In the past few weeks, OpenAI has been laying a foundation. Although most users have just begun to really explore the ChatGPT task (a new feature that allows users to arrange and trigger tasks), the company is preparing for more important things.
The Operator released yesterday is another clear signal of the development direction of artificial intelligence: from a model of simple processing information to an agent that can actively work with us.
Every day, we spend countless times to browse websites, fill in forms, book services and manage digital tasks. Most artificial intelligence is on the spot, and it is limited to providing suggestions or processing texts. Operator and other recently released agent announcements (such as the Computer Use of Anthropic and Project Mariner of Google) completely changed this dynamic.
The technical achievements here are significant. Openai creates an artificial intelligence that can view the network interface and interact with it like humans. It captures the screenshot, understands the visual layout, and decides to click position, input content, and how to navigate.
The following is the information about Operator Agent you need to understand: Although many AI tools are essentially subject to API and specialized integration, Operator can use the network like you. It sees the screen, understands the context, and takes action directly.
Observe the real performance of the operator carefully
When an artificial intelligence company releases the base on time, it is very important to carefully study the actual meaning of these numbers. The performance of operators shows different situations in different test environments.
The most impressive indicator is the success rate of Operator in the webvoyager reference test of 87%. This is important because Webvoyager tests the real world website -the actual platform we use every day, such as Amazon and Google Maps. This is not a control laboratory test. This is a performance in the wild.
But when we look at other benchmarks, we will see a more delicate situation:
- Webarena reference: The success rate is 58.1%. Test tasks such as shopping and content management of simulation websites. The lower performance here actually reveals some important parts of artificial intelligence agents to handle structured and non -structured environment.
- Osworld benchmark: The success rate is 38.1%. This will test complex multi -step tasks, such as PDF in merging emails. The significant decline in performance shows us the limitations of the current artificial intelligence agent when the task needs multiple context switching.
What these numbers are interested are how they reflect the model of human learning. We usually perform better in the familiar realistic environment than in artificial testing scenes. Operator performed well on the actual website, but it is not performed on the simulation website. This fact shows that its training is preferentially considered practicality rather than theoretical performance.
These benchmark tests set new records in the automation of the browser, but the different success rates of different tests tell us some important information about the OpenAI strategy.
Think about your own web browsing situation. Most tasks are simple: fill in forms, buy, and make appointments. This is the flash point of Operator 87%. More complicated tasks (decreased performance) are usually valuable tasks of human supervision.
These data show that OpenAI is making a thoughtful choice: first improve the common tasks, and then gradually expand to more complicated operations. This is a practical method that gives priority to real -time utility rather than theoretical ability.
AI proxy benchmark (Openai)
The cooperation between Openai and Operator reveals the strategy of careful planning.
First, consider the timing. Recently launched functions such as ChatGPT TASKS are not just added functions, but for users to prepare for independent agents.
But what is really interesting is: OpenAI plans to open the CUA model through the API. This means that developers will be able to create their own computer use agents.
Its significance is great:
- Integrated potential
- Direct merger into the existing workflow
- Customized proxy that meets specific business needs
- Automation solutions in specific industries
- Path of future development
- Extend to Plus, Team and Enterprise users
- Directly ChatGPT Integrated
- Geographical expansion (although Europe needs longer time due to regulatory requirements)
Strategic partnerships can also explain the problem. Openai is trying to create a complete ecosystem. They not only cooperate with companies such as DOORDASH, Instacart and OpenTable, but also organize cooperation with public sectors such as Stockon.
This indicates that the future artificial intelligence agent will be not only an assistant, but also an indispensable part of our interaction with the digital system.
What does this mean to you?
We are entering a stage. Artificial intelligence is not only answering questions, but also becoming an active participant in our digital life.
Think about your daily online task. It does not require the complex and strategic work of your professional knowledge, but the task of repetitiveness. I am talking about studying travel selection, filling in standardized forms across multiple websites, collecting data from various network sources, and managing daily booking. This is where Operator first eliminated digital and complicated work. But this is not there. Over time, artificial intelligence agents will be able to complete increasingly complex workflows.
Early performance data also tells us some important information: Operator is good at performing conventional Web tasks with a success rate of 87%. Learn to effectively integrate its early adopted people will have a significant productivity advantage.
The integrated timetable reveals the cautious approach of Openai. They started from PRO users in the United States, and then extended to Plus, Team, and Enterprise users, and finally integrated directly into ChatGPT.
We are observing the fundamental change of the working method of artificial intelligence tools. You should ask your real question whether to adapt to this change, but how to strategically adapt to this change. Technology will continue to develop, but the principle is still unchanged: artificial intelligence is shifting from answering questions to taking action. Those who understand this transformation as soon as possible will have significant advantages in shaping how these tools are integrated into their workflows.
