The Future Of AI, According To Former Google CEO Eric Schmidt (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:03) the key thing that's going on now is (00:00:05) we're moving very quickly through the (00:00:08) capability ladder steps and I think (00:00:11) there are roughly three things going on (00:00:13) now that are going to profoundly change (00:00:15) the world very quickly and when I say (00:00:18) very quickly the cycle is roughly a new (00:00:21) model every year to 18 months the first (00:00:26) is basically this question of context (00:00:29) window and for non-technical people the (00:00:32) context window is the prompt that you (00:00:34) ask so you know study John F Kennedy or (00:00:36) something right but in fact that context (00:00:39) window can have a million words in it (00:00:42) and this year people are inventing a (00:00:44) context window that is infinitely long (00:00:47) and this is very important because it (00:00:50) means that you can take the answer from (00:00:52) the system and feed it in and ask it (00:00:55) another (00:00:56) question so I want a recipe let's say I (00:00:59) want a recipe to make a drug or (00:01:00) something so I say what's the first step (00:01:02) and it says buy these materials so then (00:01:05) you say okay I've bought these materials (00:01:07) now what's my next step and then it says (00:01:09) buy a mixing pan and then the next step (00:01:11) is how long do I mix it for you see it's (00:01:13) a recipe that's called Chain of Thought (00:01:16) reasoning and it generalizes really well (00:01:19) we should be able in five years for (00:01:22) example to be able to produce a thousand (00:01:24) step recipes to solve really important (00:01:26) problems in science in medicine in (00:01:30) Material Science climate change that (00:01:32) sort of thing that's the first one (00:01:34) second one is Agents an agent can be (00:01:38) understood as a large language model (00:01:40) that knows something new or has learned (00:01:42) something so an example would be um read (00:01:46) all of chemistry learn something about (00:01:50) chemistry have a bunch of hypothesis (00:01:52) about chemistry run some tests in a lab (00:01:56) about chemistry and then add that to (00:01:58) your agent these agents are going to be (00:02:01) really powerful and it's reasonable to (00:02:03) expect that agents will be not only will (00:02:06) there be a lot of them and I mean (00:02:07) Millions but there'll be like the (00:02:08) equivalent of GitHub for agents there'll (00:02:10) be lots and lots of Agents running (00:02:12) around and available to you and the (00:02:14) third one which to me is the most (00:02:15) profound which is already beginning to (00:02:18) happen is text to action and what that (00:02:20) is is write me a piece of software to do (00:02:24) something right you just say it and can (00:02:26) you imagine having programmers that (00:02:28) actually do what you say you want (00:02:30) and it does it 24 hours a day and (00:02:33) strangely these systems are good at (00:02:34) running writing codes such as language (00:02:37) like python you put all that together (00:02:40) and you've got infinite context window (00:02:41) the ability for agents and then the (00:02:43) ability to do this programming now this (00:02:45) is very (00:02:47) interesting what then (00:02:49) happens there's a lot of questions here (00:02:52) and now we get into the questions of (00:02:54) Science Fiction I'm sure the three (00:02:56) things I've named are happening because (00:02:58) that work is happening now but it's some (00:03:00) point these systems will get powerful (00:03:03) enough that you'll be able to take the (00:03:05) agents and they'll start to work (00:03:07) together right so your agent and my (00:03:10) agent and her agent and his agent will (00:03:12) all combine to solve a new problem at (00:03:15) some point people believe that these (00:03:17) agents will develop their own language (00:03:20) and that's the point when we don't (00:03:23) understand what we're doing you know (00:03:25) what we should do pull the plug (00:03:29) literally unplug the computer so it's (00:03:33) really a problem when agents start to (00:03:36) communicate in ways and doing things (00:03:38) that we as humans do not understand (00:03:41) that's the limit in my view and you (00:03:44) think again how how far off in the (00:03:46) future well there have been many many (00:03:48) predictions clearly agents and these (00:03:50) things will occur in the next few years (00:03:53) and it won't occur in like there won't (00:03:55) be one day where everybody says oh my (00:03:57) God it's more a question