Dario Amodei in conversation with JPMorgan CEO: Anthropic's competitive advantages, computational challenges and future bets.

CN
链捕手
Follow
20 minutes ago

Guest: Dario Amodei, Co-founder and CEO of Anthropic

Host: Jamie Dimon, Chairman and CEO of JPMorgan Chase

Translation: Jiahua, ChainCatcher

As Anthropic prepares for its IPO, CEO Dario Amodei engages in a public dialogue with JPMorgan Chase CEO Jamie Dimon.

Dimon mentioned that Anthropic has disclosed approximately $900 billion in forward commitments related to computing power, leasing, and data centers, and questioned how the company ensures a long-term supply of adequate computing power while controlling costs.

Amodei referred to computing power as the "deepest business challenge" faced by Anthropic: the company needs to procure computing power one to two years in advance, but it is challenging to predict when technological advancements can translate into revenue. Overbuying could put financial pressure on the company; underbuying could lead to insufficient ability to meet customer demand and affect the development of the next generation of models.

In addition to computing power and the IPO, the two also discussed Anthropic's corporate culture, the commercial applications of Claude, agent safety, price competition for AI models, and whether AI can drive medicine and scientific research forward.

Key Points:

  • Computing power is the biggest business challenge. Amodei stated that Anthropic’s revenue is expected to grow about tenfold in both 2024 and 2025, but the company struggles to predict the demand for the upcoming year, while computing power needs to be procured one to two years in advance.

  • Anthropic prefers a conservative approach to procuring computing power. The company locks in a portion of computing power at a lower cost in advance, and if the business growth exceeds expectations, they supplement from the spot market at higher prices. He likens this strategy to purchasing insurance.

  • One important reason for going public is to increase cash reserves. Amodei believes that the more cash a company has, the greater its buffer across different scenarios of computing power demand, making it more capable of seizing opportunities in fields like biology, robotics, and cybersecurity.

  • The competition in AI goes beyond chatbots. Amodei believes that each breakthrough in the capabilities of cutting-edge models could create new markets. Following chatbots and programming, scientific discovery, life sciences, robotics, and cybersecurity could become larger markets.

  • The key constraint for agents entering the market is safety. He stated that Anthropic chooses to serve corporate clients partly because companies value trust, long-term relationships, and reputation, which makes the commercial incentives of the company more aligned with safety goals.

  • The bottleneck in cybersecurity is shifting from discovering vulnerabilities to fixing them. Amodei stated that clients and Anthropic have each identified hundreds of thousands of vulnerabilities, but the pace of fixes does not keep up with discoveries. In the next 6 to 12 months, the application of defensive agent clusters may increase.

  • Anthropic intends to allocate 5% of the company’s equity to a public welfare fund. Amodei is concerned about the concentration of new wealth brought about by AI in a few companies; the fund may focus on issues such as workforce transformation and biosecurity, with specific directions still under research.

Talent and Culture are Anthropic's Competitive Advantages

Jamie Dimon: You previously worked at OpenAI and are an early participant in this field. Now there are many companies globally working on AI; how does Anthropic differentiate itself from others?

Dario Amodei: I have worked at several major companies in this field and have been pondering why some projects succeed while others do not. After many years, I concluded that talent and culture are the most important factors.

Talent and culture create technology, technology creates products, and products drive market expansion, sales, and other business aspects of the company. How the public views a company is also related to the people who create technology and their intentions.

When founding Anthropic, the other six co-founders and I decided to prioritize talent and culture from day one. Every candidate must undergo a cultural interview. We also maintain a firm rejection of technically capable individuals who do not align with the company culture.

We do not want individuals in our team who are disrespectful, only seeking the limelight, or unable to collaborate with others; we also wish for employees to identify with the public good. In the short term, adhering to these standards can sometimes make hiring difficult, but in the long term, it allows everyone to be on the same side.

A company's difficulties often stem from not lacking excellent people, but rather from excellent people not working together. Different teams can end up competing against each other, ultimately draining energy internally.

At Anthropic, when a significant objective arises—be it training a model, reaching safety goals, developing a product, or promoting policy work—the whole company can cooperate around that goal.

We also see positive results from employee retention. More employees from other companies joined Anthropic than left; all seven co-founders remain at the company; over three-quarters of the first 100 employees are still here. Even when someone leaves, they often come back. On one occasion, two key product personnel returned less than two weeks after leaving, stating that they didn't realize what they had before leaving.

I spend about 40% to 50% of my time on talent and culture. Every two weeks, I participate in a company-wide internal event called Dario Vision Quest, where I candidly inform employees about what's happening in the company, not hiding issues and avoiding bureaucratic language, and I encourage other managers to do the same.

The management can trust the employees, and employees believe the management will communicate openly. This trust is very valuable.

