news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

amodei took the photovoltaic industry as an example, arguing that even if a technology becomes very important and widely used in the market, it may still be difficult for a single company to generate high profits, and high commercialization will limit the company's profitability. similar to the photovoltaic industry, the market size of ai technology may also be very large, but it may also be difficult to profit from it, and the profit distribution issue among different participants is also very complicated.

hard ai

author | huang wenwen

editor | hard ai

this week, dario amodei, ceo and co-founder of "ai unicorn" anthropic, participated in the interview and discussed the development of the ai ​​industry, scaling law, etc.

scaling law mainly refers to the fact that as data and computing power increase, the capabilities of the model will also increase.amodei believes that if this pattern continues to hold true, ai will have the potential to become very powerful, bringing about new business models and technological innovations.

however, he also emphasized that this law does not always apply, and if the performance of the model cannot be further improved in some aspects, the entire ai industry may be affected.

amodei also highlighted the ai ​​industry and the photovoltaic industry. he used the example of the photovoltaic industry to illustrate that even if a technology becomes very important and widely used in the market, it may still be difficult for a single company to generate high profits.

although photovoltaic technology has shaken up the entire energy industry, its market is highly commoditized, competition is fierce, and there is a lack of significant branding, so the profitability of each company is subject to certain restrictions.

similar to the photovoltaic industry, the market size of ai technology may also be very large, but it may also be difficult to profit from it. the profit model of ai may be affected by multiple factors, such as the computing cost and inference cost of the model. the profit distribution between hardware providers, model developers and application layers is also quite complex and faces considerable challenges.

the main points of the conversation are as follows:

1. it is somewhat difficult to discuss scaling law completely separately from business issues.

2. with the market so large, my initial response is that the gains will flow to all of these places.

3. if the scaling law is correct, this will be a very large market. even if only 10% of the profits flow to a certain part of the supply chain, it is still very huge.

4. if we are building models with tens or hundreds of billions of parameters, there will probably be no more than four or five entities (and maybe some state-owned enterprises) involved. so what we are seeing may be more like an oligopoly than a complete monopoly or complete commoditization.

5. even if such a model is released, such a large model is very expensive to run inference, and most of the cost is in inference rather than model training. if you have a better inference method, even if it is only 10%, 20% or 30% improvement, it can have a big impact.

6. we have larger, more powerful models and faster, cheaper, less intelligent models. some customers have found that large models can assign tasks to a large number of small models, which then report to the large model to complete the task like a swarm.

7. no one should believe that scaling laws will continue forever, it is just an empirical observation and may stop at any time. i have observed them for ten years, and my guess is that they will not stop is based on the length of time of observation, it is just a 60-40 or 70-30 proposition.

8. if we can use ai correctly, it can increase the speed at which we make these discoveries by 10 times, maybe 100 times.

the full interview is as follows, with some parts edited:

01

google failed to become the bell labs of the ai ​​era

Noah Smith：

in this economics podcast, we prefer to talk about economics rather than pure technology. so, is google the bell labs of the ai ​​era?

they did the research that led to modern deep learning and transformers and so on, but they didn't really succeed in commercializing it, just like bell labs. they funded the research with monopoly money, and then interesting people like you worked there and then left to start companies, just like the fairchild people at bell labs. do you think that's an apt analogy?

Dario Amodei:

while nothing is a perfect analogy, i definitely think there is some truth to it. a lot of people see it as a continuation of their academic careers, and it's very similar to the industrial environment of bell labs, except that google has a lot more resources to accomplish things. as a result, there are many projects being worked on. the transformer was one of the key inventions that pushed the field forward, and it's just one of about a hundred projects being worked on.

if you're at the top of the organization, you can't reasonably distinguish it from the other 99 projects that are being developed. it's like a hundred flowers blooming and competing with each other. i think that's when i first proposed the scaling law, that we need to scale and integrate these innovations at a large scale.

in theory, google is the best place to do this; they have the largest cluster in the world, a large number of talented engineers, and all the necessary elements.however, google is organized around search, and i don't think it's necessarily organized around the idea of ​​combining all of these parts and scaling up to something radically different than what came before.

Noah Smith：

just like bell labs wasn't set up to invent computers and have everyone have one, it was set up to wire everyone.

Dario Amodei：

right, it's a phone company. so, i can't speak for google, but obviously now in addition to inventing all these amazing things, they're also one of the four largest companies with cutting-edge models that are both our partners and our competitors. i know a lot of people there, and they're very smart.

but i think you're right, there was a time when if they could have combined those elements in the right way, they could have been the only dominant force. but for whatever reason, things didn't go in that direction.

02

scaling law:

as the data size increases, the model becomes more powerful

Noah Smith：

this brings up another question that we're thinking about. actually, the idea for us to talk to you came from something we discussed in another podcast, where we were talking about the economics of internet businesses, and then someone brought up some pessimistic views on the ai ​​business, questioning how many economic moats ai companies actually have.

obviously, this is very relevant to anthropic and other companies that we would call startups, but they are already pretty large.so, tell us what you think about the economic moats of ai companies.