of capab ities (00:04:00) every month every 6 months and so forth (00:04:03) a reasonable expectation is we'll be in (00:04:05) this new world within 5 years wow not 10 (00:04:10) and the reason is there's so much money (00:04:13) and not there are also so many ways in (00:04:15) which people are trying to accomplish (00:04:17) this you have the big gu guys the the (00:04:19) three large so-called Frontier models (00:04:21) but you have a very large number of (00:04:23) players who are programming at one level (00:04:26) lower at much lesser lower cost who are (00:04:29) iterating very quickly (00:04:30) plus you have a great deal of research I (00:04:32) think there's every reason to think that (00:04:35) some version of what I'm saying will (00:04:37) occur within 5 years and maybe sooner (00:04:40) well now so you say pull the plug so two (00:04:43) questions so how do you pull the plug (00:04:46) but even before you pull the plug if you (00:04:48) know you're already in Chain of Thought (00:04:49) reasoning and you're headed to what you (00:04:53) fear don't you need to regulate at some (00:04:57) point that it doesn't get there or is (00:04:59) that beyond the scope of Regulation well (00:05:03) a group of us have been working very (00:05:05) closely with the governments in the west (00:05:07) and we've started talking to the Chinese (00:05:08) which of course is complicated and takes (00:05:11) time uh about these issues and at the (00:05:14) moment the governments with the (00:05:16) exception of Europe which is always kind (00:05:17) of slightly confused have been doing the (00:05:20) right thing which is they've set up (00:05:22) trust and safety institutes they're (00:05:24) beginning to learn how to measure things (00:05:27) and check things and the right approach (00:05:30) is for the governments to watch us and (00:05:33) make sure we don't get confused on what (00:05:35) the goal is right so as long as the (00:05:38) companies are well-run Western companies (00:05:42) with shareholders and lawsuits and all (00:05:44) that we'll be fine there's a great deal (00:05:46) of concern in these Western companies (00:05:48) about liability doing bad things nobody (00:05:50) wants to hurt people they're not they (00:05:52) don't wake up in the morning saying (00:05:53) let's hurt somebody right now of course (00:05:56) there's the proliferation problem yeah (00:05:58) but in terms of the core research the (00:06:00) researchers are trying to be (00:06:02) honest okay so that's the West so by (00:06:04) saying the West you're implying that (00:06:06) proliferation outside the West is where (00:06:08) the danger is the bad guys are out there (00:06:10) somewhere well one of the things that we (00:06:13) know and it's always useful to remind (00:06:16) the Techno optimists in my world there (00:06:19) are evil people and they will use your (00:06:21) tools to hurt people my favorite example (00:06:25) is that the face recognition stuff was (00:06:27) invented not to constrain the Wagers (00:06:30) you know they didn't say we're going to (00:06:31) invent face recognition in order to (00:06:33) constrain this the minority in China (00:06:36) called the Wagers right but it's (00:06:39) happening all technology is dual use all (00:06:42) of these inventions can be misused and (00:06:45) it's important for the inventors to be (00:06:47) honest with that so in open source which (00:06:51) is for those of you who don't follow it (00:06:54) open source is where the source code in (00:06:56) in models the weights that is the (00:06:58) numbers that have been calculated are (00:07:00) released to the public those immediately (00:07:03) go throughout the world and who do they (00:07:04) go to they go to China of course they go (00:07:07) to Russia they go to Iran right they go (00:07:10) to bellaria they go to North Korea yeah (00:07:13) uh when I was most recently in China the (00:07:15) vast essentially all of the work I saw (00:07:18) started with open- Source models from (00:07:20) the West which were then (00:07:22) Amplified so it sure looks to me like (00:07:26) these leading firms the ones I'm talking (00:07:28) about the ones that are putting 10 bill (00:07:30) you know a billion 10 billion dollar (00:07:32) eventually into this will be tightly (00:07:35) regulated I worry that the rest will not (00:07:38) you can see I'll give you another (00:07:39) example look at this problem of (00:07:41) misinformation I think it's largely (00:07:43) unsolvable and the reason is the code (00:07:46) generate (00:07:48) misinformation is essentially free right (00:07:51) any any you know person right a good (00:07:55) person a bad person has