PodClips Logo
PodClips Logo
Exponential View with Azeem Azhar
How GPT-3 Is Shaping Our AI Future
How GPT-3 Is Shaping Our AI Future

How GPT-3 Is Shaping Our AI Future

Exponential View with Azeem AzharGo to Podcast Page

Azeem Azhar, Sam Altman
·
29 Clips
·
Oct 7, 2020
Listen to Clips & Top Moments
Episode Summary
Episode Transcript
0:01
Hbr presents.
0:12
Hi there. I'm as a bizarre and you're listening to the first episode in season 5 of the exponential view podcast. This is the place where every week I come together with a brilliant mind to discuss the forces that are shaping our near future. For those of you new to this podcast, let me introduce myself. I'm a Serial founder and investor in Tech and sustainability, startups, and an advisor to some amazing companies. I am fascinated by the
0:36
Exponential Technologies on our society politics and business. Now, I profoundly believe that if we're to enjoy the benefits of these Technologies and to mitigate their worst effects conversations, like the one you're about to hear are key this clear respectful and thoughtful communication does become the Bedrock for our civilization who put simply technology is too important to be left to technologists. The impact it has on society is pretty wide-ranging. Broad conversations will help. And if you want more than this,
1:06
Cast. Please do find my newsletter at exponential view dot Co today. I'm thrilled to be kicking off this season with an entrepreneur investor. A thinker who is truly of the exponential age. Sam Altman is a prolific investor and the previous president of Y combinator the world's most
1:23
successful Tech accelerator. It
1:25
helped launch air B&B Dropbox Reddit and over
1:28
2,000 other companies into our lives today, Sam is the
1:32
CEO of open Ai and organization created, to research and develop.
1:36
It's first AGI or artificial general intelligence a technology. So powerful that it could outperform humans across a wide range of tasks, at least in theory in the summer of twenty twenty open. A i released GPT three, a small Milestone perhaps in the journey towards a GI. It is by far the most effective and sophisticated language generating software ever created. But like any powerful technology, it does come with profound risks. And in the case of this one, this includes the danger for
2:06
I love amplifying racial or gender bias or becoming a tool for malicious disinformation, Sam. And I touch on how you manage risks like this when developing these kind of breakthrough Technologies in our conversations. But we did record this discussion before. Open a I announced its exclusive distribution deal with Microsoft so that is something that we couldn't talk about. Now if as Sam predicts, the AI Revolution will be bigger than the agricultural industrial and computer revolutions. Put together these new technologies must be harnessed with sensitivity.
2:36
And deep reflection Sam. Welcome to exponential
2:39
view. Thanks for having me.
2:41
Now, do you still believe a few years on from setting up open AI? That the AI Revolution is going to be more significant than the Agricultural Revolution, which created civilization
2:52
itself on a long enough time Horizon. I definitely do. I think that Technologies get sort of to stand on the shoulders of the technologies that come before and we needed civilization itself to be able to have the industrial.
3:06
Lucian. And then the computer Revolution. Certainly we need those to be able to build AI, but everything just compounds at this phenomenal rate. And so in this next level of a merchant power I do think is going to transform everything.
3:20
So this summer, we had G PT 3, which was a really powerful milestone in a number of ways, how would you characterize G, PT 3, as a technology on that continuous curve
3:35
on the curve of a
3:36
I think it is a baby baby step. I think that it's it's still quite weak and important ways but it is a preview of what is to come. It is definitely not a GI in the sense that people mean when they use that word, but it is like it is a model that has General capabilities and some degree of what feels like intelligence or at least, what can be used in an intelligent way for lots of interesting applications.
4:06
And I think it's also sort of a preview of how these Technologies are going to be, as you mentioned, both good and bad or they have the at least the potential to be significantly helpful or harmful. And I think it as important as the technological Milestone is the sort of societal Milestone of saying, okay, this is going to happen. How do we as a society, decide what the rules should be and how we're going to use this? I don't think Silicon Valley has a perfect track record here and so, hopefully we can do it a little bit differently this time.
4:36
What was the dissonance that you were experiencing between what you thought of G, PT 3 and what you were sort of reading in the frenzied excitement around
4:43
it? Yeah. I think what you just said is exactly right. I think it just got sort of frenzied. I think I think it is if you're not looking at this every day and sort of watching the exponential curve, I see why it looked like a discontinuity. I can understand that because people have been talking about AI for so long and there's been no there's been no publicly available system that feels really General.
