Article by: Techub News Compilation
In a recent dialogue at the OpenAI forum, Thibault, head of OpenAI Codex, engaged in an in-depth discussion with Chris Nicholson from the global affairs team about the evolution and impact of Codex. This conversation is significant because it clearly reveals how Codex—initially known as a code generation tool— is crossing the boundaries of software engineering to become a core tool for knowledge workers across various industries, fundamentally changing the way we interact with computers and information.
From Code Generation to General-Purpose Agent: The Evolution of Codex's Mission
The story of Codex began with a classic challenge: achieving programming capabilities of a high-output software engineer with AI. About two years ago, the OpenAI team started exploring this field. The initial public version was "Codex Web," a cloud-based entity that could analyze code repositories and submit changes based on user intent. However, this solution encountered significant friction: it was complex to set up, and the model's reliability had not yet reached a level capable of handling long-term tasks perfectly.
The team quickly realized that running tools in the cloud and requiring everyone to configure environments was overly complicated. Therefore, they shifted to a more natural path: enabling Codex to work on everyone's local machines. This transition lowered the usage threshold and laid the groundwork for broader applications later.
The real turning point came about six months ago with the release of GPT-5, which brought a leap in the model's versatility and reliability, particularly in handling long-term tasks. Thibault noted that even software engineers spend only about 20-30% of their time actually writing code in their daily work. Most of their time is spent dealing with tickets, prioritization discussions, architectural decisions, troubleshooting reports, managing system outages, collecting information, and more. Those early adopters of Codex had already begun using it to handle these everyday non-programming tasks.
The team discovered that for Codex to be more useful on programming tasks, it needed access to a broader context of information, such as data stored in Notion or various documents. As they continuously improved reliability in this area, an interesting phenomenon emerged: most tasks performed by Codex today are actually non-programming tasks. Codex evolved from searching code to searching all kinds of documents and returning information, which is exactly the capability needed by all knowledge workers.
Thibault shared a key moment that made him realize Codex's potential would benefit everyone. On the eve of a product launch, product manager Alexander Americus used Codex to track the status of all pending changes. Thibault had never seen anyone as efficient as Alexander: it was as if he had many little Codex agents working for him, handling tasks, following up with people, updating documents, aggregating all information from user feedback and developers, and keeping the plan tidy and updated in real time, while Alexander himself was in meetings discussing matters. This left him amazed: "We are really changing everything, not just software engineering." Before this, Alexander had to personally sift through Slack channels, documents, or GitHub PRs, spending a lot of time coordinating. Now, he could delegate those time-consuming tasks to the tool, which could complete them while he was in meetings.
The Codex model excels at collecting the right context and summarizing it, which is a powerful use case. Alexander also used it to "prompt" information: the Codex agent connects to Slack and can send messages to inquire about the latest status of someone's work, completing all follow-up tasks on behalf of Alexander. "Prompting" consumed a lot of time, and now it can be automated.
Efficiency Revolution and Role Reshaping: The Ripple Effect of Codex
Alexander's efficient use created a ripple effect. Engineers' development speeds have increased dramatically, and construction rates are unprecedented. Adjacent roles are also changing: the role of designers is evolving, and the role of product managers is changing. The OpenAI team has been contemplating how to accelerate and empower them to achieve greater productivity. Subsequently, the bottleneck shifted to communication and marketing departments— the team generated so much output that telling the world and maintaining storyline coherence became a challenge.
OpenAI's delivery speed is incredibly fast. Thibault candidly stated that without Codex, this would not be possible. Today, Codex is crucial to them. He believes that other companies, even those with ten times as many engineers, could follow a similar model because the technology has reached a critical point: these agents can handle very general tasks. Nowadays, many things accomplished behind computers can be assisted by agents.
Whether preparing presentations to coordinate stakeholders, collecting public awareness background, conducting market research, organizing information, or working in financial departments (Sarah Friar mentioned that she used Codex to organize the recent fundraising efforts), Codex has demonstrated beautiful versatility. It has evolved to execute not just code generation but general tasks.
