Ex-OpenAI Scientist WARNS: “You Have No Idea What’s Coming” (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) You may not take interest in politics, (00:00:03) but politics will take interest in you. (00:00:06) So the same applies to AI many times (00:00:07) over. (00:00:08) >> Ilas Sutskver, the man behind the (00:00:10) invention of open AI, gave a pretty (00:00:12) strong speech at the University of (00:00:14) Toronto. He expressed great concerns (00:00:16) about the upcoming AI, which might (00:00:18) disrupt our entire world. Watch this. (00:00:21) >> The reason it's not going to be the most (00:00:23) conventional convocation speech is (00:00:25) because there is something a little (00:00:26) different going on right now. (00:00:30) You all leave, we all leave in the most (00:00:33) unusual time ever. And this is something (00:00:36) that people might say often, but I think (00:00:37) it's actually true this time. And the (00:00:39) reason it's true this time is because of (00:00:42) AI, right? Obviously, I mean, from what (00:00:47) I hear, (00:00:49) the AI of today has already changed what (00:00:52) it means to be a student by a pretty (00:00:55) considerable degree. (00:00:58) That's (00:01:00) I uh especially I that's what I I sense (00:01:04) and I think it's true. But of course the (00:01:06) impact of AI goes beyond that. (00:01:10) What happens to the kind of work we do? (00:01:12) Well, it's starting to change a little (00:01:14) bit in some unknown and unpredictable (00:01:16) ways. (00:01:17) And (00:01:19) some some some work may feel it sooner. (00:01:22) Some work might feel it later. With (00:01:24) today's AI, you can go on on uh on (00:01:26) Twitter and you can look at what AI can (00:01:28) do and what people say and you might (00:01:30) feel a little bit of that. You wonder, (00:01:32) hey, which skills are useful? Which ones (00:01:34) will be less useful? So, you got these (00:01:36) questions going on. And so, you can say (00:01:38) that the current level of challenge is (00:01:44) how will it affect (00:01:47) work and our careers. (00:01:50) But the thing the real challenge with AI (00:01:52) is that is really unprecedented and (00:01:54) really extreme and it's going to be very (00:01:57) different in the future compared to the (00:01:59) way it is today. Like you know we've all (00:02:01) seen AI, we've all spoken to a computer (00:02:03) and a computer has spoken back to us (00:02:05) which is a new thing. Computers would (00:02:07) not do this in the past but now they do. (00:02:10) So you speak to a computer and it (00:02:12) understands you and it speaks back to (00:02:13) you and it also does it in voice and it (00:02:15) writes some code. It's it's pretty (00:02:17) crazy, but there are so many things it (00:02:20) cannot do as well and it's so deficient. (00:02:22) So, you can say it still needs to catch (00:02:24) up on a lot of things, but (00:02:28) it's evocative. (00:02:30) It's good enough that you can ask (00:02:32) yourself, you could imagine, okay, fine, (00:02:35) in some number of years, some people say (00:02:37) it's in three, some people say it's in (00:02:39) five, 10. Numbers are being thrown (00:02:41) around. It's a bit hard to predict the (00:02:43) future, but (00:02:46) slowly but surely or maybe not so (00:02:48) slowly, AI will keep getting better. And (00:02:50) the day will come when AI will do all of (00:02:53) our all the things that we can do. Not (00:02:56) just some of them, but all of them. (00:02:57) Anything which I can learn, anything (00:02:59) which any any one of you can learn, the (00:03:01) AI could do as well. How do we know (00:03:04) this? By the way, how can I be so sure? (00:03:06) How can I be so sure of that? The reason (00:03:09) is (00:03:10) that all of us have a brain and the (00:03:14) brain is a biological computer. (00:03:17) That's why we have a brain. The brain is (00:03:19) a biological computer. So why can't a (00:03:22) digital computer, a digital brain do the (00:03:24) same things? This is the one sentence (00:03:26) summary for why AI will be able to do (00:03:28) all those things because we have a brain (00:03:30) and a brain is a biological computer. (00:03:32) And so you can start asking yourselves (00:03:34) what's going to happen. What's going to (00:03:36) happen when computers can do all of our (00:03:38) jobs? Right? Those are really big (00:03:39) questions. Those are dramatic questions. (00:03:41) And right