5:06
And so the first one of those is in some sense, a big deal, but but I think people got ahead of themselves and, you know, this like often happens with the new technology and then you use it for, I only get used to it. So, what we, what we were trying to do there is just like, remind people that look. This is a moment in time. This is a, an interesting checkpoint. This is not a GI, this is not mean, AGI is around the corner, but it does mean that like something
5:36
Powerful in this direction is going to happen and we should have a conversation about how that's going to go.
5:41
Let's talk a little bit about G, PT 3 and what it actually does. I mean if you're explaining it to somebody who is not involved in the field, what does this tool that you've built actually
5:51
do? I would say that it is a general-purpose language tool. You can have it respond to a text in any format you'd like and it's pretty good in many cases. Not all at understanding. What should come next?
6:06
But to do that. Well, it has to sort of understand what what a human would think, would come next. It's trained on looking at lots of text and predicting, what would come next and to the surprise of a lot of other people? It can do that. It can use that same model to have a chat in a answer questions. Write code summarize documents translate between languages and I think that's what's powerful about it. Is that this one model just by predicting the words that should come next.
6:36
Can do so many powerful applications,
6:38
right? The the point being that when you think about where we were say, three years ago with text based systems, you'd have had to have one system that could do customer service chat support for a mortgage company. Another to do customer service for an insurance company. A completely different one to help you with writing software code and the, and a fourth model to translate between Mandarin in French, right? These would all be distinct separate.
7:06
Protools. And what we've started to see across the industry in the last few years and GPT 3 is a very good example of this is ones that are more General, that can be used for many, many different
7:17
applications. Yeah. And that's what I think is really exciting about this little stepping stone on the path. Is this, this is the first really general-purpose AI, that were aware of that has been that has been released when these general purpose. Things come along like the iPhone spawned.
7:36
Huge number of new companies before the internet, the computer the combustion engine. They spawned a lot of new derivative Technologies and companies. And why I think this matters is we are now at the beginning of what I think will be fairly rapid development of sufficiently General AI that it counts as one of these general-purpose Technologies and you know, like obviously I'm biased here but my guess is that this will be the next technological platform that we've
8:06
All sort of been waiting for in the industry for some time to enable a lot of new Incredible services, that would even be difficult to imagine today,
8:14
as adults sort of observing this from afar, I can say, oh, this thing looks quite General. But do you have a formal way of saying, this particular model is actually General? I mean, what is the test for something to be
8:28
General? The best way is to just evaluate it all across. Lots of different sorts of tasks, and lots of different metrics. I'm a big believer in you. You make what you
8:36
So if you measure some narrow thing, you'll do well that not everything else and we really try to hold ourselves to evaluate and its performance across many, many different paths. And, you know, we kind of know the areas where it's week and then we know the areas where it's strong and some it's really strong but we're going to just keep expanding the number of tasks. We evaluate these systems on and try to do better on all of them simultaneously. So there's like mathematical definitions that I think we're kind of like reaching the edge of the utility of and
9:06
At this point, like our evaluation metric will just be like, how useful is it to the world.
9:12
And you'll be able to measure that by the amount of views. People make a vault in the number of different applications.
9:17
You know, one of the things that we learned at YC was that if you just set a growth goal and a very aggressive growth goal for yourself like 10% a week, or something on your key metric. You can't hide from that very long. And it's either like a good enough product that people really love it and base there.
9:36
Company off of it and tell their friends. It's amazing or it's not. And if you just sort of hold yourself to a growth metric, that will eventually
9:45
be an acid test of whether what you're doing is good enough or not,
9:48
you keep referring to GPT 3 as a small step on this this curve, you know, when I look at it, I think of it as one of the most complex machines that has been built in the world, it has a hundred and seventy five billion parameters, which if you're not a computer scientist, think of a parameter as a dial that you can turn like a volume button or the base or the treble or it's a it's a slider on a graphic equalizer that's a lot of
10:15
Parameters. And the previous baby step that you took, which was G Pt 2, which was a year and a half ago was 1.5 billion parameters so it was a hundred times less. And the one before that was tenth, the size of GPT to. So for you as insiders, you see these as sort of incremental steps that you're taking along the curve for people, observing from the outside these look like really big steps. I mean, they're kind of order of magnitudes is increasing each time.
