Greg Brockman, President of OpenAI, made a latest bold prediction: After the integration of ChatGPT and Agent, all interfaces will disappear.
Greg Brockman, the President of OpenAI, outlined OpenAI's ambitious plan to integrate conversational AI with autonomous agents in a recent in - depth interview. From the birth of ChatGPT to the evolution of Codex, and finally to the ultimate vision of "the disappearance of interfaces", he mapped out a clear path.
1. From "Super - apps" to "No Interfaces"
When asked about OpenAI's product direction, Brockman believed that the concept of "super - apps" was inaccurate.
"From a broader perspective, what we really want to build lies within AGI. Think about what people have been using since ChatGPT came out - it's a language model. You talk, it replies, which is great. But when it was launched in 2022, it had no memory function, was not connected to any tools, and had no context."
"So, it's just a conversation. Intelligence is just a part of it. What people really need is the ability to get work done and achieve goals."
His core point is: What you ultimately want is almost no interface.
"You don't need any product - what do you want it to look like? What's the interface? You and I can have a continuous conversation, and some form of companion entity can achieve goals for you."
"Interfaces will disappear." This is a bold prediction: Future AI interactions will no longer require buttons, menus, or toggle switches - only natural conversations, and then the system understands your intentions and executes.
2. Codex: From a Code Tool to a General Agent Platform
Codex was initially designed as a software engineering tool, but its use cases have far exceeded programming.
"The non - software workload being done by Codex has been growing exponentially. Inside OpenAI, our user penetration rate is now basically on par with Slack."
Brockman gave a vivid example: A member of OpenAI's communication team was organizing an event. After talking to ChatGPT, the system asked all the questions about the event - attendees, dietary preferences - and then automatically arranged the seating chart, basically completing all the related work.
"In this way, she can focus on the part she really wants to do - thinking about the vision she wants to achieve. I think this kind of situation will happen commonly."
The key is that Codex enables non - technical users to build agents. Users no longer need to write code. They just need to connect tools - Gmail, calendars, file systems - and then tell the AI what to do through natural language.
3. Trust is the Most Critical Product Feature
When AI starts to take actions on your behalf, trust becomes the core issue.
"Trust is an aspect of the Agent era. We need to really understand how to build trust with these systems, understand their strengths and weaknesses, and figure out what kind of work you want to delegate to them."
Brockman described the evolution of trust from "drafting" to "sending":
"Suppose you have a connector. It may not send an email directly. It will say: 'I've drafted it. You need to send it to them.' In another scenario, you've built enough trust with the system, and it will show: 'The email has been drafted and sent.'"
He believes that trust is not something that can be "granted", but rather "earned" by providing a large number of tools, control, and supervision mechanisms.
"We think this will be a very important thing - a key product feature and differentiating factor."
4. The Failure Lesson of 2023: The Model Wasn't Ready
It's worth recalling that OpenAI tried to connect ChatGPT to tools as early as 2023 - a plugin system was released at that time.
"It didn't work at all because the model wasn't ready. We could only expose three different connectors to the model at the same time, otherwise it would forget. At that time, we only had a context window of about 2000 to 4000 tokens."
Brockman compared that era to "the early days of computers": "Just like in the 1960s and 1970s, your computer had very little memory. And today, your phone is more powerful than any supercomputer of that era."
Now, the context window has reached 52 million tokens, and the model can connect to hundreds of different tools and even access the entire file system.
"You can almost have the full power of the Internet, and almost any application you want is within reach."
5. Model Evolution: No Bottlenecks, Only Breakthroughs
Regarding the question of "whether LLMs have reached a bottleneck", Brockman's answer was straightforward: No.
"I was wrong. Every time we felt like we had reached a bottleneck - 'Oh, this doesn't seem to meet the scale expectations' - it usually turned out that we had encountered a bug, or our implementation didn't match the mathematical calculations well."
He cited the history of neural networks: The first hardware implementation of the perceptron was in 1959. "For more than 70 years, people have been saying that these things will never work, will never scale, and it feels like hitting a wall. There is no bottleneck yet, and the wall is still out of sight."
He believes that basic science allows for continuous progress. "We're building giant supercomputers. It's difficult and expensive. We have teams working very hard to solve these incredibly tricky technical problems - network protocols designed by ourselves, and someone is responsible for looking at the strange fluctuations in each single layer of the stack."
"If you can assemble the right team, put the right mission in front of them, and persevere - it turns out that it's worth it and achievable."
6. Commodification of Intelligence? No, There's a Huge Differentiation Space
When asked if "models are becoming commodities", Brockman thought it was the wrong question.
"Intelligence is not a dimensionless thing."
Even if two models have a similar level of general intelligence, their depth in specific fields will be completely different. "If you've never operated a spreadsheet, you can't successfully complete some complex modeling work."
This means there is a huge differentiation space. Different companies and different models will form advantages in different fields. "We can't be excellent in all fields at the same time."
He also talked about the value of "depth": "Think about the 37th move of AlphaGo - it changed people's understanding of Go. Now, more people are playing Go than ever before. This depth will never stop growing. Every time you solve a puzzle, about 10 new puzzles will be unlocked."
7. The Computing Era: Demand Always Exceeds Supply
"We're moving towards an economy dominated by computing," Brockman said.
He believes that computing power will always be in short supply. "From a business - model perspective, all suppliers' computing resources are sold out. The number of people using these agents could reach 10 million, 20 million - we haven't reached the planetary scale yet. We have a billion users, but we haven't brought enough agents, and there isn't enough power."
When asked if "OpenAI should buy all the computing resources", he said bluntly: "The answer is always yes."
But he also admitted that the price of models will continue to decline. "Cutting - edge intelligence is always the most expensive. But a year later, that level of intelligence will become quite common and cheaper. Then there will be new and better things."
8. The Most Exciting Application of AI: Healthcare
Brockman shared some personal stories about the application of AI in the medical field.
He mentioned the story of Sid Sijbrandij, the CEO of GitLab - he input his cancer - detection data into ChatGPT and defeated cancer with the help of AI. There was also a dog named Rosie - its owner used AlphaFold to analyze its tumor mutations, designed an mRNA vaccine, and successfully shrank the tumor.
"This is definitely the standard practice. I personally have many friends who have done similar things - getting the right health - diagnosis data, using Codex correctly, and gaining insights from these models."
He revealed an astonishing figure: More than 230 million people use ChatGPT to ask health - related questions every week.
"I think we've always lived in a world where patients lack autonomy. Patients have to be their own doctors - you're the final decision - maker, and you're responsible for it. Doctors can make mistakes, and you'll pay the price."
His wife has multiple diseases. "I don't even know how we would deal with many of her conditions without ChatGPT. We're just starting this journey."
"If you can really help prevent diseases and proactively solve potential health problems, you can relieve the burden on doctors and nurses. They're exhausted now. AI can help solve all these problems."
In Brockman's vision, AI is not just a chat tool, nor just a code assistant. It's a real companion that can understand you and act for you. From conversation to action, from interfaces to no interfaces - this is the "Agent era" that OpenAI is building.
This article is from the WeChat official account "AIGC Index", author: Mark. It is published by 36Kr with authorization.