now, like you start thinking (00:03:43) about it a little bit, you go, gosh, (00:03:44) that's a little intense. But it's (00:03:46) actually only part of the intensity (00:03:48) because what's going to happen? (00:03:52) What what will we the collective V want (00:03:55) to use these AIs for? Do more work, grow (00:03:59) the economy, do R&D, do AI research. So (00:04:03) then the rate of progress will become (00:04:05) really extremely fast for some time at (00:04:08) least. These are such extreme things. (00:04:10) These are such unimaginable things. So (00:04:12) right now I'm trying to pull you into (00:04:14) that a little bit into this headsp space (00:04:16) of this really extreme and radical (00:04:19) future that the AI creates. But it's (00:04:21) also very difficult to imagine. It's (00:04:23) very very difficult to imagine. It's (00:04:24) very difficult to internalize and to (00:04:26) really believe on an emotional level. (00:04:27) Even I struggle with it. And yet the (00:04:30) logic seems to dictate that this very (00:04:34) likely should happen. (00:04:37) So (00:04:40) what does one do in such a world? You (00:04:43) know there is a quote which is like this (00:04:46) uh (00:04:48) uh which goes like this. It says you may (00:04:51) not take interest in politics but (00:04:53) politics will take interest in you. (00:04:55) So the same applies to AI many times (00:04:57) over. And in particular, I think that by (00:05:02) simply using AI and looking at what the (00:05:05) best AI of today can do, you get an (00:05:08) intuition. You get an intuition. And as (00:05:11) AI continues to improve in one year, in (00:05:14) two years, in three years, the intuition (00:05:16) will become stronger. And a lot of the (00:05:18) things that you're talking about now, (00:05:21) they will become much more real. they'll (00:05:24) become less imaginary. In the end of the (00:05:26) day, no amount of essays and and (00:05:28) explanations can (00:05:31) can compete with what we see with our (00:05:33) own senses, with our own two eyes. And (00:05:35) especially with AI, the very smart, (00:05:38) super intelligent AI in the future, (00:05:40) there will be very profound issues about (00:05:43) making sure that they are they say what (00:05:45) they say and not pretend to be something (00:05:48) else. And I'm really condensing a lot (00:05:50) into a small amount of information here (00:05:53) in time here. But overall, by simply (00:05:57) looking at what AI can do, not ignoring (00:06:00) it when the time comes, that will (00:06:02) generate the energy that's required to (00:06:05) overcome the huge challenge that AI will (00:06:07) pose. And the challenge that AI poses in (00:06:09) some sense is the greatest challenge of (00:06:12) humanity ever. and (00:06:16) overcoming it will also have the will (00:06:18) also bring the greatest reward (00:06:21) and in some sense whether you like it or (00:06:24) not your life is going to be affected by (00:06:25) AI to a great extent and so looking at (00:06:28) it paying attention and then generating (00:06:31) the energy to solve the problems that (00:06:33) will come up that's going to be the main (00:06:36) thing (00:06:37) >> so that was Ilia's view but to (00:06:39) understand the full debate I want you to (00:06:41) watch this interview clip of Eric (00:06:42) Schmidt where he talks about a much (00:06:44) broader impact of AI on human lives in (00:06:47) the coming years. Watch this. (00:06:48) >> Okay. So, we believe as an industry that (00:06:51) in the next one year, the vast majority (00:06:54) of programmers will be replaced by AI (00:06:57) programmers. (00:06:58) We also believe that within one year, (00:07:01) you will have graduate level (00:07:03) mathematicians that are at the tippy top (00:07:05) of graduate math programs. There's lots (00:07:08) of reasons to think this is going to (00:07:09) happen. This is the consensus. You go, (00:07:11) okay, well, that's pretty interesting. (00:07:13) Now, I can't do that kind of math. Very (00:07:16) few people can do that math. How can the (00:07:19) computer do that math better than (00:07:20) anybody else? To some degree, it's (00:07:23) because math has a simpler language than (00:07:25) human language. So, the way these (00:07:27) algorithms actually work is they're (00:07:29) doing essentially word prediction. So, (00:07:31) you take you take a a sentence, you take (00:07:33) a word out, and then it learns how to (00:07:35) put the correct word back in. This is (00:07:37) called the loss function, and it's (00:07:38) optimized to do that at a scale that's (00:07:40) unimaginable to us as humans. (00:07:43) So you do the same thing for math, but (00:07:45) there you use a conjecture and then a (00:07:47) proof format through a protocol called (00:07:49) lean. In programming, it's pretty (00:07:51) simple. You just keep writing code until (00:07:54) you pass the programming test. (00:07:56) So strangely, the first question I (00:07:58) always ask programmers is what language (00:07:59) do you program in? And the correct (00:08:01) answer is it doesn't matter because (00:08:03) you're trying to design for an outcome. (00:08:05) You don't care what code is generated by (00:08:06) the computer. This is a whole new world. (00:08:09) Okay. So that's one year. Okay, what (00:08:13) happens in two years? Well, I've just (00:08:16) told you about reasoning and I've told (00:08:17) you about programming and I've told you (00:08:19) about math. Programming plus math are (00:08:21) the basis of sort of our whole digital (00:08:23) world. So, the evidence and the claims (00:08:26) from the research groups in OpenAI and (00:08:29) and anthropic and so forth is that (00:08:31) they're now somewhere around 10 or 20% (00:08:34) of the code that they're developing in (00:08:36) their research programs is being (00:08:38) generated by the computer. (00:08:41) That's called recursive self-improvement (00:08:43) is the technical term. So what happens (00:08:46) when this thing starts to scale? Well, a (00:08:49) lot. (00:08:51) One way to say this is that within three (00:08:54) to five years, we'll have what is called (00:08:56) general intelligence, AGI, which can be (00:08:59) defined as a system that is as smart as (00:09:01) the smartest mathematician, physicist, (00:09:05) you know, artist, writer, thinker, (00:09:07) politician, maybe not in the same level. (00:09:10) Um, but you get the idea. Uh, just the (00:09:13) creative industries and so forth. But (00:09:15) imagine that in one computer. Okay. (00:09:17) Well, that's pretty interesting. I call (00:09:18) this, by the way, the San Francisco (00:09:20) consensus because everyone who believes (00:09:21) this is in San Francisco. It may be the (00:09:24) water. (00:09:25) What happens when every single one of us (00:09:29) has the equivalent of the smartest human (00:09:32) on every problem in our pocket? So, it (00:09:34) means you have the best architect when (00:09:35) you have an architecture problem. (00:09:37) Another thing that's going on is the (00:09:39) development of agentic solutions and (00:09:40) agents are refer to systems that have (00:09:43) input and output in memory and they (00:09:45) learn. An example here is that I want to (00:09:48) uh buy another house. Uh I happen to (00:09:50) like Virginia. I grew up in Virginia. I (00:09:52) say find me a house in the greater MLAN (00:09:55) area. Look at the that's one agent. Look (00:09:57) at all the rules. Figure out how big a (00:09:59) house I can build. That's another agent. (00:10:02) Do the transaction to buy the land. (00:10:04) That's another agent. design the house (00:10:06) with a human architect, right? But sort (00:10:09) of ignore them for most of the thing, (00:10:10) but they have to sign it off and then I (00:10:13) approve it and then find the contractor, (00:10:16) right? Hire the contractor, pay the (00:10:18) bills, and at the end sue the contractor (00:10:20) for lack of performance. (00:10:23) Okay? Now, I just gave you the stupidest (00:10:26) possible explanation. I just described (00:10:29) every business process, every government (00:10:32) process, and every and every sort of (00:10:34) academic process in our nation. (00:10:37) >> So, it isn't just the programmers that (00:10:38) are going to be out of work. We're all (00:10:40) going to be out of work. (00:10:41) >> No, that's not a consequence. I'll come (00:10:43) to that. But, but the reason I want to I (00:10:45) want to make the point here is that in (00:10:47) the next year or two, this foundation is (00:10:50) being locked in and it's not we're not (00:10:52) going to stop it. (00:10:55) gets much more interesting after that (00:10:58) because remember the computers are now (00:11:00) doing self-improvement. They're learning (00:11:02) how to plan and they don't have to (00:11:05) listen to us anymore. We call that super (00:11:08) intelligence or ASI