10:44
Sure. But if you think on a lot,
10:45
Log scale it's like,
10:47
right? So do you think on a log scale? Is that how it's done with an opening I
10:50
do? Yeah, I think that is that's basically like been the story of the field for for some time
10:56
now. So in a product like G PT 3 which is quite complicated. What have you had to learn in order to build something like G PT 3. So why is it May 20 2010? Not May 2017 to bring
11:15
Sout. One of the things about opening I that is unusual is that we are a combination of a, of a research lab of like, a very good system. Scaling startup Engineering Group, and sort of a policy and safety something pink. And the thing that we've learned is, you have to execute as well as anybody in each of those three areas the same time. So there's like a lot of complicated research. There's very complicated system scaling the that the
11:45
The policy issues around how we allow access to this technology who gets to use it for what how we make? Sure people are not misusing. It are very complicated how you have the model itself, help with that and we have had to learn how to do well in all of those axes and I think the thing that is tricky is they're very different cultures. And so getting one group of people to execute is challenging but that's why we've been to whatever degree. We've been successful so far. It's not that we just get the system up. It's not that
12:15
Just that research is not that. We just figured out sort of the policy rules here. It's doing all of them at the same time.
12:21
So what would you say is culturally different about the way that open AI runs relative to the many hundreds of startups that you will have seen and experienced through your time of Running, Y combinator,
12:34
actually feels quite different. I had a steep learning curve here and I'm thankful for people's Indulgence with me but I think research culture is like quite different than than startup culture and
12:45
Like, even how like, what the company values? How you set the goals? How you measure? Progress, what motivates people. It's a very different world. Like an example of something that was hard for me to understand is how important publication is two people, right? Like I just didn't really have a mental framework for that and it took me a while to understand. Now I think I get it but how you get that to coexist with people who are like, we're going to measure ourselves off of a product growth goal and people say we're going to measure ourselves off of like the
13:15
Number of citations, we get, it sounds like those should be complementary and I'd say, it, turns out you have to do more work than you think to get those to co-exist getting the company to exist, happily in the middle of these three, sort of Clans, or cultures has been, has been an interesting challenge. I think we're doing pretty well on it now though,
13:33
but you've also discussed this idea that, you know, you have to now start to measure, open my eyes value in terms of the usage,
13:42
right? I certainly don't think that's the only way to
13:45
To evaluate our progress, in addition to the academic benchmarks. We now have a new one, which is the real world. We will continue to measure ourselves on lots of metrics. I think that's a key to getting something general. But but like, what we really care about, are we, what I think most people doing this kind of work, she really care about is like, how do you do something useful? Do you make people's lives better? And now that we can directly measure sort of impact in the world? I think that's like a wonderful new Avenue. That's
14:15
Exposed itself to us
14:16
as they suck entity. That is part research and development and also putting products out there. It seems like the the way in which you need to make decisions about what projects to go forward with and what Avenues you don't is perhaps different to what it might be in a traditional product based company. So, what is the framework that you might use to internally to Desert to decide whether to initiate an internal project and whether to progress it?
14:44
So we're like one thing that helps with that is we it's a difficult Mission but we have like a very clear Mission which is to build safe AGI and maximally share the benefits of humanity and we basically evaluate everything we do on that question. If we should do a piece of research at all, if we should scale it up, if we should release it and we have a small enough company where we can even though we're doing something new and hard and sort of with this kind of quirky structure. And as far as I know, no other way.
15:14
Organization in the world runs quite like this, but we have to make these hard decisions. We have a small enough team that we can get everybody in the room or now on the zoom call and or the Google Hangouts call and talk through it. And, you know, I think it's like it's that's a that's a real advantage of being a being small. If we also realized that we don't have enough input from broader Society,
15:42
how do you know that you don't have?
15:43
Have the right external
15:45
input. I think there's like a hundred and fifty people. That work at open II. There's no way we have enough input from Ryan sense of the world as a whole.
15:54
Okay, fair enough. Then I'm curious about one of the things that you learned with GPT to which was a precursor to the latest model. I think there was there was a particular statement around not wanting to release this into the into the wild into the open because you felt it was too risky. There were a whole set of risks with the with my
16:14
To risk that it might be too risky, right? And when you got to GPT 3 which was a much more sophisticated, more powerful technology, you did open it up. If there's an access program with a big long queue of people where researchers and developers can access this through an API and build their own applications. So I'm curious about what you learnt such that, the decision was different for the more powerful technology.
