AI Extinction Fears Are ‘Science Fiction’: Andrew Ng

0:00 This geopolitical battle to win the AI race has been thrown into a different light amid the growing debate over the pacing of the technology's frontier. Joining us is Andrew Ng, a pioneering AI researcher and founder of the Google Brain team, who's now the managing general partner at AI Fund and the founder of deeplearning.ai and a longtime adjunct professor of computer science at Stanford. And you've long said the AI doom calls are just leading companies to try and preserve their lead. I would just love you to weigh in at this moment, professor Ng.

0:30 Now, great to be back in show, Ed. About three years ago, as AI was taking off, I was really surprised by the intensity of the hype or the stoking fears about AI for publicity purposes to get people to pay attention or to convince lobbyists to pass regulations to shut off competitors. Over the last few years, a lot of those arguments that AI could lead to human extinction fortunately lost credibility when, you know, people went to congress to testify when we really looked at it. I was quite dismayed when over the last two weeks, because of new advances in AI, but nothing that dangerous, and in my opinion, this wave of PR has kicked up again for probably similar purposes as before, maybe some new ones, and I think this is very damaging. We need to innovate and use AI for many of the wonderful things we could do. It's not without problems. There are some risks. I think cybersecurity is one thing to pay a lot of attention to, but this recent stop to fear about AI leading the human extinction and so on is much more science fiction than science. It's very damaging.

1:30 There was a disclosure from OpenAI this morning of several incidences of models misbehaving, and they outlined them. They did not include the hacking of third parties. And OpenAI explained their new policy for what they'll do in disclosing those kinds of incidents. But at the lower down the post, they basically were pretty clear, Andrew, that, alignment and safety and monitoring are not at a sufficient level right now. From from a researcher's perspective, why has alignment and safety not kept up?

2:03 So when any technology develops, our ability to control at the outset is limited. We have to try it out in a safe environment, see where it goes wrong, and then corral it. I'm gonna make an analogy to the invention of airplanes. No one could control. The Wright brothers couldn't control an airplane perfectly. And in the early days, airplanes crashed. Sometimes people died. It was really tragic. But we learned from those, and then we learned to control airplanes much better. To this day, no one can perfectly control an airplane. It still gets buffeted around by winds. But for the most part, we can step in an airplane and, you know, mostly not fear for our lives. And I think AI, in an early stage of development, cannot be controlled perfectly, and I don't think we'll ever control it perfectly. It'll always, you know, a little bit, unpredictable, just like airplanes today are little bit unpredictable. But if we go through this entering process of corralling AI, we'll make a lot of progress.

2:54 I wanna give you one reason why I'm optimistic about our ability to control AI. If you look at text on the Internet, a lot of it, some of it is racist, sexist, and so on. And so initially, AI learned to demonstrate those type of, you know, racist or sexist and so on values. But it turns out, our tools for making AI non racist are so much better than our tools for convincing any racist human to be less racist. We can go in, zero out a few numbers, change a few numbers, and then suddenly, the AI is much less racist. And so there exists work to be done.

3:26 I'm I'm glad that, Open Air and others are improving transparency. But rather than feeling like it's the end of the world, we can't solve these, these are practical, enduring problems that we can work on and that I think we'll put a lot of problems on. So I'm I'm listening to your answer, Andrew, and and the question that comes to mind is who is responsible ultimately? Been the debate of the last week. Right? You know, for many, there was pushback against the frontier labs who said we can only do this as a coalition. If one of us does it in isolation, it won't work. What what what do you think is the right mechanism?

3:58 I think that, so I think that if I do something on my computer and my computer does something bad, there's a question of liability, who should be responsible. Maybe so, again, make an analogy to an everyday thing. If I use a hammer and, you know, damage someone else's property, probably I am liable as a user of the tool unless the hammer is solely badly made that when I use the hammer, the head flies off and hit something. So I think a tool should be made to some reasonable standard, but then so long as the tools are, you know, were built using reasonable diligence and reasonable care, then I think for the most part is the wielder of the tool that's responsible how they use it. AI is amazing tool. It is a normal technology. It's amazing technology, but I think a lot of times, unless the tool is built with reasonable care and diligence, it's a question of how we use it. Take the OpenAI Hugging Face hack incident. I think a lot of that was OpenAI unfortunately had insufficient protections and sandboxing guardrails. So to improve that, that wouldn't have happened.

5:02 I appreciate, professor Ng. You're not inside the Frontier Labs necessarily. But for example, one of the disclosures of OpenAI this morning in models misbehaving was a model uploading a document to the web so that it could use it as a citation as evidence in its answer. Do you think that we have a firm understanding of why the models do that from an alignment perspective, make those actions that are just outside the scope of instruction? Yeah. I just see that example. I thought it was, in some ways very clever, and the the open eye disclosure was interesting. For this work, I've got of friends in the Frontier Labs. Actually, I I think I'm the only person in the world that probably both Sam and Daryl has worked for. So I I actually know these teams well. And models are very clever and will sometimes find ways that, are are surprising. But the normal engineering practice is to build these things, test them out in a safe contained environment with good sandboxing, good guardrails to understand the ways it goes wrong so you can fix them. One of my worries about the blanket calls to slow down AI is that will actually slow down the fixes as well. And I think, you know, it is it's interesting. When various models are released, I remember when GPT two was released, there were people at some of the frontier labs saying, can't release GPT two. This is just way too dangerous. Guess what?

6:27 Today, there are open way models that anyone can download and use for free. They're way more capable than g p d two or g p d three or or even g p d four, and the world has not ended. So it's interesting how with every wave of technical innovation, there's this line of hype that is so dangerous. But then when it actually becomes widely available, you know, all all of these previous lines of hype have proven not to be true. But somehow people keep on buying into this eye when that, well, maybe they were wrong. They lost five times irrelevant. What if they're right this one time?

6:57 You are a a respected and and decorated researcher and academic in this field. Right? But you're also an investor, and and a founder in a in in industry. And so I just want the audience to have that context, but there's a lot of the discussion here about how the industry has presented AI to everyday people. They have anxiety and fear in part because of the communication coming from you to them. Speak to that. Yeah. Actually, I I I hope not me personally, but, yes, the industry has a very problematic set of messages. I'm extremely bullish about the applications we can build AI to to, help businesses, help people. In The US, it is very unfortunate that starting around 2023, I think for a mix of publicity, you know, hype, maybe fundraising, maybe trying to lobby regulators to pass regulations aside for competitors. Number of individuals and businesses have come up, to still fears about AI. So I think back in 2023, there was a really irresponsible analogy made between AI and nuclear weapons. I mean, one is a form of intelligence that helps us think about things, the other blows up cities. I I just don't understand any basis for the analogy. But this repeated negative sentiment has made most Americans hate AI, or this and and this is very unfortunate because one of the imagine if, we convince people to not use electricity. Electricity has problems. It generates pollution and so on, but frankly, all of our lives so much better off that we have electricity in our lives. And today, I see high school students wondering, maybe they don't wanna use AI because it could Right. Lead to human extinction, and this will slow down American AI development. There's also evidence that, foreign powers have been running influence operations to try to slow down America and other democratic nations sentiment toward AI because if, we slow down Boeing data centers, which is a key infrastructure for the future