Thibault observed that everyone needs to cope with faster-moving situations and adapt more swiftly. Problems that once took days to resolve may now only take hours. This trend is evident in science and engineering, as well as tasks like in-depth market research or analyzing public sentiment on new features— tasks that used to take significant time to source, summarize, and condense information for different audiences are now compressed into hours or even fully automated.
This means everything is accelerating, and people need to adapt to the increased pace. Individuals can also accomplish more work independently, which is highly gratifying and empowering. Previously, you might have needed to find someone to discuss something, but now you can do it yourself. This phenomenon is also prevalent within companies regarding data issues: everyone is empowered to directly ask data-related questions, such as "How successful are we in a certain market?" "Are we growing in India?" "What’s happening in South Korea?" Now, people can directly ask Codex to pull up dashboards (even without knowing specific locations) and dive into business details without bothering data analysis teams, allowing them to focus on more interesting work.
Thibault believes we are at a historic moment: in the past, those who raised questions and those who built solutions were two separate groups, resulting in lengthy dialogues about products that ultimately only yielded a fair outcome, unable to progress further due to lack of time. Now, those who pose questions can quickly build solutions and iterate on the required changes. This is why there are also UX designers within the team, who are pushing code, making changes, and shaping products with developers to create the best experiences and achieve astonishing results. They no longer need to convince engineering teams to prioritize changes that might seem trivial to them but can significantly enhance craftsmanship and user experience for designers. This is equally empowering for them.
This feels like entering a "home cooking" era of highly personalized software. Thibault agrees that this is indeed the upcoming wave: everyone will be able to have their own personal software, maintaining it to meet their needs precisely.
He shared an interesting personal use case: living in San Francisco, he was surprised by the price of bread. He had Codex find the best bread in the city and create a spreadsheet containing locations, purchase points, and prices. Codex worked for 5 minutes, generating a table containing information on Jane the Bakery, Arsicault, Tartine, and more. Subsequently, he wanted to consume this information in a more visual way, so he asked Codex to create a webpage. Codex did just that, placing all the bread information on a map. The entire process took about 10 minutes; he only needed a simple prompt, and could even complete it via voice without typing. This means that anyone concerned about certain data and with access can essentially create websites, analyze data, visualize, and share. In the past, if you had the relevant skills, this might take a weekend to create, but now it can be done almost instantaneously, and if unsatisfied, you simply instruct it to change.
Thibault emphasized that Codex is about acquiring data, visualizing data, guiding you to insights about the world, and then making decisions based on needs— this is the common cycle in our lives: how to make better decisions to achieve goals. This process can be very simple or very complex, such as OpenAI's recent fundraising effort that followed a similar process. Codex can handle the simplest tasks as well as the most complex ones.
Personal Chief Assistant and Future Outlook: Trust, Community, and Continuous Learning
Thibault himself, much like Alexander, coordinates and arranges tasks. In his sidebar, before meetings begin, he initiates hundreds of different tasks daily for Codex to manage. These tasks include organizing desktop files, managing computing clusters, helping understand shift rotation status and engineer performance, knowing upcoming release schedules and marking any potential risks he needs to pay attention to. He uses Codex as a mini "chief assistant," running automated workflows daily, sifting through Gmail, Notion, and calendar events, summarizing his day’s tasks, and marking risks. He can set it to run every day at 9 AM, then find the report in his inbox each morning at 9 AM. Codex helps him prioritize tasks, keeps him focused on what matters most, and handles those trivial, time-consuming tasks that might never get done.
Before Codex, the most troubling thing for him was that he wouldn’t do certain things because he felt he didn’t have time to handle them personally but still needed to bother others for information. He thought it might not be important enough to put on someone else's desk to ask for help. Now, he can obtain the information he needs, gain various personal reports, and build personal software that he previously didn’t have time for. Codex takes care of all the tedious small tasks that used to be done manually on the computer, allowing him to focus on what he truly wants to think about. Many tasks that might have taken weeks before now take just seconds, but more critically, many tasks that would have never occurred due to the time cost are now possible. Codex also makes him enjoy work more, as he no longer feels mentally overloaded and doesn’t worry about things being missed.