16:38
So one thing that we learned is that you just need to like stick with your convictions, even if you get panned on Twitter and like
16:43
Lately ignore it, which, you know, everyone knows. But it's good to be reminded like when we did that for GPT to, we were just sort of continuously in ruthlessly mocked saying, like, oh, this is so silly. Like this is never going to be dangerous, blah, blah, blah, and like then. G PT 3 comes along those same people are like, I can't believe I open my eyes releasing this at all. It's so dangerous blah blah blah. So like it was a good moment for the company to like really stick with the courage of our convictions and that we would, you know, the I think the general mistake Silicon Valley has made has been be cautious.
17:14
Too late and we'd rather err on being cautious too early. So that was like a good overall learning. One specific learning is that, if you, if you just release the model way, it's like we did eventually with GPT to on the stage process, it's out there and that's that you can't do anything about it and if you instead release things via an API which we did with G, PT 3, you can turn people off. You can turn the whole thing off. You can change the model, you can improve it to continually like do less bad things you
17:43
An rate limit it you can you can do a lot of things, you can do a lot of things. So this idea that we're going to have to like have some Access Control to this Technologies seems very clear and this current method may not be the best but it's a start. This is like a way where we can enforce some usage rules and continue to improve the model so that it does more of a good and less of the bad. And I think that's going to be some something like
18:13
God is going to be a framework that people want. Is these Technologies get really powerful? I mean, again I think we're like the baby baby
18:20
stage. I think what we've seen with G PT 3 of course, is that the first thing that happens with any new piece of software that is that is programmatic, is people try to push it to its limits, right? That's what you, you know, you get access to the Twitter API in 2009 2010 and the first thing you try to do is what can I break down. So you must have seen people trying to break G PT 3 in different ways.
18:43
Ways to what extent is that helpful for you and other surprising ways in which people have tried to stress the system. It's
18:50
totally helpful. Look at some level. You could say that. It's like unfair that most of the negative stuff you see on Twitter as like, people turning off the content toxicity filter really trying hard to beat it into giving offense of outputs, but then on another level like,
19:06
that's the kind of stuff that can happen with technology like this and it is important for us to get that input in that feedback so that we can build
19:15
Models that do much less of that or someday just don't do it. And I think it's important with a technology like this to get enough exposure to the real world that you find some of the misuse cases you wouldn't have thought of so that you can build better tools. There's a whole set of hard questions about. Like if someone really deliberately wants to use the model to say offensive things in a context where that might be okay, or even good. Should that be allowed? And my personal belief is yes but it's a tricky.
19:45
Action but I think we can all agree. On is no one who is like people developers and users should have controls on these models such that they behave in a way that the actual user using them once and I think seen where that doesn't happen and getting the feedback is what lets us get there. I think
20:07
this is a this is a challenging set of questions here though that relate in general to the question of governance of powerful
20:15
Knowledge. He's one of the things I think that's tricky with AI and software Technologies in general, is it? They get progressively easier to develop and deliver, right? It's very different to nuclear power, right? Which is still a state level entity and you can be a big country like Iran and still struggle you know militarize in those Technologies. But with software every two years it gets ten times cheaper and the breakthroughs that you've made this year will be you know quotidian within three or four years. So the the sort of March of wine.
20:45
Law and Moore's Law. So how do you think about whether we need some kind of broader governance framework for these Technologies? I mean I think Silicon Valley has been brilliant at many things but it hasn't been brilliant around. The questions of ought there, be some sort of governance around the Technologies in the way in which they get used. Have you thought through that problem with, with, with AI in general? I
21:10
think we do need new governance. I think that more generally
21:15
I would say our existing governance models did not account for the idea that technology would eventually become so powerful in such that companies like open a, I could eventually do. What only nation-state actors could do? You know, if we go back and think about your example, of, let's say, nuclear weapons first, and then power or something like the Apollo program, those required an amount of capital, and just resources that.
21:43
You know, like a private company just wasn't gonna do it. I forget the exact number for the amount of electricity in 1943 that went to enriching uranium but it was something like 23 percent of all of the US has electricity. Like just took massive resources and so I think we need new governance models. And we also need new economic models.
22:05
And part of what I think opening II is good about, is that we are able because of our sort of unique structure and this capped profit model that we came up with to
22:17
Advocate loudly for those. I think your economic model is super interesting and I would love us to spend a bit of time explaining it. Let's talk about the governance question though because I think that seems to me to be a really difficult question. So the way that I look at, you know, technology governance at the moment is that especially internet technologies have not.
