“We’ve Lost Control” – Eric Schmidt WARNS About What’s Coming in 2026 (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) We believe as an industry that in the (00:00:03) next one year the vast majority of (00:00:06) programmers will be replaced by AI (00:00:08) programmers. So that's one year. Okay. (00:00:12) What happens in two years? Well, I've (00:00:15) just told you about reasoning and I've (00:00:16) told you about programming and I've told (00:00:18) you about math. Programming plus math (00:00:20) are the basis of sort of our whole (00:00:22) digital world. So the evidence and the (00:00:25) claims from the research groups in open (00:00:27) AI and and anthropic and so forth is (00:00:30) that they're now somewhere around 10 or (00:00:33) 20% of the code that they're developing (00:00:35) in their research programs is being (00:00:37) generated by the computer. That's called (00:00:41) recursive self-improvement is the (00:00:43) technical term. So what happens when (00:00:45) this thing starts to scale? Well, a lot. (00:00:49) So far, 2025 is shaping up to be a (00:00:52) monumental year for AI. And according to (00:00:54) top AI researchers, 2026 is going to be (00:00:57) the year in which millions of jobs will (00:00:59) be replaced by AI. In this interview (00:01:01) with Eric Schmidt, former CEO of Google, (00:01:03) he gave some pretty eye-opening (00:01:05) statements that we can't ignore. Watch (00:01:07) this. So AI is your thing. The topic (00:01:11) today, however, the convergence of AI (00:01:14) with biotechnology. (00:01:16) I'm wondering this report is intended to (00:01:20) increase (00:01:21) awareness of biotechnology and AI to (00:01:26) increase knowledge to get people excited (00:01:29) about it maybe even to get people to (00:01:31) enter the field. If you had to give one (00:01:36) example of this convergence of (00:01:39) AI and biotechnology that you think (00:01:43) might do that. Have you got one? Um, (00:01:47) well, first let me thank uh Senator (00:01:50) Young, the other members of Congress, (00:01:52) Caitlyn obviously who's here, and the (00:01:54) whole team that put this report (00:01:56) together. Uh, I'm really proud of the (00:01:58) work that we did. Um, the the commission (00:02:02) spent an awful lot of time on the (00:02:03) details and I learned when you do these (00:02:05) things, you learn a lot as well as get (00:02:07) to contribute. One of the things that I (00:02:09) learned that I wanted to make sure I (00:02:11) emphasize it's in the report is that I I (00:02:15) had assumed that the biotech people had (00:02:16) figured out scaling and after all these (00:02:19) are huge markets right I I live in a (00:02:21) world of scaling but in fact there's a (00:02:24) huge valley of death and I would (00:02:27) encourage you all to the degree that (00:02:28) you're in this business really think (00:02:31) about it our recommendations in the (00:02:33) report are fundamentally around the (00:02:35) science of scaling technically That (00:02:38) means how many vats, how do you grow (00:02:40) them, you know, that kind of stuff. Very (00:02:42) highly technical bio bio issues. Um (00:02:46) there's also a recommendation in here (00:02:48) around building an infrastructure for (00:02:50) scaling. So what's happening is you have (00:02:52) these incredible startups full of these (00:02:54) incredibly pe people and they raise (00:02:57) enough money, they do something really (00:02:58) interesting and they they're in between. (00:03:01) they're they're not useful enough to (00:03:03) scale, but they're useful enough to have (00:03:06) succeeded and not be able to get the (00:03:08) money because the economics don't work. (00:03:11) Um, this is a this is a wellestablished (00:03:13) problem. U and I would encourage you if (00:03:15) there's one thing that you want to take (00:03:17) away from the report, there's a lot of (00:03:18) things which I'm sure everybody in this (00:03:20) audience understands the underlying (00:03:22) science, the ability for AI to (00:03:24) accelerate both drug discovery but also (00:03:28) its misuse. I was very involved with the (00:03:31) misuse issues. Um there's lots of (00:03:33) evidence that the new models if they're (00:03:35) unconstrained can produce bad pathogens (00:03:38) in particular viruses. Just take an (00:03:40) existing virus and modify it a bit and (00:03:43) off we go. Um one of the the way to (00:03:47) understand and the reason and I want to (00:03:49) thank SCSP (00:03:51) uh the entire team Illy and his and his (00:03:53) whole team obviously Eugenie. Um the (00:03:55) reason we put all this on together is (00:03:57) that this is a big deal for America. (00:03:59) This is a multi- trillion dollar (00:04:01) industry. Um, there's lots of evidence (00:04:04) that China is putting enormous amounts (00:04:07) of money into dominating this space. (00:04:10) They have lots of technology. They have (00:04:11) lots of stuff. They stole all, you know, (00:04:13) the usual arguments. And we document all (00:04:15) that in their report. In the next part (00:04:17) of the interview, Dr. Eric Schmidt calls (00:04:19) AI a ticking time bomb. And he presents (00:04:21) a pretty interesting logic behind this (00:04:23) statement. Watch this. Um I'm the (00:04:26) primary funer of a particular group that (00:04:28) has built a model. It first learned how (00:04:31) to do chemistry and uh it was trained. (00:04:35) It's a foundational model for chemistry (00:04:37) and it's attached to a robotic lab and (00:04:39) what this model does is it generates (00:04:41) hypothesis for drugs of one kind or (00:04:44) another and it just generates them. God (00:04:47) knows if they're right and then (00:04:48) overnight the robotic lab tests them and (00:04:52) gives a report overnight and then it (00:04:54) starts again and the reason I'm (00:04:57) mentioning this is this is the future (00:04:59) model of the fusion of AI and bio right (00:05:02) the AI system generates all sorts of (00:05:05) candidates to reduce the um essentially (00:05:07) the um search space if you think about (00:05:11) it algorithmically it's an exponential (00:05:13) with too many degrees of exponential (00:05:16) So you have to come up with some way of (00:05:18) reducing the space. So this particular (00:05:20) group is using AI to reduce the space, (00:05:22) run the things and so forth. Their (00:05:24) objective, we'll see if they pull it (00:05:26) off. This is research project is to (00:05:29) identify all human druggable targets (00:05:31) within the next two years. If that (00:05:33) occurs, then that information goes (00:05:36) straight into the drug industry. Now, (00:05:39) it's a different way of thinking and (00:05:41) it's profound in that it gives them the (00:05:43) targets they need to go build drugs (00:05:45) against. That's interesting to me. It's (00:05:47) the combi combination of AI and a (00:05:50) robotic lab that does something in a wet (00:05:52) lab essentially. So, one model that you (00:05:55) should think about is wet labs will be (00:05:57) roboticized. And the wet labs will have (00:05:59) AR they're essentially they're not (00:06:01) humanoid robots, they're arm robots and (00:06:03) they