all of us doing knowledge
[5:52] work have to do. That's right.
[5:55] So
[5:56] So maybe
[5:57] What is Do you Do you remember the first
[5:59] moment when you were like, "Oh, this is
[6:01] actually
[6:02] going to be for everybody."
[6:03] Right? When when that when the ambition
[6:05] changed.
[6:06] Yes, we were in the middle of a launch
[6:08] for Codex and
[6:10] Alexander
[6:12] Americus, who is like the the lead
[6:14] product manager on Codex, was was using
[6:18] Codex to keep track of the state of
[6:21] all of the changes that we were trying
[6:23] to land Mhm. um in in in in the run-up
[6:26] to the launch.
[6:27] >> Mhm. And I had I had never seen someone
[6:29] be as productive as Alexander at that
[6:32] moment where it seemed like he had many
[6:34] little Codex agents, you know, doing his
[6:37] work on his behalf and chasing people up
[6:39] >> Mhm. and then updating a document with
[6:41] the latest state of things of like, "Oh,
[6:43] you know, it's like we actually still
[6:45] need to polish like this specific
[6:46] feature. We have this user feedback. We
[6:47] have that." And it was So, it was
[6:49] coalescing all of this information from
[6:51] all the user feedback that we have, the
[6:53] information from the developers, and
[6:55] keeping like a plan, you know, neat and
[6:57] up-to-date while Alex was having a
[6:59] discussion with me. Um and you know, I
[7:02] felt like, "Wow, you know, it's like
[7:03] we're really changing everything here,
[7:04] not just software engineering." So So,
[7:06] in the world before Codex, Alexander
[7:08] would be either himself digging through
[7:11] Slack channels or documents or GitHub
[7:13] PRs and just like everybody like spend
[7:16] so much time with coordination and then
[7:17] you saw he had somehow delegated that
[7:20] really time-consuming work to these
[7:22] tools and they're doing it while he's in
[7:24] a meeting with you. That's right. Um
[7:27] Our models are very good at genera-
[7:30] gathering the right context and
[7:31] summarizing as well. And so, that's a
[7:33] strong use case that we see. And he was
[7:35] doing that and then also using it to
[7:37] chase up information. So, we have it
[7:40] connected to Slack. It's able to send in
[7:42] to send messages. The Codex agent can
[7:44] send messages to people to, you know,
[7:46] also ask like what's the latest status
[7:48] on, you know, this thing and so doing
[7:49] all the chasing on behalf of Alexander.
[7:52] Chasing takes so much time. So so So
[7:54] there's Alexander. He's one of the most
[7:56] productive people you've ever seen. He's
[7:57] using Codex in this way. What's the
[8:00] ripple effect? Like what's the
[8:02] in concentric circles around the people
[8:03] around him, the team, what gets shipped?
[8:06] The what The ripple effects that we've
[8:08] seen is
[8:09] engineers have been greatly accelerated
[8:12] and are now building faster than ever.
[8:15] Adjacent jobs are are changing like the
[8:18] role of the designer is changing, the
[8:20] role of like the product manager is
[8:21] changing and then, you know, we looked
[8:23] at how can we accelerate and and
[8:26] um give them also the means to, you
[8:28] know, just be more productive.
[8:30] And then we saw bottlenecks shift to
[8:34] communications and marketing
[8:36] where
[8:37] we were suddenly producing so much work.
[8:40] Um and you know, it was a challenge to
[8:42] sort of like
[8:44] tell the world and, you know, keep the
[8:45] story coherent and um so the bottleneck
[8:48] was shifting to other departments at the
[8:50] company. Hm.
[8:51] >> [clears throat]
[8:51] >> So OpenAI is shipping really fast. Do
[8:53] you
[8:54] would it be possible without Codex?
[8:57] No, I don't think so. At this point
[8:58] Codex is critical for us. Yes.
[9:00] >> Do you think other companies
[9:02] >> with 10 times more engineers. Hm. Do you
[9:04] think other companies are similar enough
[9:06] to OpenAI and like the problems they're
[9:08] trying to solve for the world that they
[9:10] can follow a similar playbook?
[9:12] Yes, I think now we the state of the
[9:14] technology is uh
[9:17] at has reached a point where these
[9:19] agents are capable of very general work.
[9:22] Um there's really many things, you know,
[9:25] that you can do behind your computer
[9:27] that the agent isn't able to also help
[9:29] with.
[9:30] And so, you know,
[9:32] a lot of like, you know, the work that
[9:33] needs to go into like preparing a
[9:35] presentation to, you know, align, you
[9:37] know, stakeholders, for example,
[9:38] gathering context on public perception,
[9:41] doing some marketing research,
[9:43] organizing it's used a lot in our
[9:45] finance department. Sarah Friar talks
[9:47] about this a lot that she organized the
[9:48] latest fund raise with Codex with the
[9:51] help of Codex and it was a great
[9:52] accelerant there as well. And so there's
[9:54] this beautiful generality to it.
[9:57] And it's evolved like you know, much
[9:58] more to like it's not for generating
[10:00] code. It's just to perform general
[10:02] purpose tasks really. Mhm. So you
[10:04] mentioned that roles are changing.
[10:06] What kind of changes do you see?
[10:11] The change that I see is that
[10:15] everyone has to deal with
[10:20] things moving much more quickly
[10:23] uh and adapt much more quickly as well.
[10:26] Um hard hard problems, things that used
[10:29] to take days now maybe take like a
[10:31] couple of hours.