access to them (00:07:57) it doesn't cost anything and they (00:07:58) produce very very good (00:08:00) images uh there are regulatory solutions (00:08:03) to that but the important point is that (00:08:05) that cat is out of the bag or whatever (00:08:07) metaphor you want it's important that (00:08:09) these more powerful (00:08:11) systems especially as they get closer to (00:08:14) general intelligence have some limits on (00:08:17) proliferation and that problem is not (00:08:19) yet solved yet to follow up on on your (00:08:22) point about the funding Faith Lee at (00:08:25) Stanford argues that's the biggest (00:08:27) problem is that there's so much money (00:08:28) going into the private sector (00:08:30) and who's their competition to look at (00:08:32) what the red lines are or whatever it's (00:08:35) the universities which don't have a lot (00:08:38) of money um (00:08:40) so you really trust these companies to (00:08:43) be transparent enough to be regulated by (00:08:47) government that doesn't know what're (00:08:50) talking about really the correct answer (00:08:52) is always trust but verify yeah and the (00:08:56) truth is you should trust and you should (00:08:57) also verify and at least in the West the (00:09:00) best way to verify is to use private (00:09:02) companies that are set up as verifiers (00:09:05) because they can employ the right people (00:09:06) and so forth so in all of our industry (00:09:09) conversations it's pretty clear that the (00:09:12) way it will really work is you'll end up (00:09:14) with AI checking AI it's too hard think (00:09:20) about it you build a new model it's been (00:09:23) trained on new data you worked really (00:09:25) hard on it how do you know what it knows (00:09:27) yeah now by you can ask it all the (00:09:29) previous questions but what if it's (00:09:32) discovered something completely new and (00:09:33) you don't think about it right and the (00:09:36) systems can't regurgitate everything (00:09:38) they know you have to ask them chunk by (00:09:40) chunk by chunk so it makes perfect sense (00:09:42) that an AI would would be the only way (00:09:45) to police that people are working on (00:09:48) that with Fay's argument she's (00:09:49) completely correct we have the rich (00:09:52) Private Industry companies and we have (00:09:54) the poor universities who have (00:09:56) incredible Talent it should be an major (00:10:00) national priority in all of the Western (00:10:02) countries to get research funding for (00:10:05) the hardware if you were a um physicist (00:10:09) 50 years ago you had to move to where (00:10:11) the cyclon cyclotrons were because they (00:10:14) were really hard and expensive and by (00:10:15) the way they still are really hard and (00:10:17) inexpensive you need to be near a (00:10:18) cyclotron to do your work as a physicist (00:10:20) we never had that in software our stuff (00:10:23) was Capital cheap not Capital expensive (00:10:27) the arrival of heavyduty training in our (00:10:30) industry is a huge economic change and (00:10:34) what's happening is that companies are (00:10:36) figuring this out and the really rich (00:10:38) companies I'm thinking of Microsoft and (00:10:40) Google as an example are planning to (00:10:42) spend billions of dollars because they (00:10:45) have the cash they have big businesses (00:10:47) the money's coming in that's good where (00:10:50) does the Innovation come from they don't (00:10:51) have that kind of hardware and yet they (00:10:53) need access to that (00:10:55) yeah um okay let's go to China so uh you (00:10:59) just um you on Kissinger's last trip to (00:11:03) China you went with him and he had a (00:11:05) discussion with Luan Ping On exactly (00:11:09) this set of issue you your your idea was (00:11:12) to set up a high level group to discuss (00:11:15) the potential and catastrophic (00:11:17) possibilities of (00:11:19) AI (00:11:21) uh where do the Chinese fit in on this (00:11:24) on the one hand I've heard you say and (00:11:26) not only you that we need to go all out (00:11:28) to compete with the Chinese (00:11:30) uh for some of the reasons you just said (00:11:31) because there could be bad players or (00:11:34) bad intentions but where is it (00:11:36) appropriate to cooperate and (00:11:39) why well first place the Chinese should (00:11:42) be pretty worried about (00:11:44) generative Ai and the reason is that (00:11:47) they don't have um free speech and so (00:11:51) what do you do when the system generates (00:11:53) something that's not permitted in their (00:11:56) country right right who do you jail yeah (00:11:59) right (00:12:00) the computer the user the developer