# AI's Rising Risks: Hacking, Virology, Loss of Control — With Dan Hendrycks

- **Show:** Big Technology Podcast
- **Published:** 26 Mar 2025
- **Source:** https://pdst.fm/e/tracking.swap.fm/track/t7yC0rGPUqahTF4et8YD/pscrb.fm/rss/p/traffic.megaphone.fm/LI3085827432.mp3?updated=1743032996
- **Transcript:** https://podcasts.opennash.com/shows/big-technology-podcast/ais-rising-risks-hacking-virology-loss-of-control-with-dan-hendrycks-f1bfbd/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Dan Hendrycks argues that the most urgent AI dangers are misuse by people, not rogue machines. Expert-level virology help could let outsiders make dangerous viruses, and autonomous hacking could attack infrastructure. The longer-term danger is a race to automate AI research itself, which could cause a loss of control. He favors narrow safeguards and deterrence between states over halting AI development.

The central tension is that everyone involved is incentivized to speed up. Companies and governments fear falling behind, so safety gets little budget and human review gets skipped. Hendrycks compares this to nuclear weapons: a security dilemma that leaves everyone worse off. His answer is to restrict specific dangerous capabilities, let trusted users keep access, and have states credibly threaten to sabotage anyone racing toward superintelligence.

## Key insights

1. **Misuse is a nearer threat than rogue AI** Hendrycks says AI today can't make PowerPoints or work autonomously for hours, which limits the damage it could do on its own. The risks he rates most urgent in the next year or two are bioweapons and cyberattacks directed by bad actors.
2. **AI now matches expert virologists at lab troubleshooting** His team had Harvard and MIT virologists photograph their lab work and ask what to do next. The newest reasoning models scored around the 90th percentile against those experts, so they could guide a novice step by step.
3. **Safeguards should be narrow, not a development slowdown** He proposes refusing topics like reverse genetics (building viruses from their genetic code) for anonymous users. Verified researchers, such as those at Genentech, could ask for access, so the benefits stay.
4. **Automated AI research is the real loss-of-control risk** If one AI can do an AI researcher's job, it can be copied 100,000 times, compressing a decade of progress into a year. Competitive pressure would push labs to drop human review, and he says this can't be tested safely in advance.
5. **Deterrence, not a superintelligence race, is his strategy** He argues that if one state nears superintelligence, rivals may threaten to disable its data centers or power with cyber means. He says that makes a race irrational, much as the nuclear standoff did.

## Questions this episode answers

### Is today's AI actually dangerous enough to worry about?

Hendrycks says AI poses no extinction risk today because it can't yet do long, multi-step tasks on its own. He does think misuse for bioweapons and cyberattacks could become serious within a year or two.

### How could an AI chatbot help someone make a dangerous virus?

Hendrycks says the newest models can look at photos from a wet lab (a hands-on biology lab) and guide a person step by step. In his team's tests they scored around the 90th percentile against expert virologists.

### What is an 'intelligence explosion' or automated AI research loop?

It means AI doing the work of AI researchers, copied many times and running fast. Hendrycks says 100,000 copies could turn a decade of progress into a year, with almost no human oversight.

### Can AI companies be trusted to keep their own safety promises?

Hendrycks says voluntary commitments are a distraction because most companies would likely break them if they clash with staying competitive. He says over 90% of lab energy goes to staying competitive, such as paying for bigger supercomputers.

### What does he think should be done about open-source AI models?

He has not historically opposed releasing model weights, but says the line should be drawn once models have strong hacking skills or expert-level virology knowledge. At that point he wants an international norm against open release, like the Biological Weapons Convention, and API access for trusted users.

## Facts

- **Dan Hendrycks** [0:11:40]: "the most recent reasoning models are getting around 90th percentile compared to these expert-level virologists in their area of expertise"
- **Dan Hendrycks** [0:11:07]: "we have Harvard and MIT expert-level virologists sort of taking pictures of themselves in the wet lab and asking, what steps should I do next?"
- **Dan Hendrycks** [0:15:47]: "the people with these capabilities has been a very small number, maybe a few hundred top virology PhDs"
- **Dan Hendrycks** [0:19:57]: "we did, for instance, a paper on the weapons of mass destruction related knowledge that models would have together last year"
- **Dan Hendrycks** [0:21:09]: "Humanity's Last Exam, where we put together various professors and postdocs and PhDs from all over the world"
- **Dan Hendrycks** [0:22:19]: "They're in the ballpark of like 10 to 20 % overall. They're the very best models."
- **Dan Hendrycks** [0:24:43]: "the new reasoning paradigm that has emerged in the past year, which is where you train models on math and coding types of questions with reinforcement learning"
- **Dan Hendrycks** [0:32:14]: "Then you could just run 10 ,000 of them simultaneously. And then some rogue actor could have it target critical infrastructure."
- **Dan Hendrycks** [0:32:49]: "And then you sort of turn them up and down simultaneously. And this can just like ruin like transformers and like blow them. And then they take multiple years to replace."
- **Dan Hendrycks** [0:36:22]: "there the software is not updated rapidly. So even if you identify various vulnerabilities, there will not necessarily be a patch because the system needs to always be on"
- **Dan Hendrycks** [0:35:50]: "in bio, I think that is offense-dominant. if somebody creates a virus, there's not necessarily a cure that it will immediately find for it"
- **Alex Kantrowitz** [0:47:40]: "It's called Superintelligence Strategy. It's with you, Eric Schmidt, who we all know, former CEO of Google."
- **Dan Hendrycks** [0:49:03]: "having more secure supply chains for robotics, for instance, and for AI chips having reduced sole source supply chain dependence on Taiwan for making AI chips"
- **Dan Hendrycks** [0:50:33]: "then we will disable your data center or the surrounding power infrastructure so that you cannot continue building this"
- **Dan Hendrycks** [0:53:39]: "making conditional threats and by means that are less escalatory, such as cyber sabotage on data centers or surrounding power plants"
- **Dan Hendrycks** [0:55:31]: "Maybe OpenAI has got a few hundred, maybe it'll say 300 AI researchers."
- **Dan Hendrycks** [0:57:40]: "maybe you've got 100 ,000 of these things operating at 100x the speed of a human"
- **Dan Hendrycks** [0:56:06]: "This is something that Alan Turing and the founders of computer science had pointed out, that this is a natural property of when you get AIs at this level of capability"
- **Dan Hendrycks** [0:58:13]: "Maybe it could find some breakthrough for anti-ballistic missile systems, which would make nuclear deterrence no longer work"
- **Dan Hendrycks** [1:03:51]: "we've done research with methods such as called circuit breakers and things like that. And those seem to improve the situation quite a bit and make it actually prohibitively difficult to do that jailbreaking."
- **Dan Hendrycks** [1:06:04]: "Then some of them will like lie like 20 % of the time, some of them like 60 % of the time."
- **Dan Hendrycks** [1:07:14]: "The main funder would be Jan Talon and Jan Talon, others who's a Skype co-founder."
- **Dan Hendrycks** [1:07:14]: "My appointment at XAI, I get $1 a year. At Scale AI, I've increased my salary exponentially to where I get $12 a year, $1 per month from Scale."
- **Dan Hendrycks** [1:10:20]: "I ended up leaving Berkeley before graduating just because of the sort of relatively suffocating atmosphere"
- **Dan Hendrycks** [1:16:16]: "Russia and or the Soviet Union and the US got together for the Biological Weapons Convention."

## Transcript

**[0:00:00] Alex Kantrowitz:** Now that artificial intelligence has tried to break out of a training environment, cheat at chess, and deceive safety evaluators, is it finally time to start worrying about the risks that artificial intelligence poses to us all? Here to speak with us about it is Dan Hendrycks. He's the director of the Center for AI. How do you price intelligence as a service? How do you charge for delivering end-to-end work, not just access? Those questions remain front and center for AI founders and operators and getting to the right fit business model, whether that's selling usage credits, outcomes, seats, or a hybrid combination requires constant iteration.

**[0:00:37] Alex Kantrowitz:** Chargebee is the billing and monetization infrastructure that powers just that for ambitious, fast scaling AI businesses. Visit chargebee.com slash Alex to learn more. That's C-H -A-R -G-E -B-E -E.com slash Alex. AI policy, and also an advisor to Elon Musk's XAI and Scale AI. Dan, it's so great to see you. Welcome to the show. Glad to be here. It's an opportune moment to have you on the show because I'm recently doom curious.

**[0:01:09] Alex Kantrowitz:** And I'll explain what that means. So I had long been skeptical of this idea that AI could potentially break out of its training set or out of computers and start to potentially even harm humans. I still think I'm on that path, but I'm starting to question it. We've recently seen research of AI starting to try to export its weights and scenarios where it thinks it might be rewritten, trying to fool evaluators and even trying to break a game of chess by rewriting the rules because it's so interested in winning the game.

