Andrew Feldman, CEO of Cerebras, describes his company's AI chip as 56 times larger than previous chips and 20-40× faster than conventional processors, used by OpenAI and GlaxoSmithKline. He argues AI should augment thinking, not replace it, and addresses risks from bias and deepfakes while highlighting breakthroughs like personalized medicine at Mayo Clinic.
Chapters
Andrew, firstly, thank you for agreeing to an interview. It's actually very difficult to get people with that lanyard to actually speak here at Davos because everyone's just run away. But can you tell me briefly about what you guys do at Cerebras with AI?
Sure. We build compute for AI. So we've designed a chip and I have it here. We've designed the largest chip ever built. It's 56 times larger than any previous chip, and we built it specifically to do the computational work that underpins AI. And because it's so big, it is 20, 30, 40 times faster than any other processor. And what this means is when you use AI running on our chip, you get results faster. It takes less time. Your experience is better. And that's why we have customers like OpenAI, like Cognition, like GlaxoSmithKline, like leaders around the world are choosing to run their AI on our equipment.
But one of the concerns, I was going to say my generation, like I'm an old person, but like people my age and getting a bit older, one of the things we— I'm now 30, but I find that the younger generation especially, the more people use this, they use their brains less. Is that not a concern that it's almost dumbing down the population through laziness?
I think it's, it's interesting you said that. I think if you look back at the commentary in the, in the mid-'60s around the rise of television, there were the exact same concerns. I think parents should control the amount of screen time their kids have, whether it's using AI or others. I think we have to think very hard at the high school and at the college level about how AI doesn't replace thinking but augment thinking, right? This gives— has the potential to give the thoughtful student access to insight that is very different. It's not designed to replace thinking, it's designed to augment thinking and to allow your thinking to move more quickly and to discover new things. And I think It's brand new and we have as a community some learning still to do about how to teach with it, how it's going to change our education system, how to parent with it, right? It's the same as any new technology, right? You don't want your kid watching TikTok all day. You want them outside. You don't want them playing video games all day. You'd like them playing with other kids, doing whatever it is, whether it's ballet or taekwondo or soccer. I think in all these cases, we have to learn how to engage with the new technology.
What are the concerns that— because obviously I'm guessing when you're going through the technology, when you're creating these things, you're talking about, oh, what if it's used for this? What if it's used for that? Are there concerns that you guys talk about around the table when you're developing all this stuff?
Of course. I think when you build powerful technology, right? You know it can be used for good, and you also know that bad people can use it for bad, right? There are no powerful technologies that are free of that, that challenge. And so controls and safeguards are very much on our mind as we build the fastest computer in the world. I mean, that's what we set out to do, is build the fastest computer in the world for AI. And so working with our partners, working with the the model makers, is something that's fundamental to us. And I think one of the ways speed can play a role there is that you can, if you're very fast, right, you can actually check your answers, right, and still not interfere with the user experience. If you make the user wait 10 or 15 seconds each time to check their answers, right? They'll use a different, a different technology. Nobody likes to wait. But if you're so fast that sub-second, you can get a good answer, you can run it through safeguards, you can check it to make sure it's reasonable, it stays within the rules. All right, perfect. Then that's the way that speed, speed can be converted into safety.
And finally, do you think it should be— it with your product? Is it going to be something that's very almost open to all? I, I was reading an article the other day where it was about how certain companies are going to use with their AI stuff, when people ask it questions, some things that they call misinformation will not be published, even though there's some things which have been like proven. Take vaccine harms, for example. I was reading the article today that vaccine misinformation, it will not publish things that what they say is misinformation but isn't. Is that, is that a concern, if that makes sense?
Yeah, sure. So there are certain models, um, say some models that came out of, out of China, that if you ask about Taiwan give you the Chinese sort of political response. And that's not exactly, uh, the way much of the rest of the world views the history. So I think the ability to strategically bias models is real. But I think the good thing about having a community that's engaged is they can call you on it, right? They can retrain the model. They can move it away from things that are patently untrue. The other thing is that when you're fast, you can also put down the citations. And we can let the user choose, right? If you have a reference from the Journal of American Medical Association, right, a leading or a top refereed publication, that's different than somebody's blog, right? All information isn't equal. And so, by being really fast, you can get good answers and you can have the references. And if you're unsure, you can double-check, right? You can go and read the primary sources rather than relying on the answers that the AI is giving you.
And finally, you know, in the UK, they're looking at banning Grok. Do you think that there is a worry, I suppose from your personal belief or the company, that AI could go too far?
I think All extraordinary technologies, right? All meaningful technologies can go too far. I think we have a great deal of thinking to do around what, what good regulation is. Right. And Grok has chosen to be at the aggressive edge. But I think pioneers often are. I think they create conversations in society in which we have to address. Right. So Grok has a feature where You can take a photo of someone and you can ask it to undress them. I don't think we need this. I don't think that's cool. We can do better than that. Whether you want to block that is a decision for a democratic society to decide. This is not something we approve. I think deepfakes are something that is a real worry. I think remember at the start of the internet, you were getting emails from Nigerian princes.
They owe me so much money. They owe me so much money.
They owe me so much money. It's like, me too. How is that possible, right? And obviously, that was all crap. And there will be also piles of crap that are created through AI. On the other hand, look at what the internet gave us. With Teams and it changed the way— and Zoom— it changed the way that we interact business, the way we conduct business, with the way we communicate. Gave us Slack, which changed the way we communicate within organizations. It's allowed us to expand the scope of control any given manager has because they can see, they can communicate, they can engage with vastly more people. I think AI will be the same. I mean, sure, it will generate some crap. And if you look at TikTok or YouTube right now, it's filled with AI-generated crap. But also it's our best chance to better understand the solar system. It's our best chance to cure important diseases. At Cerebrus, we work with Mayo Clinic and we're working on personalized medicine that would be impossible to do without AI. And so I think it's always, it's easy in the press, It's easy to say, oh, it's all great, or oh, it's the end of the world. The truth is powerful technology can do some of each, and we have to think carefully and work together as an industry and as a community to get the best and foreclose the worst.