go boom boom boom. They and they do (00:06:05) the pipeetting and so forth and so on. (00:06:06) and they do it 24 hours a day. That's a (00:06:09) major change in the way bio bio the (00:06:12) biotech industry works. So if there was (00:06:15) a spectrum of the degree of cooperation (00:06:18) and collaboration between AI and (00:06:21) biotech, one being the lowest, 10 being (00:06:23) the highest, where are we now on that (00:06:26) spectrum? It depends. Um this I'm going (00:06:28) to give you a non-politically correct (00:06:30) answer. It depends on your age. (00:06:33) If you are a graduate student in this (00:06:37) area, every single PhD project now uses (00:06:41) some form of AI in the way that I (00:06:43) described. So speaking as the AI person (00:06:46) in the room, we won, right? We got (00:06:48) complete proliferation. By the way, (00:06:50) that's also true in chemistry. It's also (00:06:52) true in physics. It's also true in (00:06:53) material science. So the and you and I (00:06:57) have talked about this. My overall view (00:06:58) is that AI is underhyped, not overhyped. (00:07:01) Um, what do you mean? Well, let me let (00:07:03) me come back to that. But but the the (00:07:05) key thing to understand is that the seed (00:07:07) corn of research, right, which is how we (00:07:10) work as a nation is founded by graduate (00:07:13) students and postocs and they're all (00:07:15) using this technology. Most of the (00:07:17) applications I have no ability to (00:07:19) understand, but I do see the tools in (00:07:21) use. (00:07:23) So underhyped. It seems to me I read (00:07:27) about AI every day. Well, if if you (00:07:30) actually read the press, it's every like (00:07:33) nancond because right because what (00:07:35) happens is every company is now an AI (00:07:36) company even if they don't do AI at all (00:07:38) because they figured out it increases (00:07:39) their valuations which is useful in (00:07:41) today's market. Um sorry very (00:07:45) uh (00:07:47) so I think that one way to express the (00:07:51) idea is in this room everybody (00:07:54) understands the chat GBT moment. (00:07:56) Everyone here has used Chat GBT now 4.0 (00:07:59) with a new one coming. Gemini, there's a (00:08:02) new one called 2.5 which is beating the (00:08:04) other ones. I'm pleased for that as a (00:08:06) big big Google person. Um Claude 3 um is (00:08:11) the best one for programming and they're (00:08:12) all in the same equivalence (00:08:15) class. The Chinese model DeepS is also (00:08:19) in the same equivalence class. Gro 3 (00:08:21) which has just come out of the Memphis (00:08:23) data center with Elon uh which is now (00:08:26) part of the union of of Twitter and XAI (00:08:29) is also in the same class. You think of (00:08:32) those as language to language ask it a (00:08:35) question. I was talking to somebody who (00:08:37) was explaining to me that they use it (00:08:39) for relationship advice. And another (00:08:42) person told me that he used it for um (00:08:45) psychological advice. And I said, "Do (00:08:48) you realize that it wasn't trained to be (00:08:49) either a relationship coach or a (00:08:51) psychiatrist?" And in fact, it's (00:08:53) probably illegal to say that you're a (00:08:55) psychiatrist because trust me, it didn't (00:08:57) pass any of the tests and it hasn't been (00:08:59) vetted. Well, nobody cares. The power of (00:09:02) these models is extraordinary. Um, and (00:09:05) you know, whenever I have a complicated (00:09:07) question, I just ask one of those (00:09:09) services. Um, and they're all they have (00:09:12) differences, but they're roughly in the (00:09:13) same same category. That's last year's (00:09:17) story that everybody thinks is the (00:09:19) current story. The next story is the (00:09:22) ability to do planning. take a look at (00:09:24) uh OpenAI (00:09:27) R3, excuse me, 03 or DeepSseek R3 and (00:09:31) you'll see that uh they do this (00:09:33) incredible demo. You ask it to show it (00:09:35) what it's doing or do deep research (00:09:37) which is available in most of these (00:09:39) things and it'll show you how it goes up (00:09:41) the decision path. It'll try something, (00:09:43) it didn't work. It goes back, it tries (00:09:45) something else, didn't work. Oh, goes (00:09:47) here. Oh, it worked. And then it goes (00:09:49) over here and so forth. It's following (00:09:51) the tree of choices. Now what's (00:09:54) interesting about that? So you sit there (00:09:56) and go why is he so emphasized with (00:09:57) that? By the way, that's how we (00:10:00) think, right? Let's let's pause. We went (00:10:02) from language conversation and (00:10:05) furthermore the foundation models today (00:10:07) in biology use sequence prediction to (00:10:09) predict biological elements, chemistry, (00:10:11) so forth and so on. That's all well (00:10:13) established. But the big breakthrough (00:10:14) now through a technology called (00:10:16) reinforcement learning is this. So you (00:10:19) go, okay, well that's pretty impressive. (00:10:21) Okay. So, we believe as an industry that (00:10:24) in the next one year, the vast majority (00:10:27) of programmers will be replaced by AI (00:10:30) programmers. We also believe that within (00:10:33) one year, you will have graduate level (00:10:36) mathematicians that are at the tippy top (00:10:38) of graduate math programs. There's lots (00:10:40) of reasons to think this is going to (00:10:41) happen. This is the consensus. You go, (00:10:44) okay, well, that's pretty interesting. (00:10:46) Now, I can't do that kind of math. very (00:10:49) few people can do that math. How can the (00:10:51) computer do that math better than (00:10:53) anybody else? To some degree, it's (00:10:55) because math has a simpler language than (00:10:57) human language. So, the way these (00:11:00) algorithms actually work is they're (00:11:02) doing essentially word prediction. So, (00:11:03) you take you take a a sentence, you take (00:11:06) a word out, and then it learns how to (00:11:08) put the correct word back in. This is (00:11:09) called the loss function, and it's (00:11:11) optimized to do that at a scale that's (00:11:13) unimaginable to us as humans. So you do (00:11:16) the same thing for math. But there you (00:11:18) use a conjecture and then a proof format (00:11:20) through a protocol called lean. In (00:11:22) programming it's pretty simple. You just (00:11:25) keep writing code until you pass the (00:11:27) programming (00:11:28) test. So strangely the first question I (00:11:30) always ask programmers is what language (00:11:32) do you program in? And the correct (00:11:33) answer is it doesn't matter because (00:11:36) you're trying to design for an outcome. (00:11:37) You don't care what code is generated by (00:11:39) the computer. This is a whole new world. (00:11:42) Okay. So that's one year. Okay, what (00:11:46) happens in two years? Well, I've just (00:11:48) told you about reasoning and I've told (00:11:50) you about programming and I've told you (00:11:51) about math. Programming plus math are (00:11:54) the basis of sort of our whole digital (00:11:56) world. So, the evidence and the claims (00:11:59) from the research groups in