This week, Dario Amodei, CEO and co-founder of "AI unicorn" Anthropic, participated in an interview to discuss the development of the AI industry and the Scaling Law.

The Scaling Law mainly refers to the enhancement of model capabilities with the increase in data and computing power. Amodei believes that if this law continues to hold, then AI could become very powerful, bringing new business models and technological innovations.

However, he also emphasized that this law is not always applicable. If the performance of models cannot be further improved in certain aspects, the entire AI industry may be affected.

Amodei also highlighted the comparison between 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 to bring high profits to individual companies.

Although photovoltaic technology has almost shaken the entire energy industry, its market is highly commoditized, fiercely competitive, and lacks significant brand effects, so the profit capabilities of various companies are limited.

Similar to the photovoltaic industry, the market size of AI technology may also be very large, but the difficulty of profiting from it may also be high. The profit model of AI may be affected by multiple factors such as model computing costs, inference costs, complex profit distribution among hardware providers, model developers, and application layers, presenting considerable challenges.

The key points of the conversation are as follows:

It is somewhat difficult to completely separate the discussion of Scaling Law from business issues.

In such a huge market, my preliminary answer is that the profits will flow to all these places.

If Scaling Law is correct, this will be a very large market. Even if only 10% of the profits flow to a part of the supply chain, it is still very significant.

If we are building models with billions or trillions of parameters, there may not be more than four or five entities (perhaps some state-owned enterprises) involved. So what we see may be more like an oligopoly rather than complete monopolization or commodification.

Even if such models are released, the operating costs of these large models are very high in terms of inference, with the majority of the costs being spent on inference rather than model training. If you have a better inference method, even a 10%, 20%, or 30% improvement can have a significant impact.

We have larger and more powerful models and faster, cheaper, less intelligent models. Some customers have found that large models can delegate tasks to a large number of small models and then report back to the large models, completing tasks like a bee colony.

No one should believe that Scaling Law will continue indefinitely. It is just an empirical observation that could stop at any time. I have observed for ten years, and the basis for my guess that they will not stop is only based on the length of the observation period, which is a proposition of 60-40 or 70-30.

If we can use AI correctly, it can accelerate our discovery process by 10 times or even 100 times.

The full text of the interview is as follows, and some contents have been deleted:

Google failed to become the Bell Labs of the AI era.

Noah Smith:

In this economics podcast, we prefer to talk about economic aspects rather than purely technical aspects. So, is Google the Bell Labs of the AI era?

They have made research on modern deep learning and Transformers, but they have not successfully commercialized them like Bell Labs. They funded these research projects with monopoly funds, and then interesting people like you worked there and left to start their own companies, just like people from Bell Labs' Fairchild company. Do you think this is an appropriate metaphor?

Dario Amodei:

While there is no perfect analogy, I certainly think there is some truth to it. Many people see it as an extension of their academic careers, which is very similar to Bell Labs' industrial environment, except Google has more resources to achieve their goals. As a result, people are researching many projects. Transformers are one of the key inventions driving this field, and it is just one of around a hundred projects being researched.

If you are in the upper echelons of the organization, you will not be able to reasonably differentiate it from the other 99 projects in development. It's like a competition of numerous innovations. I think it was at that time that I first proposed the Scaling Law, which states that we need to scale and integrate these innovations on a large scale.

In theory, Google is the best place to do this. They have the world's largest cluster, a large number of talented engineers, and all the necessary elements. However, Google's organizational structure is for search services, and I do not think it is necessarily meant to combine all these parts and expand on a large scale into something completely different from the previous business.

Noah Smith:

Just as Bell Labs was not established to invent computers and give everyone a computer, it was to connect everyone.

Dario Amodei:

That's right, it was a telephone company. So, I cannot speak for Google, but obviously, in addition to inventing all these amazing things, they are also one of the four companies with cutting-edge models, both our partners and our competitors. I know many smart people there.

But I think you are right, there was a time when if they could combine these elements in the right way, they might become the only dominant player. But for whatever reason, things did not develop in that direction.

Scaling Law: As the data scale increases, the model's capability becomes stronger

Noah Smith:

This raises another question we are thinking about. In fact, the idea of talking to you comes from the content we discussed in another podcast, where we were mainly discussing the economics of internet business, and then someone raised some pessimistic views on the AI business, questioning how much economic moat AI companies really have.

Clearly, this is closely related to companies like Anthropic and other companies we call startups, but they have become quite large. So please tell us your views on the economic moat of AI companies.

Dario Amodei:

I would say I want to divide this issue into two branches. I think it is somewhat difficult to completely separate the discussion of Scaling Law from business issues. So, let's consider the case where Scaling Law holds in a very strong form, and then consider the case where it may partially hold or not hold at all. If it holds in a very strong form, the situation is this:

Now, you train a model with a billion parameters, and its capabilities are equivalent to those of an excellent college freshman;

Then you train a model with ten billion parameters, and its capabilities are equivalent to those of an excellent undergraduate student;

You train a model with a hundred billion parameters, and its capabilities are equivalent to those of a top graduate student;

When you train a model with a trillion parameters, its capabilities are equivalent to that of a Nobel Prize winner.

