Written by: Techub News Compilation

Introduction

Recently, Dario Amodei, co-founder and CEO of Anthropic, gave an in-depth interview at its San Francisco headquarters. As a former core member of OpenAI, the head of the GPT-3 project, and the leader of Anthropic, an AI star company valued at several tens of billions, Amodei's views have always attracted attention. Recently, his comments on AI potentially replacing a large number of white-collar jobs and calls for stronger chip export controls have sparked widespread discussion both within and outside the industry, even drawing criticism from “doomsters.” In this conversation, Amodei addressed these controversies comprehensively, detailing Anthropic's technological path, business logic, and core judgments about the future development of AI.

Summary

• Dario Amodei strongly refuted the “doomster” label, emphasizing that he profoundly understands the immense benefits of AI, but for that very reason, risks must be warned about.
• He firmly believes that AI capabilities are following the “law of scaling” and are increasing exponentially, while the market generally underestimates the disruptive impact this exponential growth is about to bring.
• Anthropic focuses on enterprise-level AI applications, believing this is the business model that best reflects the value of the models and promotes the technology for good; its revenue has achieved 10-fold growth for several consecutive years.
• Amodei believes that the safety and capability development of AI are inextricably linked. Anthropic is committed to leading a competition of “excellence” rather than one of “inferiority” through responsible scaling policies and investments in explainability research.
• He criticized the misinterpretation of his motives by some industry leaders as “shameless lies,” calling for a more serious and honest attitude towards the high-risk, high-reward technology of AI.

Refuting “Doomsters”: To Truly Understand Benefits is to Seriously Warn of Risks

At the beginning of the interview, the host pointed directly to the core issue: Dario Amodei's recent numerous warning statements regarding the economic impact and safety of AI have led him to be labeled a “doomster.” Amodei reacted strongly, even using the word “angry.”

He recounted a personal family experience in response: his father suffered from a serious illness and passed away in 2006. However, just three or four years after his father's death, the cure rate for that disease jumped from about 50% to 95%. “I understand better than anyone how urgent the benefits of this technology are,” Amodei said, “When I wrote the article 'The Loving Machines,' I listed all the ways this technology can make life better for billions of people.”

Amodei believes that some people cheering for “acceleration” on social media do not genuinely care about the humanistic benefits of the technology but are simply immersed in the adrenaline rush. When these people label him as a “doomster,” they completely lose their moral credibility in his view. “I understand the benefits of this technology. You heard what I just said… I understand the benefits of this technology.”

He explicitly stated that his stance is not about slowing down technological development; rather, it is to ensure that technology can progress sustainably and safely. “I warn of risks so that we do not have to slow down. This way, we can invest in safe technologies and continue advancing the field.”

Firm Belief in Exponential Growth: The Market Generally Underestimates the Speed of AI Disruption

Amodei is regarded by many industry insiders as one of the most radical leaders among mainstream AI laboratories regarding estimates of the timeline for technological development. His explanation stems from a firm belief in the “law of scaling.”

He avoids using terms such as “AGI” (Artificial General Intelligence) or “superintelligence,” which he considers “meaningless” and “like marketing jargon.” However, he is convinced that AI capabilities are rising along a clear exponential curve: every few months, by putting in more computational power, data, and new training methods, we can obtain a model that is more powerful than the previous one. Initially, this was merely scaling during the “pre-training” phase; now it includes a second phase of “reinforcement learning” or “reasoning,” both of which are scaling simultaneously.

“People are not very good at understanding exponential growth,” Amodei pointed out, “If something doubles every six months, then in the two years leading up to it, it appears to have only completed 1/16 of the progress.” He believes we are currently in mid-2025, and the economic impact of model capabilities has begun to explode. He cited Anthropic's revenue as an example: growing from zero to $100 million in 2023, reaching $1 billion in 2024, and by the first half of 2025, growing from $1 billion to about $4.5 billion. “If this exponential growth continues for another two years (I am not saying it will definitely), you would reach hundreds of billions in scale.”

He likened this to the internet wave of the 1990s: at that time, the speed of the network and the underlying performance of computers rapidly improved, creating a global digital communication network in just a few years, while virtually no one foresaw its impact and speed. “That's my starting point… I think people have been fooled by exponential growth and do not realize how fast it can happen.”

Regarding the industry discussion about “diminishing returns to scale,” Amodei denies it based on the performance of Anthropic's models. He pointed out that in coding capabilities, from Claude 3.5 Sonnet to Claude 3.7 Sonnet, and then to Claude 4.0 series, each generation of models has made substantial improvements in coding, with benchmark scores and actual usage rising exponentially. “We see no signs of diminishing returns.”

Anthropic's Business Model: Betting on Enterprise-Level Applications, Aligning with Exponential Growth

When asked whether generative AI is a true business, Amodei answered affirmatively based on Anthropic's performance. He emphasized that Anthropic is the fastest-growing software company in history at that scale.

Unlike OpenAI's focus on consumer products (like ChatGPT) and Google's emphasis on integrating AI into existing product suites, Anthropic chose a different path: focusing on enterprise-level AI applications. Amodei explained that this includes enterprises, startups, developers, and advanced users who improve productivity using models. “Our perspective is that the business use of AI (which may even be larger than consumer use) will be greater.”