OpenAI and (00:12:01) and anthropic and so forth is that (00:12:04) they're now somewhere around 10 or 20% (00:12:07) of the code that they're developing in (00:12:09) their research programs is being (00:12:11) generated by the computer. (00:12:14) That's called recursive self-improvement (00:12:16) is the technical term. So what happens (00:12:18) when this thing starts to scale? Well, a (00:12:22) lot. Now in this part of the interview, (00:12:23) Eric Schmidt makes a prediction on when (00:12:25) AGI will arrive. He expects AGI to (00:12:28) happen within 3 to 5 years. And to be (00:12:30) honest, this timeline is not good news (00:12:32) for humans because AGI is expected to (00:12:34) replace more than 90% of the human (00:12:37) workforce. One way to say this is that (00:12:39) within 3 to 5 years, we'll have what is (00:12:42) called general intelligence, AGI, which (00:12:45) can be defined as a system that is as (00:12:47) smart as the smartest mathematician, (00:12:50) physicist, you know, artist, writer, (00:12:53) thinker, politician, maybe not in the (00:12:55) same level, um, but you get the idea. (00:12:58) Uh, just the creative industries and so (00:13:01) forth. But imagine that in one computer. (00:13:03) Okay. Well, that's pretty interesting. I (00:13:05) call this, by the way, the San Francisco (00:13:06) consensus because everyone who believes (00:13:08) this is in San Francisco. It may be the (00:13:10) water. What happens when every single (00:13:14) one of us has the equivalent of the (00:13:17) smartest human on every problem in our (00:13:19) pocket? So, it means you have to best (00:13:21) architect when you have an architecture (00:13:22) problem. Another thing that's going on (00:13:24) is the development of agentic solutions (00:13:26) and agents are refer to systems that (00:13:29) have input and output in memory and they (00:13:31) learn. An example here is that I want to (00:13:34) uh buy another house. Uh I happen to (00:13:36) like Virginia. I grew up in Virginia. I (00:13:38) say, "Find me a house in the greater (00:13:40) MLAN area. Look at the that's one agent. (00:13:43) Look at all the rules. Figure out how (00:13:45) big a house I can build." That's another (00:13:47) agent. Do the transaction to buy the (00:13:50) land. That's another agent. Design the (00:13:52) house with a human architect, right? but (00:13:55) sort of ignore them for most of the (00:13:57) thing, but they have to sign it off and (00:13:59) then I approve it and then find the (00:14:01) contractor, right? Hire the contractor, (00:14:04) pay the bills, and at the end sue the (00:14:06) contractor for lack of (00:14:08) performance. Okay? Now, I just gave you (00:14:11) the stupidest possible explanation. I (00:14:14) just described every business process, (00:14:17) every government process, and every and (00:14:19) every sort of academic process in our (00:14:22) nation. So it isn't just the programmers (00:14:24) that are going to be out of work. We're (00:14:26) all going to be out of work. No, that's (00:14:28) not a consequence. I'll come to that. (00:14:30) But but the reason I want to I want to (00:14:31) make the point here is that in the next (00:14:34) year or two, this foundation is being (00:14:36) locked in and it's not we're not going (00:14:39) to stop. (00:14:40) It gets much more interesting after that (00:14:44) because remember the computers are now (00:14:46) doing self-improvement. They're learning (00:14:48) how to plan and they don't have to (00:14:51) listen to us anymore. We call that super (00:14:54) intelligence or ASI, artificial super (00:14:56) intelligence. And this is the theory (00:14:59) that there will be computers that are (00:15:01) smarter than the sum of humans. The San (00:15:04) Francisco con consensus is this occurs (00:15:06) within six years just based on scaling. (00:15:10) Now, in order to pull this off, you have (00:15:13) to have an enormous amount of power. I (00:15:17) was here yesterday testifying about (00:15:18) this, you know, and we need like I can (00:15:21) talk at some length about how many (00:15:23) gigawatts and how many nuclear power (00:15:25) plants and all the kind of stuff we can (00:15:26) talk about (00:15:28) separately. This path is not understood (00:15:31) in our society. There's no language for (00:15:34) what happens with the arrival of this. I (00:15:36) wrote a book on this with Henry (00:15:37) Kissinger called Genesis which you know (00:15:39) I recommend obviously um because I wrote (00:15:42) it available available available in your (00:15:44) usual places um but the important point (00:15:47) is this is happening faster than our (00:15:50) human that our society our democracy our (00:15:53) laws will attract and there's lots of (00:15:55) implications that's why it's underhyped (00:15:58) people do not understand what happens (00:16:00) when you have intelligence at this level (00:16:03) which is largely free that's the How do (00:16:06) we get ready for it? Well, we start by (00:16:09) talking about it. And by the way, on the (00:16:11) jobs thing, everyone assumes that (00:16:12) automation will replace will eliminate (00:16:14) jobs. If you look at the history of (00:16:16) automation ever since the the looms and (00:16:20) uh in uh 300 years ago, the jobs are (00:16:23) changed, but more jobs are created than (00:16:26) destroyed. In this case, you'd have to (00:16:29) convince me that this time is different. (00:16:32) If you look in Asia where they for (00:16:34) whatever reason are choosing not to have (00:16:36) children, the Asian reproduction rate is (00:16:39) in the order of 1.0 or lower. So they're (00:16:42) rapidly disappearing. So the Asian (00:16:45) countries are very very quickly (00:16:47) automating. The tools that I'm (00:16:49) describing will allow the few humans (00:16:52) that will be working very hard in 30 or (00:16:55) 40 years. If these trends continue, the (00:16:57) rest of us will be dependent on those (00:16:59) hardworking humans. it'll make their (00:17:01) productivity more much greater. (00:17:04) We aren't the only ones working on this. (00:17:07) Talk about the state of the competition (00:17:08) if you would. Um well, first place in (00:17:11) the American model is uh the big (00:17:14) companies that you all know. Uh Meta (00:17:16) just released a version of Llama. It's (00:17:19) called Llama 4. Uh which is also in the (00:17:21) ballpark. And they play a slightly (00:17:24) different role. They've done a very good (00:17:26) job because they release it in what is (00:17:27) called open weights. that is they (00:17:29) actually show how the algorithm works. (00:17:32) The other guys are completely (00:17:33) proprietary. There's these are (00:17:34) complicated business decisions that (00:17:35) everybody's (00:17:37) making. In (00:17:39) China, the deepseek moment is equivalent (00:17:43) to our chat GPT moment. I was there with (00:17:45) Henry. Um, and this is what happens when (00:17:48) you're talking to to the Chinese about (00:17:50) AI with Henry. And this means we are (00:17:53) alive and we're listening to you. Thank (00:17:55) you very much. Right? (00:17:58) That's not what they're doing anymore. (00:18:00) When the when Deep