the (00:12:04) training (00:12:05) data it's not at all obvious and the (00:12:09) Chinese Regulators so far have been (00:12:11) relatively intelligent about this but (00:12:13) it's obvious if you think about it that (00:12:16) the spread of these things will be (00:12:17) highly restricted in China because it (00:12:20) fundamentally addresses their (00:12:21) information Monopoly right that makes (00:12:24) sense so in our conversation with China (00:12:27) both Dr Kissinger and I when we were (00:12:29) together (00:12:30) um and unfortunately he passed away and (00:12:32) the subsequent meetings have been set up (00:12:34) as a result of his inspiration to do (00:12:36) them everyone agrees that there's a (00:12:39) problem but we're at the moment with (00:12:41) China we're speaking in (00:12:43) generalities there is not a proposal in (00:12:46) front of either side that's actionable (00:12:49) and that's okay because it's complicated (00:12:52) and a lot of this because of the stakes (00:12:54) involved it's actually good to take your (00:12:57) time to actually explain what you view (00:12:59) as the problem so many Western computer (00:13:02) scientists are visiting with their (00:13:03) Chinese counterparts and trying to say (00:13:06) if you allow this stuff to to (00:13:09) proliferate you could end up with a (00:13:11) terrorist Act Right the misuse of these (00:13:14) for biological weapons the misuse of (00:13:16) these for cyber um the long-term worry (00:13:20) is is much more existential but at the (00:13:22) moment I think the Chinese conversations (00:13:25) are largely very constrained by bio by (00:13:27) concerns about biothreats and and uh (00:13:29) cyber threats the long-term threat goes (00:13:32) something like (00:13:34) this it's when I talk about AI I talk (00:13:37) about it as human generated so you or I (00:13:42) give it at least in theory a command and (00:13:46) you may it may be a very long command (00:13:48) and it may be recursive in the sense but (00:13:50) it starts with a human judgment right (00:13:54) there is something technically called (00:13:55) recursive self-improvement right where (00:13:58) the model actually runs on it own and it (00:14:00) just learns and gets smarter and smarter (00:14:03) right when that occurs or when agent to (00:14:07) agent interaction that's heterogeneous (00:14:09) occurs we have a very different set of (00:14:12) threats which we're not ready to talk to (00:14:14) anybody about because we don't (00:14:16) understand them but they're coming do (00:14:19) you see I guess I'm trying to think (00:14:20) about what a kind of dialogue with the (00:14:22) Chinese could mean would it be something (00:14:24) like nuclear proliferation I mean where (00:14:27) if they understand the existential (00:14:29) threat (00:14:30) to start at that level maybe an iaea (00:14:34) type of thing for proliferation do you (00:14:36) think that's possible on on the (00:14:38) political (00:14:40) Horizon it's going to be very difficult (00:14:42) to get any actual treaties with China um (00:14:45) what I'm engaged with is called a track (00:14:47) two dialogue which means that it's (00:14:49) informal it's not it's it's educational (00:14:52) it's interesting it's very hard to (00:14:54) predict by the time we get to real (00:14:56) negotiations between the US and China (00:14:59) yeah what the political situation will (00:15:01) be what the threat situ would be a (00:15:04) simple requirement would be that if (00:15:07) you're going to do training for (00:15:09) something that's completely new you have (00:15:11) to tell the other side that you're doing (00:15:13) it okay so that you don't surprise them (00:15:17) so it's like the open Skies during the (00:15:18) Cold War so so an example would be a no (00:15:21) surprises rule when a missile is (00:15:23) launched anywhere in the world all the (00:15:25) countries acknowledge that they know (00:15:28) it's coming that way they don't jump to (00:15:30) a conclusion and think it's targeted at (00:15:32) them that strikes me as a basic rule (00:15:35) right furthermore that if you're doing (00:15:37) powerful training there needs to be some (00:15:40) agreements around safety um in biology (00:15:43) there's a broadly accepted set of layers (00:15:46) BSL one to four right for bios safety (00:15:49) containment which makes perfect sense (00:15:51) because these things are dangerous (00:15:53) eventually there will be a small number (00:15:55) of extremely powerful computers that I (00:15:59) want you to think about they'll be in an (00:16:01) army base and they'll be powered by a (00:16:04) some nuclear power source in the