Dario Amodei：

i would say i'm going to split this question a little bit into two branches. i think it's a little difficult to completely separate the scaling law from the business question. so, let's consider the case where the scaling law holds true in a very strong form, and then consider the case where it may not hold true partially or at all. if it holds true in the very strong form, the situation is like this:

now you have trained a model with 100 million parameters, which is equivalent to the capabilities of an excellent college freshman;

then you train a billion-parameter model that has the capabilities of a top undergraduate student.

you trained a model with 10 billion parameters, which is equivalent to the capabilities of a top graduate student;

when you train a model with hundreds of billions of parameters, its capabilities are equivalent to that of a nobel prize winner.

then you put this model into use and it basically serves everyone, it becomes your colleague, it becomes your personal assistant, it helps with national security, it helps with biological research.

i think in that world, this system and the products based on it will be a large part of the economy. there is still a question of where the benefits will go? will it go to nvidia on the one hand, or to the ai ​​companies on the other, or to downstream applications? with the market being so large, my initial answer is,the benefits flow to all of these places.

03

the future of ai will be similar to the photovoltaic industry

Noah Smith：

but think about solar, which is obviously going to be very important. the more energy we need, the more solar will be used. and yet, it's hard to say which solar company is making a lot of profit. solar is a very commoditized product, and while there's a lot of innovation in it, without the branding, without the network effects, without any lock-in effects, it's very hard for any solar company to make a profit on this thing that's revolutionizing the world right before our eyes.

so i'm not entirely sure that just because everything is going to boom like it is right now with solar, that that will necessarily lead to companies making profits. but i'm certainly open to that possibility. i'm just wondering,what do you think the source is? why is the development of artificial intelligence different?

Dario Amodei：

solar? i think there are two points here because i think this is an important issue in most of the world. maybe i'm just saying that if the scaling law is correct, this is going to be a very huge market. even if only 10% of the profits go to one part of the supply chain, it's still very huge.

just as you make the pie bigger, that becomes the most interesting question, although those who decide how to distribute dollar bills will certainly care a lot about where the trillion goes. but let's get back to your question, because i think it's important all over the world.the key question is how big the "pie" you are dividing is.

first, in terms of the model, it depends on the scaling law. if we are building a model with tens or hundreds of billions of parameters, there may not be more than four or five entities (maybe some state-owned enterprises) involved. so what we see may be more like an oligopoly rather than a complete monopoly or complete commoditization.

i guess one question here is, will anyone release an open source model with a billion or ten billion parameters? i doubt it, and even if they do, models this big are very expensive to run on inference, and most of the cost is in inference rather than model training. if you have a better way to do inference, even if it's just a 10, 20, or 30% improvement, that can have a big impact. the economics are a little weird, there's a big fixed cost that you have to amortize, but there's also a unit cost of inference, and under that assumption, if it's deployed widely enough, the difference can be huge. i'm not sure how this is going to play out.

Noah Smith：

this is actually similar to the economics of heavy industry, such as the way steel is made.

Dario Amodei：

yes, a little bit. and the other thing i would say, interestingly, is that within these models, we're starting to see models have different personalities. so commoditization is a possibility, but even in an oligopoly, some of the ways in which models are deployed may be commoditized, although i'm not sure.

but one of the forces that's working against that is: hey, i made a model that's good at programming, you made a model that's good at creative writing, and a third person made a model that's good at entertainment. these are choices, and once you start making those choices, you start building infrastructure around them, and that seems to create the preconditions for some level of differentiation.

another area that could lead to differentiation is product building based on models.in theory you can separate the model layer from the product layer, in practice they are interrelated and working across organizations can be challenging. so while there is a common logic on the model side, many companies are moving in the same direction, adding multimodal capabilities, making models smarter, making inference faster, the products are so different.

if you look at this artifacts project that we did, it's a way to visualize models in real time while writing code. we do this, openai has their own way, google has their own way, and i think that's one of the sources of differentiation between the companies.

we have found that the economics of selling applications based on models, even relatively thin applications, are getting thicker.

Erik Torenberg：

if the scaling law holds true and things get as big as we think they will, do you foresee these companies being nationalized at some point? or what do you think?

Dario Amodei：

we can divide it into two cases: one is that the scaling law is right, and the other is that the scaling law is wrong. if it is wrong, then this is just a technology, like the internet or solar energy, which may be more important than most technologies, but it is not unprecedented. based on the current development, i don't think it will be nationalized.

if it's correct, and we're building models that are as good as nobel prize-winning biologists and top industry coders or better, i'm not sure if it's really nationalized, we'll care that much about whether adversaries can keep up with us, or whether we can deploy them as fast as our adversaries can.