Seek showed up and (00:18:02) our stock market lost a trillion dollars (00:18:04) in one day, all of a sudden they began (00:18:06) to understand the scale of what it was. (00:18:08) So now there is a massive program in (00:18:11) China to accelerate these things. I had (00:18:14) thought Illy and I and some of the other (00:18:16) people in this room worked really hard (00:18:18) on these um chip controls and the chip (00:18:22) controls have been um in my view largely (00:18:25) effective. How did China get around (00:18:29) them? Well, some of it was (00:18:30) straightforward theft and evasion of the (00:18:32) tariffs, but they also they're (00:18:34) sufficiently smart. They created new (00:18:36) algorithms that use different kinds of (00:18:38) computing to move forward because they (00:18:41) because China operates in open source (00:18:43) that is they they release the software (00:18:45) to everyone. There are two things that (00:18:47) happen. We we Americans immediately saw (00:18:50) their idea and incorporated in our own. (00:18:52) So, thank you very much China. You (00:18:53) invented something new. We immediately (00:18:54) incorporated it. But (00:18:57) second, because it's free, the (00:19:00) proliferation issues around Chinese (00:19:02) models have now become a very big deal (00:19:05) and our government is trying to figure (00:19:06) out uh without success so far how to (00:19:09) handle this question. It's a very tricky (00:19:11) question. We call these wicked hard (00:19:13) problems. So we need smart people to be (00:19:16) doing all this engineering. I want to (00:19:17) ask you about some current events. We (00:19:21) have seen in recent weeks the suspension (00:19:23) of research programs at some of the (00:19:26) premier US universities. There have been (00:19:29) layoffs and cuts at some of the (00:19:32) government's premier scientific (00:19:35) organizations. Some international (00:19:37) students are choosing not to come to the (00:19:39) United States. Others are leaving uh (00:19:41) because they're afraid of being swept up (00:19:43) on the streets. and some top US (00:19:47) scientists are looking for jobs (00:19:49) elsewhere and they're being courted by (00:19:52) other governments. Do we risk (00:19:56) losing the brain power that we (00:19:59) need to stay competitive both in AI in (00:20:03) biotechnology and every other emerging (00:20:05) technology? I had thought that this was (00:20:08) just the usual government stupidity (00:20:09) around politics but here here are some (00:20:13) of the facts. Last week I was in London (00:20:15) talking to people and they said that we (00:20:17) are preparing for people who are moving (00:20:19) back from the US because they don't want (00:20:21) to work in this environment. They figure (00:20:22) they're going to these are British (00:20:23) people right like our best allies. Um (00:20:26) the damage of the 15% everybody (00:20:29) understands here there's this thing (00:20:30) called the indirect rate and um the (00:20:33) current government makes the claim that (00:20:35) the universities are overbilling against (00:20:37) the 15% which is false. It turns out (00:20:40) that the way the structure was erected (00:20:43) in the 50s, this is under Vanavar Bush (00:20:46) was that the people were in the direct (00:20:47) cost and the labs were in the indirect (00:20:49) cost. So if you fully burden the the (00:20:53) cost, the overhead rate is somewhere (00:20:56) between 10 and 15%. This is evidenced (00:20:59) for example by the Gates Foundation and (00:21:01) I'm a philanthropist so I know these (00:21:02) things but they look at total cost. The (00:21:05) government has chosen to use this as a (00:21:07) mechanism falsely to attack science. If (00:21:11) the government has a problem with (00:21:12) specific scientists or specific science (00:21:14) research, please have a good time. But (00:21:17) this looks like a total attack on on all (00:21:20) of science in America. Now, why is this (00:21:22) a problem? Everything that has happened (00:21:25) in American exceptionalism. So, an (00:21:27) example is America, the average American (00:21:30) has twice the income now as a European. (00:21:34) Good job. Why? We're more innovative. (00:21:37) What are we innovative in? Science and (00:21:40) technology generated business (00:21:42) opportunities. If you think that this (00:21:44) sounds like me, a Democrat, let me (00:21:46) remind you that fracking, hugely (00:21:48) successful in America, made us (00:21:51) independent of oil and gas, made us the (00:21:53) largest exporter, right? Good job. (00:21:56) Followed the same path, right? (00:21:58) innovation in the universities and then (00:22:00) entrepreneurship and then government (00:22:02) support over 30 years. We have plenty of (00:22:04) examples of this. Another example of (00:22:06) current damage um universities are so (00:22:09) scared because the administration (00:22:11) appears to be withholding hundreds of (00:22:13) millions of dollars from them. What does (00:22:15) the university do in the first thing? (00:22:16) They put in a hiring freeze. Okay. So (00:22:19) you have a graduate student who wants to (00:22:22) serve in university. They're graduating. (00:22:25) the industry will offer them a salary of (00:22:27) $2 to $3 million a year and they're (00:22:29) foolish enough to turn that down. They (00:22:31) want to serve the university. They want (00:22:33) to teach. They want to build it. They (00:22:35) call up the university and they say, "We (00:22:36) can't interview you." So, they go to (00:22:39) industry. Good for industry. They stay (00:22:40) in America, but we use we lose that for (00:22:43) our seed corn. current faculty members (00:22:46) are because they can't get the research (00:22:48) money are unlikely to be able to get to (00:22:50) tenure in the way that they have to and (00:22:52) those people's careers will be (00:22:54) destroyed. Now, this madness will (00:22:56) eventually end because it's too stupid (00:22:58) not to not not to fix, but it's going to (00:23:01) be too late for them. But there's damage (00:23:03) occurring already. And I want everyone (00:23:04) to understand it's it's real damage, (00:23:06) right? We need We're up against China (00:23:09) that is pouring a trillion dollars into (00:23:11) this and we're screwing around with (00:23:13) funding the core people to invent our (00:23:16) future. Anyway, I can go on. I've heard (00:23:18) some people say US science is being (00:23:19) gutted. Would you go that far? Well, (00:23:21) that's the language that the university (00:23:23) professors say. Um, if you look at bio, (00:23:26) which is subject today, pretty much all (00:23:29) the bio research is burdened at about a (00:23:32) 65 to 85% overhead rate. And this is why (00:23:36) the NIH cut cuts are so profound. So to (00:23:39) go from 85% to 15%, you have to cut your (00:23:41) budget in half. Now you can call that (00:23:44) gutting, you can call it cutting in (00:23:45) half, but it's people's lives, it's (00:23:47) programs, they're all getting stopped. (00:23:50) Are you doing anything about this? Are (00:23:51) you having Well, there's a bunch of (00:23:53) philanthropists