Then you put this model into use, basically serving everyone. It will become your colleague, your personal assistant, helping national security and biological research.

I think in such a world, this system and the products based on this system will have a large share in the economy. There is still a question, where will the profits flow? Will they flow to NVIDIA, artificial intelligence companies, or downstream applications? In such a large market, my initial answer is that the profits will flow to all of these places.

The future of AI will be similar to the solar energy industry.

Noah Smith:

But think about solar energy, which will obviously become very important. The more energy we need, the wider the application of solar energy will be. However, it is difficult to say which solar energy company is making a lot of profit. Solar energy is a highly commoditized product, despite the many innovations, there is no brand effect, no network effect, and no lock-in effect. It is difficult for any solar energy company to make profits from this, and this is completely changing the world right in front of us.

Therefore, I'm not entirely sure that just because everything will thrive like solar energy does now, it will necessarily lead to companies making profits. However, of course, I am open to this possibility. I just want to know, what do you think the source is? Why is the development of artificial intelligence different?

Dario Amodei:

Solar energy? I think there are two points here, because I think this is an important question in most of the world. Maybe I just want to say that if the Scaling Law is correct, this will be a very large market. Even if only 10% of the profits flow to a part of the supply chain, it is still huge.

Just like you enlarging the "cake", this has become the most interesting issue, even though those deciding how to allocate the dollar bills will certainly care very much about where the trillion dollars are going. But let's go back to your question, because I think it's important all over the world. The key to the question is how big the "cake" you are distributing is.

First, in terms of models, it depends on the Scaling Law. If we are building models with billions or tens of billions of parameters, there may not be more than four or five entities (perhaps some state-owned enterprises) involved. So what we see may be more like an oligopoly rather than a complete monopoly or complete commodification.

I have a question about whether someone will release an open-source model with billions or tens of billions of parameters. I am skeptical about this, even if such models are released, the operating cost of these large models for inference is very high. The majority of the cost is in inference rather than model training. If you have a better inference method, even if it's only 10%, 20%, or 30% improvement, it can still have a big impact. Economically, this is a bit strange, it's a huge fixed cost that you have to depreciate, but there's also a unit cost for inference, and assuming widespread deployment under this assumption, the difference will be significant. I'm not sure how this will develop.

Noah Smith:

This is actually similar to the economics of heavy industry, such as the way steel is manufactured.

Dario Amodei:

Yes, a bit. Interestingly, one other thing I want to mention is that in these models, we are already beginning to see models having different personalities. Therefore, commodification is a possibility, but even in an oligopoly, the deployment of certain models may also be commodified, although I'm not certain.

But a force that opposes this view is: Hey, I made a model that is good at programming, you made a model that is good at creative writing, and a third person made a model that is good at entertainment. These are choices, once you start making these choices, you start building infrastructure around them, which seems to create the conditions for a certain degree of differentiation.

Another factor that could lead to differentiation is product based on model construction. In theory, you can separate the model layer from the product layer, but in reality, they are interconnected, and cross-organizational work may pose certain challenges. Therefore, although there is a common logic in the model aspect, many companies are moving in the same direction, increasing multimodal functions to make the model smarter and inference faster, but the products are so different.

If you look at the 'Artifacts' project we are doing, it is a way to write real-time visualized model code. We do it this way, OpenAI has their own way, and Google has their own way. I think this is also one of the sources of differentiation between companies.

We have found that the economics of model-based sales applications is becoming thicker even for relatively thin applications.

Erik Torenberg:

If Scaling Law holds true and things get as big as we think they will, do you expect these companies to be nationalized at some point? Or what do you think?

Dario Amodei:

We can divide it into two scenarios: one is if Scaling Law is true, and the other is if Scaling Law is false. If it is false, then it is just a technology, like the internet or solar energy, which may be more important than most technologies, but not unprecedented. Based on the current development, I don't think it will be nationalized.

If it is true, the models we are building will be as excellent as Nobel Prize-winning biologists and top industry coding personnel, or even better. I'm not sure if they will really be nationalized. We will be very concerned about whether competitors can keep up with us or if we can deploy them as quickly as our competitors do.

The Scaling Law affects AI in creating new business models.

Noah Smith:

I have a question about the impact of artificial intelligence on business models. You know the story of electrical utilities, basically when they first got electricity, manufacturers tried to dismantle their steam generators because they were inefficient. But later someone discovered that electricity could be run in parallel to multiple workstations, which changed the way manufacturing worked, shifting from one large assembly line to multiple small workstations, leading to significant productivity improvements over the decades.