[10:32] Across the field we're seeing this in
[10:34] science, we're seeing this in
[10:35] engineering, but we're also seeing this
[10:36] of like
[10:38] you know, I would have expected like
[10:40] doing an in-depth marketing research or,
[10:43] you know, analyzing the public sentiment
[10:45] around a new feature, you know, that
[10:47] would have taken a lot of time to go and
[10:49] like find all the sources and summarize
[10:51] and then
[10:52] uh condense it in a way where, you know,
[10:53] like different people could consume
[10:55] that. Now that's that's just really
[10:58] being compressed to like a couple of
[10:59] hours of work and, you know, being
[11:01] completely automated away.
[11:03] It does mean that everything is slowly
[11:05] like moving much faster as well and you
[11:06] need to adapt to things, you know, just
[11:09] the the the pace having picked up.
[11:11] People are able to do much more by
[11:12] themselves as well and, you know,
[11:14] there's something delightful about that
[11:16] too. It's like it's very empowering. And
[11:18] so maybe before you're like, "Oh, I need
[11:19] to go and talk to someone about this." I
[11:21] don't know, you can just do it. Um we're
[11:23] seeing this a lot on uh data questions,
[11:26] you know, within within companies where
[11:28] it's like you have a question about,
[11:29] "Oh, you know, how successful are we in
[11:31] this market?" Um you know, are we
[11:33] growing uh in India? Um, you know, what
[11:36] what's happening in Korea? And then now
[11:37] everyone is just empowered to ask
[11:39] critics the question of like, "Hey, can
[11:41] you pull up the dashboard? I don't
[11:42] really know where it is, but I go figure
[11:44] that." And then you can go and like dig
[11:46] into, um, you know, all the details of
[11:47] the business. And you don't need to go
[11:49] and talk to like data analyst teams.
[11:50] Like you can leave them alone. They can
[11:52] do more interesting stuff than answering
[11:53] your, you know, little questions. Yeah,
[11:55] I I remember the days when data analyst
[11:57] teams had long queues cuz they're a
[11:59] shared resource of things they owe to
[12:00] all the other teams. And as long as
[12:02] those other teams have access to the
[12:03] underlying data, that's very often they
[12:05] can extract insights by themselves. I
[12:07] feel like we're in this historical
[12:09] moment where it used to be that the
[12:10] people who had problems and the people
[12:11] who built solutions were two separate
[12:14] groups of people and the conversation
[12:16] they had about the product were really
[12:17] long and you you eventually got to
[12:19] something that was kind of okay, but you
[12:20] couldn't push it further because nobody
[12:21] had time. And now now like this the
[12:24] person who has the problem can build the
[12:26] solution very quickly and iterate
[12:27] through the changes they need. Yeah,
[12:29] that's why we have UX designers as well
[12:31] on the team. They're pushing code.
[12:33] Uh, they're making changes. They they're
[12:36] in there with the developers, uh,
[12:38] shaping the product as they want to
[12:41] like, you know, to kind of bring the
[12:42] best out of it and create an amazing
[12:43] experience. They don't have to go and
[12:45] like convince, you know, the engineering
[12:47] team, you know, to prioritize like, you
[12:48] know, what the engineering team probably
[12:50] feels like, you know, fairly minor
[12:51] changes, but, you know, for our amazing
[12:54] designers, you know, they're like, "Oh,
[12:55] no, this is this is like this really
[12:57] elevating the craft and the experience,
[13:00] you know, for millions of users out
[13:01] there." And so it's very empowering to
[13:02] them, too. It feels like we're kind of
[13:05] in this era of home-cooked, personalized
[13:07] software on so many levels.
[13:10] Yes, it feels that this is the wave
[13:12] that's coming where everyone will be
[13:13] able to have their own personal software
[13:16] and maintain [clears throat] it and it
[13:17] just does exactly what you need it to
[13:19] do. Yeah. You you were telling me before
[13:21] we started about something you built for
[13:23] yourself that visualizes some data. Do
[13:25] you do you want to do you want to share
[13:27] that with Yeah, sure. I just It's just
[13:28] really a fun uh, exercise. So, I have it
[13:30] on my screen here.
[13:32] And [snorts] I'm in I'm in I live in San
[13:33] Francisco. So do you, I think.
[13:36] >> [laughter]
[13:37] >> I think the price of bread is
[13:40] uh outrageous in San Francisco. I moved
[13:42] from Europe. Um and um so, I was like,
[13:45] "Hey Codex, can you find like the best
[13:47] loaves in the city and create me a
[13:49] spreadsheet uh so that, you know, I can
[13:51] understand like where are those loaves,
[13:53] uh where should you buy them, and uh you
[13:55] know, what's the associated price?"
[13:57] And so, it went and it worked for 5
[13:59] minutes and it created this uh this
[14:01] spreadsheet, which you can see here. Can
[14:03] we see that? Can everybody see that?
[14:04] Great. Um and so, you can see you have
[14:07] Jane the Bakery, you have Arsicault, you
[14:09] have Tartine. It has found like the
[14:11] different loaves, a little description,
[14:13] and the price. So, it seems like if I
[14:14] want to go and find a good loaf of
[14:16] bread, the sourdough loaf, I should go
[14:18] to Neighbor uh Big House, which is in
[14:20] Dogpatch. Mhm. And then, I was like,
[14:22] "Okay, it's great. I have the
[14:23] spreadsheet." I could have It traded on
[14:25] the spreadsheet and asked for more
[14:26] information, but instead, I wanted to
[14:28] consume it in a bit of a more of a
[14:30] visual way. So, I said, "Hey, how about
[14:32] you create me a webpage?"
[14:35] Um and so, it did and it just put all of
[14:37] these on uh on a map here so I can
[14:41] uh just And this is just really personal
[14:42] stuff that was created for me. This took
[14:44] 4 minutes for Codex to just generate.
[14:47] >> I can go and I can now click around and
[14:49] uh I could probably ask it to show me a
[14:51] picture of these loaves of bread as
[14:52] well.