**[0:01:43] Alex Kantrowitz:** So I'm just going to put this to you right away. is what I'm seeing in these early moments of AIs trying to deceive evaluators or trying to change the rules that it's been given, is that the early signs of us having AI as an adversary and not as a friend?

**[0:02:02] Dan Hendrycks:** I think the easier way to see that it could be adversarial is just if people maliciously use these AI systems against us. So if we have an adversarial state trying to weaponize it against us, That's an easier way in which it could cause a lot of damage to us. Now, there is an additional risk that the AI itself could have an adversarial relation to us and be a threat in itself, not just the threat of humans in the forms of terrorists or in the forms of state actors, but the AIs themselves potentially working against us.

**[0:02:36] Dan Hendrycks:** I think those risks would potentially grow in time. I don't think they're as substantial now compared to just the malicious use sorts of risks. But yeah, I think that as time goes on and as they're more capable, if some small fraction of them do decide to deceive us or try to self-exfiltrate themselves or develop an adversarial posture toward us, then that could be extraordinarily dangerous. So it depends. So I want to distinguish between what things are particularly concerning in the next year versus somewhat more in the future.

**[0:03:13] Dan Hendrycks:** And I think in the shorter term, it is more of this malicious issue. But that's not to downplay the fact that AI has to be threats later on.

**[0:03:19] Alex Kantrowitz:** Of course. And we're going to cover the short term and the immediate term as we go on through this conversation. But I don't want to lose this thread here because what I'm trying to figure out is what I've seen with, again, AI trying to rewrite the rules of chess, AI trying to exfiltrate its weights, copy itself. Now, it runs a code saying, if I run this code, I copy myself to another server. It was a fake code.

**[0:03:41] Alex Kantrowitz:** It was tricked by its evaluators into thinking it could do that. But it did run the code. And then also manipulating, showing evaluators that it was actually modeling one behavior, whereas when it thought they weren't watching, it was doing something else. This is all stuff that's happened. Are these the early signs of what could go wrong with AI? Or is this just benign activity that we shouldn't read too much into?

**[0:04:04] Dan Hendrycks:** So I think that it is suggestive of some loss of control scenarios. However, it is not the type of thing that I'm most concerned about with loss of control. So that's because we still can control these AI systems somewhat reasonably right now. And maybe we'll get better methods for doing so. However, the loss of control mechanism that I'm most concerned about is when we have automated AI research and development. So imagine at some point in the future, it could do what AI researchers do.

**[0:04:39] Dan Hendrycks:** And if you have one AI that can do that, and that would involve automating a lot of software engineering, if you could have one AI do that, then you could make 100 ,000 copies of these and have them perform research simultaneously. So that could lead to some very substantial acceleration in the rate of development. You might get a decade's worth of AI developments within the course of a year. And this would be highly automated where there's very little human oversight.

**[0:05:07] Dan Hendrycks:** And I think in those situations or in that scenario, I think a loss of control is much more likely. The sort of AIs that we have right now kind of being a little nefarious here and there, That is a concern, but that seems more tractable as a thing to research and improve and reduce the risks of. Meanwhile, some automated AI research and development loop that goes extremely quickly with very little human oversight, that seems hard to de-risk and get the risks to be at a negligible level.

**[0:05:44] Dan Hendrycks:** Just because it's so fast and there's so little human involvement. You couldn't involve humans that much because of competitive pressures. The larger geopolitics of this would make it harder to slow down in such a scenario. So I think that's more the type of loss and control risk that I'm concerned about. And I think that those other papers are suggestive, but there's a lot more hope for it being empirically tractable to counteract it.

**[0:06:10] Alex Kantrowitz:** Now, from what I understand it from your first answer, you are concerned both the way that humans use AI and AI itself sort of taking its own actions, our loss of control of artificial intelligence. So can you just rank sort of where you see the problems in terms of most serious to least serious and what we should be focusing on?

**[0:06:35] Dan Hendrycks:** That's a really good question. So I think the risks and their severity sort of depend on time. Some become much more severe later. So I don't think AI poses a risk of extinction like today. I don't think that they're powerful enough to do that because they can't make PowerPoints yet. They don't have agential skills. They can't accomplish tasks that require many hours to complete. And so since they lack that, this puts a severe limit on the amount of damage that they could do or their ability to operate autonomously.

**[0:07:17] Dan Hendrycks:** So I think there's a variety of risks. I think there's in malicious use in the shorter term. When AIs get more agential, I'd be concerned about AIs causing cyber attacks on critical infrastructure, possibly as directed by a rogue actor. There would also be the risk of AIs facilitating the development of bioweapons, in particular pandemic-causing ones, not smaller-scale ones like anthrax. And those are, I think, the two malicious use risks that we'll need to be getting on top of in the next year or two.

**[0:08:01] Dan Hendrycks:** At the same time, there's loss of control risks, which I think primarily stem from people, an AI company, trying to automate all of AI research and development. And they can't have humans check in on that process because that would slow them down too much. If you have a human do a week review every month of what's been going on and trying to interpret what's happening, this would slow them down substantially. and the competitor that doesn't do that will end up getting ahead.

**[0:08:31] Dan Hendrycks:** What that would mean is that you'd have basically AI development going very rapidly where there's nobody really checking what's going on or hardly checking. And I think a loss of control in that scenario is more likely. But that comes a bit later. So it depends, and it will depend on what we do. So these risks aren't something that exists out there and are immutable. Like maybe we can do more research to make the malicious use risks go down substantially.

**[0:09:09] Dan Hendrycks:** Maybe states can deter each other from pursuing this automated AIR &D loop so that we don't have this risk of loss of control. So it depends not just on technical research. It depends on the politics and geopolitics as well. And those will keep changing. And so the risk sources will keep changing.

**[0:09:28] Alex Kantrowitz:** Right. And with the Center for AI Safety, we're going to talk today about risks, but we're also going to talk about solutions. And with the Center for AI Safety, what you're doing is basically pointing out the risks and trying to get to solutions to these problems. You told me you were just at the White House yesterday, the day before we were talking. So this stuff is something that you're actually working towards mitigating. And I think we're going to get to that in a bit.

**[0:09:48] Alex Kantrowitz:** But first, let's talk a little bit through some of the risks that you see with AI and how serious they actually are. One of them that just jumped out for me right away was bio, creating bioweapons. Let me run you through what I think the scenario could be in my head and you tell me what I'm missing. With bioweapons, you'd basically be prompting an LLM to help you come up with new biological agents, effectively, that you could go unleash against an enemy.

**[0:10:19] Alex Kantrowitz:** And I think, wouldn't that be predicated on the AI actually being able to come up with biological discoveries of its own? Right now, current LLMs, they don't really extend beyond the training set. Maybe there's an emergent property here or there, but they haven't made any discoveries and sort of been the big knock on them to this point. So I am curious if you're talking about immediate risks and one of them being, OK, there could be bioweapons that are created with AI.

**[0:10:48] Alex Kantrowitz:** Doesn't that suppose that there's going to be something much more advanced than the LLMs that we have today? Because with current LLMs, to me, it's basically like Google. It's a search for what's on the web, and it can produce what's on the web. But it's not coming up with new compounds on its own.

**[0:11:07] Dan Hendrycks:** Yeah, so I think that for cyber, that's more in the future. But I think virology, expert-level virology capabilities, are much more plausible in the short term. So, for instance, we have a paper that will be out maybe in some months. We'll see. but most of the work before it's been done. And in it, we have Harvard and MIT expert-level virologists sort of taking pictures of themselves in the wet lab and asking, what steps should I do next?

**[0:11:40] Dan Hendrycks:** So can the AI, given this image and given this background context, help guide through step-by-step these various wet lab procedures in making viruses and manipulating their properties? And we are finding that with the most recent reasoning models, quite unlike the models from two years ago, like the initial GPT-4, the most recent reasoning models are getting around 90th percentile compared to these expert-level virologists in their area of expertise. So this suggests that they have some of these wet lab type of skills.

**[0:12:18] Dan Hendrycks:** And so if they can guide somebody through it step by step, that could be very dangerous. Now, there is an ideation step, but that seems like a capability. Them doing brainstorming to come up with ways to make viruses more dangerous, I think that's a capability that they've had for over a year. The brainstorming part, but the implementation part seems to be fairly different. So I think in bio, actually, I would not be surprised if in a few months there's a consensus that they're expert level in many relevant ways and that we need to be doing something about that.

