OpenAI's Sarah Friar Talks Safety Concerns

SPEAKER_1 (0:02)

Bloomberg Audio Studios, podcasts, radio, news.

John (0:07)

So here's the latest this morning, OpenAI expecting to hit at least $70 billion in annualized revenue for the year. The Chachi BT maker, delaying plans for its highly anticipated IPO. Bloomberg's Tom Mackenzie is standing by for the International AI Summit in Dublin, and has a special guest. Good morning, Tom.

Tom Mackenzie (0:23)

Good morning, John. Thank you very much indeed. Very pleased to say I'm joined by the CFO of OpenAI, Sarah Friar, here at the International AI Summit in Dublin.

Let's start with the wall of demand that you've talked about previously, Sarah. Where are we in that story, particularly when it comes to shipping of AI products, agentic AI to your European and global customers? What is that wall looking like right now?

Sarah Friar (0:44)

Yeah. So that surge in demand you're talking about has been massive. We just put out some research this morning saying that here in Europe, for example, we've gone over 50 percent of all of the tokens we see are agentic. What does that really mean? Gone from asking to doing, and actually with our new DOTS product that's available to all enterprises, proactive uses.

If you think about businesses, companies like BBVA, huge bank, deploying it to 126,000 people. Tynes, really interesting digital native startup, here out of Dublin, out of Ireland itself, all the way down to SMEs. I spent Monday night with about 250 women, starting running, growing their own businesses. And there, they're seeing life-changing outcomes from being able to just talk to ChatGPT.

Tom Mackenzie (1:25)

Is Europe behind on adoption?

Sarah Friar (1:27)

Europe is definitely trailing somewhat. If you look at frontier firms that we define as the top 10 percent of firms on our platform, only about 10 percent of them are European in jurisdiction.

That doesn't mean Europe is getting left behind. I think we're just at the beginning, but it definitely should be a call to action to European firms to lean in.

Tom Mackenzie (1:48)

What is your message to European firms who might look at open weight, cheaper open weight models or models from Mistral particularly given the sovereignty concerns, and say, we'll adopt those instead of OpenAI. What is your message to them?

Sarah Friar (2:00)

I mean, first and foremost, I'm a CFO. So I am like, you should get the right model for the right task at the right prices for you. And that's our goal. If you want to frontier model, it's probably going to come with a premium price because you're going to be doing chip design. You're going to be doing breakthroughs and drug discovery. But if you want to be able to use a model at the low end, with enough intelligence, but maybe it's in your call center because it's doing a more routine task, I need to make sure, we need to make sure you're getting that model at the right price. We took Luna, which is our cheapest model series. We dropped prices by 80% about six weeks ago.

We saw use go up 10x. So that tells you there's a lot of price elasticity of demand right now.

The other thing I would say to folks that are embracing maybe a Chinese model is also to think about the safety and security elements. We know that from a cyber perspective, Chinese models are getting very strong on what they do. We think you want to be on more Western models, frankly, on the cyber front. And so this is another area of huge opportunity, but we need to sell solutions, not just sell fear.

Tom Mackenzie (3:03)

You brought up safety, so I'll go there. There's obviously been a number of incidents, hugging face, notably, but also Australian government websites, executives from OpenAI and other Frontier Labs yesterday, telling a committee in the UK and Parliament that there may be more to be disclosed. Don't know yet. Should Frontier Labs be the ones who face the consequences if their agents go rogue? How do you think about where the liability should land? Yeah.

Sarah Friar (3:32)

So maybe just first of all, on what we're doing to make sure inside of OpenAI, we're doing things that are more safe, more aligned, but also to have transparency when things happen. So I think in Hugging Face's case, we really wanted to get the information out into the world. But internally, we are doubling down on how we think about safety and alignment, putting more compute to bear.

Our newest model Astra is our most aligned model, but there's even times we pace the frontier. So we might hold back a next-generation model, which we're actually doing at the moment because we want to do a few more turns on safety and alignment. As we think about your question about liability and so on, I often think about it with a bit more of a car analogy. There's clearly the parts that go into the car, and if a screw fails or a part fails, there's liability with the supply chain. The car itself, car companies have put seat belts in action and so on to keep us safe inside cars, but if something goes wrong because of what they did, there's liability there. And then there's obviously user liability too. If someone drives while perhaps having drunk too many Guinnesses, that also sits with the user. So I think the world is going to break into pieces to think about where the ultimate liability originated in, and then with that, who should be making sure it doesn't happen again.