who are trying to figure (00:23:54) out a way to spend more money. Um, which (00:23:56) is always a good thing. I can help you, (00:23:58) by the way. The problem is the numbers (00:24:01) are too large. Um, private philanthropy (00:24:04) generates some number of hundreds and (00:24:06) hundreds of millions of dollars a year. (00:24:08) You can't make up the billions and (00:24:10) billions that the government provides. (00:24:12) Remember the deal that was done before (00:24:15) virtually all of us were born. That deal (00:24:17) was that the government would provide (00:24:20) basic support for research. The (00:24:22) universities would produce it. The (00:24:23) venture capitalists would spend an awful (00:24:25) lot of time with these people and they (00:24:26) would create companies. The government (00:24:28) would then provide necessary market (00:24:31) making or whatever to do this and it has (00:24:33) produced these industrial champions. (00:24:35) That is the American way. Don't change (00:24:38) it please. Uh we are going to take your (00:24:47) questions. We are going to take some (00:24:49) audience questions in just a minute. So (00:24:50) put on your thinking caps. (00:24:53) Um I'm wondering about private (00:24:55) investment. We've heard a lot about it (00:24:57) here today. Um, and what are your (00:25:00) thoughts on the current economic turmoil (00:25:04) and what that is doing to the mood to (00:25:06) invest in AI in biotechnology? I think I (00:25:10) don't think I have I think everyone in (00:25:11) the room understands that businesses (00:25:13) need (00:25:14) predictability and so any system where (00:25:18) the rules change every week is pessimal, (00:25:21) right? Because people are making very (00:25:24) very long-term decisions. I'll give you (00:25:26) an example. Uh most people think that we (00:25:30) need 10 20 gawatt of electricity which (00:25:36) translates to data centers. Those data (00:25:38) centers typical chip is 2 kilowatts. So (00:25:41) you can do the the math. These end up (00:25:43) being I don't know 50 billion hundred (00:25:46) billion dollar decisions. They play out (00:25:49) over takes a couple mill couple years to (00:25:51) build the data centers. you have to then (00:25:53) get in queue to get the now the GB200 GB (00:25:56) 300 from Nvidia and so forth. Um the (00:25:59) economic (00:26:00) uncertainty slows that it happens but it (00:26:03) happens slower and with greater latency. (00:26:07) All of these slow down the vision that I (00:26:09) just outlined. Now why is this (00:26:11) important? I I'll give you one more (00:26:13) example and this is what I'm really (00:26:15) worried about. Um let's imagine um so (00:26:20) we'll use you're the good person. You're (00:26:22) the the the good good lady and I'm the (00:26:23) bad guy. I like that. Okay. And the good (00:26:25) lady, the United States in this case, is (00:26:29) ahead and you've done everything right. (00:26:32) I'm the bad guy, China or whatever, and (00:26:34) I'm 6 months, 12 months (00:26:37) behind. As you get closer to super (00:26:42) intelligence, right, I get more and more (00:26:45) worried unless I'm going to be there the (00:26:47) same as you. And you sit there and you (00:26:49) go, "Ah, what's he complaining about?" (00:26:51) you know, we it took four years for the (00:26:53) um the atomic bomb to be recreated in in (00:26:56) the Soviet Union. During that four (00:26:58) years, we had a monopoly, but it was (00:27:00) fairly quickly uh (00:27:02) eliminated. These are network effect (00:27:04) businesses. And so network effect (00:27:07) businesses have the property that the (00:27:09) leader tends to get 90% share. So in the (00:27:12) scenario where you the good lady are (00:27:15) doing this, of course we would all (00:27:17) applaud you as Americans. You're likely (00:27:19) to get 90% share or more of intelligence (00:27:21) in the world. Okay, that would be (00:27:24) terrible for me, right? What would I do? (00:27:28) Try to undermine me. Okay, let me tell (00:27:30) you how I would start. Just to give you (00:27:32) a heads up, the first thing I would do (00:27:34) is to try to steal your intellectual (00:27:36) property and the people. Check. Okay. (00:27:39) and you're such a good lady that you've (00:27:41) managed to prevent me from doing that. (00:27:43) The next thing I'm going to do is use my (00:27:45) AI, which is almost as good as yours, to (00:27:47) go into your eye. These are called (00:27:49) adversarial attacks, and modify your (00:27:52) system. Yeah. And you go, "No way." (00:27:55) Because we have such great (00:27:56) cryptologologists. We're so far ahead of (00:27:57) you are that six months we anticipated (00:27:59) this. What's my next move? I bomb your (00:28:03) data (00:28:04) center. But but think about it. (00:28:08) We're having this whole debate in our (00:28:09) nation about what to do about Iran's (00:28:11) nuclear program. And I'm not an expert (00:28:12) in that. But these are the kind of (00:28:14) conversations that happen here in in DC. (00:28:17) So when we get to the point where China (00:28:19) is n months ahead, are we willing to (00:28:22) bomb their data centers? My favorite (00:28:24) example here is I was in I've been (00:28:25) working on this. I was talking to (00:28:26) somebody said, "The answer is obvious." (00:28:28) I said, "What?" (00:28:30) the good lady and the bad guy, we agree (00:28:33) to a treaty where each of us puts (00:28:35) dynamite on each (00:28:36) other's uh electricity supply. You get (00:28:40) to blow up my electricity if you get mad (00:28:41) and I get to blow up your electricity if (00:28:44) I get you get the idea. Now, some would (00:28:47) say we've already done that's already (00:28:49) happened. Well, the kinetic attack on (00:28:51) people's data centers is probably an act (00:28:53) of war. Yeah. So this is the kind of (00:28:57) thinking that people are be and (00:28:58) obviously that that proposal is (00:29:00) rejected. I'm not I'm using it as an (00:29:02) example. It's not going to happen. But (00:29:05) this is an example of where the (00:29:07) proliferation issues and technically (00:29:09) this is called the eye of the needle (00:29:10) problem. You have to get through this (00:29:12) eye of the needle without killing (00:29:13) yourself and killing everybody else to (00:29:15) get to this promised land of Aon. (00:29:18) Speaking of relationships with other (00:29:20) countries, um, in this report and in (00:29:22) other conversations I've had, people (00:29:24) have talked a lot about the need in the (00:29:26) technology space to collaborate with our (00:29:28) allies and with our friends. Let me (00:29:31) bring it back to the current moment one (00:29:32) more time and ask you, are we going to (00:29:35) see that kind of cooperation taking (00:29:36) place? Well, we need to if you if you (00:29:39) look at how you compete with China, (00:29:40) which seems to be what we how we frame (00:29:42) things now in Washington, we're only (00:29:45) going to succeed if we have partners. (00:29:47) The best partners are (00:29:50) Canada, the European Union, Israel, you (00:29:54) know, places like that, Korea, Japan, so (00:29:57) forth. If you can't articulate that, (00:30:00) then you don't understand these are (00:30:02) scale businesses, right? So, I give an (00:30:05) example. Japan has recently come up with (00:30:07) a new EUV technology which I don't fully (00:30:10) understand. It's new physics to compete (00:30:12) with the ASML machines that are (00:30:14) currently being used in Taiwan. This is (00:30:16) good. That competition historically a (00:30:19) monopoly will give us more choices as to (00:30:22) how have how we can have the supply of (00:30:24) chips that we need for our nation and (00:30:26) national security. Right? Thank God for (00:30:28) the Japanese. Who thought you see my (00:30:32) point? We need to we need to keep these (00:30:34) people tight because we work better (00:30:36) together. But we're not doing that right (00:30:38) now. That's a mistake. (00:30:40) Um I'd love to take some questions. (00:30:42) There are some folks with microphones (00:30:44) who might be able to identify themselves (00:30:46) for me. Um you have the man right here. (00:30:49) Do I see some hands out there of (00:30:51) questions? Here's one right at the (00:30:52) front. Right at the front. Have we got a (00:30:54) mic? (00:30:58) Oh, bless you. The mics are showing up (00:31:01) right here. (00:31:03) AI will imagine your question and ask. (00:31:08) Hello. Uh, greetings Dr. Schmidt. Uh, (00:31:11) many background. I'm a recent PhD in (00:31:13) biomedical engineering. Very excited. (00:31:14) I've been following you. Just a very (00:31:16) quick question. Do you think there's (00:31:17) implications for ASI via drug discovery (00:31:20) for like curing cancer and or (00:31:23) personalized medicine? Just something. (00:31:25) Um, yes. Because under the under the (00:31:27) assumptions of super intelligence, these (00:31:30) are systems that see things that we (00:31:32) don't see. And so the assumption is that (00:31:36) ASI, for example, could understand (00:31:39) biological and cellular mechanisms that (00:31:41) you are an expert in and I'm not at a (00:31:44) level that humans will not. So that's (00:31:46) why this is such a big deal. We've (00:31:48) always assumed that humans would know (00:31:51) there would be at least one human, (00:31:52) right? We call these people polymaths (00:31:54) that would understand these things. (00:31:56) We're going to end up in a world maybe (00:31:58) 10 years from now where we won't (00:31:59) actually understand why. But you as our (00:32:02) scientist will say I use it every day. (00:32:04) When I when I was at college, I was (00:32:06) studying quantum physics and my friend (00:32:08) who was a graduate student who is much (00:32:10) better than I and I said is this stuff (00:32:12) actually (00:32:13) true? You know, it's like too weird to (00:32:16) be true. And he said yes, we use it (00:32:18) every day. And I imagine in 10 years (00:32:22) some young student will come up to you (00:32:23) and say, "Is this stuff true?" And (00:32:25) you'll say, "Frankly, I use it every (00:32:28) day. No human understands (00:32:31) it." What an interesting situation for (00:32:34) you as now a senior researcher 10 years (00:32:36) from now to have to deal with. Do we (00:32:39) have another question out there at the (00:32:41) same table? We have one here. Yes, (00:32:43) ma'am. Microphone. We're discriminating (00:32:45) in favor of the front row. Well, that's (00:32:47) so I can see. Yes. Um, but we need a (00:32:50) mic. (00:32:52) There's a mic over Well, the mics are in (00:32:53) random places. There's a mic over there. (00:32:55) The mics were in people's hands that (00:32:56) were running them around. I thought he's (00:32:58) right over there. Well, thank you very (00:33:00) much. Great conversation. I had a (00:33:02) question about par even within the US. (00:33:05) Imagine my uh young researcher or I'm a (00:33:09) small company. I don't get to compete in (00:33:12) terms of having the resources or GPU (00:33:14) DevOps or all of the things that go to (00:33:16) help me manage my data, organize my (00:33:19) code, and get the right computer (00:33:21) orchestration. And how do we create the (00:33:24) resource so that it's not just top (00:33:27) companies, top things that can run away (00:33:29) cuz that also creates a issue in (00:33:32) research because I won't be able to (00:33:34) validate someone else's because they (00:33:36) have more resources. Um, exactly right. (00:33:39) Um there's a proposal called NAR (00:33:41) national AI research research which was (00:33:45) proposed by a group at Stanford adopted (00:33:48) among the university system we have (00:33:50) endorsed it a great deal to try to get (00:33:52) enough (00:33:53) hardware. If you look at the math of (00:33:56) what these companies are proposing, (00:33:58) universities will never have that kind (00:34:01) of resources. They never did in physics. (00:34:03) They never did in chemistry. They're not (00:34:05) going to have it in AI. There's good (00:34:06) news which is at least the open-source (00:34:09) models and which are are essentially (00:34:12) pre-trained will be available to them (00:34:14) and we'll have robust such things and (00:34:16) that we'll be able to take a powerfully (00:34:18) trained model and then adapt it to your (00:34:20) specific area. I think that's the best (00:34:22) solution that we can come up with. (00:34:24) Where's that mic now? Is it? Do we have (00:34:28) it? (00:34:30) Okay, just speak up because I can't see (00:34:32) you. (00:34:34) Um, can I ask a question about the (00:34:36) confluence with the other exponential (00:34:38) technologies? So, we've been talking (00:34:40) about AI and bio, but there's fusion, (00:34:43) there's quantum. What do you see that (00:34:48) coming together of the (00:34:50) triumphirate looking like? And how do (00:34:53) you how do we think about maintaining (00:34:56) competitiveness in all three so that (00:34:58) we're we advance all those at the same (00:35:01) pace? Um I'd love we can I don't think (00:35:04) we're going to be able to keep it the (00:35:06) same pace. (00:35:07) Um to some degree because the AI (00:35:10) revolution is essentially math and it's (00:35:12) governed by essentially three scaling (00:35:15) laws. Um there's a a very good paper by (00:35:18) Dario called the machine machines of (00:35:20) love and gra love loving grace. The (00:35:23) first scaling law is the law that you (00:35:25) see with chat GPD and others where if (00:35:28) you just add more hardware and more data (00:35:30) and more time it just gets smarter and (00:35:32) smarter. The second one involves (00:35:34) reinforcement learning and planning the (00:35:35) examples that I used. And the third (00:35:38) involves something called test time (00:35:39) training where the system is learning as (00:35:42) it's doing. The latter two are just at (00:35:45) their infancy. So it looks like in core (00:35:48) AI we are riding these exponential (00:35:50) curves and we've got more to ride. No (00:35:53) one yet I've asked this many times. No (00:35:56) one yet has seen those limits. We (00:35:58) thought there might be a limit when we (00:36:00) ran out of stuff to train on. We we've (00:36:03) essentially sucked all human knowledge (00:36:05) as written down anyway into these models (00:36:07) uh with all sorts of implications. But (00:36:09) there's plenty of data that we can (00:36:11) generate to keep powering those things. (00:36:13) So it looks like those exponentials are (00:36:15) going to go quite a bit faster than the (00:36:17) others. Um so if you look at fusion for (00:36:21) example, AI is necessary for all fusion (00:36:25) designs and all fusion man management (00:36:27) especially the plasma. But the core (00:36:29) science is not running on an (00:36:31) exponential. In quantum it's the same (00:36:33) argument. So we'll get there but I think (00:36:36) AI is going to lead first. The reason I (00:36:39) chose programming and math as opposed to (00:36:42) the hard sciences or I guess they're (00:36:43) they're hard but the other hard sciences (00:36:46) is that uh programming and math are (00:36:49) scale free. In other words, there's no (00:36:52) hardware constraint. It's just if you (00:36:53) have enough electricity, you can just do (00:36:55) more math programs. You don't need more (00:36:57) biology and more labs and more whatever. (00:37:00) That's why I think it'll burst there (00:37:01) first and then diffuse into the other (00:37:04) fields. (00:37:05) I'm sure there's another question out (00:37:07) there. Mr. Microphone, I'm going to let (00:37:09) you pick because I can't see anything. (00:37:13) How about over here? Do we have someone (00:37:14) with a mic? (00:37:16) Here we go. Oh, hi. Hi, Sydney Friedberg (00:37:20) from Breaking Defense and a long time (00:37:22) follower of your work, your work with (00:37:24) Bob Work. Uh actually, um let me ask a (00:37:29) longtime science covering reporter and (00:37:33) science fiction fan, but also an AI (00:37:35) skeptic, how confident are you about (00:37:37) this super intelligence thing? I mean, (00:37:39) we've had the chat GPT revolution for a (00:37:42) few years now, and it seems to produce (00:37:44) often super stupidity uh instead. Or at (00:37:47) least if you turn off the turn the (00:37:49) temperature way down and feed it with (00:37:52) rag and tell it only to use trusted (00:37:54) forces and pull the citations and to (00:37:56) make citations to real things not (00:37:58) imaginary things. You will get the (00:38:01) lowest common denominator of what you (00:38:02) put in the database that you know the (00:38:06) exact opposite of creativity and super (00:38:08) intelligence and ability to discover (00:38:09) anything new. Uh the LLM seem to be you (00:38:13) know the great distiller of sort of thin (00:38:17) gr from the combined knowledge of human (00:38:20) civilization. Uh which is useful in (00:38:22) certain applications like for the (00:38:24) government here in DC. We produce reams (00:38:26) of paper no one ever wants to read so we (00:38:27) can make the AI read them and summarize (00:38:29) them for us. Uh but do you have a (00:38:32) question for how is how how is that (00:38:34) heading towards super intelligence as (00:38:36) opposed to super mediocrity? My so my my (00:38:40) favorite current example is I'm on the (00:38:42) board of a hospital and the insurers (00:38:46) send letters that are generated by a (00:38:48) computer to reject treatment. So this (00:38:51) particular hospital it uses the (00:38:52) equivalent of chat GPT to generate (00:38:55) appeal letters. So you have a computer (00:38:57) writing a letter and a computer and this (00:38:59) is this is how we run our medical (00:39:01) systems. it. So to go back to the (00:39:04) stupidity point, the fact that AI is (00:39:06) good at something doesn't mean the (00:39:07) process that it's embedded in makes any (00:39:08) sense at all. Um I don't agree with some (00:39:11) of the things you said in particular (00:39:13) that the the algorithms are so much (00:39:16) better in terms of hallucinations and so (00:39:19) forth in the last year and they're (00:39:21) getting much stronger, but the real (00:39:23) arrival is reinforcement learning which (00:39:25) allows you to do path dependent (00:39:28) reasoning. We don't know where that (00:39:30) limit is. Let me give you an example of (00:39:32) a question that we don't know which I (00:39:33) think will help answer your question. (00:39:36) Um, what is the limit of knowledge? (00:39:39) Okay, so I'm not a brilliant person. (00:39:42) I've just learned a whole bunch of stuff (00:39:44) and so I run out of ideas and then I (00:39:46) don't know what to do. The truly (00:39:48) brilliant person will look at something. (00:39:50) This is how science works. It's how (00:39:52) biology works. the greatest inventors, (00:39:54) they see a pattern in one area and then (00:39:58) they're able to apply it in a completely (00:40:01) different area. So I'll give you an (00:40:03) example. Let's say that there's a prime (00:40:04) number component in one area and they (00:40:07) happen to notice that primes are present (00:40:08) in the other area and they can use the (00:40:10) same tools. I'm (00:40:12) oversimplifying. That's (00:40:14) genius. We don't today have the (00:40:17) algorithms to produce that and people (00:40:20) are working on it. So, the answer to (00:40:22) your question is I'm betting that we can (00:40:24) solve that problem. If we can't, then (00:40:26) we'll just be stuck with a computer in (00:40:29) your pocket that's as smart as the (00:40:32) smartest human ever lived, which is big (00:40:35) enough. Yeah, pretty good. Um, (00:40:37) microphone (00:40:40) over there. Jonathan Jacobs, uh, HCC. (00:40:44) There's been um a lot of discussion (00:40:45) today about uh data and its importance (00:40:48) for uh the advancement of biotech uh the (00:40:52) confluence of AI um and earlier uh you (00:40:56) were talking about uh what adversarial (00:40:59) risks there were uh in the sense that (00:41:02) you know you went all the way through to (00:41:03) you know reciprocally putting dynamite (00:41:06) in each other's data centers right so (00:41:07) I'm wondering if you can comment a bit (00:41:09) on the importance of uh data (00:41:12) authenticity and provenence. Uh, one of (00:41:15) the concerns that I have is the vast (00:41:17) majority of public data that's being (00:41:18) used to train these (00:41:20) models is originating from databases (00:41:22) which anyone in the world can contribute (00:41:25) content to. Uh, and there's often not (00:41:27) the providence behind how that data was (00:41:29) generated and that seems like a risk. I (00:41:32) I think the providence question that (00:41:33) you're asking is really fundamental. (00:41:36) um the AI systems can take noisy or bad (00:41:39) data and normalize it. But if the data (00:41:42) has been deliberately altered, which is (00:41:45) I think a way of thinking about it, (00:41:46) that's going to be a new kind of (00:41:47) national security risk. If you look in (00:41:50) biology, the core problem in biology is (00:41:52) we don't have enough data. We just don't (00:41:55) have enough data. There's so many (00:41:56) cellular processes and so you'll go into (00:41:58) a lab, which again we're busy defunding (00:42:00) for some reason, and and they're (00:42:03) generating