[14:54] And this is just
[14:55] Yo, this was just done like in in in in
[14:57] 10 minutes. I didn't even look at it. It
[14:59] was just a single prompt for me. And
[15:01] then, the way that I like to engage uh
[15:04] with Codex, I would like to
[15:05] um you know, show you I can I do
[15:07] everything over voice. So, "Hey, I'm in
[15:10] San Francisco. I'm really into bread. Uh
[15:12] I want like a map of all the loaves uh
[15:14] the loaves that I can buy and um just
[15:17] pre-structure it uh with a little
[15:18] explanation of uh where I can buy it and
[15:21] like the price of the loaf.
[15:23] And then I don't even need to type
[15:25] anymore and it's just going to go and do
[15:26] that. Um
[15:28] and then I can do go and and and and and
[15:30] just like do other stuff on my computer.
[15:32] So this is not a thing where, you know,
[15:34] I'm even, you know, actively spending at
[15:35] the attention on it.
[15:36] >> Mhm. So you're saying anybody can do
[15:38] that if they care about some data, they
[15:40] can be and they have access to it, they
[15:42] can basically make a website, analyze
[15:44] it, visualize it, share with their
[15:47] friends, family, coworkers.
[15:49] >> Uh and and we used to talk about writing
[15:51] code. It seems like such an anachronism
[15:53] now. Now you're you're kind of speaking
[15:54] code, right?
[15:56] It's it's using code under the hood, but
[15:58] it's more of a tool
[15:59] >> Yeah. for the agent. You don't even need
[16:01] to know about it. Um it it's just, you
[16:04] know, it's editing and creating
[16:06] spreadsheets and documents and slide
[16:08] decks for you. It's creating websites
[16:10] for you. You don't need to understand
[16:12] what's, um, you know, how it how it
[16:14] actually like performs its work under
[16:15] the hood. It's just using code as a tool
[16:17] and this is why there's this beautiful
[16:19] generality to it.
[16:20] >> Yeah. So So in the old days, if you had
[16:22] the skills, this might have taken you a
[16:23] weekend
[16:25] to to create something at at college or
[16:27] something like that.
[16:28] >> quite a bit of time to go and find all
[16:29] that information.
[16:30] >> Yeah. Right. And and now it's just an
[16:32] instant and then if you don't quite like
[16:34] it, you can just tell it to change it,
[16:36] right?
[16:36] >> That's right. If I'm like, "Oh,
[16:37] actually, you know, I wanted to
[16:39] you know, do the same but for coffee."
[16:41] Mhm. Um
[16:43] "Hey, uh do the same analysis for
[16:45] coffee, um, not just bread."
[16:49] That's it. That's all I need to do. It
[16:51] will go and work with it minutes and
[16:53] it'll end up with like likely will end
[16:55] up with, um, you know, a website just
[16:57] about coffee in San Francisco. And
[16:58] you're looking for you're optimizing for
[17:00] price and quality, right? It's actually
[17:02] more complex than just give me cheap
[17:03] bread. Yes. Yes. So you you can express
[17:06] your own preferences,
[17:08] um, you know, Codex will just go and
[17:10] like take that into account and, um,
[17:12] just really work to satisfy, you know,
[17:14] your your your prompt. Right. So it it
[17:16] gets the data, it visualizes the data,
[17:18] it gets you to an insight about the
[17:20] world, then you can make a decision to
[17:22] suit your needs, which is kind of like
[17:24] the universal loop we're all in in our
[17:26] lives. How do I actually decide, make a
[17:28] better decision Yes.
[17:29] >> to reach my goals. Yes. And this can be
[17:31] very simple, this can be like as
[17:33] complicated as, you know, like the same
[17:35] process was used for the latest, you
[17:37] know, fund raise at OpenAI. So, it's
[17:39] it's
[17:40] it's just Codex, it's just in
[17:42] capable of like the smallest things and
[17:43] then the most complicated things
[17:45] >> as well. Yeah. Amazing. So, so are you
[17:47] now using it like Alexander? Are you are
[17:50] you kind of coordinating and
[17:51] orchestrating things?
[17:52] >> You can see here in my uh
[17:54] sidebar is this is before we started to
[17:56] talk. I fire uh off like uh you know,
[17:59] little tasks like all the time like you
[18:01] know, maybe over a hundred different
[18:02] tasks like per day that I just like hand
[18:04] off to Codex and it just goes and like
[18:06] does it for me. This goes from like
[18:08] organizing my desktop files, uh you
[18:10] know, managing uh compu- the compute
[18:12] fleet, um helping me understand
[18:15] uh how the on-call rotation is doing and
[18:17] like how the engineers are doing,
[18:19] helping me understand, you know, the
[18:20] launch schedule that's coming up uh and,
[18:22] you know, flagging anything that's like
[18:24] at risk that I maybe I need to look
[18:25] into. Um I use it as a little chief of
[18:28] staff where
[18:29] uh and I I run this every day. So, for
[18:31] example, you can do a little automation
[18:34] and like go through Gmail,
[18:38] uh Notion, calendar,
[18:41] and uh flag um
[18:43] and give me give me a summary
[18:47] of my day
[18:49] and flag anything uh that is at risk.
[18:53] And then I can say, "Okay, like do this
[18:55] daily.
[18:56] Um
[18:57] do this daily at uh 9:00 a.m."
[19:00] And then, you know, I can just go and
[19:01] schedule that and then it will run it
[19:03] and then I will find it in my inbox um
[19:06] you know, at 9:00 a.m. every day. So,
[19:07] it's like you've got your own chief of
[19:08] staff. Yeah. Hey,
[19:10] do you don't forget to do these These
[19:12] are It's It's even helping you
[19:14] prioritize. And that's right. It's
[19:16] helping me prioritize. It's helping me
[19:18] spend my attention where it matters the
[19:20] most. It's and it's helping me also by
[19:24] you know, I don't have to go and like do
[19:25] the small things.
