ChatGPT Transforms into a Life Operating System: Altman Previews the Next-Generation Top-Notch AI
OpenAI plans to release a very powerful open-source model.
It enables people to run extremely powerful models locally and redefine the possibilities of "local deployment."
During a conversation at the AI Startup School in San Francisco, Altman announced the above news. The conversation was hosted by Garry Tan, the CEO of Y Combinator.
During the conversation, they also discussed OpenAI's development process, future direction, the origin of ChatGPT, and the construction of AI hardware.
Altman mentioned that the operating cost of the o3 model last week was five times that of this week, and the price decline trend continues. Moreover, the cost of the API will continue to drop significantly, and the open-source model will be excellent. The memory function of ChatGPT is not just a small piece of hardware but more like an AI companion.
In the future, GPT - 5 and various inference models will develop towards integrated models and connect with AI intelligent terminals and robots to become people's operating systems for daily life.
He also said that if users subscribe to the highest - level ChatGPT plan, they will be given a robot for free.
Based on the original meaning, the following is a summary.
The Future of GPT Inference Models
Tan: What surprised you the most about the latest o3 model? What emerging behaviors or use cases have impressed you?
Altman: I think we are in a very interesting era. Although we haven't seen inference model products that reach a new level of innovation yet, the capabilities of models in the world have entered a new realm, and there are still a lot of new things for us to build.
We will soon launch an open - source model that will surprise you. I think it will be much better than you expect, and you will be able to run very powerful models locally. Meanwhile, the cost of the API will continue to drop significantly, and the open - source model will be excellent.
Tan: For me, the memory function even feels like having a conversation with someone who knows me well. It's quite interesting.
Altman: Yes, the memory function is my favorite feature we launched this year.
I think it points to the direction we want our products to go. That is, you will have an entity that can understand you, connect to all your things, and proactively help you.
It won't be like you send a message and then it replies. It will run continuously, check your things, know when to send you a message, and know when to do something for you.
You will have special new devices that will be integrated into every service you use and accompany you throughout your life.
When this system runs continuously in the background and keeps pushing content to you, the interaction area will become richer. And when we launch the first new device, there will be more functions or content added.
But I think the key is not the small piece of hardware. This thing has evolved to the point where it can run in the background and feel like an AI companion.
Tan: I think we've seen the powerful function of integrating LMS with real - world data. I heard that MCP is coming to OpenAI. What's surprising about the actual integration? You know, at YC, we actually have an internal agent infrastructure and we've been using it.
Altman: Undoubtedly, people are starting to use ChatGPT as an operating system and integrating their entire lives into it.
Integrating as many data sources as possible, along with devices that always accompany you, such as new - type web browsers, connections to all data sources, memory, and continuously running models. If you put all these together, it will be a very powerful platform.
Tan: Do you think it will be in the cloud, on our desktops, or both in the future?
Altman: It will be a combination of all these. People will definitely run local models for certain things. If we can push half of the chat workload to local devices, no one will be happier than us. As for the cloud part, I think we will soon operate the largest and most expensive infrastructure in the world. Tan: Are you surprised by the high computational difficulty of running in the cloud? Altman: We are good at starting from scratch. Just like we didn't have ChatGPT.com two and a half years ago, now it has become the fifth - largest website in the world.
The Vision of GPT - 5 and Multimodal Super - Models
Tan: What will happen when the o3, o4mini inference models and multi - models like 4o develop in parallel and these two threads converge? What is the vision for GPT - 5 and beyond?
Altman: We can't achieve all our goals with GPT - 5, but ultimately, we do want an integrated model that can perform inference when needed and generate real - time videos when needed.
If you ask a question, you can imagine it will think very hard, do some research, and write a lot of code, like a brand - new application for you to use. I think it's like a real new - type computer interface. AI has already achieved this to some extent, but when we get a truly complete multimodal model, such as perfect video, perfect coding, and all - around in - depth inference, it will feel very powerful.
Tan: This seems to be a step towards materialization. You know, robots with vision, language, and reasoning abilities are a leap forward from what we currently have.
Altman: Our strategy is to solve this problem first and then ensure that we can connect it to robots.
But the era of robots is coming soon. I think if you subscribe to the highest - level ChatGPT plan, we will give you a robot for free.
Tan: It's such a crazy future to have robots doing real - world work.
Altman: I think we're not far from the goal. Although the mechanical engineering of robots is very complex and the cognitive abilities of AI are also challenging, overall, we're gradually mastering these key issues.