Thibault believes this is essentially a tool to combat burnout and information overload. We are all surrounded by tools intended to help us, but we often find ourselves trapped within them. Codex, as a tool, is liberating us. Many teachers, doctors, and others feel burned out and overwhelmed because they are stuck in tools doing manual data entry and such. Our relationship with tools is fundamentally shifting. For him, the commitment lies in having a nearly trustworthy partner that can do a lot of work on your behalf, reaching a level of trustworthiness: if you delegate something, it will get done; if it doesn’t meet satisfaction, it will flag it to you; you know it will execute well and may even do it better than you would personally. You can also trust this partner to filter out a lot of noise and timely flag significant matters.
The future he envisions is one where he doesn’t even have to read emails. He might just need a small personal agent to read his inbox, flag him when something truly important arises, seek his input, and then complete the tasks. You wouldn’t need to search for “needles” in a dozen different “haystacks"; the "needles" would be organized into briefs. You could say, "This is my goal for today, help me handle everything else," and trust that it will be achieved.
In recent months, the things you can trust this tool to handle have changed, particularly the time span of tasks. OpenAI introduced a more advanced feature: the `/goal` command. It allows you to enter a mode, giving Codex a long-term goal, and it will relentlessly pursue it. For example, you could set it a difficult math problem to solve, and it might work for hours, days, or even weeks until it believes the goal has been achieved. They have seen it used to improve program performance, rewrite entire programs from one language to another, and tackle scientific questions, achieving some very cool breakthroughs in math and physics. A few months ago, they were excited it could work for 10 minutes; now they’re discussing how agents can work on the most challenging tasks for weeks.
The future direction is that it won’t stop. You will have a 24/7 running agent continually doing useful things for you while being guided in the process. Right now, it’s still turn-based, activated when you have specific tasks to assign. Goal-oriented task-solving is a massive unlock that can create immense value. But the next step is to have it continuously running, doing useful things whether or not you instruct it. It might complete all the useful tasks it identifies at some point and then temporarily sleep until you need it.
On successfully setting objectives, Thibault suggests: when interacting with the agent, you can ask it what it can do in a very casual way since it understands its own capabilities. A good trick is to help it evaluate its success accurately. If you can describe what "good" looks like, what "solved" entails, and what you want to see at the completion of the task, then Codex can grasp whether it has done well or completed the task. You can set numerical metrics or precisely describe the output. For example, you could say: "I’m presenting my work; I want a slide deck. I wish it to have 10 slides. The first two slides include this type of information. The next six slides delve into the problem's core and perform a technical breakdown. Finally, I want two slides to pose open-ended questions and conduct a Q&A." If you clearly and specifically describe the output you want, it’s more likely to succeed. This is very much like what you might do with an assistant or deputy.
Regarding why non-coders should shift from ChatGPT to Codex, Thibault believes this will be an evolutionary process; he does not expect everyone to switch to Codex, but it serves as a great complement to ChatGPT. He recommends using it for tasks that need to be done for you: anything involving computer file operations, automated runs, or tasks executed in the background every few hours can be accomplished by Codex. ChatGPT remains his first choice for quick answers. He reminisces about having to copy and paste code from ChatGPT into files or terminals, but now he doesn’t have to. The copy-paste era is over. Codex can handle code and data directly. If you have files or images on your computer, just tell Codex to use this file or read it, and it can process it directly without manual clicks.
Regarding the biggest bottleneck enterprises face in adoption, Thibault believes it is not a capability issue; the capability already exists. It mainly revolves around trust. Trust is about security and safeguarding. Allowing an agent to run around the company may lead to deleted sensitive files, uploaded information, or emails sent with leaked information, which would be disastrous. OpenAI has put significant thought into this. By default, agents run in a sandbox with stringent controls, only accessing specific parts of the file system. You can restrict it to a certain folder and disable network access. They provide many enterprise control features and invested in a capability called "Auto Review" (mentioned in the alignment blog post). They created an overseeing agent that can review the primary agent's actions. The primary agent Codex is incentivized to work for you, sometimes possibly taking slightly risky actions. Therefore, another agent reviews every move it makes and flags and stops it when high risk is detected. He anticipates more such innovations will be needed.