22:35
He had a governance framework, they emerged from a west coast of the free addressable end-to-end internet which in a way is a governance model of of its own. And recently as you rightly point out, these technologies have got more and more powerful and the biggest companies like the Microsoft and the Google's have state like capabilities in certain areas. You know, when there's a Cyber attack, the local military can't do anything about it. And then the trouble with this this governance
23:05
Model is that these large companies don't have any accountability to Citizens and we are ultimately at the whim of the senior Executives and the board of directors to make good decisions about the chain of accountability has been broken. So I'm curious about what you think Global compact might look like, in the next few years where we feel, we can get these Technologies working for Humanity but in a way that as well governed,
23:33
it's a super important question.
23:35
We can study history and see how that goes when there is no public oversight. You know, when you have these companies that are as powerful or more powerful than most Nations and are sort of governed by unelected and to be perfectly honest like fairly unaccountable leaders
23:57
I think that should give all of us. Pause as these Technologies, get more and more powerful. And I think we need we need to figure out some way. Not just that economics, get shared and that we deal with going to Quality problems. That Society is facing. But that governance about these very big decisions. You know what do we want our future to look like in the war with very powerful? A I did that is not a decision that a hundred and fifty people sitting in San Francisco, make collectively like that.
24:27
Not going to be
24:27
good. I haven't found good models right now that do this. I mean, I think that people talk about the idea of mini lateral is MM, which is that if you can just bring enough people to talk often enough, perhaps you create the kernel of a framework that other people can buy into, and then people copy that.
24:47
Yeah, I think that is helpful for sure. And I think, I think, but then the glares, like, always a question of how much teeth does it really have?
24:57
Let's say we really do. Create a GI mom like em shirt. Like right now if I'm doing a bad job with that, our board of directors which is also not publicly accountable can fire me and say, let's try somebody else but doesn't it feels to me like at that point at some level of power like the world as a whole should be able to say like hey maybe we need a new leader of his company just like we are able to vote on you know the person who runs the country and so this question of what democratic process for companies that
25:27
Have sufficient impact looks like I am very far from an expert here and much smarter people than me have been thinking about it. But it occurs to me that we've got these models and we can look at the different ones in the around the world and talk about some countries, where it works better, and where it doesn't work as
25:44
well. Is that one that Springs to mind for you?
25:47
We are working on this actively right now, I wouldn't say we're at a point where were like, okay, this is like, definitely the right model for us or for the industry as a whole.
25:56
But it's something that we're trying to turn our Focus to
26:00
the other one. The other piece that's interesting is that you know, historically technologies have always created inequality. That's why we see the arrival of status and hierarchy and private property. You're tend to have 15,000 years ago. This technology is going to be no different but it's clearly been something on your mind as you've gone through to design with your colleagues open. A I one of the particular things
26:26
You've built within open AI that start to address that.
26:30
So I think the biggest thing that we did is this idea of a cap, The Profit structure. So we said like look we started as a non-profit, we are hoping that we could make it work there. Making a GI looks like it's just going to take so much Capital that we thought we couldn't do that as a non-profit. But we didn't want to be like a full for probably either because of the potential to drive like truly, massive inequality. And so we said, is there some new structure?
26:56
Where we can be, you know, we can have the benefits of capitalism and all of the wonderful aligning power that produces attract the capital and talent we need in Fairly reward them, but then figure out a way to mostly share the benefits here. And I think that as technology keeps going, these companies get bigger and bigger and more powerful and have more leverage. We will need something new as a more General structure. And I think like, either companies need to somehow either.
27:26
Make less money because they pay people more or like somehow there's got to be some degree of societal ownership of Corporations or like a direct sharing in the equity value. Otherwise inequality will just run away and open my eyes, like one way of sharing the equity value
27:45
and I think just maybe put some historical framing on some of those observations. I mean, the the labor share of income in most developed countries has declined by about 10 points over the last
27:56
80 years with most of that gain going back to companies themselves. In fact, not even to shareholders. And the other thing is that within an AI, kind of product, you have a rich get richer phenomenon because your you've got the data Network effect that is making your model better than the next person's. And every query that comes into G. PT 3 is a not new form of data that you can use to improve the next the next model, right? So you have these Dynamics and you know, I think it's fantastic that you've identified that this is one of the
28:27
The challenges you talked about there being a problem with the inequality that would cause. So do you have an internal mental model about how much inequality is too much and how much inequality is kind of just enough to keep people
28:40
motivated. That is such a hard question. You some degree of inequality is the price of capitalism. I still think capitalism has like, these wonderful benefits that have been replicated by no other system. And if you stop rewarding people for innovating or even,
28:56
Officially allocating Capital, that'd be really bad. So, perfect. Equality is wrong. I think but this the current level of inequality we have is like way too much and in terms of like the multiple of how much should the average person have relative to the riches or whatever else? I don't really know. What I will say is that I think people are most sensitive to just being on a very steep curve of their lives. Improving and the fundamental thing that has gone most wrong with inequality in the u.s. is not the spread between like the rich.