[19:27] You know, that would otherwise just take
[19:28] me time and maybe I would never get to.
[19:30] What were the things that annoyed you?
[19:33] Like most hideous things
[19:35] in the pre-Codex time that you spent
[19:38] hours every week on?
[19:39] Um the thing that annoyed me most was
[19:41] like I would not do certain things
[19:43] because I felt like oh, I don't have the
[19:44] time to do it myself, but I need to go
[19:46] and annoy someone.
[19:47] You know, to ask for like you know, this
[19:49] information. Um but you know, I felt
[19:51] like it was you know, maybe not
[19:53] important enough to go and you know, put
[19:55] this on someone's desk and be like you
[19:57] know, hey, can you can you help me with
[19:58] this question? Um so now I feel like you
[20:01] know, I can just you know, get the
[20:02] information that I want, have the little
[20:04] reports that I want. Um I can build all
[20:06] sorts of personal software that you
[20:08] know, I didn't have the time to do. So
[20:11] you know, it's just like a really uh it
[20:13] gives me joy because um it it is doing
[20:16] all the little tedious, you know, things
[20:18] that were like a very manual on my
[20:20] computer before. And now I'm like okay,
[20:22] I'm just in the Codex app and um it's
[20:24] doing all these things for me and I can
[20:26] just focus on you know, the things that
[20:28] I feel I actually want to think about.
[20:30] Yeah. Uh so I know people often talk
[20:32] about it as like
[20:34] this would have taken weeks, now it
[20:35] takes seconds, but it actually feels
[20:37] like this never would have happened. It
[20:39] would have taken infinity time because
[20:40] we never would have gotten to it. And
[20:42] that And that was actually
[20:43] >> of that. There's a lot of that. It also
[20:45] sounds like
[20:47] are you enjoying your job more because
[20:48] these things are getting done that used
[20:50] to annoy you? Yes. Yes, I am. There's
[20:52] like so much that I don't um
[20:55] you know, like for example,
[20:57] every day I'm like oh, you know, what's
[20:58] the latest news?
[21:00] Um and there's like
[21:02] it has very personal instructions for
[21:05] the kind of news that I care about.
[21:07] And so,
[21:08] what that is an example like I would
[21:09] have never
[21:11] received such a report about the news.
[21:14] Um
[21:15] and it gives me great joy like to just
[21:17] like
[21:17] read this in the morning and, you know,
[21:19] feel a little bit more informed.
[21:22] A lot of things on um when it comes to
[21:24] like investigating, for example, we have
[21:27] like a bug report
[21:28] uh from a user. I see it like on Twitter
[21:31] and it's like, "Oh, you know, that that
[21:33] is a bummer because I would have usually
[21:35] put this as like this is like a low
[21:37] priority thing. This is probably
[21:38] affecting like three users on the
[21:40] planet. Um [laughter] this doesn't feel
[21:42] like a thing." Like now, I'm like, "Oh,
[21:44] let's just put this in Codex and you
[21:45] know, it will get it done." And, you
[21:48] know, I can feel like, you know,
[21:49] just much better because I don't feel
[21:51] like, you know, things are falling
[21:52] through the cracks anymore.
[21:54] So, it's reducing the cognitive load for
[21:56] me.
[21:56] >> Yeah. I I
[21:58] It makes me think that this is actually
[21:59] a tool to address burnout and
[22:01] over-information overwhelm. Like people
[22:03] we we're all kind of surrounded by tools
[22:06] that are supposed to help us, but in the
[22:07] end we're kind of trapped within our
[22:08] tools. And and this is a thing that's
[22:11] kind of
[22:12] liberating us as a tool. Um and and
[22:15] maybe decre- like so many people,
[22:16] teachers, doctors, so many people feel
[22:19] burnout and overwhelmed because they're
[22:21] trapped in those tools, manual data
[22:22] entry, that sort of thing. And do do you
[22:24] feel like
[22:25] it almost feels like our relationship to
[22:26] tools is changing. Is there a
[22:28] fundamental shift in how we relate to
[22:29] tools?
[22:30] Yeah, to me that's the promise. The
[22:32] promise is you have
[22:34] a trustworthy partner almost that you
[22:39] can go through and, you know, will do a
[22:41] lot of work on your behalf to a level of
[22:43] trustworthiness where you're like, "Hey,
[22:45] it's like if I delegate something to
[22:47] you,
[22:47] um you know, you'll you'll get it done.
[22:50] If you don't you don't get it done to
[22:51] the level of satisfaction, you know,
[22:53] you'll flag it to me, but I know that it
[22:55] will be done well. I know that it will
[22:56] done be done, you know, maybe to a level
[22:58] of quality that's superior to like now
[23:00] if I were doing it myself. I can also,
[23:03] you know, trust this partner to
[23:06] um shield me from like a lot of the
[23:08] noise and you know, to flag important
[23:10] things, you know, when in in a timely
[23:12] fashion. And so, I don't have to go and
[23:16] you know, consume like information in
[23:18] like, you know, seven different apps and
[23:20] be bombarded by things that, you know, I
[23:22] don't actually care about. Um what I
[23:25] imagine the future to be like is, you
[23:27] know, I don't even have to read my
[23:28] email.
[23:29] Um
[23:31] I maybe I just have, you know, I have
[23:33] like my little personal agent that is
[23:37] just like reading my inbox for me,
[23:39] flagging me like whenever there's
[23:40] something really important, asking me
[23:42] for input, and then just you know, doing
[23:45] the work. So, you don't have to search
[23:47] for needles in a dozen different
[23:48] haystacks. Needles are kind of arranged
[23:51] in a newsletter. I can I can say like,
[23:53] here are my goals for today.