I've always suspected that AI is similar in this regard. I think the internet is similar too. The similarity between AI and the internet is that initially everyone seemed to think of AI as one person. Some people actually compared the number of AIs to the number of human employees, which doesn't make sense to me because AI cannot be divided into individuals.

You can create an agent-based system to mimic this way, but why bother? I see everyone considering AI as a direct replacement for humans, and my argument is that this is the first stage, just like how electricity directly replacing steam boilers was not a good idea. I think people will be somewhat disappointed because this direct replacement of humans is only effective in a few cases, such as customer service and some other well-defined tasks.

But I think this direct replacement of humans is only effective in a few cases, and then we will experience the bursting of the Gartner hype cycle.

Some creative entrepreneurs will say that we are not just using artificial intelligence as a substitute for humans, but using it to create new business models. Then we will see a period of renaissance and prosperity, which is my prediction. My Gartner-style prediction, am I crazy?

Dario Amodei:

So I think this is a mixture of things I agree with and things I may disagree with. First, I basically agree that if you freeze the quality of the current model, what you said is correct. We basically observe similar things in business activities. We provide models that can be conversed with, but also sell models to many customers through APIs. It took people a long time to figure out the best way to use models.

There are many issues regarding the reliability of models, which I think are reasons for concern, such as a model giving the correct answer 95% of the time, but not giving the correct answer 5% of the time. How to detect these situations and how to handle error handling is very important. This is very different from being theoretically useful and practically useful.

We had a feature early on that allowed the model to write some code, and then you could paste the code into a compiler or interpreter to create JavaScript video games. When problems occurred, you could go back to the model and make corrections. We also saw large models coordinating with small models, which is very different from viewing the model as a person's idea.

We have larger, more powerful models and faster, cheaper, and less intelligent models. Some customers find that large models can assign tasks to a large number of small models, and then report back to the large model to complete tasks like a swarm.

We are also exploring the best ways to use models. As models become more intelligent, their ability to solve these problems also becomes stronger. So ultimately, it comes back to whether the Scaling Law will continue. If they do, it will be a process you described. If they stop, innovation will also stop, and the process you described will end.

No one should believe that the Scaling Law will continue forever, it is just an empirical observation that could stop at any time. I have observed for ten years, and I guess the basis for them not stopping is just based on the length of observation time, it's just a 60-40 or 70-30 proposition.

Erik Torenberg:

What would change your mind? What would change your chances there?

Dario Amodei:

First of all, if we just train a model and then try the next scale of model, but the effect is very bad. We tried several times to solve the problem, but still failed, I would feel that, oh, I guess this trend is stopping.

If there is a problem with running out of data, and we cannot generate enough synthetic data to continue the process, at some point I would say, hey, this actually looks difficult, at least this trend will pause, maybe stop, but maybe not stop. I still guess that these things won't happen, but you know, it's a very complex problem.

AI can speed up the discovery of biology by 100 times and compress the progress of the century.

Noah Smith:

If the bottleneck of AI resources is more in terms of computing power rather than energy, then we will have more comparative advantages in using AI. Do you basically agree with this view?

Dario Amodei:

Yes, I think that makes sense. What you mean is, using a somewhat absurd analogy, if AI is like replicants and the process of manufacturing and nurturing them is very similar to humans, then we're in trouble. But if it's just a cluster of servers in some place, with completely different inputs, then we're fine.

I have not thought deeply about this issue, but at first glance it seems to make sense. If we are in a situation where AI is reshaping the world and the economic structure has changed, then we may be discussing some different things. However, if the conventional rules of economics still apply and I think they will apply for some time, then this sounds very reasonable.

Noah Smith:

But my other question is, is it necessary to consider an extremely abundant world? AI is so powerful that it provides us with amazing biology and manufacturing, making everything we want ten times, a hundred times, and so on.

Dario Amodei:

I think we have really underestimated the potential of AI in biology. Ten years ago, when I was in this field, the attitude was that the quality of data we obtained from biology was questionable, the amount of data we could obtain was limited, and experiments were often disrupted. Of course, more data analysis, big data, and AI are great, but at most they are just supporting roles. Maybe with the emergence of Alpha Fold, this situation is changing.

But my view is that AI models can play the role of a biologist or a co-biologist. If we consider truly advanced biology, it is really disproportionately powered by a few technologies. For example, genome sequencing, the ability to read genomes, is the foundation of modern biology. The recent CRISPR technology, the ability to edit genomes. 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 as an example, its assembly comes from the bacterial immune system and it took 30 years to invent. I think if we can greatly speed up the pace of these discoveries, we will also greatly speed up the pace of curing diseases.

My idea is, can we compress the progress of the 21st century? Can we achieve all the biological progress in the 21st century with a 10-fold acceleration from AI? If you think about all the progress we have made in biology in the 20th century and compress it into five to ten years, for me, this is a good thing. I think it's possible. We can cure diseases that have plagued us for centuries, which will greatly increase productivity, expand the economic pie, and extend human life.

Editor/ping