I think robots will be able to do some very useful things in a few years, but it will still take some time to manufacture a billion robots. I also don't know how many robots are needed to automate the supply chain and whether they can run the entire supply chain, such as driving mining equipment and container ships.
Tan: Talking about Level 3 AGI and AI agents, Greg Brockman said this is the year of agents. With tools like the operator code interpreter, what types of workflows do you think will disappear or emerge that we're not prepared for yet?
Altman: For a long time, ChatGPT has been like an advanced version of Google Search, still more of a substitute.
But now you can really assign a task to codeex for in - depth research. It can provide relevant suggestions, just like a junior employee can handle a task in a short time.
I think you can achieve this with the current o3, not to mention our next model. You can have many such experiences.
Tan: How do you view the future of human - computer interaction and interfaces? What limitations of these interfaces make you think this way?
Altman: Just like today's voice interfaces, I think they're a bit bad because they don't work very well.
If you can say to your computer: "This is what I want to accomplish today. If I'm delayed or there are any changes, I trust you to handle everything without disturbing me."
Unless it's an excellent human assistant, the interface will disappear. I hope we can show people a different way of using computers.
AI for Science: Sam's Personal Bet
Tan: Looking ahead ten to twenty years, what personally excites you the most? What should people build now to achieve such a future? Altman: This is a world of super - intelligence that's hard to imagine. I'm looking forward to seeing its development rather than giving a vague answer. I think applying AI to science is what personally excites me the most.
I believe that all long - term sustainable economic growth in the world, just like all things that can improve people's lives, basically lies in discovering new science and having a fairly good governance mechanism so that this science can be developed and shared with the world.
If we can significantly increase the speed of AI in scientific exploration, I believe it can bring improvements and miracles to everyone.
OpenAI Development: Talent Attraction and Competitiveness
Tan: Are you one of the best in the world at gathering the smartest people? What's the toughest lesson you've learned in recruitment?
Altman: Recruiting people who are truly smart, motivated, and good at high - performance teamwork can make you 90% successful. I'm always surprised by how much people focus on other things during the recruitment process.
Recruiting someone with a good track record, strong curiosity, who actively engages in work and aligns with the company's vision will yield good results.
Tan: Does having a good track record mean someone who has held management positions in top - tier institutions for twenty years and has a top - notch reputation? Altman: At the beginning of the startup, I don't recommend recruiting such people. Frankly, YC recruited people with such management experience at the beginning, but the results weren't ideal.
We still choose to recruit young and energetic people who can get the job done rather than those with extremely glorious resumes. I'll ask, "What's the most impressive thing you've ever done?"
Tan: As the CEO of OpenAI, what's the toughest lesson overall?
Altman: We have to do many things simultaneously, and many large companies are challenging us in various ways. We have to spend more energy dealing with these problems, and we need to switch from one major decision - making mode to a completely unrelated but equally important one to handle them.
Tan: For many software engineers who want to create B2B or SAS products, how can they handle a complex and troublesome task now?
Altman: Now is the best time to start a business in the history of technology. The key to success is that startups can iterate more easily and at a lower cost than large companies. Large companies have many advantages, but their iteration speed is very slow. However, cheap things are also easily replaceable, so there are many perspectives to look at this issue.
I suggest looking at it this way: Everyone will face the same challenges and opportunities, but when the industry cycle changes so significantly, startups are almost always the winners. We may have never seen such a huge change.
Taking action in this direction, I think you'll be in a very favorable position. Maybe you can invite me to talk about, for example, what defensive areas you can build over time. I think this is an internal issue.
Tan: What would you say to your student - self?
Altman: I wish someone had taught me the importance of long - term belief and resilience. Many people give up after one failure. Learning how to persevere is really important.
Cultivate trust in your intuition and continuously improve your decision - making and intuition over time to strengthen this trust. The good parts are really much better than you think, and the difficult parts are hard to express in any way you can understand. You have to persevere.
One More Thing
Coincidentally, a few days later, he publicly launched a poll asking in which year a model at the o3 - mini level will be able to run on mobile phones?
As of the time of writing, more than 40% of the voting netizens believe that this model will be able to run in 2025.
Well, we're just waiting (Doge).
Reference links:
https://www.youtube.com/watch?v=V979Wd1gmTU
https://x.com/WesRothMoney/status/1937148640575009176
This article is from the WeChat official account "QbitAI", author: Shiling, Yiran. Republished by 36Kr with permission.