[23:55] Help me out, manage everything else.
[23:58] And trust that that will happen. That's
[23:59] amazing. You must have seen over the
[24:01] past few months like the
[24:03] what you can trust this tool with is
[24:05] changing. It's like you mentioned time
[24:07] horizon, like the time it takes to do a
[24:09] complex task. That's that's changing.
[24:11] Yes, so we we actually launched a a new
[24:14] thing there that is like more of an
[24:15] advanced feature, uh which is uh {slash}
[24:18] goal. So, in in Codex you have {slash}
[24:21] commands, uh which allow you to uh for
[24:24] example, enter like different modes.
[24:27] {slash} goal allows you to enter a mode
[24:29] where you give a long-term to Codex, and
[24:32] it goes and is really relentless at it.
[24:36] And so, for example, you can give it the
[24:37] goal of solving a very hard mathematical
[24:40] problem, and it will go at it for hours,
[24:43] maybe days, maybe weeks until um it has
[24:47] decided that it has satisfied the goal.
[24:50] Um we're seeing it being used to do like
[24:53] amazing things on like improving the
[24:55] performance of programs, rewriting
[24:57] entire programs from one language to
[24:58] another.
[25:00] Um it's used in science problems as
[25:02] well. I've seen like really cool
[25:04] breakthroughs on uh mathematics and and
[25:06] and physics breakthroughs with it.
[25:08] Um
[25:10] and uh this is fascinating to me because
[25:13] maybe like a couple of months ago we
[25:14] were oh you know we were this is
[25:16] exciting it's working for 10 minutes.
[25:18] Now we're talking about you know agents
[25:20] working for like weeks on like the
[25:22] hardest tasks. Right. Right. Sometimes I
[25:25] think genius is just the ability to
[25:26] think about the same thing longer.
[25:29] Yes. I mean if you know if you have
[25:30] infinite time like probably like
[25:33] uh you know a sufficiently smart human
[25:34] would be able to come up with everything
[25:36] that we have done you know in humanity
[25:37] so far. So I feel threatened to me. So
[25:39] so goal you can use that in the in the
[25:41] CLI can you use that in this Codex app
[25:44] that you're showing us right now? Yeah
[25:45] it it's it's it's um
[25:47] it's coming. It's not yet uh launched
[25:49] but it will be available. Okay. We'll
[25:51] we'll let the community know when
[25:53] [laughter] when goal is available on all
[25:54] surfaces. Okay. Awesome. So so it's
[25:57] running for days and it's just going
[25:59] going going in the background and it
[26:00] comes and then it comes back and reports
[26:02] to you hey I'm I'm done or I'm I'm stuck
[26:03] on something. Mhm. Yeah. That's right.
[26:05] Um and where the future is headed I
[26:08] think is where it just it's not going to
[26:09] stop. Um where [clears throat] you're
[26:11] just going to have an agent that runs
[26:14] like 24/7 and you know does useful
[26:16] things for you and gets [clears throat]
[26:18] steered along the way.
[26:19] >> Yeah. Right now it's like very
[26:20] turn-based. Yep. Uh where you know you
[26:23] go and you have a specific task and
[26:25] you're like oh you know it's like I want
[26:26] you to go and solve that task. Like go
[26:27] find all the loaves you know in San
[26:29] Francisco and like tell me the price.
[26:31] That's great. Solving tasks is super
[26:33] important. Um
[26:34] goal-oriented task solving is also a big
[26:37] unlock um in terms of like the value
[26:39] that you it can create for you.
[26:42] But the next step I think is just it
[26:43] just runs continuously and it does
[26:46] useful things whether you instructed or
[26:48] not. And it may be at some point like
[26:49] I've done all the things that I think
[26:51] are useful. I'm just going to sleep for
[26:53] a little bit until I talk to you. Um but
[26:56] it's not like um it's not like it's just
[26:59] working when you have like a very clear
[27:01] instruction. Is there any trick to
[27:04] specifying the goal you want it like how
[27:07] do you make goals succeed?
[27:09] Yeah. So, one of the one of the shifts
[27:12] with agents
[27:14] um
[27:15] compared to
[27:16] you know, maybe if if you haven't
[27:18] experienced
[27:19] you know, if you haven't installed the
[27:20] Codex app and you know, it's like the
[27:22] first time you're you're using it
[27:24] just just explore a little bit. You can
[27:25] engage um in a very chatty way about you
[27:28] know, what it what it's capable of. Like
[27:30] it it it is it understands its own
[27:32] capabilities. So, you can just you know,
[27:33] ask it like what it can do.
[27:35] But a good trick is to be precise on
[27:39] helping it evaluate its own success.
[27:42] And so, if you're able to describe what
[27:44] good looks like
[27:45] and you know, what's solved looks like
[27:47] and what you want to see you know, at
[27:49] the end of the completion of the task
[27:52] then you know, Codex will be able to
[27:54] sort of like
[27:55] understand whether it's doing well and
[27:57] whether it has completed the task or
[27:59] not. So, can you put can you put a
[28:00] number on it? Can you be precise?
[28:03] Yeah. So, if for example, you know, you
[28:05] can say like hey, I want uh I'm
[28:07] presenting
[28:09] my work.
[28:10] Um I would like a slide deck. You know,
[28:12] I would like it to have like 10 slides.
[28:14] The first two slides I want it to
[28:16] contain this kind of information. The
[28:18] next six slides I want it to you know,
[28:20] just really go into the meat of the
[28:22] problem and do a technical breakdown and
[28:24] then I want you know, two slides at the
[28:26] end with open questions and Q&A. And if
[28:29] you're a clear about it and you know,
[28:32] you're you're specific about what you
[28:33] want as the output then it's much more
[28:35] likely to succeed. So, very similar to
[28:37] what you might do with an assistant or a
[28:39] lieutenant.