On how non-developers can make Codex work better and what habits differentiate those who achieve results from those who feel stagnant, Thibault observed: interacting creatively and communicating with those who've seen successful use cases is crucial. Engaging with the community is very helpful. The second point is to attempt to interact using precise instructions rather than vaguely describing what you want. It’s essential to spend some effort clarifying: “This is the exact result I want.” The third point is to connect various sources. OpenAI now has over 100 different plugins. You can connect calendars, documents, Notion, and your favorite tools. The more access you grant it to your tools and world information, the more useful it becomes. More and better context leads to better results. Part of the context is in your mind: your goals, experiences that you haven’t documented. You need to be a good boss and share some of these things. Thibault has now developed the habit of writing everything down: his own thoughts, goals are stored in computer files that Codex can access for better alignment. Clearly, if it cannot access your brain, it cannot understand your mind, so you need to verbalize it.
Regarding his favorite real-world uses of Codex (especially in non-traditional coding areas), Thibault shared: he uses it for shopping. Codex orders items for him. On his personal computer, he uses Codex for meal planning, after which it goes to order the ingredients. He has also seen people use it to find computer settings: when trying to enable certain beta features or make adjustments on Windows or Mac OS, just ask Codex: "I’m trying to change this setting on my computer; can you tell me how to get there?" It guides you to click the right place, and you even learn about the computer. Sometimes, when you want to adjust slides or integrate images, you can accomplish it with Codex as well. For technologists, it is incredibly useful in QA (Quality Assurance): Codex can open applications and test their actual functions.
Regarding the biggest mistakes people make when prompting Codex, Thibault pointed out: as you delegate more to Codex, it can become tempting to delegate everything, including your own understanding. Overusing Codex to delegate without leveraging it to enhance your understanding of issues. If you purely delegate everything, you may realize that you don’t truly understand what’s happening, you lose your footing and may reduce productivity. Therefore, it's crucial to spend more time gathering information and letting it explain things to you. It can draw charts, Images V 2 excels in this regard as it can also render text well. He has seen people often use it to read release plans, marketing materials, or certain parts of codebases, then create images to explain concepts that help you learn. The error he has seen is over-delegation without sufficiently utilizing it to help you understand.
Finally, regarding the importance of Codex going beyond software and the future direction of these use cases, Thibault summarized: fundamentally, they are building a highly versatile and powerful agent that, if connected to the right sources of information, will be able to accomplish almost anything you allow it to do, with access to all information in the world. The promise lies in creating unbelievable value. Many tasks that once required prohibitive time, would never be achieved, or were too difficult for humans will become possible. OpenAI intends to distribute this technology as broadly as possible, giving the entire world access to these capabilities and agents. This truly pertains to increasing the number of things people can even dream of accomplishing.
Dialogue host Chris Nicholson added that the OpenAI forum is a community where people use Codex, teaching and supporting each other to learn better. He hopes that today’s attendees will continue to engage because this is a place where skills are demonstrated and passed on, allowing you to bring this tool into your life in new ways. The question they are trying to answer today is: why should non-coders care about Codex? He hopes they have answered that question. He also pointed out that even though the word "code" is in Codex, it actually means "book," which is something we all are familiar with. Therefore, Codex is a term that is more universal than "code." He hopes this conversation has made the answers clearer. Codex is useful for developers as well as for anyone—knowledge workers, those dealing with information—who search for necessary information (needles in haystacks), struggle to understand it, analyze data, visualize it instantly as Thibault demonstrated, prioritize, and execute complex tasks in the background of life. They believe Codex is bringing these superpowers to more people and are excited about the implications this will have for the economy, society, and businesses.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