Which Industries is Claude Entering?

Jamie Dimon: JPMorgan Chase is one of Anthropic's major clients. Claude has brought us significant changes; currently, 250,000 employees use it multiple times a week in various ways. How are other Fortune 100 companies using Claude?

Dario Amodei: Claude is being applied across various industries. One of the most surprising characteristics of this technology is its versatility. You can think of it as a type of "Moore's Law" for cognition or intelligence.

Through interactions with clients, I have been exposed to many industries I previously didn't understand, seeing what Claude can do in these fields.

The earliest and fastest-adopting area is programming. Most of the co-founders and I have written a lot of code in the past, and seeing the tasks the models can now accomplish is still astonishing. Almost every Fortune 100 company has a large demand for programming, so programming is often where they first use Claude.

Subsequently, the application range has expanded to a broader scope of knowledge work. In the financial services sector, Claude can handle simple clerical tasks and assist with more complex strategic work. In insurance and legal fields, it can draft contracts and identify issues within them.

Life sciences are a particular area of focus for me, with applications including writing clinical research reports, assisting in basic research, predicting protein interactions, improving manufacturing efficiency, and aiding in foundational biology research.

In e-commerce, some companies have deployed Claude as a direct consumer-facing agent. The manufacturing sector is similar, with an increasing number of businesses using Claude to enhance efficiency; applications include semiconductor manufacturing and physical construction.

Agent Entry Requires Resolving Safety Issues First

Jamie Dimon: You mentioned agents. What preparations need to be made before allowing agents to access sensitive information in regulated companies?

Dario Amodei: The most important issue is safety. The capabilities of agents are rapidly advancing. Whether they can complete life sciences research or write complex software architecture is not the only issue; the real limitation on technology development, internal deployment, and customer use is how to use these capabilities safely.

We do not want agents to be uncontrollable. Thus, Anthropic has invested a significant amount of resources to ensure that appropriate safety and alignment mechanisms are in place from the model training stage.

We have a large alignment team that remains vigilant on these issues. Every few months, we revisit safety, security, and alignment work and increase our input.

Anthropic has focused on corporate clients from the start. We know that companies often value safety more than consumers, and corporate businesses align with our safety goals.

We value safety because we believe technology should be developed the right way. At the same time, we are also thinking about what business models can give us a stronger incentive to pursue safety. The answer is corporate business: companies value trust, long-term cooperation, and their own reputation, not wanting models to cause severe damage.

This alignment makes our commercial incentives more in line with the societal goals we wish to achieve.

Jamie Dimon: Can we truly control the permissions of agents? For instance, allowing them access to sensitive systems while ensuring they can only execute operations A, B, C, D, and nothing else?

Dario Amodei: This is somewhat akin to the development of autonomous driving technology.

Autonomous driving holds immense potential value and may ultimately be much safer than human driving. But first, the technology must reach a sufficiently reliable level, establish correct protective mechanisms, and design systems in the proper way.

We can never provide a 100% guarantee. You cannot say that an autonomous vehicle will never have an accident, regardless of how many people use it. However, we can strive to make systems considerably safer than human operation, achieving safety levels multiple times, hundreds of times, or even higher than human capabilities.

The $900 Billion Commitment: The Dilemma of Over or Under Purchasing Computing Power

Jamie Dimon: Anthropic has disclosed approximately $900 billion in forward commitments related to computing power, leasing, and data centers. I also saw a claim that about 80% of the computing power is used to train new models, with only around 20% used for inference, while we customers mainly utilize inference services.

How do you ensure you obtain the needed computing power in the future while maintaining cost efficiency?

Dario Amodei: Computing power is the deepest business challenge faced by Anthropic.

The technological development in this field is very rapid, but we do not know how quickly technological advancements will translate into commercial value, hence there is significant uncertainty.

In 2024 and 2025, Anthropic's revenue is expected to grow about tenfold each year. Yet each year we ask ourselves: What will happen next year? We do not know. As the company scales, the amounts involved become larger.

Meanwhile, computing power needs to be procured one to two years in advance. The amount to purchase depends on how much cost the company can bear.

Buying too much can place an excessive financial burden on the company; buying too little means insufficient inference services for customers. More critically, insufficient computing power can impact research and development, which determines what models we can release in six months or a year, and ultimately the company's future revenue generation capabilities.

We mainly expand our buffer space through two aspects.

The first is gross profit margin. The higher the gross profit margin, the greater the adjustment space when buying more or less computing power. Simply put, it reflects the difference between the cost of computing power and the revenue the company generates from utilizing that computing power. The prospectus shows that our gross profit margin is quite strong, especially compared to other companies in the industry.

The second is cash on the balance sheet. We currently have a substantial amount of cash and hope to continue increasing our cash reserves.