[28:39] >> That's right. Yeah.
[28:40] >> That's right. Okay. So, we take
[28:42] questions from the community
[28:43] and I'm going to read some of those. So,
[28:45] we've got one here um
[28:47] from and excuse me if I mispronounce
[28:49] this, Gal Gen Karya,
[28:51] principal engineer in financial
[28:52] services, um
[28:54] is asking what would make non-coders
[28:57] move away from ChatGPT to Codex?
[29:01] I think it's going to be an evolution
[29:04] and it's really
[29:08] I I don't expect that
[29:11] um
[29:12] you know, everyone will want to move to
[29:14] Codex, but it is a nice complimentary to
[29:16] ChatGPT.
[29:18] Um
[29:19] what I recommend to use it for is to do
[29:24] things for you.
[29:25] Um so, anything that you know, where
[29:29] it's like manipulation of files on your
[29:31] computer,
[29:32] um running things on like automation and
[29:37] doing things in the background every
[29:39] couple of hours, you know, you can do
[29:40] that Codex,
[29:41] um
[29:42] you know, and and ChatGPT for me is
[29:44] still I go to for like, you know, quick
[29:46] answers.
[29:47] >> Mhm. Yeah, I I remember the old days for
[29:48] me, I would be copying and pasting code
[29:51] from ChatGPT into a file or a terminal,
[29:54] but you don't have to do that anymore.
[29:55] The copy and paste era is over. Yes,
[29:57] Codex can just like work on the code and
[29:59] the data, right?
[30:00] >> Yeah, and if you have you know, if you
[30:02] have like a a file on your computer, you
[30:04] have pictures on your computer, you
[30:06] know, just tell Codex like, hey, you
[30:07] know, use this file, read it and it can
[30:09] just do it. You don't need to go and
[30:11] like, you know, click on it manually and
[30:13] yeah.
[30:14] Okay, we've got one from Anastasia
[30:16] Chueva.
[30:18] What do you think becomes the biggest
[30:20] bottleneck in enterprise adoption
[30:23] uh and then the choices are model
[30:24] capability, human trust, or
[30:26] organizational process design?
[30:32] I don't think it's a capability thing.
[30:34] Uh the capability's there, that's not
[30:36] what what's holding back adoption uh in
[30:38] enterprise. I I it's primarily, um,
[30:42] trust and
[30:45] uh, for trust, it's it's about is it
[30:47] safe and is it secure?
[30:49] Um,
[30:51] you know, if if you have an agent just
[30:53] run around, uh, in in your in your in
[30:55] your company and like, you know, maybe
[30:56] delete sensitive files or, you know,
[30:58] upload information and you know, send an
[31:01] email that contains, um, you know, an
[31:02] exfiltrates like information that it
[31:04] shouldn't.
[31:04] >> People are going to use it.
[31:05] >> That's that's that's terrible. Um, so
[31:07] we've been thinking a lot about that. By
[31:09] default, agents run inside of sandbox
[31:12] with like tight controls, you know,
[31:13] exactly
[31:15] that they are able to only access, um,
[31:17] you know, a specific portion of the file
[31:19] system. And so, you know, like you can
[31:21] constrain it to a folder. You can
[31:23] disable network access. We're giving a
[31:25] lot of enterprise controls. I think this
[31:27] is very key to continue to expand from
[31:29] there. There's also investments that
[31:31] we've done. It's on our alignment blog
[31:33] post. Uh, we call this feature auto
[31:35] review, where we have created an agent
[31:37] that is capable of reviewing the actions
[31:39] of the primary agent.
[31:41] And the way it works is that it the
[31:44] primary agent codex is incentivized to
[31:46] do the work for you. At times, it might
[31:48] take an action that is, you know, maybe
[31:50] a little bit risky. And so, you have
[31:51] another agent that sort of reviews every
[31:54] action that it's taking and flags and
[31:56] stops it if it's like a high risk thing.
[31:58] And, you know, I expect like more
[31:59] innovations like this will be necessary.
[32:01] >> referee or an umpire. That's against the
[32:03] rules. Yes, yes. And be like, "Hey, stop
[32:04] right there. You know, this is like too
[32:06] risky of an action. I would like you to
[32:08] take like a less risky action."
[32:09] >> So, just to help the non-engineers in
[32:11] the audience, a sandbox, is that just
[32:12] like a folder in like, in a file system?
[32:15] Like, do you say, "Don't go outside of
[32:17] this folder. Just do things within
[32:18] this?" Yeah. So, that's that's right.
[32:21] It's, um, it's similar to how, um, you
[32:23] can think about, you know, if you're in
[32:25] a company like,
[32:26] oh, you know, maybe you only get access
[32:28] to information for one part of the
[32:30] organization that you're in.
[32:31] Um, and you know, people have like
[32:33] different silos. This is like the
[32:34] equivalent of a sand sandbox for an
[32:36] agent. You can sort of like restrict
[32:38] what it can have have access to and like
[32:40] the set of actions that you can take.
[32:41] Yeah. And And so a rule might be you can
[32:43] read this data that IT gave you access
[32:45] to, but you can't write back to it. You
[32:47] can't delete it. You can't like change
[32:48] it. The data is preserved, but you can
[32:51] kind of extract and analyze it. Yeah. So
[32:52] yeah, for example, you could say, "Oh,
[32:54] you can only have read access to
[32:55] something." Yeah, [clears throat] cool.
[32:56] Um okay, we got one uh
[32:58] Jason DeLuca, owner and principal
[33:00] consultant Crossing Point IT Solutions.
[33:03] What are non-developers doing that makes
[33:05] Codec Codec's work well for them? What
[33:07] habits separate the people who are
[33:09] getting results from the ones who feel
[33:11] stuck?