29:26
Person and the median person. But it is that all of the gains, all of the relative progress has gone to the people at the top. And that I think is the most toxic of all. So, rather than a spread, I would say, everybody needs to participate in a massive updraft year-over-year. I do think AG, I will do that.
29:47
So, let's talk about the mechanism by which a GI then helps address the inequality
29:53
problem, the cost of the
29:56
Most critical goods and services should in in like some sort of, you know, real dollars sense. Just go down and down and down as technology can do more and more of what human labor has done, which drives the cost of most goods and services. And that means that like what you really want is not a certain number of dollars but a certain amount of wealth and we're saying an increasing amount of wealth. And as a g, I can drive the cost of goods and services down dramatically if we even hope.
30:26
Just redistribute a little bit of
30:28
money, I hope we do more than that. Everybody should be able to afford what they need for, really good life and more of it. Every year,
30:35
one of the problems that we've seen in the u.s. in particular, is that the price of things that I suppose a kind of consumer-oriented maybe a little bit trivial, whether it's, you know, it's televisions or games, and so on has gone down. But the things that you actually need to have a fulfilling flourishing, Safe, Life, Health and education has has shot through the roof.
30:56
Oof,
30:57
it's really disappointing to see. I think this is like a huge problem and not well understood. But basically, we're technology and more Technologies been able to do its thing and where people have had to have like a competitive market. It's been great and then where you've had these problems where there's not very good Market competition. And particularly where services are sort of in some way or other government subsidized, or government protected, but not government-run, you
31:26
These like wild and efficiencies. So I think it like it clearly works and other countries that besides the us where the government just provides the healthcare, but they also pay for it and they make the decisions about it. But what we have in the u.s. is an absolute disaster and if you look at kind of like the three Horsemen of the cost apocalypse housing, health care, and higher education for slightly different reasons. And all three cases, we are just facing an absolute catastrophe and that, you know, technology can help those somewhat for sure.
31:56
So I think those are policy failures more than anything else and until we address, those can do a basic income, you can build a GI, you can do a lot of other things, it's still going to be bad and I think those are so critical to. As you said, a good life, a fair life and the American dream that if we can't fix those basically as a society. Very it would be very limited impact on everything else that we can
32:23
do. So you talked about being a non-profit and the cost of doing a GI
32:26
I was clearly going to be yeah. Hi. I mean to the to the average listener I think your initial funding was around a billion dollars is what's been put in the Press. So how much will a GI actually cost a
32:39
lot more than that? I don't know the exact number, we're going to spend as responsibly as we can and as effect efficiently as we can but we'll also spend whatever it takes. So I don't know what that number is, but it's going to be vastly more than a billion
32:53
dollars. So somebody with a cheeky
32:56
Sheet might try and run some numbers. They might say well you've got a hundred and fifty headcount, we know that g PT 3 cost x million dollars of computer time to train the model and there are some numbers that are floating out there. We know that the price per compute cycle is declining at this rate because of the great work of you know Nvidia or graphic or whoever's doing it. So they're going to put that in very quickly and say well if it's more than a billion dollars and we've got these declining
33:26
Kind of price curves, we have a sense of the amount of compute that is going to be needed in order to build a GI. So I feel like I have to ask you this question, which is how much
33:38
computers me. Not even Oracle could tell me exactly how much compute we're going to need. It would save me like a lot of hand-wringing and I'd love to know. So if you figure that out, please tell me it would be really useful information. I, first of all, I the estimate that I've seen for how much it costs to train gbt three were all way under, but I'm sure I haven't seen all of
33:56
Them. None of this stuff is cheap. When people have run those spreadsheets, they come up with wildly different numbers. I have seen people say 10 billion, I've seen people say a hundred billion, I have seen people say a trillion I have seen people say a hundred million we just need the right algorithm and the true answer of course is no one knows and everyone spreadsheet is wrong in some important way. We ourselves, I think we do some of the best work in the world trying to project this and we ourselves have huge error bars.