**[0:12:54] Alex Kantrowitz:** Wow. That's crazy to me because I would think it would be the opposite, right? That cyber would be the thing that we need to be worried about because these things code so well, not virology. So I just want to ask you.

**[0:13:04] Dan Hendrycks:** But on that, biology has been such an interesting subject because they just know the literature really well. They know the ins and outs, but they've got a fantastic memory. And they have so much background experience. It's been, for some reason, their easiest subject historically, biology and virology, in earlier forms of measurements. Like if you see how they do on exams. But now we're looking at their practical wet lab skills, and they have those increasingly as well.

**[0:13:34] Alex Kantrowitz:** So what about the evolution of the technology? Because this is all with large language models, right? reasoning is just something that's taking place within a large language model, like the GPT, which powers ChatGPT. So what is it about the current capabilities that have increased to the point where they're now able to guide somebody through the creation or manipulation of a virus?

**[0:13:57] Dan Hendrycks:** That seems to be like a step-changing capability. Where did that come from? Well, now they have this image understanding skills. So that's a problem that they didn't used to have. That makes it a lot easier for them to do guidance or sort of, you know, be in a print or sort of a guide on one's shoulder saying, now do this, now do that. But I don't know where that came from, that skill. They've just trained on the internet and maybe they read enough papers and saw enough pictures of things inside those papers to have a sense of the protocols and how to troubleshoot appropriately.

**[0:14:32] Dan Hendrycks:** So since they've read basically every academic paper written. Maybe that's the cause of it. But it's a surprise. I mean, I was thinking that this practical, tacit knowledge or something wouldn't be something that they would pick up on necessarily. It'd make a lot more sense for them to have academic knowledge about knowledge of vocab words and things like that. So I don't know where it came from. It's there.

**[0:15:01] Alex Kantrowitz:** Right. But this is still all stuff that is known to people. It's not like the AI is coming up with nude viruses on its own. You can't prompt whatever GPT it is and say, create a new coronavirus.

**[0:15:16] Dan Hendrycks:** So if you're saying, I'm trying to modify this property of the virus so that it has more transmissibility or a longer stealth period, then I think it could, with some pretty easy brainstorming, make some suggestions. And then if it can guide you through the intermediate steps, that's something that can make it be much more lethal. I don't think it needs a... You don't need breakthroughs for doing some bioterrorism, generally. The main limitation for risks, generally, risks will be capability and intent.

**[0:15:47] Dan Hendrycks:** And historically, our bio risks have been fairly low because the people with these capabilities has been a very small number, maybe a few hundred top virology PhDs, and then a lot of them just don't intend to do this sort of thing. However, if these capabilities are out there without any sorts of restrictions and extremely accessible, then as it happens, then your risk surface is blown up by several orders of magnitude. A solution for this, to let people keep access to these expert level virology capabilities, is that they can just speak to sales or ask for permission to have some of these guardrails take it off.

**[0:16:26] Dan Hendrycks:** Like if they're a real researcher at Genentech or what have you, wanting these expert level virology capabilities, then they could just ask and then like, oh, you're a trusted user. Sure, here's access to these capabilities. But if somebody just made an account a second ago, then by default, they wouldn't have access to it. So for safety, a lot of people think that the way you go about safety is slowing down all of AI development or or something like that.

**[0:16:52] Dan Hendrycks:** But I think there are very surgical things you can do where you just have it refuse to talk about topics such as reverse genetics or guide you through practical intermediate steps for some virology methods. Wait, those safeguards don't exist today?

**[0:17:09] Alex Kantrowitz:** At XAI, they do. You're an advisor at XAI? Yeah, yeah, yeah. But what were the models that you were testing to try to find out whether they would help with the enhancement of virologists?

**[0:17:19] Dan Hendrycks:** So we tested pretty much all of the leading ones that have these sort of multimodal capabilities. And they'll have some sort of safeguards, but there are various holes. And so those are being patched. We've communicated that, hey, you know, there are various issues here. And so I'm hopeful that very quickly some of these vulnerabilities will be patched with it. And then if people want access to those capabilities, then they could possibly be a trusted third-party tester or something like that or work at a biotech company.

**[0:17:51] Dan Hendrycks:** And then those restrictions could be lifted for those use cases. But random users, we don't know who they are asking how to make some virus more lethal or something. Sorry, animal-affecting virus. It's just have the model refuse on that. That seems fine.

**[0:18:05] Alex Kantrowitz:** Yeah, we do see the benchmarks come in through each model release. and it's like, oh, now it's scored 84th or 90th percentile or 97th percentile on this math test or on this bio test. And for us, it's like, oh, that's the model doing it. But what you're trying to say is, and correct me if I'm wrong, if it's getting 90 % of the way that an expert virologist might get, then it could take a crafty user a number of prompts effectively to find their way towards that 100%.

**[0:18:32] Alex Kantrowitz:** Because if they try it enough times, they might accidentally get to the – Not accidentally, but they might end up getting the bad virus that we're trying not to have the public create.

**[0:18:42] Dan Hendrycks:** Yeah, yeah. So this is what concerns me quite a bit, and I'm being more quiet about this just to, you know. Well, you're talking about it on the podcast. Yeah, I guess what I'm talking about now, but I'm not, you know, there's this orders of magnitude. And it's being taken care of at XAI, and this is sort of in our risk management framework there. And other labs are taking this sort of stuff more seriously or finding some vulnerabilities.

**[0:19:06] Dan Hendrycks:** and then they're patching them. So I'm being non-specific about some of the vulnerabilities here, but hopefully can provide more precision once they have that taken care of. But yeah.

**[0:19:17] Alex Kantrowitz:** Okay. I look forward to reading the paper. You're an advisor to Scale AI. They are a company that will give a lot of PhD level information to models in post-training, right? So you've trained up the model on all of the internet. It's pretty good at predicting the next word. and then it needs some domain-specific knowledge. Scale, from my understanding, has PhDs and really smart people writing their knowledge down and then feeding it into the model to make these models smarter.

**[0:19:46] Alex Kantrowitz:** How does a company like Scale AI approach this? Do they have to say, all right, if you're a virology PhD, we shouldn't be fine-tuning the model with your information. What's going on there and how are you advising them?

**[0:19:57] Dan Hendrycks:** So I've largely been advising on measuring capabilities and risks in these models. So we did, for instance, a paper on the weapons of mass destruction related knowledge that models would have together last year. And for that, we were finding a lot of the academic knowledge or knowledge that you would find in the literature. Like, does it really understand the literature quite well? And we were seeing that in biology and for bioweapons related papers that they did.

**[0:20:35] Dan Hendrycks:** However, this just tested their knowledge, not their know-how. So that's why we did the follow-up paper to see what's their actual wet lab know-how skills. And those were lower, but now they're higher. And so now those vulnerabilities need to be patched. And those patches are, I gather, underway. way. So we've also worked on other sorts of things together, like in measuring the capabilities of these models, because I think it's important that the public have some sense of how quickly is AI improving?

**[0:21:09] Dan Hendrycks:** What level is it at currently? So a recent paper we did together was Humanity's Last Exam, where we put together various professors and postdocs and PhDs from all over the world, and they could join in on the paper if they submit some good questions that stump the AI systems. And I think this is a fairly difficult test. So it was think of something really difficult that you encountered in your research and try and turn that into a question.

**[0:21:40] Dan Hendrycks:** And I think each person, each researcher probably has one or two of these sorts of questions. So it's a compilation of that. And I think when there's very high performance on that benchmark, that would be suggestive of something that has, say, in the ballpark of superhuman mathematician capabilities. And so I think that would revolutionize the academy quite substantially because all the theoretical sciences that are so dependent on mathematics would be a lot more automatable. You could just give it the math problem and it could probably crack it or crack it better than nearly anybody on earth could.

**[0:22:19] Dan Hendrycks:** So that's an example capability measurement that we're looking at. We excluded in humanity's last exam, no virology related skills. So we were not collecting data for that because we didn't want to incentivize the models getting better at that particular skill through this benchmark. And how's the AI doing today on that exam? They're in the ballpark of like 10 to 20 % overall. They're the very best models. So, you know, it'll take a while for it to get to 80 plus percent.

**[0:22:51] Dan Hendrycks:** But I think once it is 80 plus percent, that's basically a superhuman mathematician is one way of thinking of it.

**[0:22:55] Alex Kantrowitz:** But the thing is they're at 10 to 20 % now. And many experts within the AI field, the practitioners, we had Jan on a couple weeks ago, talking about how we're getting to the point of diminishing returns with scaling, right? that the current growth trajectory of, or the current trajectory of generative AI in particular, is limited because basically the labs are maximizing their ability to increase its capabilities. So I'm curious what you think, whether you think that's right, because you're obviously working with these companies, you're working with XAI, you're working with scale.