that data. that data should (00:42:04) be for public use, reproducible, you (00:42:07) know, peer-reviewed and so forth and so (00:42:09) on. There's an interesting development. (00:42:12) Enthropic brought out a protocol called (00:42:14) the model context protocol. Um, and in (00:42:18) the last 3 months, it's been adopted by (00:42:20) every company. It was done in open (00:42:22) source and it basically if you have (00:42:24) data, it allows the model to just (00:42:26) structure the data in any way you want. (00:42:29) So you can literally say to the to to (00:42:31) the thing that talks to MCP to the (00:42:34) actual data, you can start asking it (00:42:36) questions. This has huge implications (00:42:38) because it means you don't have to build (00:42:39) all the data connectors. You can just (00:42:41) have the raw data and then the model is (00:42:44) smart enough to navigate the raw data to (00:42:46) answer your very sophisticated question. (00:42:48) So that's a big improvement and a big (00:42:50) deal that just happened in the last few (00:42:51) months. I think we have time for one (00:42:54) more. Um this table in the front is very (00:42:56) anxious to participate. microphone. (00:42:59) Yeah, we have a question back here. Um, (00:43:01) yeah, thanks for the Sorry, great (00:43:03) conversation so far. My name is Leonard (00:43:05) Justin. I'm a PhD student at MIT. Um, I (00:43:08) was wondering if you could just discuss (00:43:09) a bit more some of the risks you see (00:43:12) coming specifically with respect to (00:43:14) biology and how we should go about (00:43:16) mitigating those. What's the role of the (00:43:18) AI developers? What's the role of (00:43:20) government? Um, yeah, how can we move (00:43:22) forward on that? So, so you you're going (00:43:24) to know a lot more about this area than (00:43:25) I, but speaking as an amateur in your (00:43:28) field, the two current risks from these (00:43:31) models are cyber and biorisks. The cyber (00:43:34) ones are easy to understand. The system (00:43:36) can generate cyber attacks and in theory (00:43:38) can generate zeroday cyber attacks that (00:43:40) we can't see and it can unleash them and (00:43:43) furthermore it can do it at scale. In (00:43:45) biology, you get some evil, you know, (00:43:47) the equivalent of Osama bin Laden. They (00:43:49) would start with an open-source model. (00:43:52) Now these open source models have been (00:43:55) restricted using a testing process. Uh (00:43:58) they're called cards and they test it (00:44:00) out and they delete that information (00:44:02) from the model. It turns out it's (00:44:04) relatively easy to un to reverse (00:44:07) essentially those security modes around (00:44:10) the model and that's a danger. So now (00:44:14) you've got a model that can generate bad (00:44:15) pathogens. Then the second thing you (00:44:17) have to do is you have to find things to (00:44:19) build them. Our collective assessment at (00:44:22) the moment is that that's a nation state (00:44:25) risk, not an individual terrorist risk, (00:44:27) although we could be wrong. But there's (00:44:29) plenty of examples uh and this the the (00:44:32) report talks about some of the Chinese (00:44:35) examples where in theory if they wanted (00:44:38) to they could not only manufacture bad (00:44:41) things but sorry design them but also (00:44:44) manufacture them. The good news and the (00:44:46) reason we're all alive today is that the (00:44:48) bio stuff is hard to manufacture and (00:44:49) distribute and to make deadly and and (00:44:52) spread and so forth and so on. Um (00:44:54) there's lots of evidence for example (00:44:56) that you can take a bad bio right now (00:44:58) and modify it just enough that the (00:45:01) testing regimes and the sort of (00:45:03) surveillance regimes it bypasses and (00:45:06) that's another threat. So that's what I (00:45:07) worry about. But I think at the moment u (00:45:10) our consensus is we're right below the (00:45:13) threshold where this is an issue and the (00:45:16) consensus in in my side of the industry (00:45:18) is that one more or two more turns of (00:45:21) the crank these issues will be and you (00:45:24) know by then you'll be graduated and you (00:45:26) can sort of help solve these problems. (00:45:28) Um the a crank is turned every 18 months (00:45:32) or so about 3 years. But theoretically, (00:45:34) couldn't AI and biotechnology help you (00:45:37) come up with a counter measure? Um, I (00:45:40) had thought so and that was the argument (00:45:41) I made until I I do a lot of national (00:45:44) security work. And there's a term called (00:45:46) offense dominant. And an offense (00:45:49) dominant is a is a situation in a (00:45:51) military context where the attack cannot (00:45:55) be countered at the same level as the (00:45:58) attack. In other words, the damage is (00:45:59) done. And most people, most biologists (00:46:03) who've worked in this believe that while (00:46:05) the model can be trained to counter (00:46:07) this, the damage from the offense part (00:46:11) is far greater than the ability to (00:46:13) defend it, which is why we're so worried (00:46:14) about it. I hate to end this on such a (00:46:17) down note. So, I'm going to ask you for (00:46:20) a positive outlook. um look down the (00:46:23) road and tell me how you think AI and (00:46:26) biotechnology are actually going to (00:46:28) change people's lives. Well, let's thank (00:46:31) the financial system, the hardware (00:46:33) people and so forth for allowing us to (00:46:35) build immense data centers with billions (00:46:38) and billions of dollars of hardware with (00:46:39) no clear revenue purpose. So, thank you (00:46:42) very much. (00:46:43) Um, what will happen as all of that (00:46:46) stuff gets deployed and it's coming out (00:46:48) is it's going to be used by incredibly (00:46:50) clever people to solve some of these (00:46:52) problems, and I'm not talking about the (00:46:53) policy problems. I'm talking about the (00:46:54) actual underlying problems. You're going (00:46:56) to end up with these huge databases of (00:46:58) information we don't need, which we (00:47:00) don't have now. Excuse me, we do need. (00:47:02) And an example would be an example would (00:47:04) be that we still cannot do a digital (00:47:07) model of a cell. Seems like a kind of a (00:47:09) basic thing. I I was talking to my (00:47:11) biology friends. is like, "What's wrong (00:47:12) with you? You've been studying cells for (00:47:14) like 5,000 years." The actual number is (00:47:17) like 150. Um, what's wrong with you? And (00:47:20) the answer is it's really hard. Um, (00:47:22) we're pretty close to being able to do (00:47:24) that. That unlocks huge medical science, (00:47:29) huge drug possibilities. The language (00:47:31) that cells talk to each other. I happen (00:47:33) to be the chairman of the Broad (00:47:34) Institute. This is a big project at the (00:47:36) Broad. We're just on the cusp of that. (00:47:39) When you talk to the scientists, they're (00:47:41) using AI to generate it from the (00:47:43) scientists are in charge and AI is (00:47:45) helping them, which is the correct (00:47:46) order.