34:26
But the more we learn, the more confident, we are, and then we can go often raise more specific amounts of capital, and make more specific computer orders. But all I can say with confidence, it's going to be a lot.
34:37
Yeah, but I think the important thing as well as it's not just a dollar. Some there's something that you said earlier in our conversation about you have these different teams who are have competing interests and competing constraints. And essentially, the capabilities of each one has to have to grow like an ecosystem with the other teams.
34:56
So it's not as if you could get someone to write you a huge check today and that would solve the problem tomorrow. It's more that there's some undiscovered capabilities that need to be learned. For
35:07
sure. One thing that I certainly believe is that if you do you could throw like a trillion dollars at the problem without any new ideas and you still wouldn't get to where we'd like to get. I'm dying. I'm very confident that we're making Fast progress on the new ideas and certainly on kind of the policy and safety work. Question that like we get a lot is
35:27
wat, you know, open II is
35:29
Smaller. It has less compute than some of these other organizations in labs and like, how are you outperforming them or why? You are performing them. And I think one of the answers that sounds a little bit silly, but if somehow true is, it's not in spite of that. It's because of it and because we can stay so small and focused and only do a few things at a time. And really have to align a lot of people with these different cultures into thinking about, you know, where are we going to bet are very extremely Limited.
35:59
Horses and talent. That's actually the best way to make progress towards these
36:04
problems with that approach. I mean, it sounds to me, like, it's the, it's the startup approach, right or sure. So it's come out of your experience of not having been a researcher, but having nurtured, a thousand startups or more. Absolutely, it's a very distinctly different way that our lab would work, but then I'm curious also about whether you can do basic science in that way, right? Sometimes basic
36:29
It's just requires someone to be able to ask the question and just spend their time scratching their forehead and throwing things up on a
36:37
whiteboard. Well, it seems like we're doing. Okay at that. It we really try to build an organization that supports that and I view a big part of my job is not letting all of the pressures that come with success and offering a product to get in the way of what got us here in the first place, which is someone scratching their head in front of a white board.
37:00
I still think we have a long, a lot of that, that we need to do, obviously some of the most talented people in the world to do it. And we've got to make sure that we keep a culture where that used to happen. I
37:11
want to zoom back out a little bit. We talked about the importance of AI and AGI as heralding, a new epochal revolution in human development. And when we look at these revolutions, one of the things that's interesting is I think of it as
37:29
You can't tell a goldfish about the water that they're in, right? Because they're in the water, right. And a new technology comes out and takes the gold fish out of the water and say, hey, look, you're in water and there's a, there's a bowl and the Goldfish has to learn about that these sort of emergent properties, that come out of society. So whenever we see these breakthrough general-purpose Technologies, we see that change happening and it's not just an industrial change, it is a kind of political economic change. It is a values change. So it took Galileo 350
37:59
Has to get his pardon from the Catholic church. And yet here we are with a new set of ways of thinking about the world. That are you can't disentangle them from the technologies that went before in the potentials that they created. So it's a really hard question for me to say to you when you build a GI, which are essentially is like a five dimensional telescope, right? What are we going to see in The Fifth Dimension, but when you ponder that question and you think where are the point the fracture lines?
38:29
Lines in our existing world that will look so primitive. When we are living in a society that has as its Bedrock technology, I think
38:39
it's going to call into question, all of sort of the deep philosophical issues that have been on Humanity's mind for Millennia and they will be newly relevant. Like what it really means to be conscious and what it means to sort of, like be a separate entity, if we really try to like, look through that five-dimensional,
38:59
Scope about what this is going to tell us about our place in the universe and what intelligence and Consciousness and awareness. And all of that mean, like, that's hard to think about like that's a big, big thing. And that's the one that I would say like when I wake up at 3:00 in the morning is the thing that's on my mind it's like what's going to happen. Like what are these? You know, when this thing gets kicked off
39:23
And these like intelligent entities that are much smarter than any humans and sort of self-aware in some way. But maybe in a very alien way, maybe we might build a very alien intelligence. What's that going to mean for like the next ten to the Hundred Years of the Universe? On a more like pedantic level for society? I think it will be a big shift because we have for so long.