**[0:23:29] Alex Kantrowitz:** If we are getting to this data wall or some wall or some moment of diminishing marginal return on the technology, is it possible that all this fear is somewhat misplaced? Because if the AI is not going to get much better than it is right now, at least with the current methods, we may not be a year or two away from AGI, right? We may not be getting AGI at the end of 2025, like some people are suggesting.

**[0:23:53] Alex Kantrowitz:** And so then maybe we shouldn't be as afraid because, again, the stuff is limited.

**[0:23:59] Dan Hendrycks:** Yeah. So if we were trapped at around the capability levels that we're at now, then that would definitely reduce urgency and, you know, means one could chill out a bit more and take it easy. But I'm not really seeing that. I think maybe what he's referring to is the sort of pre-training paradigm, sort of running out of steam. So if you take an AI, train on a big blob of data, and have it just sort of predict the next token, do what basically gave rise to older models like GPT-4, that sort of paradigm does seem like it's running out of steam.

**[0:24:43] Dan Hendrycks:** It has held for many, many orders of magnitude, but the returns on doing that are lower. That is separate from the new reasoning paradigm that has emerged in the past year, which is where you train models on math and coding types of questions with reinforcement learning. And that has a very steep slope. And I don't see any signs of that slowing down. That seems to have a faster rate of improvement than the pre-training paradigm, the previous paradigm had.

**[0:25:23] Dan Hendrycks:** And there's still a lot of reasoning data left to go through and do reinforcement learning on. So I think we have quite a number of months or potentially years of being able to do that. and so personally I'm not even thinking too specifically about what AIs will be looking like in a few months. They'll be I think quite a bit better at math and coding but I don't know how much better so I'm largely just waiting because the rate of improvement is so high and we're so early on in this new paradigm that I don't find it useful to try and speculate here.

**[0:26:00] Dan Hendrycks:** I'm just going to wait a little while to see. But I would expect it to be quite better in each of these domains, in these STEM domains.

**[0:26:09] Alex Kantrowitz:** Right. I guess reasoning does make it better at the areas that you're mostly concerned about. Doing math, doing science. Yeah, that's right. Because when it goes, tell me again if I'm wrong, when it goes step by step, it's much better at executing and working on these problems than if it's just printing answers.

**[0:26:25] Dan Hendrycks:** Yeah. And there is a possibility, and this is sort of a hope in the field. I don't know whether it will happen, is that these reasoning capabilities might also give these agent type of capabilities where it can do other sorts of things like make a PowerPoint for you and do things that would require operating over a very long time horizon. Potentially that skill set would fall out of this paradigm, but it's not clear. There has been a fair amount of generalization from training on coding and mathematics to other sorts of domains like law, for instance.

**[0:26:59] Dan Hendrycks:** And maybe if those skills get high enough, maybe it will be able to sort of reason its way through things step by step and act in a more coherent, goal-directed way across longer time spans. I'm going to try to channel Jan here a

**[0:27:11] Alex Kantrowitz:** little bit. I think he would say that this is still going to be constrained by the fact that AI has no real understanding of the real world. Well, I don't know. That sounds like almost a

**[0:27:22] Dan Hendrycks:** no true Scotsman type of thing. It's like, what's real understanding mean? Well, let me give you an example. If it is predictive ability, if it's sort of like, if it can do the stuff, that's what I care about. But if it doesn't satisfy some strict philosophical sense of something, some people might find that compelling, but I don't.

**[0:27:41] Alex Kantrowitz:** I'll give you an example, like with the video generators. If AI really understood physics, then when you say, give me a video of a car driving through a haystack, it will actually be a car driving through a haystack, as opposed to what I've done is give it that prompt, and it's just hay exploding onto the front of a car with perfectly intact hay bales in the background.

**[0:28:03] Dan Hendrycks:** I think that for a lot of these sorts of queries, at least with images, for instance, we'd see a lot of nonsensical arrangements of things and things that don't make much sense if you look at it more closely. But then as you just scale up the models, then they tend to just kind of get it increasingly. So we might just see the same for images, or excuse me, for video. I think as well they have some good world model stuff.

**[0:28:30] Dan Hendrycks:** They'll have vanishing points being more coherent. If I were drawing or anything like that, I'd probably be lacking an understanding of the physics and geometry of the situation and making things internally coherent relative to them. So I don't know. Yeah, they seem pretty compelling and have a lot of the details right, including some of the more structural details. But there'll be gaps that one can keep zooming into. But I just think that that set will keep decreasing, as was sort of the case with images and text before.

**[0:29:03] Dan Hendrycks:** I mean, text back in the day, the same argument. It doesn't have a real understanding of causality. It's just sort of mixing together words and whatnot. when it was barely able to construct sentences coherently. Now it can. Yeah, now it can. So I don't know if it then got a real understanding in the sort of philosophical sense that he's thinking for language, but it was good enough, and that might be the case with video as well.

**[0:29:28] Alex Kantrowitz:** There were points where I was like, oh, but it is getting the guy sitting on the chair when I say, you know, do a video of a guy sitting on a chair and kicking his legs, and those legs are kicking, and they are bending at the joints. So there must be some understanding there.

**[0:29:42] Dan Hendrycks:** Yeah, in some ways. But if you ask them to do like gymnastics, then I'll just have lips flailing all over. No, the person just disappears into the floor.

**[0:29:49] Alex Kantrowitz:** Okay, like you said at the beginning, chat chapti isn't going to kill us yet. Let's talk about hacking. I do think that we glanced over a little bit before, but in terms of we're now going through, I think, the humans plus AI problem, right? And hacking to me is one that I think we should definitely focus on. You mentioned that we're still not quite there, but it does seem to me, again, I'm just going to go back to the point I made earlier.

**[0:30:16] Alex Kantrowitz:** You can really code stuff up with these things, and they enable pretty impressive code. Already, you could think that ChatGPT could produce pretty good phishing emails if you just kind of creatively, and not just ChatGPT, but all of these GPT models. If you creatively prompt it, it will give you an email that you can send and try to phish somebody. Or even, let's say you just take an open source model like DeepSeek, download it, and then run it without safeguards.

**[0:30:43] Alex Kantrowitz:** So where's the risk with hacking? I know you said it's a little bit further off. Why is it further off? And what should people be afraid of? Or what should people be concerned of? Yeah, yeah.

**[0:30:53] Dan Hendrycks:** So the risk from it, more of the risk comes from when they're able to autonomously do the hacking themselves. So trying to break into a system, finding an exploit, escalating privileges, causing damage from there, things like that. And that requires multiple different steps and these agential skills that I keep referring to that they currently don't have. So although they could facilitate in like ransomware development and other forms of malware for them to autonomously execute and infiltrate systems, that is something that will require the new agential skills.

**[0:31:36] Dan Hendrycks:** And I don't see – it's very unclear when those arrive. Could be a few months from now. Could be a year from now. It's a little less – I'm a little more suspicious. Maybe it would even take two years for that. So that's something for us to get prepared for, figure out how we're going to deal with that, try and make safeguards increasingly robust to people trying to maliciously use it in those ways. But, yeah, I think much of the risk source comes from being able to take one of these AIs, let's say one of these DeepSeek AIs, let's say it's DeepSeek agent version, and it's able to actually do these cyber attacks.

**[0:32:14] Dan Hendrycks:** Then you could just run 10 ,000 of them simultaneously. And then some rogue actor could have it target critical infrastructure. Then this is causing quite severe damage. So for critical infrastructure, this could be like have it reduce the detector or the filter in a water plant or something like that. Then the water supply is ruined. Or you could target these thermostats in various homes because they're often – some of the more advanced ones are connected to Wi-Fi.

**[0:32:49] Dan Hendrycks:** And then you sort of turn them up and down simultaneously. And this can just like ruin like transformers and like blow them. And then they take multiple years to replace. Things like that. But they aren't capable of doing that sort of thing currently. So it's more of an on-the-horizon type of thing. But I'm not feeling the urgency with that currently. I'm more concerned about it. I think there's more the geopolitics of this, like making sure that states are aware of what's going on in AI.

**[0:33:31] Dan Hendrycks:** like they're at least able to follow the news and things like that in some capacity. I think things like that feel somewhat more urgent to me than trying to address cyber risks. There are things to do, though, and I think we should create incentives beforehand.

**[0:33:46] Alex Kantrowitz:** Maybe I'm too much of an optimist for my own good, but when I hear you talk about this, I also get a little bit excited about the capabilities of these programs because, for instance, if AI can enhance the function of a virus, AI can probably create a vaccine, make medical discoveries. If AI can hack into the infrastructure of some country, find exploits and turn the thermostats up and down, then AI could probably do incredible amounts of very beneficial coding and computer work for humanity.