04

scaling law

influencing ai to create new business models

Noah Smith：

i have a question about the impact of ai on business models. you know the story of electricity, basically in the beginning when they got electricity, manufacturers tried to dismantle their steam generators, and the generators had losses. and then someone figured out that you could run electricity to multiple workstations in parallel, and that changed the way manufacturing worked, instead of one big assembly line going to multiple small workstations doing the work, and that led to huge productivity gains over the decades.

i've always suspected that ai is similar. i think the internet is similar,the similarity with ai is that at first everyone seemed to think that ai was a person. some people actually compared the number of ais to the number of human employees, which doesn't make sense to me because it can't be divided into individuals.

you could make an agent-based system that mimics that, but why? i see everyone thinking about ai directly replacing humans, and my argument is that this is phase one, just like electricity directly replacing steam boilers is not a good idea. i think people will be a little disappointed because there are only a few cases where this direct replacement of humans works, like customer service and a few other well-defined things.

but i think there will only be a few cases where this direct replacement of humans works, and then we’ll experience a bust of the gartner hype cycle.

some creative entrepreneurs will say, let's not just use ai as a human replacement, but use it to create new business models. then we'll see a renaissance boom, that's my prediction. my gartner-style prediction, am i crazy?

Dario Amodei：

so i think it's a mix of some things i agree with and some things i might disagree with. first, i basically agree that everything you say is true if you freeze the quality of the current model. we observe basically a similar thing in business. we provide models that you can talk to, but we also sell models to many customers through apis. it took a long time for people to figure out how to best use the model.

there are a lot of questions about the reliability of models, and i think that's the reason for some of the concerns, like a model gives the right answer 95% of the time, but 5% of the time it doesn't, and how to detect those situations and how to handle error handling is very important. that's very different from being useful in theory and being useful in practice.

we had an early feature where you could have the model write some code, and then you could paste that code into a compiler or interpreter to make a javascript video game, and if something went wrong you could go back to the model and fix it. we also see large models coordinating small models, which is very different from the idea of ​​a model as a person.

we have larger, more powerful models and faster, cheaper, less intelligent models, and some customers have discovered that large models can distribute tasks to a large number of small models, which then report back to the large model, completing the task like a swarm.

we're still figuring out the best way to use the models, and the models are getting smarter and smarter and more capable of solving these problems. so it ultimately comes back to whether the scaling laws will continue. if they continue, it will be a process that you describe. if they stop, innovation will stop, and the process that you describe will end.

no one should believe that scaling laws will continue forever, it is just an empirical observation and may stop at any time. i have observed them for ten years, and my guess is that they will not stop is based on the length of time of observation, it is just a 60-40 or 70-30 proposition.

Erik Torenberg:

what would change your perspective? what would change your odds there?

Dario Amodei：

i think, first of all, if we just train one model and then try the next size model and it works terribly, and we try to solve the problem a few times and it still doesn't work, i would be like, oh, i guess this trend is stopping.

if there's a problem with data exhaustion and we can't generate enough synthetic data to continue this process, then at some point i would say, hey, this actually looks like it's going to be hard, at least this trend will pause, maybe pause, but probably not. i'm still guessing that these things won't happen, but you know, it's a very complex problem.

05

ai could speed up biological discoveries 100 times

compressing the time of progress in the century

Noah Smith：

if the bottleneck of ai resources lies more in computing power rather than energy, then we will have more comparative advantages in utilizing ai. do you basically agree with this view?

Dario Amodei：

yes, i think that makes sense. you mean, to use a slightly ridiculous metaphor,if ais are anything like cyborgs, and the process of making and breeding them is very similar to humans, then we're in trouble.but if it's just a cluster of servers somewhere, with completely different inputs, then we're fine.

i haven't thought about this in depth, but it sounds plausible at first glance. if we're in a situation where ai is reshaping the world, and the economic structure has changed, then we might be talking about something different. but if the normal rules of economics still apply, and i think they will for some time, then this sounds very plausible.

Noah Smith：

but my other question is, is it necessary to think about a world of extreme abundance, where ai is so powerful that it gives us amazing biology and manufacturing that makes everything we want ten times, a hundred times, and so on, better?

Dario Amodei：

i think we've really underestimated the potential of ai in biology. ten years ago, when i was in the field, the attitude was that the quality of the data we were getting from biology was questionable, the amount of data we could get was limited, and experiments were often perturbed. of course, more data analysis, big data, and ai are great, but at best they are supporting roles. maybe with alpha fold, that's changing.

but the way i see it is that ai models can act as biologists or co-biologists. if we think about really advanced biology, like it's really disproportionately there are a few technologies that power everything. for example, genome sequencing, which is the ability to read the genome, is the foundation of modern biology. more recently, crispr technology, which is the ability to edit the genome. if we can use ai correctly, it can increase the speed at which we make these discoveries by 10 times, maybe 100 times.

take crispr, for example, which is assembled from the bacterial immune system, and it took 30 years to invent. i think if we can greatly increase the rate of these discoveries, we will also greatly increase the rate of curing diseases.

my thought process is, can we compress the progress of the 21st century? can we make all the progress in biology in the 21st century but accelerate it 10 times using ai?if you think about all the progress we've made in biology in the 20th century and then compress it into five to ten years, to me that's the bright side. i think it's probably true. we could cure diseases that have plagued us for millennia, which of course would greatly increase productivity, expand the economic pie, and extend the human lifespan.