[33:13] Um
[33:15] I think what I've seen is
[33:20] engaging with creativity and talking to
[33:23] others who have um seen success. It's
[33:26] just I I
[33:28] I am surprised myself at like some of
[33:30] the use cases.
[33:32] Um
[33:33] So being part of the community. Being
[33:34] part of the community is great.
[33:36] Um the second thing is
[33:40] trying to engage with it um with
[33:43] precise instructions instead of uh you
[33:47] know, being being quite vague about what
[33:48] you want.
[33:50] And so putting some effort into like
[33:52] just really being like, "Hey, you know,
[33:54] this is
[33:56] this is um this is precisely the outcome
[33:59] that I would like." We talked about this
[34:00] a little bit, but I would just really
[34:02] consider, you know, agents to be, you
[34:04] know,
[34:05] someone new who just joined the company
[34:07] has like no context about you, doesn't
[34:09] know what you like, doesn't know about
[34:11] your preferences. Like how would you
[34:12] engage with them? You would probably be
[34:15] uh you know, quite uh
[34:17] um
[34:18] you you would probably give them quite a
[34:20] bit of information about like where to
[34:22] find things, how to think about, you
[34:23] know, relevant documents to read. Mhm.
[34:25] Uh and so often people like forget about
[34:27] that. Uh and then the the third thing is
[34:32] connect a lot of like different sources.
[34:33] So, we have plugins for like almost
[34:35] everything now. We have more than 100
[34:36] different plugins. You can collect, you
[34:38] know, your calendar, docs, you know,
[34:40] notion. You can connect like your
[34:42] favorite tools. The more you give it
[34:44] access to your like tools and
[34:46] information from your world like the
[34:47] more useful it will be. Mhm. So, the
[34:49] more and better context gives you more
[34:51] and better results. And some of that
[34:53] context is in your head. And so,
[34:55] [snorts] it's your goals. It's things
[34:56] you've experienced that are not in the
[34:57] docs and you and and you you got to be a
[35:00] good boss and actually share some of
[35:01] these things. That's actually
[35:03] interesting. So, I
[35:04] I have taken now the habit of like
[35:06] writing everything down.
[35:08] Um so, a lot of like my own thoughts, my
[35:10] goals, it's like it's it's actually in
[35:12] files on my computer and uh Codex has
[35:15] access to it as well so that it can like
[35:17] align itself a little bit more. So,
[35:19] obviously yes, if it doesn't have access
[35:21] to, you know, it doesn't have access to
[35:22] your brain so
[35:24] it can't read your mind.
[35:25] So, you have to verbalize it something.
[35:27] >> Yeah. Neat. Okay, we've got one more
[35:29] from the community. Daniel Green, uh
[35:31] Kansas City AI Collective. What are your
[35:33] favorite real-world uses of computer use
[35:37] so far, especially outside traditional
[35:38] coding?
[35:40] Um
[35:41] I I use it for shopping.
[35:43] Um where it it will it will go and um
[35:47] order stuff for me so
[35:49] uh I use it What I do is on my personal
[35:53] computer I have Codex do meal planning
[35:55] and then um
[35:57] and then it will go and actually like
[35:59] order the ingredients. So, that's the
[36:01] thing that I use it for um
[36:03] another thing that I've seen uh people
[36:05] use uh it for is uh it's all it's it's
[36:09] almost impossible to find like which
[36:11] settings
[36:12] you know, which settings pop up to open
[36:14] up, you know, like beta on Windows on or
[36:16] Mac OS to to do something on your
[36:17] computer. And so, it's just being like,
[36:19] "Hey Codex, um I'm trying to
[36:22] uh change this on my computer. Just Can
[36:24] you show me how to get there?" And it
[36:27] will go and you know, it will actually
[36:28] like you know, take you around and like
[36:29] you know, click on the right things. Um
[36:32] and then you know, you learn something
[36:33] about your computer.
[36:34] Um sometimes as well, um
[36:37] there's like something that you want to
[36:38] kind of want to adjust, um
[36:41] you know, some slides or integrate like
[36:43] an image and then your computer uses
[36:45] that for that as well. For technical
[36:47] people, it is very, very useful for QA.
[36:50] Um where, you know, you Codex can then
[36:52] like open the app and click around and
[36:55] test that it actually functions. Yeah,
[36:56] amazing.
[36:58] Um okay, let's see what we got. Sai Sri
[37:00] Hamsini in at AI/ML engineer at
[37:04] AdeptMind. What are the biggest mistakes
[37:06] people make when prompting Codex? I want
[37:08] to know the answer to that one.
[37:09] >> [snorts]
[37:10] >> Oh, um
[37:15] As you progress in delegating to Codex,
[37:20] it can be
[37:22] at some point I see
[37:25] it becoming very tempting to just
[37:27] delegate everything including your own
[37:29] understanding. Mhm. [clears throat]
[37:31] And not
[37:33] using Codex enough to
[37:36] elevate
[37:38] your own understanding of the problem.