**[0:34:15] Alex Kantrowitz:** So if we do get to that point, it seems to me like there's going to be these maybe two poles here, right? One is the potentially scary and destructive stuff that you can mitigate with some of the controls that you talked about, but also amazing opportunity.

**[0:34:28] Dan Hendrycks:** Yeah, so it's – and the thermostat thing was for messing with the electricity and that causing strain on the power grid and destroying transformers. Just for clarification in case – but yeah, I think you're pointing at that it's dual use. So I'm not saying AI is bad in every single way. It's like other dual use technologies. Bio is a dual use technology. It can be used for bioweapons. It can be used for healthcare. nuclear technology is dual use there's civilian applications for it as well and chemicals too and we have managed all of those other ones by selectively trying to you know limit some particular types of usage and restricting the capabilities of rogue actors to some of these technologies

**[0:35:14] Dan Hendrycks:** and making sure there are good safeguards for the civilian applications and then we can actually capture the benefits. So it's not an all or nothing type of thing with AI. It's what are surgical restrictions one can place so that we can keep capturing the benefits. And so, for instance, with virology, that's a matter of you add the safeguards and then the researchers who want access to those can speak to sales. That's basically a resolution of that problem, provided that you have the models kept behind APIs.

**[0:35:50] Dan Hendrycks:** And so now on this dual-use part, though, there's an offense-defense balance. So for some applications, it can help, it can hurt. And maybe it helps more than it hurts, or maybe it will hurt more than it will help. So in bio, I think that is offense-dominant. if somebody creates a virus, there's not necessarily a cure that it will immediately find for it. If it would help a rogue actor make a somewhat compelling virus, now that could be enough to cause many millions to die.

**[0:36:22] Dan Hendrycks:** And it may take months or years to find a cure. There are many viruses for which we have not found cures yet. And for cyber, in most contexts, there's a balance between offense and defense, where if somebody can find a vulnerability with one of these hacking AIs, then they could also use that to patch the vulnerability. There is an exception, though, where in the context of critical infrastructure, there the software is not updated rapidly. So even if you identify various vulnerabilities, there will not necessarily be a patch because the system needs to always be on, or there are interoperability constraints, or the person, the company that made the software is no longer in business, these sorts of things.

**[0:37:04] Dan Hendrycks:** So our critical infrastructure is a sitting duck. And so in that context, cyber is offense dominant. But in normal context, there's roughly a duality. And for virology, I think that's largely offense

**[0:37:15] Alex Kantrowitz:** dominant. So before we go to the nation state element of this, I need to ask you a question about the actual research houses themselves. Every research house has their concern with safety. From open AI to XAI, everything in the middle. Maybe not DeepSeq. We'll get to DeepSeq. Yet, they're the ones that are building this technology. And I find it a little strange that you have companies that are saying, it's weird. We have to build this and advance this technology so we can keep people safe.

**[0:37:48] Alex Kantrowitz:** I never really understood that message.

**[0:37:50] Dan Hendrycks:** Yeah, I don't know if it's to say that we need to keep people safe. I think it's more that the main organizations that have power in the world now are largely companies. And so if one's trying to influence the outcomes, one basically needs to be a company is how many of them will reason. They'll think that, yeah, you could be in civil society or you could protest, but this will not determine the course of events as much.

**[0:38:16] Dan Hendrycks:** So there's sort of many of them are buying themselves the option to hopefully influence things in a more positive direction. But most of the effort will be to stay competitive and stay in this arena. So I think over 90 % of the intellectual energies that they're going to spend is actually, how can we afford the 10x larger supercomputer? And that means being very competitive, speeding this up, and making safety be some priority but not necessarily a substantial one.

**[0:38:48] Dan Hendrycks:** So I do think there is sort of an interesting contradiction or something that looks like a contradiction there. But I think if we think back to nuclear weapons, nuclear weapons, nobody wants nuclear weapons. If there'd be zero on Earth, fantastic. That would be a nice thing to have if that would be a stable state. But it's not a stable state. One actor may then develop nuclear weapons and they could destroy the other. So this encourages states to do an arms race and it makes everybody all collectively less secure.

**[0:39:20] Dan Hendrycks:** But that's just how the game theory ends up working. So you get a classic, what's called a security dilemma. Everybody's worse off collectively. But – and even if you took it seriously, you say, yes, nuclear technology is dual use and potentially catastrophic. And we need to be very risk conscious about it. You can agree with all those things. But you still might want nuclear weapons because other parties will also have nuclear weapons. And unilateral disarmament in many cases just didn't make game theoretic sense.

**[0:39:53] Dan Hendrycks:** in the way that an individual company pausing their development while others race ahead doesn't make game theoretic sense. So I think this just points to the fact that game theory is kind of confusing, and so you're getting some things that are seeming contradictions that if you use a nuclear analogy, go, yeah, I suppose that makes sense, and it's just kind of an ugly reality to internalize.

**[0:40:15] Alex Kantrowitz:** Doesn't that discount the fact that these companies, if they want to influence the way things are going, they are going to be it's like you're one and the same. Yes, you're influencing but without you this wouldn't be moving as fast as it is. It is interesting for instance, think about Elon Musk right? Obviously he has you in two days a week to work on safety inside XAI but he's also putting together what, million GPU data centers to build the biggest, baddest LLM ever I don't understand

**[0:40:42] Dan Hendrycks:** Well, if he didn't then he would be having less influence over it so it's it's not something that I would envision everybody would just sort of voluntarily pause. So subject to companies not sort of voluntarily rolling over and dying, then what's the best you can do subject to those constraints? But the competitive pressures are quite intense, such that they do end up prioritizing, focusing on competitiveness and other priorities, like what's the budget for safety research, it will be generally lower than would be nice to have

**[0:41:23] Alex Kantrowitz:** if this were a less competitive environment. Do you think Elon is more interested in restoring this original vision that he had for OpenAI, making everything open source, making it safe, I would imagine? He founded OpenAI with Sam Altman as sort of a beachhead against Google. because he was afraid of what Google was going to do with this technology. So I'm curious if you think that XAI is along that mission, or is he more interested in the sort of soft cultural power that comes with having the world's best AI?

**[0:41:58] Alex Kantrowitz:** For instance, like you can change the way that it speaks about certain sensitive political issues. It can be anti-woke, which we all know is sort of where Elon stands. So what do you think his true interest lies in building XAI?

**[0:42:11] Dan Hendrycks:** Well, I think the – and I won't position myself as sort of speaking on behalf of them here. We won't put you as a spokesperson, but you are in there. Yeah, so I think that the mission is to understand the universe. And so this means having AIs that are honest and accurate and truthful to improve the public's understanding of the world. So we will be getting in a very fast-moving, trying situation with AI if it keeps accelerating.

**[0:42:42] Dan Hendrycks:** And so good decision-making will be very important. And us understanding the world around us will be very important. So if there are more features that enable truth-seeking and honesty and good forecasts and good judgment in institutional decision-making, Those would be great to have with – the hope is that Grah could help enable some of that so that civilization is steered, steering itself more prudently in this potentially more turbulent period that's upcoming. That's one read on the mission statement.

**[0:43:18] Dan Hendrycks:** But I think that's the objective of it is understand the universe and their different sub-objectives that that would give rise to. I think its ability to help culture process events without censorship or political bias one way or the other is a stated objective. And I think that would be indispensable in the years going forward.

**[0:43:47] Alex Kantrowitz:** Do you buy that that's what they're doing? Because we also heard the same thing from Elon when it came to buying Twitter, now X. but I think community notes has been quite good. But that was something that was built under Jack Dorsey. I'm not going to take sides here. I'm going to just observe empirically what I've seen. We know that Substack links have been deprioritized because it was seen as a competitor with Twitter. We know that Musk, I think according to reporting, changed the algorithm to have his tweets show up more often and his tweets took a strong stance towards supporting Donald Trump in the election.

**[0:44:23] Alex Kantrowitz:** So to me, the idea that like hearing again from Elon, and again, look, I respect what Elon's done as a business person, but hearing again that he has a plan to make a culturally relevant product that's free of censorship and politically unbiased, I don't know if I believe that anymore. So I don't know about some of the specific things

**[0:44:43] Dan Hendrycks:** such as the waiting thing or something like that, profile things, for instance. I think that overall in terms of cultural influence and people being more disagreeable and doing less self-censoring has been successful. I think that was the main objective of it. So I think that X had a large role to play there. So I don't know. I think in terms of shaping discourse norms in the U.S., that seems to have been successful in my view.