artificial super (00:11:10) intelligence. And this is the theory (00:11:13) that there will be computers that are (00:11:15) smarter than the sum of humans. The San (00:11:18) Francisco convent consensus is this (00:11:20) occurs within six years just based on (00:11:22) scaling. Now, in order to pull this off, (00:11:26) you have to have an enormous amount of (00:11:29) power. I was here yesterday testifying (00:11:32) about this, you know, and we need like I (00:11:35) can talk at some length about how many (00:11:37) gigawatts and how many nuclear power (00:11:39) plants and all the kind of stuff we can (00:11:40) talk about separately. (00:11:42) This path is not understood in our (00:11:46) society. There's no language for what (00:11:48) happens with the arrival of this. I (00:11:50) wrote a book on this with Henry (00:11:51) Kissinger called Genesis which you know (00:11:53) I recommend obviously um because I wrote (00:11:55) it (00:11:56) >> available (00:11:57) >> available available in your usual places (00:12:00) um but the important point is this is (00:12:02) happening faster than our human that our (00:12:05) society our democracy our laws will (00:12:07) address and there's lots of implications (00:12:10) that's why it's underhyped people do not (00:12:12) understand what happens when you have (00:12:15) intelligence at this level which is (00:12:17) largely free that's the How do we get (00:12:20) ready for it? (00:12:22) >> Well, we start by talking about it. And (00:12:24) by the way, on the jobs thing, everyone (00:12:25) assumes that automation will replace (00:12:27) will eliminate jobs. If you look at the (00:12:29) history of automation ever since the the (00:12:32) looms and uh in uh 300 years ago, the (00:12:37) jobs are changed, but more jobs are (00:12:39) created than destroyed. In this case, (00:12:42) you'd have to convince me that this time (00:12:44) is different. If you look in Asia where (00:12:47) they for whatever reason are choosing (00:12:49) not to have children, the Asian (00:12:52) reproduction rate is in the order of 1.0 (00:12:54) or lower. So they're rapidly (00:12:57) disappearing. So the Asian countries are (00:13:00) very very quickly automating. The tools (00:13:02) that I'm describing will allow the few (00:13:05) humans that will be working very hard in (00:13:08) 30 or 40 years. If these trends (00:13:10) continue, the rest of us will be (00:13:12) dependent on those hardworking humans. (00:13:14) It'll make their productivity more much (00:13:16) greater. (00:13:17) >> Now, here's another clip of Eric (00:13:19) Schmidt. Here he shares much deeper (00:13:21) concerns about upcoming AI technology. (00:13:24) Watch this. (00:13:25) >> One way to think about the AI that you (00:13:27) all know is think of it as language to (00:13:30) language. You ask a question, the answer (00:13:32) comes back. You ask a question, it can (00:13:34) even write code. Nowadays, the models (00:13:37) are multimodal. So, for example, you can (00:13:39) take a picture and say, tell me what's (00:13:41) in this picture. (00:13:43) uh technically there are APIs which (00:13:45) allow uh one firm to call an open AAI or (00:13:48) Gemini API or anthropic for etc and do (00:13:53) the classification of the picture and so (00:13:55) forth. These are all tactics that (00:13:57) increase the intelligence of the (00:13:58) underlying system. There are three (00:14:00) things going on right now this year. So (00:14:04) less than the time frame you gave are (00:14:06) really interesting. One is called (00:14:07) infinite context windows. Infinite (00:14:10) context men windows means that you can (00:14:12) keep feeding the answer back in is the (00:14:14) question. So it allows you to do (00:14:16) stepby-step planning. You know, how do I (00:14:19) uh how do I build a house? Well, the (00:14:21) first is I have to find a contractor. I (00:14:22) found a contractor. What do I have to (00:14:24) talk to them? Then I have to have an (00:14:25) architect. How do I find an architect? (00:14:27) Then I have to tell the architect what (00:14:28) to do and then design me the house. I'll (00:14:29) give it to the architect. He can (00:14:31) redesign it. You know, it's a series of (00:14:32) steps. Uh the next one are called (00:14:36) agents. (00:14:38) And agents is a generally overused term (00:14:42) and most people think that agents will (00:14:44) essentially act as memory sources. So an (00:14:46) agent can be