army (00:16:06) base and they'll be surrounded by even (00:16:09) more barred wire and machine guns (00:16:11) because their capability for invention (00:16:15) for power and so forth exceeds what we (00:16:18) want as a nation to give either to our (00:16:21) own citizens without permission as well (00:16:23) as to our competitors makes sense to me (00:16:26) that there will be a few of those and (00:16:28) there'll be a lot of other system (00:16:29) systems that are more broadly available (00:16:31) but you're saying that you would notify (00:16:33) the Chinese that those systems exist yep (00:16:37) again it's possible that that would be (00:16:40) an answer and vice versa and vice versa (00:16:43) all all of these things are mutual but (00:16:46) the you want to avoid a situation where (00:16:49) a runaway agent in China ultimately gets (00:16:54) access to a weapon and launches it (00:16:57) foolishly thinking that that's some game (00:17:00) without because remember these are not (00:17:01) human humans they don't necessarily (00:17:04) understand the consequence these systems (00:17:06) are all based on a simple principle of (00:17:08) predicting the next word right so we're (00:17:10) not talking about High Intelligence here (00:17:12) we're certainly not talking about the (00:17:14) kind of emotional understanding and (00:17:16) history that that humans have and human (00:17:18) values so when you're dealing with a a (00:17:21) non-human intelligence that does not (00:17:23) have the benefit of Human Experience (00:17:27) what bounds do you put on it and maybe (00:17:29) we can come to some agreements on what (00:17:31) those are are they moving as (00:17:35) exponentially as we are in the west with (00:17:38) the billions going into generative (00:17:40) AI uh is trying to have the commensurate (00:17:43) billions coming in from government or (00:17:48) companies it's not at the same level in (00:17:50) China for reasons I don't fully (00:17:52) understand my estimate having now (00:17:54) reviewed it at some length is that (00:17:56) they're about two years behind two years (00:17:59) is not not very much by the way but (00:18:00) they're definitely behind there are at (00:18:03) least four companies that are attempting (00:18:05) to do large scale model training which (00:18:07) is similar to what I've been talking (00:18:09) about um and they're the obvious big (00:18:11) tech companies in China right they're (00:18:14) hobbled because they don't have access (00:18:16) to the very best (00:18:17) hardware um which is restricted from (00:18:20) export by the Trump and now Biden (00:18:22) administrations those restrictions are (00:18:24) likely to get tougher not easier and so (00:18:27) as the Nvidia and their competitor chips (00:18:30) go up in value China will be struggling (00:18:32) to stay relevant right because their (00:18:35) stuff won't move at the same Chinese you (00:18:38) agree with not letting those chips flow (00:18:42) China the the chips the chips are (00:18:46) important because they enable this kind (00:18:48) of learning it's always possible to do (00:18:50) it with slower chips you just need more (00:18:52) of them and so it's effectively a cost (00:18:56) tax um for for Chinese development (00:18:59) that's a way to think about it and Is It (00:19:02) ultimately dispositive does it mean that (00:19:04) China can't get there no but it makes it (00:19:06) harder and makes it means that it takes (00:19:08) them longer to do so and we should do (00:19:11) that as the West Well the West has (00:19:12) agreed to do it I think it's fine yeah (00:19:15) uh it's a fine strategy I'm I'm much (00:19:17) more (00:19:18) concerned about the proliferation of (00:19:20) Open Source and the reason is and I'm (00:19:23) sure the Chinese would have the same (00:19:25) concern so again these are the kinds of (00:19:26) things that we'll be talking to them (00:19:28) about is do you understand that these (00:19:30) things can be misused against your (00:19:32) government as well as ours so the (00:19:33) scenario is open source folks basically (00:19:37) do something called basically guard (00:19:39) rails and they fine-tune and they use a (00:19:41) technology called rhf to eliminate some (00:19:44) of the bad answers there's plenty of (00:19:47) evidence that it's relatively easy if I (00:19:50) gave you all of the weights all of the (00:19:52) stuff so forth it' be relatively easy (00:19:54) for you to back them out and see the raw (00:19:57) power of the model and that's a great (00:19:59) concern that's problem's not been solved (00:20:01) engineering yeah reverse engineer and (00:20:02) that's not been solved yet