**[0:45:19] Alex Kantrowitz:** Yeah, I'm not saying pre-Elon Twitter didn't censor, which is probably the wrong word because that's usually from the government, didn't sort of shape the definition of speech to its own liking. It obviously had a progressive approach and moderated speech on a progressive approach. I just don't think Elon is not using his own influence when it comes to how he runs X. But you and I could speak about this forever.

**[0:45:44] Dan Hendrycks:** Yeah, this isn't even my sort of wheelhouse as much. But yeah, I mean, it's sort of like since I'm doing the next as most things. You brought it up, so. Oh, okay. All right. Sure. All right. I mean, just the non-biased and truthful things, so it's worth it. So, I mean, if there are ways in which it's extremely biased one way or the other, that's useful to know. This is a thing that is continually trying to be improved, at least for XAI's GROC.

**[0:46:20] Dan Hendrycks:** And I think that all of the sort of product offering could get quite better at this. But I'm not speaking as a sort of representative there or anything like that. But I guess maybe in my – I guess right now in my personal capacity, I think that there's things to improve on for all these models. in terms of their bias.

**[0:46:39] Alex Kantrowitz:** All right. We agree on that front. You hinted at it previously, but you talk a little bit about how companies, basically how you don't think it's a good idea for there to be an arms race here. And certainly there is one between the U.S. and China. We know that U.S. has put export controls on China. China has in some ways gotten around them through very creative procurement processes that go through Singapore. We can probably say that with a pretty good degree of confidence.

**[0:47:09] Alex Kantrowitz:** Then, of course, we see the release of DeepSeek and some other AI applications from China. And everyone's trying to build the better AI so that they have the soft power, like we spoke about, to effectively, A, control, to influence culture across the world. But also it's an offensive capability and defensive, like you're saying. if your country has the ability to manipulate viruses or to do cyber hacks, you become more powerful and you get to sort of, you know, potentially put your view of the world, implant your view of the world on the way that it operates.

**[0:47:40] Alex Kantrowitz:** You have a paper out that's sort of arguing against this arms race. It's called Superintelligence Strategy. It's with you, Eric Schmidt, who we all know, former CEO of Google. I think he just started. He's taking over a drone company. So you Alexander Wang, the current CEO of ScaleAI, who's been formerly on this show. Talk a little bit about why you don't think it's a good idea for countries to pursue this arms race. You say it might be leading us to mutually assured AI malfunction, not mutually assured nuclear destruction.

**[0:48:13] Alex Kantrowitz:** I think that's what you get that from.

**[0:48:14] Dan Hendrycks:** Yeah, so the strategy has three parts, one of which is competitiveness. but we're saying that some forms of competition could be destabilizing and that you may be irrational to pursue it because you couldn't get away with it. So in particular, making a bid for superintelligence through some automated AI research and development loop could potentially lead to one state having some capabilities that are vastly beyond another state's. If one state gets to experience a decade of development in a year and the other one is the year behind, then this results in a very substantial difference in the states' capabilities.

**[0:49:03] Dan Hendrycks:** So this could be quite destabilizing if one state might then start to get an insurmountable lead relative to the other. So I think that form of competition would be very dangerous and because there's a risk of loss of control and because it might incentivize states to engage in preventive sabotage or preemptive sabotage to disable these sorts of projects. So I think states may want to deter each other from pursuing superintelligence through this means. And this then means that AI competition gets channeled into other sorts of realms, such as in the military realm of having more secure supply chains for robotics, for instance, and for AI chips having reduced sole source supply chain dependence on Taiwan for making AI chips.

**[0:49:59] Dan Hendrycks:** So states can compete in other dimensions, but them trying to compete to develop superintelligence first, I think that seems like a very risky idea, and I would not suggest that because there's too much of a risk of loss of control and there's too much of a risk that one state, If they do control it, it uses it to disempower others and affects the balance of power far too much and destabilizes things. But the strategy overall, think of the Cold War.

**[0:50:28] Alex Kantrowitz:** Before you go on the strategy, my reaction to that is good luck telling that to China.

**[0:50:33] Dan Hendrycks:** So I think it's totally – so for deterrence, I think if the U.S. were pulling ahead, both Russia and China may have a substantial interest in saying, hey, cut this out. pulling ahead to develop superintelligence, which could give it a huge advantage and an ability to crush them. They'd say, you don't get to do that. We are making a conditional threat that if you keep going forward in this because you're on the cusp of building this, then we will disable your data center or the surrounding power infrastructure so that you cannot continue building this.

**[0:51:08] Dan Hendrycks:** I think they could make that conditional threat to deter it. And we might do the same, or the U.S. might do the same to China or other states that would do that. So I don't see why China wouldn't do that later on. Right now, they're not just thinking about superintelligence and advanced AI.

**[0:51:43] Alex Kantrowitz:** Your laptop will help keep you locked in, even when it's bring your dog to work day. Built for those who stay in the flow. That's Dell Pro. Built for you. Dell.com slash Dell dash pro. Everyone's talking about AI, but there's a big difference between running a promising pilot and actually making AI work across an enterprise. That's why we're launching the Big Technology Guide To series presented by DXC Technology, available on YouTube. Throughout the series, we're tackling some of the biggest questions facing businesses today.

**[0:52:16] Alex Kantrowitz:** How do you take enterprise AI from pilot to production? How do you navigate an increasingly complex cybersecurity landscape? What's the real opportunity with AI agents? And how do you calculate the return on your AI investments? On my Alex Kantrowitz YouTube channel, you can hear from experts like DXC Technologies Chief Digital Information Officer Russell Jukes. When you ask a question to somebody in your company, do you trust them? And what made you trust them? Do you trust but verify every single time?

**[0:52:43] Alex Kantrowitz:** Russell and I discussed the challenges and opportunities of building production-grade AI and what it really takes to move beyond the hype. Check out the Big Technology Guide To series presented by DXC Technology on my Alex Kantrowitz YouTube channel today.

**[0:52:57] Dan Hendrycks:** This is more of a description of what the dynamics later on when AI is more salient. But it would be surprising to me if China were saying, yes, the United States, go ahead, do your Manhattan Project to build superintelligence, come back to us in a few years, and then tell us, you can boss us around because now we're in a complete position of weakness and we'll be at your mercy and we'll accept whatever you say or tell us to do.

**[0:53:23] Dan Hendrycks:** I don't see that happening. I think they would just say move to preempt or deter that type of development so that they don't get put in that fragile position.

**[0:53:33] Alex Kantrowitz:** Are you in like the Eliezer Yudkowsky camp of bombing the data centers if we get to superintelligence?

**[0:53:39] Dan Hendrycks:** Well, so I think I'm advocating or pointing out that it becomes rational for states to deter each other by making conditional threats and by means that are less escalatory, such as cyber sabotage on data centers or surrounding power plants. I don't think one needs to get kinetic for this. And I think that if discussions start earlier, I don't see any reason things need to be escalating in that way or unilaterally actually doing that. We didn't need to get in a nuclear exchange with Russia to sort of express that we have a preference against nuclear war.

**[0:54:20] Dan Hendrycks:** Thank goodness. So indicating or making conditional threats through deterrence seems like a much smarter move than, hey, wait a second, what are you doing there? And then bomb. That seems needless. Yeah, I'm not into that.

**[0:54:41] Alex Kantrowitz:** But what you're talking about is sort of assuming that there will be a lead that will be protectable for a while. But everything we've seen with AI is that no one protects a lead. Right?

**[0:54:51] Dan Hendrycks:** Well, if there's – so one difference is that when you get to a different paradigm like automated AI R &D, the slope might be extremely high such that if the competitor starts to do automated AI R &D a year later, they may never catch up just because you're so far ahead and your gains are compounding on your gains. Sort of like in social media companies, Eric will use this analogy, where if one of them starts blowing up and growing before you started, it's often the case that you won't be able to catch up and they'll have a winner-take-all type of dynamic.

**[0:55:31] Dan Hendrycks:** So right now, the rate of improvement is not that high or there's less of a path for a winner-take-all dynamic currently. But later on, when you have the ability to run 100 ,000 AI researchers simultaneously, this really accelerates things. Maybe OpenAI has got a few hundred, maybe it'll say 300 AI researchers. So going from 300 AI researchers to orders of magnitude more world-class ones create quite substantial developments. This is something that isn't new.

**[0:56:06] Dan Hendrycks:** This is something that Alan Turing and the founders of computer science had pointed out, that this is a natural property of when you get AIs at this level of capability, then this creates this sort of recursive dynamic where things start accelerating extremely quickly and quite explosively.

