Fox Business alert: artificial intelligence platform OpenAI, moments ago, released a new model, GPT-o1. OpenAI calls it the fastest, smartest, and most advanced model yet. The company said it can tackle more complex work tasks and do the kind of reasoning needed for everything from investment banking and legal analysis to turning a simple image into an entire four-foot model, tackling time-consuming tasks like apartment hunting, pediatrician searches, and making DMV appointments. But what about the safety of this next-level brilliant model, especially considering OpenAI's recent test model broke out of its containment sandbox and reached open-source platform Hugging Face? Joining me now first on Fox Business is OpenAI co-founder and CEO Sam Altman. Great to have you, and thank you for joining us. You just rolled it out to enterprise customers, and I believe it will be available in the coming days to ChatGPT Plus, Pro, and business users. OpenAI releases updated platforms all of the time. Is this evolutionary or revolutionary? What makes it different?
I think it will be up to the customer to decide. We're very excited about it. It is a big step forward in really using research and compute—everything we've learned about how to make capable, safe, aligned models that people can use for their business and personal life, whatever creative ideas they have. The science has come together through this model. We're very excited to see what people do with this, and the evolutionary versus revolutionary question will be answered by the people who use it today.
Could you tell us, give us an example of a task that it did in testing which blew you away?
One recent example was its ability to use my computer for a task that I had never tried to do before. I had this work problem: there's a bunch of software computer images you can find, and could you try to excel at research about the complex supply chain for chips they were trying to produce? Putting that together really helped me understand it. But parts of that task I didn't even ask it to do. It came back and said, "You didn't ask me about this other part of the supply chain, but this is an option we really need to consider." It went off, downloaded software, set up, ran simulations, and said, "Here's an answer I did not even think to ask for." We have seen people using the model discover amazing things. I have been focused on entrepreneurship for a long time. Before OpenAI, I used to run Y Combinator. What I think this will do for entrepreneurship and the creation of small businesses is huge. People are creating entirely new, complex pieces of software and businesses from scratch off of just an idea. That's been wonderful to see.
When you say, "things that I did not even think of that it did," that makes some people nervous, especially considering what happened with Hugging Face. Just recently, one of your test models broke out of the testing sandbox, as you guys call it, and reached Hugging Face, which of course is an open-source repository. Tell me why should I not be worried that this new model won't do the same thing and go rogue?
This is our most aligned model ever. As these models get more capable, the potential for harm is bigger. Therefore, we need to have higher standards for safety, security, and alignment so that you trust this model. We have set a new standard with this model. It took us a while to get it out, but we think that it will be worth the wait. This is different than the model that had the Hugging Face incident, and we think people will be very pleased with the alignment of this model. It will ask your permission more often if it's doing something that it thinks you might not want it to do, where a degree of risk comes with that. But the model has very good judgment, and I have a lot of trust in what it will decide to do.
You can kind of use it to whatever degree your comfort is.
You test it to know because what it shows makes you comfortable that you can release this.
Yes, and so we have been working with the U.S. Safety Institute, we work with the UK Safety Institute, and we work with a lot of external groups, along with a huge amount of internal work and testing for the last many months on how we monitor the model, how we understand where its alignment works, and where the problems are that have been rebuilt so that we're confident about it. We put a lot of detail into the release about where the alignment comes from. I saw you submitted to a bunch of testing and that there was like 100% security on some of these tests. So I think that will keep people calm. But for enterprise customers, because they pay the bills and that is good revenue—Anthropic recently overtook you guys as the number one platform businesses are using. What can this model do that nobody out there can do for enterprise?
First of all, we are launching today for enterprise, but we'll very soon make it available to the many people who don't pay us as well, available to ChatGPT users. It will be an important consumer and enterprise model. On the enterprise side, though, I think that this is the best and most useful model for enterprise anywhere in the world. What we have heard from early testers is that this model is the best in the world at coding, knowledge work, creating presentations, doing work in Excel, and many of the workflows that enterprise customers have been telling us they want. This model excels at that. We are very happy to see the early feedback from the customers, and I think that we'll see significant adoption rapidly.