However, even with these buffers, the company's forward commitments remain substantial. Although these commitments span multiple years and revenue is rising rapidly, the amounts are still significant.

Therefore, our strategy is to relatively conservatively procure computing power: purchasing a portion of computing power at lower costs in advance. If business performance falls below expectations, we have controlled costs; if performance exceeds expectations significantly, we can purchase additional computing power from the spot market.

Spot computing power is more expensive, but we will only buy it when business performance is especially good. This is somewhat like buying insurance: to ensure coverage under various circumstances, one must pay for additional flexibility.

We believe this strategy allows the company to continue investing in research and development when needed and to seize growth opportunities while also providing protection when business performance does not meet expectations.

More Affordable Models Are Emerging; Will AI Become a Commodity?

Jamie Dimon: More and more companies are purchasing tokens. People quickly realize that many tasks can be done without expensive models; they can be accomplished with more affordable models and tokens. This segment of services may gradually become standardized commodities. What do you think about this assessment?

Dario Amodei: First, every time the capabilities of a cutting-edge model expand, it may create entirely new markets.

The market development path people are familiar with is this: early on, there are many startups willing to pay high prices; as the market matures, products gradually commoditize, and discussions about moats and cheaper alternatives begin.

In a specific application of AI, this may indeed occur. Chatbots are an example. But AI is not just one application or one market. As smart capabilities continue to improve, new markets will constantly emerge.

In 2023, everyone was focused on chatbots. Anthropic, upon seeing the model's capabilities, believed that programming would become a larger market, as the model was beginning to become smart enough to handle programming tasks.

Thus, we are focusing on the cutting edge of capabilities, rather than competing for the chatbot market or building moats around it.

Now, AI programming is gradually maturing, but there is still a long way to go. Looking at the penetration rate among Fortune 500 companies, there remains a lot of work to be done.

Meanwhile, I believe the market's focus will shift toward new areas. These areas may be larger than programming, just as the programming market is larger than the chatbot market. Scientific discovery, life sciences, physics, materials, robotics, and cybersecurity all belong to these areas. They face more complex challenges than ordinary programming, hence we will continue to focus on the cutting edge of capabilities.

Second, even for tasks that do not require cutting-edge models, we can leverage large models to train smaller, more efficient models.

We improve the efficiency of large models and can also apply these technologies to small models. For example, Haiku 5.5 is priced at $0.10 per million input tokens and $0.50 per million output tokens. We are also capable of producing low-cost models and competing with other vendors in terms of efficiency.

Can AI Drive Breakthroughs in Medicine and Science?

Jamie Dimon: Can AI help humanity cure cancer and other diseases or drive major breakthroughs like nuclear fusion and quantum technology?

Dario Amodei: We repeatedly observe similar technological advancement processes across different fields. Take mathematics, for example; in 2023, models struggled with even high school math problems. People generally believe chatbots are good at communication but poor at math and logic.

By 2024, models began to perform reasonably well on high school math problems. By 2025, they started solving some previously unsolved mathematical problems. Initially, there may be one or two cases, and people might say these problems could appear in the training data or that the models were just lucky.

But soon, the models began to solve more and more problems, also tackling some more significant and challenging questions.

Now, we are seeing early signs of this similar process in new fields like biology and robotics. Experts tend to be very skeptical, especially in the pharmaceutical industry. They have seen many technologies trying to solve industry challenges in the past and know how difficult these problems are.

In the life sciences field, people may say that bottlenecks exist in clinical trials, manufacturing, or FDA approvals. But we see that AI is gradually helping to address issues in these processes. I believe this change may become more evident in the coming years.

We cannot determine the outcomes in advance. In a specific field, skeptics could be right. But I believe the potential for AI to drive significant progress in biology is enormous.

Finding Vulnerabilities is Just the First Step; Fixing Them is Harder

Jamie Dimon: Prior to the emergence of Mythos and agents, cybersecurity was one of the biggest risks JPMorgan Chase faced. We have thousands of people working on cybersecurity, and we may need even more to fix identified vulnerabilities.

Now, ongoing agents may attack various security boundaries of enterprises. How should companies, vendors, and the entire industry prepare for this?

There are still some key industries that lack sufficient protective capabilities, and we need to assist them in catching up.

Dario Amodei: First, discovering vulnerabilities is just the first step.

After the launch of Mythos, we invested heavily in finding vulnerabilities. Clients discovered hundreds of thousands of vulnerabilities, and we identified hundreds of thousands as well. The next bottleneck became how to fix these vulnerabilities.

Fixing is harder than discovering. We have identified many vulnerabilities, but neither our clients nor we can fix them quickly enough. Therefore, Claude is gradually improving its ability to fix vulnerabilities, and we are also starting to offer more related services.