**[0:56:27] Alex Kantrowitz:** Okay. We managed to spend most of our conversation today talking about present risks or risks in the near future. we should focus a little bit more on intelligence explosion and loss of control. And we're going to do that right after the break. And we're back here on Big Technology Podcast with Dan Hendrycks. He is the director and co-founder of the Center for AI Safety. Dan, it's great speaking with you about this stuff. Let's talk a little bit.

**[0:56:51] Alex Kantrowitz:** You've been sort of talking about it in the first half, but I want to zero in here on this idea of intelligence explosion or what you talk about as basically having AI autonomously improve itself. Just talk through a little bit about how that might happen and whether you see that being something that is actually probable in our future.

**[0:57:11] Dan Hendrycks:** Yeah. I mean, the basic idea is just imagine automating one AI researcher, one world-class one. Then there's a fun property with computers, which is there's copy and paste. So you can then have a whole fleet of these. It's, well, you know, with humans, you know, if you just have one of them, you know, it's maybe they'll be able to train up somebody else who has a similar level of ability. So this adds a very interesting dynamic to the mix.

**[0:57:40] Dan Hendrycks:** And then you can get so many of them proceeding forward at once. So and, you know, AIs also operate quite quickly. They can code a lot faster than people. So maybe it's maybe you've got 100 ,000 of these things operating at 100x the speed of a human. How fast will that go? Maybe conservatively, let's say it's just overall 10xing research. But 10xing research would mean, say, like a decade's worth of developments in a year. So that telescoping of all these developments makes things pretty wild.

**[0:58:13] Dan Hendrycks:** It means that one player could possibly get AIs that go from very good world class to being vastly better than everybody at everything and to super intelligence, something that towers far beyond any living person or collective of people. So if we get an AI like that, this could be destabilizing because it could be used to develop a super weapon potentially. Maybe it could find some breakthrough for anti-ballistic missile systems, which would make nuclear deterrence no longer work, or other types of ways of weaponizing it.

**[0:58:55] Dan Hendrycks:** So that's why it's destabilizing. And so states then, if they're seeing, oh, they're, you know, don't run this many AI researchers simultaneously in these data centers working on to build a next generation or super intelligence, because if you do so, then that will put us in, that will make our survival be threatened. And so them saying, them deterring that would help them secure themselves. And they can make those threats very credible currently. And I think we'll continue to be able to have these threats be credible going forward.

**[0:59:33] Dan Hendrycks:** So this is why I think it might take a while for superintelligence to be developed because there'll be deterrents around it later on. And then maybe in the farther future, there could be something multilateral. But that's speaking quite far out in very different economic conditions. In the meantime, with the AIs that we'd have in the future, those could still automate various things and increase prosperity and all of that. So we'd still have explosive economic growth if you had something that was just at the average human level ability running for very cheap.

**[1:00:10] Dan Hendrycks:** So I think that those are some of the later stage strategic dynamics. And I don't think we can get away with or I don't think any state could get away with trying to build a super intelligence, go build a big data center out in the middle of the desert, 1 trillion dollar cluster, bring all the researchers there and just not invite the other states to go. What do you think you're doing here? You were at the White House yesterday.

**[1:00:35] Dan Hendrycks:** Well, this is largely just sort of speaking about some of these strategic implications. Are they receptive? Yeah. I mean it's a – this isn't a – there's always interest in thinking what are some of the later term dynamics, what things should happen now and whatnot. But this is – yeah, I think when people think White House, it sounds – Well, where the president lives. So there's the – well, yeah. So there's the Eisenhower building, which is part of the White House, kind of not.

**[1:01:17] Dan Hendrycks:** But that's where everybody works in White House. And I think some of the things we're speaking about here, like virology advancements, things like that, But there's just a lot of things to speak about and think what things make sense or what things to keep in mind going forward.

**[1:01:32] Alex Kantrowitz:** Yeah, I guess I'd rather an executive branch paying attention to this stuff than not.

**[1:01:36] Dan Hendrycks:** Yeah, that's right. Yeah, and what are the sort of ways that help maintain competitiveness? Because how people normally think about this, they'll think it's all or nothing and good or bad thing. And instead, we're saying no, it's dual use. So that means there are some particular applications that are concerning and there are other applications that are good. And you want to stem the particularly harmful applications and what are ways of doing that while capturing the upside.

**[1:02:03] Dan Hendrycks:** Right.

**[1:02:04] Alex Kantrowitz:** Okay. So the intelligence explosion part of this conversation inevitably brings up the loss of control part. where to me, I think the thing that when people think about AI harm, they're always worried that AI is going to escape the simulation or whatever it is and act on its own and try to basically ensure that it preserves itself. We've seen it recently, I think I brought this up at the beginning of the show, where Anthropic has done some experiments where the AI has run code to try to copy itself over onto a server if it thinks that its values are at risk of being changed.

**[1:02:39] Alex Kantrowitz:** Is this – so it's fun to think about, but it's also like probably just probability. Like if you run it enough times because it's a probabilistic engine.

**[1:02:48] Dan Hendrycks:** If it was like, oh, it's only one in a thousand of them intend to do this. Well, if you're running a million of them, then you're basically certain to get many of them to try and self-exfiltrate. And so are you worried that this self-exfiltration is going to be a thing? I think from a recursive automated AR &D thing, I think that has really substantial probability behind it of a loss of control in that situation.

**[1:03:16] Alex Kantrowitz:** So you're worried about this?

**[1:03:17] Dan Hendrycks:** So there's that. But I would distinguish between that and these sort of things that are not super intelligences or things that are not coming from that sort of really rapid loop. And like the currently existent systems. I think that the currently existing systems are relatively controllable. Or if there is some very concerning failure mode, we have been able to find ways to make them more controllable. For instance, for bioweapons refusal, we used not to be able to make robust safeguards for them two years ago.

**[1:03:51] Dan Hendrycks:** But we've done research with methods such as called circuit breakers and things like that. And those seem to improve the situation quite a bit and make it actually prohibitively difficult to do that jailbreaking. And so maybe we'll find something similar with self-exfiltration. So I think people generally want to claim that, like, oh, current AIs are not controllable. And I think that they're not highly reliably controllable. They're reasonably controllable. Maybe we could get some – or it seems plausible that we'll get to have increasing levels of reliability.

**[1:04:23] Dan Hendrycks:** and so I'm sort of reserving judgment. It'll depend more on the empirical phenomena so I think everybody should research this more and we'll sort of see what the risks actually are but there are some that seem less empirically tractable or things that can't be empirically solved like this loop thing like how are you going to you can't run this experiment 100 times or something like that and make it go well. you're making a huge attempt at building a superintelligence and has destabilizing consequences.

**[1:04:57] Dan Hendrycks:** This isn't something that's totally unprecedented. And for that, you have more of a one chance to get it right type of thing. But with the current systems, we can continually adjust them and retrain them and come up with better methods and iterate. So it is concerning. It would not surprise me if this would really start to make AI development itself extremely hazardous. instead of just the deployment, but instead inside the lab, like you need to be worried about the AI trying to be breaking out sometimes.

**[1:05:28] Dan Hendrycks:** That's totally in the realm of possibility. But, yeah, I could see it going either way.

**[1:05:33] Alex Kantrowitz:** Yeah, I mean, this personally freaks me out because, yeah, if you see the AI trying to deceive evaluators, for instance, or you see the AI trying to break out, you really can't trust anything it's telling you. And we had Demis Asabis on the show a little while ago, And he's basically like, listen, if you see deceptive behavior from AI, if you see alignment faking, you really can't trust anything in the safety training because it's lying to you.

**[1:05:59] Dan Hendrycks:** There is truth to that.

**[1:06:02] Alex Kantrowitz:** Are you seeing deceptiveness at Grok, by the way?

**[1:06:04] Dan Hendrycks:** Oh, yeah, yeah. So we have a paper out last week. We're just measuring the extent to which they're deceptive. and in the scenarios we have, like all the models were in these sorts of scenarios under slight pressure to lie, not being told to lie, but just some slight pressure. Then some of them will like lie like 20 % of the time, some of them like 60 % of the time. So they don't really have this sort of virtue sort of baked into them, the virtue of honesty.

**[1:06:31] Dan Hendrycks:** So I think we'll need to do more work and we'll need to do it quickly. So I'm sort of speaking in a more nonchalant way about this, but I can't get worked up at every single risk because they're also just, ooh, be at 11 all the time. So there are some that I'm putting in different tiers than other risks. And this is a more speculative one. We've seen these sometimes get surprisingly handleable. But, yeah, it could end up making things really, really bad.