[37:40] Um
[37:41] if you just if you're just like purely
[37:43] like, oh, I'm going to delegate
[37:44] everything, at some point, you know, you
[37:46] might realize like you don't actually
[37:48] understand, you know, what's happening
[37:50] anymore. Um and then, you know, you've
[37:53] kind of like, you know, you're not as
[37:54] grounded and you're, you know, maybe
[37:56] going to lose a little bit, um of your
[37:59] your productivity and like that's not
[38:01] good. So,
[38:02] spending more time to also gather
[38:04] information and have it like explain
[38:05] things to you. It can draw diagrams like
[38:08] images V2 is like amazing at this, um
[38:11] because it's very good at rendering text
[38:12] too. And so, I see people use it all the
[38:15] time to go and read, um you know, read
[38:17] like
[38:18] um our plans for launches or, you know,
[38:21] like
[38:23] and things from marketing, or like parts
[38:24] of the code base, and then create images
[38:26] to like, you know, just explain a
[38:27] concept to you if you're trying to learn
[38:29] something. So, I would say like the
[38:30] mistake I'm seeing is just like too much
[38:32] delegation, and you know, not enough
[38:34] like using it to just helping you
[38:37] understand. I totally agree. Whoever
[38:38] does the work does the learning,
[38:39] basically. So, I I built myself a little
[38:42] skill, a tutor skill, flips the tables,
[38:44] and it asked me to give
[38:45] it answers about what it's taught tried
[38:48] to teach me already, so that I can force
[38:50] myself to do that learning of active
[38:52] recall. Yeah, super cool. Let's see,
[38:54] what else do we have? Um
[38:56] Why does Codex matter beyond software,
[38:58] and and where do you see these use cases
[39:00] going next? So, we're really talking
[39:02] about the future. We've talked a lot
[39:03] about how things are changing up till
[39:04] now. I think we're the future for a lot
[39:06] of people, but there's a future for us,
[39:08] too, and for everybody together.
[39:10] Yes.
[39:11] So,
[39:13] fundamentally, what we're building is a
[39:15] very general and very powerful agent
[39:18] that
[39:19] if connected to the right sources of
[39:22] information, and you know, with
[39:23] increasing access to all the information
[39:25] in the world, and given the ability to
[39:29] take actions upon the world, we'll be
[39:31] able to do pretty much anything that you
[39:34] know, you allow it to do.
[39:36] And so, the promise there is just
[39:38] incredible value creation.
[39:40] Um
[39:41] A lot of things that, you know, would
[39:43] have taken like prohibitive amounts of
[39:45] of time, you would never have gotten to,
[39:47] or like just like too hard to accomplish
[39:49] for humans, will become possible.
[39:52] Everyone we will distribute this
[39:54] the the intention is to distribute this
[39:56] as like a broadly broadly as possible,
[39:58] and to give, you know, the entire world
[40:00] access to these capabilities, and these
[40:03] agents.
[40:04] And so, it's just really about, you
[40:06] know, just lifting, you know, the the
[40:09] the amount of things that people can
[40:11] just, you know, even dream about
[40:12] accomplishing. Yeah. Yeah. I've spoken
[40:14] everybody from
[40:15] prize-winning physicists to
[40:18] industry project managers, they've come
[40:19] to me and said, "I had a stack of good
[40:21] ideas and I've been able to start work
[40:24] on them and they never would have been
[40:25] realized in the world before. It feels
[40:27] like we're entering a golden era." It
[40:29] feels like it. Yes. Yeah, amazing. Okay,
[40:32] well, Thibault, this was a great
[40:34] session. Thank you.
[40:35] >> Thank you. Um
[40:36] I and I just want to follow up uh
[40:39] Thibault mentioned that having a
[40:40] community will help you learn how to use
[40:43] Codex better and I just want to make the
[40:44] point that Open AI Forum is a community
[40:46] where people are using Codex and we can
[40:49] teach each other and support each other
[40:50] in learning that better and and I hope
[40:52] everybody in the session today will
[40:53] continue to engage because this is a
[40:55] place where skills are modeled and
[40:56] transferred uh so that you can bring
[40:58] this tool into your life in new ways. Um
[41:00] the question we tried to answer today is
[41:03] is why um why should people who don't
[41:05] code care about Codex? I hope we
[41:07] answered that. I'd also like to make the
[41:08] point that although the word code is in
[41:10] Codex
[41:12] Codex actually means book, which is
[41:14] something we all are used to,
[41:15] right? So so so Codex is a more general
[41:18] word than just code. Um but I but I hope
[41:20] this conversation um made that answer
[41:22] concrete. Uh
[41:24] Codex is uh useful for developers, um
[41:27] also useful for people anybody knowledge
[41:29] workers, people working with information
[41:32] go you know, searching for the
[41:33] information they need, the needles in
[41:34] the haystacks, uh struggling to make
[41:37] sense of it, data analysis, visualize it
[41:40] with kind of instant visualization like
[41:43] Thibault showed it showed us um and
[41:46] prioritization [clears throat]
[41:46] and performing complex tasks in in the
[41:48] background of their lives. So, we think
[41:51] that Codex is uh bringing those powers
[41:54] those superpowers to lots more people
[41:56] and we're really excited about what
[41:58] that's going to mean for our economy and
[42:00] society and their businesses. Um
[42:03] as we close I'd like to ask everybody
[42:05] listening to think of one workflow in
[42:07] their life or in their organization
[42:08] where code Codex might help. Um and that
[42:11] could be a research brief or a plan or
[42:13] an onboarding process or reporting or a
[42:15] decision memo. And just give it a try
[42:18] because uh you've heard a lot from us
[42:20] today, but the the clearest message is
[42:22] just going to be your experience with
[42:24] the tool. Right? You don't have to trust
[42:26] anybody else about it. You can just
[42:27] experience that tool. It's there,
[42:29] available to you.
[42:30] So,
[42:31] thank you again, everyone, for being
[42:33] part of Open AI Forum. Uh we appreciate
[42:35] your time and your curiosity and the
[42:37] thought everybody brought here, and
[42:39] especially the the folks who got their
[42:42] questions in. Those were great
[42:43] questions. So, thank you. Um we have an
[42:45] upcoming event. It's
[42:47] Tuesday, May 26th, "How AI Could Detect
[42:50] and Prevent the Next Pandemic" with the
[42:53] Gates Foundation. We're super excited
[42:55] about that. So, thank you, everyone, and
[42:56] thank you, Tegan.