The third step is to write safer code from the beginning. As Claude's programming capabilities improve, we are also specifically training it to excel in writing secure code.

The fourth step is real-time defense. Cyberattacks can occur in hours, minutes, or even seconds. If we can deploy a cluster of defensive agents to continuously patrol enterprise servers and systems and resist attacks in real time, it will be very valuable.

This approach has not yet been widely adopted, but I believe that in the next 6 to 12 months, there will be an increasing number of applications. Because attackers will also use clusters of agents, defenders also need clusters of agents.

Why Choose to Go Public Now?

Jamie Dimon: Anthropic is about to go public. Why now? Going public has its pros and cons; do you think it will make the company better or make management more difficult?

Dario Amodei: There are several reasons for going public, the most important of which is the investment in computing power.

There are huge economic opportunities in the fields of biology, robotics, cybersecurity, and broader knowledge work. We want to invest enough resources to ensure we can seize these opportunities.

The more cash we have, the greater the buffer space between lower and higher computing power demands. This is also a critical reason for going public.

As for the specific month, day, or market conditions under which we go public, we did not particularly agonize over that.

Before starting the public listing process, I had never experienced anything like this. Anthropic took about six years to evolve from a startup to where it is today. I sought advice from many people, and one that resonated with me the most was: Do not let the company's valuation exceed its true value by even a dollar.

This is very close to one of Anthropic's principles: make fewer promises and deliver more.

We typically do not speak in advance about what capabilities a model will have before it is completed. We first make the model, confirm it is good enough, and then release it. The same applies to safety systems. I hope the public offering will adhere to this principle as well.

Allocating 5% of Equity to Establish a Public Welfare Fund, Concerned about Wealth Concentration in AI

Jamie Dimon: Anthropic also announced the establishment of a public welfare fund. Many people may not understand why you are doing this. What are your thoughts on the fund?

Dario Amodei: We have decided to allocate 5% of the company’s equity to a public welfare fund.

In addition, the co-founders have also pledged to donate 80% of their personal wealth. Since the company's inception, we have also set up an employee donation matching program: if employees donate a certain percentage of shares to charitable organizations or reserve them for public welfare purposes, the company will provide additional shares for matching.

These arrangements reflect a common idea: AI is an incredibly powerful technology. It has already created substantial economic value for Anthropic and the entire industry, and we hope future opportunities will be even greater, but we cannot be certain.

I hope AI can help humanity cure cancer and Alzheimer’s disease, bring about cheaper and more abundant energy, better materials, and provide more people with enriched material lives.

But we are also concerned that if a large portion of the new wealth originates from a single technology and concentrates in a few companies, society might simultaneously experience rapid growth and concentrated wealth.

We hope to alleviate the problem of wealth concentration as much as possible, and the public welfare fund is part of that effort.

The specific direction of the fund is still under research. We won’t announce a long list of problems we intend to solve just yet; rather, we want to first clarify which issues are worth investing in.

Potential directions include helping society smoothly cope with workforce transformation brought about by AI and ensuring biosecurity. AI could bring significant advancements to biology, but like network technologies and other technologies, it could also be misused. We need to prevent misuse and establish effective defensive mechanisms.

In 15 Years, I Hope Diseases Become History and AI Does Not Go Out of Control

Jamie Dimon: One last question. If you look back at today in 15 years, what would you most like to have happened, and what would you most like to have not happened? Is there anything we haven't discussed today but you want to convey to everyone?

Dario Amodei: The most important thing is to ensure that this technology is developed correctly.

I have discussed many technical risks, including their potential use for biological or cyber-attacks, causing economic shocks, and humanity losing control over agents. These risks are very serious and not mere marketing phrases. Anthropic has taken these issues seriously since its inception and invested resources in addressing them.

We hope to set an example for the entire industry on how to deal with risks. But doing well solely within Anthropic is not enough; we also want other companies, clients, and suppliers in the AI industry to act responsibly.

Secondly, we need to genuinely realize the positive value of AI.

In biology and medicine, we are just beginning to see the potential of this technology. There are still many skeptics among pharmaceutical companies, early-stage biotech firms, and academia. We cannot know the results in advance; skeptics may indeed be right, but I hope to ultimately prove that their concerns were unfounded.

I hope that one day, diseases like heart disease, Alzheimer’s, and cancer can become things of the past, just like the Black Death and smallpox today.

This may sound a bit idealistic or even crazy, but I truly believe it is achievable. I don't know whether it will take five, ten, or fifteen years, but we will do our utmost to promote the realization of this goal as early as possible.

免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到[email protected],本平台相关工作人员将会进行核查。

Share To
APP

X

Telegram

Facebook

Reddit

CopyLink