**[1:07:02] Dan Hendrycks:** We'll see. We'll do things about it to make that not be the case.

**[1:07:06] Alex Kantrowitz:** Okay. Thank you. Two more topics for you, then we'll get out of here. Center for AI Safety. Who's funding it?

**[1:07:14] Dan Hendrycks:** Well, so there's not sort of one funder. It's largely just various philanthropists. The main funder would be Jan Talon and Jan Talon, others who's a Skype co-founder. There's a variety of other philanthropies or philanthropists. It's the generally for – so, for instance, Elon doesn't – I've never asked him to fund the center. So that isn't to say I don't get any money from Elon. My appointment at XAI, I get $1 a year. At Scale AI, I've increased my salary exponentially to where I get $12 a year, $1 per month from Scale.

**[1:08:03] Dan Hendrycks:** But I'll try not to – or I'll try to avoid having some complicated relations with them just so that I can – not feel on behalf of any of them in particular.

**[1:08:16] Alex Kantrowitz:** So you're basically doing the work for them for free?

**[1:08:18] Dan Hendrycks:** Well, but it's useful. Right. It's useful to do. And I mean, yeah, I think the main objective is just try and generate some value here and have as best as one can. by reducing these sorts of risks. Yeah, I think it's a good arrangement because it enables me to, like, you know, do have a choose-your-adventure type of thing. Right. Oh, now I think the politics or geopolitics, this is more relevant. So now I can go off and learn about this for some months and then work on a paper there compared to if it's, like, no, you've got to be coding 80 hours a week.

**[1:08:56] Dan Hendrycks:** That's your job. Yeah. That would be quite restrictive. I couldn't be speaking with you.

**[1:09:01] Alex Kantrowitz:** I'm glad you're here. Thank you, Alex Wayne. So let's talk a little bit about this funding, because I think that after Sam Altman was fired and then rehired at OpenAI, there was a sort of skepticism around effective altruism's impact on the AI field. Even Jan Tallon, I'm reading from his statements, right after the OpenAI governance crisis highlights the fragility of voluntary AI-motivated governance schemes. So the world should not rely on such governance working as intended.

**[1:09:34] Alex Kantrowitz:** Now, Jan is, of course, associated with EA. EA is basically leading the conversation around AI safety. Is that good?

**[1:09:43] Dan Hendrycks:** So I think that in terms of Jan, I think he's funded organizations that are EA-affiliated. I don't know if he'd call himself that, but whatever. People can ascribe labels how they'd like. I think that the, I mean, I've tweeted that EA is not equal to AI safety. I think that EA community generally is insular on these. So I lived in Berkeley for a long time when I was doing my PhD. And there was sort of a school, a sort of AI risk school that was, had very particular views about what things are important.

**[1:10:20] Dan Hendrycks:** So malicious use, for instance, when I was talking about malicious use in the beginning of this thing, you know, they're historically really against that. Yeah, yeah, only loss of control. Don't talk about malicious use. That's a distraction. And so that was annoying because I'd always been working on robustness as a Ph.D. student where the main thing was malicious use. so yeah I ended up leaving Berkeley before graduating just because of the sort of relatively suffocating atmosphere and the sort of central focus on there'd be some new fad and you'd have to get interested in that elk eliciting late knowledge this is the important thing that you have to focus on you have to focus on inner optimizers there's lots

**[1:11:05] Dan Hendrycks:** of these speculative empirically fragile things so for instance this alignment faking stuff that you're seeing. There's some concern there, but I'm not totally sold that this is a top-tier type of priority. But in these communities, this is all that matters currently, roughly speaking. This involuntary commitments from AI companies. I think voluntary commitments from AI companies are also a distraction because the companies will, you should expect most of them by default to just break those sorts of commitments if they end up going up against economic competitiveness.

**[1:11:43] Dan Hendrycks:** So I think it's a distraction, relatively. And so I think there are many people who think that EA's broadly, their influence on this sort of thing, has not been overall positive. I think at least for me and other sorts of researchers in this space who've been interested in AI risks, the amount of pressure to adopt some particular positions on this. It would be extraordinarily high and, I think, quite destructive. So I'm very pleased now that in the past year or so, there's been a lot more diversity of opinion, which has been quite important.

**[1:12:25] Dan Hendrycks:** And I think this is just because the broader world is getting more interested in AI. So a lot of these fixation on this is the one particular risk, this is the most important risk and everything else is distraction. That just doesn't work when you're speaking with the – or interfacing with the real world. There's a lot of complications. And AI is so multifaceted. So you can't – in your risk management approach, you can't just be focusing on one of them.

**[1:12:51] Dan Hendrycks:** Right. So you're not an effective altruist. I don't think of myself as that. I don't particularly get along with this school of thought, this sort of Berkeley AI alignment monolith. And I'm pleased that people can be more independently operating in this space now, which I don't think was the case for many, many years, including basically the entire time I was doing my PhD. And there'll be many people like Dylan Hadfield-Manel, a professor at MIT, who was also at Berkeley at the time, very suffocating.

**[1:13:29] Dan Hendrycks:** Rohan Shah, researcher, very suffocating. They all feel this way.

**[1:13:33] Alex Kantrowitz:** Okay. Let's bring it home. We've been talking for more than an hour about AI safety as if it's controllable. But open source is really putting up a pretty valiant effort in this field, keeping pace with the proprietary labs. And, of course, open source is not controllable. What do you think about that? I mean, we just saw DeepSeek, not to go back to it all the time, but it effectively equaled the cutting edge at the proprietary labs.

**[1:14:02] Alex Kantrowitz:** Get your business ready for clients faster with Base44. Describe what you need in your own words and build the website and software behind it with database, backend, and hosting included. Go to Base44.com. That's Base44.com. You know, put the weights on its website. Right. So how can we possibly have a relationship of safety with AI if open source is out there exposing everything that's been done?

**[1:14:28] Dan Hendrycks:** So I've been I haven't been endorsing open source historically, but I've thought that releasing the weights of models didn't seem robustly good or bad. So I sort of was like, it's fine. Seems to have complicated effects. There's an advantage to it, which it helped with diffusion of the technology so that more people would have access to it and sort of get a sense of AI and sort of increase literacy on this topic and just increase public awareness and get the world more prepared for more advanced versions of AI.

**[1:15:02] Dan Hendrycks:** So that's been my historical position, but this depends on – it should always proceed by a cost-benefit analysis. So if, for instance, they have these cyber capabilities later on, yeah, I think that would be a potential place to be drawing the line on open-weight releases personally, in particular the ones that could cause damage to critical infrastructure. You could still capture the benefits by having the models be available through APIs. And if they're like software developers, they have access to these more cyber offensive capabilities.

**[1:15:45] Dan Hendrycks:** But if they're a random user, they don't. If they're a random faceless user, they don't. And likewise for virology. Once there's consensus, once the capabilities are so high that there's consensus about it being expert level in virology, I think that would be a very natural place to be having an international norm. not saying a treaty because it takes forever to write and ratify, but to a norm against open weights if they are expert level biologists for the same reasons that we had the Biological Weapons Convention.

**[1:16:16] Dan Hendrycks:** Russia and or the Soviet Union and the US got together for the Biological Weapons Convention. US and China did as well. We also coordinated on chemical weapons with the Chemical Weapons Convention and the Nuclear Nonproliferation Treaty. States find it in their incentive to work together to make sure that rogue actors do not have extremely hazardous, potentially catastrophic capabilities like chem, bio, and nuclear inputs. So I think something similar might be reasonable for AI when they get at that capability threshold.

**[1:16:56] Alex Kantrowitz:** Dan, I am at once kind of reassured that people are thinking about this stuff, but also more freaked out than I was when we sat down. But I do appreciate you coming in and giving us the full rundown of what to be concerned about and what maybe not to be as concerned about as we think about where AI is moving next. So thank you so much for coming on the show.

**[1:17:16] Dan Hendrycks:** Yep, yep. Thank you for having me. This has been fun.

**[1:17:18] Alex Kantrowitz:** Super fun. If people want to learn more about your work or get in touch, how do they do that?

**[1:17:21] Dan Hendrycks:** I guess this paper or strategy you've been speaking about is at nationalsecurity.ai. And then I'm also on Twitter or X or whatever it's called. You should know. X.com. X.com slash Dan Hendrycks would be another way of following the goings-ons as the situation evolves. So we'll keep trying to put out work and seeing what's going on with these risks. And if we come up with technical interventions to make them less, then we'll also put that out too.

**[1:17:50] Dan Hendrycks:** So, yeah, that's where you can find me.

**[1:17:53] Alex Kantrowitz:** Well, Godspeed, Dan, and we'll have to have you back. Thanks again. All right, everybody, thank you for listening, and we'll see you next time on Big Technology Podcast.