understood as it's watching (00:14:48) something and when it sees it, it takes (00:14:50) an action. It does that by knowing what (00:14:52) to do based on what it's seen. The specs (00:14:56) for how agents work are completely (00:14:58) undefined in the industry. The dominant (00:15:00) companies want to have their own agents (00:15:03) and they don't want the agents to (00:15:04) interact because they want control for (00:15:05) obvious reasons. Many people think that (00:15:08) there will be an agent store that you (00:15:10) will download like we see with apps but (00:15:12) not this year. And the third one is text (00:15:15) to text to code. Now I don't know about (00:15:17) you all but I've programmed and managed (00:15:19) programmers for more than 40 years and (00:15:21) they never do what I want. So can you (00:15:23) imagine if the computer you said write (00:15:25) me a program to do this and it actually (00:15:27) writes the code. In our case, uh the (00:15:30) program would be (00:15:33) search through all the literature, find (00:15:34) out who is working on energy policy, who (00:15:37) has a technological background or a role (00:15:39) in which they have to be technologically (00:15:41) liberate literate. Identify them, rank (00:15:43) them, score them based on whatever our (00:15:45) goal is. Um and and then automatically (00:15:49) invit send them an invitation. If they (00:15:52) say yes, say congratulations. If they (00:15:53) say no, why not? and call them and with (00:15:56) a synthetic voice tell them that they're (00:15:57) idiots for not coming. That's the kind (00:15:59) of program I would write. Thank god I'm (00:16:01) not doing that. But but you see how easy (00:16:03) it would be to automate tasks. So that's (00:16:06) I think the first step. The next step is (00:16:09) not as clear. There are uh there's sort (00:16:13) of huge contest um there's a huge set of (00:16:16) contests going on now which are at a (00:16:18) scale that's unimaginable. You have the (00:16:20) big three in the US. Anthropic uh which (00:16:22) is allied with Amazon, Gemini obviously (00:16:25) from Google, OpenAI, Microsoft and let's (00:16:28) assume they all do really well. It looks (00:16:29) that they're doing really well. I can (00:16:30) talk about what their problems are but (00:16:32) fundamentally they're they're doing (00:16:33) well. You have Facebook which has chosen (00:16:36) open- source path for the 400 billion (00:16:39) model that has a lot of implications (00:16:41) strategically, right? Which we can (00:16:43) discuss. um all of these are vying for (00:16:46) the best reasoning, the best answers and (00:16:48) then the best predictive analytics, the (00:16:50) best image classifiers and the best (00:16:52) multimodal. (00:16:54) That technology then diffuses or the (00:16:56) technical term is distilled into more (00:16:58) specialized models. And I think that's (00:17:00) the action you'll see in the next one to (00:17:02) two years. You did not mention (00:17:04) artificial general intelligence. First, (00:17:07) for those of us who aren't necessarily (00:17:09) um totally up to speed on AI, what is it (00:17:13) and where are we? (00:17:14) >> There are multiple definitions of AGI, (00:17:16) but the it's the term has been around (00:17:18) for 15 years. The basic idea is what is (00:17:21) the point where you have the flexibility (00:17:23) of a human in your intelligence system. (00:17:25) So, one way to understand it today is (00:17:27) that we these are called narrow AI (00:17:29) approaches, although they're not (00:17:30) certainly not narrow. you basically (00:17:32) they're they're initiated by a human at (00:17:36) what point is the question can the (00:17:38) computer generate its own objective (00:17:41) function its own goal and how will that (00:17:44) emerge uh the there's what I call the (00:17:47) San Francisco school because they're all (00:17:49) in San Francisco uh which is a separate (00:17:52) set of issues and they all talk to each (00:17:54) other and they've all convinced (00:17:55) themselves that if within two to three (00:17:57) cranks of the systems the crank is about (00:17:59) 18 months you get to AGI (00:18:02) And they define AGI as intelligence (00:18:04) greater than the sum of human (00:18:06) intelligence. I personally think that (00:18:08) that's likely but not in three years, (00:18:11) not in (00:18:12) >> what is the time frame, do you? We don't (00:18:13) know.