39:46
We Humanity have identified and gotten our identity, based off of being the smartest things on the planet. Hmm. And when that's no longer true, how will we Define our success? How will we get our sort of intrinsic motivation and happiness? What will we do with our days? And also, like, how will this system working in concert with us? Help us govern ourselves in ways better than we could ever think of ourselves.
40:12
How do you avoid the challenge of taking?
40:15
Taking a kind of anthropocentric view of what that intelligence is. I mean, I think you hinted at this where you said, this intelligence may be more alien, like it may be that the tools you end up building have as much in common with our intelligence as a helicopter, has with a, with an eagle or a hawk. If that's the case, do these things become tools. Do they ever have agency that, that we could recognize as agency?
40:41
I suspect they will on a long enough time Horizon but that might be a very long.
40:45
Horizon and it might be in a different way than we think now. But I suspect that they will, I still sort of think that some version of emerge is what's most likely to happen and it we won't be talking about these things and us, but it'll all sort of kind of co-evolved in one very powerful Direction. And I think that's honestly the future. That seems the safest and the most exciting to me.
41:09
Well, one of my favorite examples of the merge is what happened to our lower intestines, as we started to use.
41:15
Flint's which as I'm sure you know, we started to externalize our digestion and so our gut proportions are very different to hominids who have to digest those fibers internally. So we use this,
41:30
we've been working with technology for a long time, and it doesn't have to happen with like, plugging electrodes into our brain. There's a lot of there's a lot more pleasant ways than that to happen. And I expect it will, but it won't like, I don't think it's going to seem like this super differentiated, right?
41:45
Fine thing, it's US versus them. I hope it doesn't. I think that would be bad.
41:50
I think every year is it right? That you and the team get together and discuss forecasts for when you get to the Milestone that we've been discussing, which is a GI. So where is a current forecast consensus View and how do those discussions take place? I mean take us to that room sure.
42:07
So this is the year where we realized it was a bad question and I think we've all as part of just
42:15
Seeing G PT 3 out in the world. Not all of us, but most of us have updated towards thinking. That instead of this moment in time, it's going to be this continual exponential curve, like everything is. And this question of like, when we get a GI is somehow framed wrong because it's going to be a many year, I think smooth exponential curve of progress. And so, then the question is, like, do you want to talk about one that curve starts when it ends? Like what does it even mean to end? Maybe it never ends. Maybe it just keeps going. So,
42:45
Like and certainly for me, this was the heroes that I can't give a number because it's going to be this like thing that's already started and it's going to go on for a long time and I can't say we're done because it's just going to keep getting more powerful. But the way that it happens is we do an annual off-site. Every year we give talks, we have discussions, we like drink around a fire, whatever. And then at some point, we do a company pole and it's very unscientific, you don't have to respond. Not everybody does. And
43:15
the question is sort of a gi's you want to Define it how many years and it was sort of going down or coming. It was
43:22
The number of years away was going down faster than one year per year. So again, some sort of
43:27
exponential curve, yeah, small team ISM. And
43:30
optimism could be the wrong kind of optimism could be, like, we just sort of hired people who are too optimistic. So I don't want to make too much of that, but but this year, many people would say things like it's really hard to Define because it's not a point in time and that's definitely what I feel. So I think next
43:51
next year, we'll figure out how to ask the question in a very different way.
43:55
It sounds like you and open a I are always learning some
43:58
we are trying I think one thing we do do we do a lot of things badly by think. One thing we do well is like get data and adapt
44:05
that has been Evolutions answer for the miracle of survival and thriving which is get data and adapt
44:12
it will keep working. I'm pretty sure.
44:14
Thank you very much Simon. Thank you. Well, thanks for listening if you enjoyed this conversation. Be sure to subscribe because we have some of
44:21
Most powerful figures in the AI and Tech World lined up for you later in this season. In the meantime, you should check out some of my previous discussions with many of the world's top AI researchers, including Faithfully, Gary Marcus, and Jürgen schmidhuber, I would encourage you to listen to those three in particular as they provide some perspectives on some of the issues that I discussed with Sam to stay in touch, subscribe to my podcast or my newsletter at exponential view dot Co. This podcast was produced by Maria gavrilovna Fred Casella
44:51
Ilan. Goodman is our researcher boyens. Happy a cello, the sound editor, special, thanks to exponential view members, Elizabeth Ling, Paola bonomo and Jionni, Jack O'Malley for help with this episode. Exponential view is a production of e to the I pi plus 1
45:04
Limited.
ms