He believes that focusing on business use cases provides better incentives for model improvement. “A thought experiment: Suppose I have a model with a biochemical level equivalent to an undergraduate. Then I improve it to reach a PhD level. If I'm targeting consumers, maybe only 1% of consumers care about that. But if I go to Pfizer and say I’ve improved the model from undergraduate biochemical level to graduate level—that would be the most important thing in the world, and they might pay ten times for that.”

Therefore, making models smarter and more capable of solving global problems (such as biomedicine, geopolitics, economic development, etc.) and directly linking them to the value creation for enterprises is not only a more promising market but also better drives technology for good. “What I want to say is that we are betting on the business use of AI because it aligns best with the trend of exponential growth.”

Regarding Anthropic's pricing strategy and profitability, Amodei acknowledged that pricing schemes and rate limits are very complex, with early impacts on some usage patterns being underestimated and adjusted. He emphasized that from the perspective of the value created by the models, the cost of providing a specific level of intelligence will decrease, while the cost of providing cutting-edge intelligence may remain stable, but the economic value it creates will significantly increase.

Regarding the company's current unprofitability, Amodei provided a unique perspective: each model should be viewed as an independent “risk project.” The company might invest huge amounts of money every year to train the next generation of models (leading to accounting losses), but the models of the previous generation that have been deployed are profitable. “If the model stops getting better, or the company stops investing in the next model, you might have a viable business… but everyone is investing in the next model.” This reflects the industry's expectations and investments in exponential growth.

Intertwining Safety and Capability: Advocating for “Excellence” Rather than “Inferiority”

As a former key member of OpenAI and a driving force behind the GPT-3 project, Amodei's motivation for leaving and founding Anthropic has always been a focal point of interest. He revealed that the original construction of GPT-2 and GPT-3 was actually an extension of AI alignment work, aimed at large-scale research and application of technologies such as “reinforcement learning based on human feedback” to enable models to better align with human intentions.

This experience gave him a deep understanding that the “alignment” (safety) of AI systems is tightly intertwined with the development of “capability,” making it difficult to separate the two. What can truly guide the field towards a more positive direction is organizational-level decision-making: when to release, what to research internally, and what type of work to do on the system. This was one of the reasons he decided to start from scratch.

Amodei refuted the criticism that “he only believes he can safely build AI and thus wants to control the entire industry,” calling it “shameless lies.” “I have never said anything close to that… This is an incredible and malicious distortion.”

On the contrary, he proposed that Anthropic's goal is to lead a competition of “excellence.” “In ‘inferiority,’ whoever wins, everyone loses because you create unsafe systems… In ‘excellence,’ whoever wins, everyone is a winner.” Anthropic's approach is to set an example: to be the first to release “responsible scaling policies,” publicly share results of explainability research, share the “Constitutional AI” method, and publish assessments of dangerous capabilities, encouraging other companies to follow suit. “We try to set an example for the field… but there is an interplay in which being a strong commercial competitor is helpful.”

Resources and Talent: Firm Belief that Mission and Culture Cannot be Bought

Facing competitors like xAI and Meta that have the backing of trillion-dollar parent companies making massive investments in computational clusters, Amodei acknowledged that Anthropic has raised nearly $20 billion, which is “not insignificant,” while its data center scale in collaboration with Amazon is not inferior to other companies.

However, he believes that compared to pure computational resources, “talent density” is a more critical factor. The reason Anthropic can maintain capital efficiency (creating better models with fewer resources) lies in the gathering of top talent and efficient collaboration. He mentioned that despite facing poaching by giant companies offering high salaries, Anthropic's turnover rate is low, and many even refuse to meet with Mark Zuckerberg.

Amodei attributed this to the company's culture and mission alignment. “We refuse to compromise our principles… because we are confident that people join Anthropic because they truly believe in this mission.” He believes that competitors are trying to buy “consistency with the mission” with money, but that is precisely what cannot be purchased. “I am quite pessimistic about what they are trying to do.”

Views on Open Source and the Future of the Industry

Regarding the “disruption” risks that open-source models may bring, Amodei holds a different view. He believes that in the field of AI, the operation of open-source (more accurately, “open weights”) is different from traditional software open-source. The weights of models are not as easily understood and collaboratively improved as source code. Ultimately, large models need to perform inference in the cloud, which itself has cost and efficiency thresholds.

“When I see a new model, I don’t care if it’s open-source… I only ask, is it a good model? Is it better than us at certain things? That’s the only important thing.” He believes that considering “open-source” as the main dimension for thinking about competition is a “disruption factor.”

At the end of the interview, Amodei reiterated the severity of the current situation: trillions of dollars in capital stand on one side accelerating AI, while Anthropic, despite its high valuation, remains quite small relative to this force. He continually voices warnings about risks, even if it invites criticism from peers, government officials, and capital parties.

“We need more careful thought, more honesty, and more people willing to act against their own interests. We need people to truly invest in understanding the situation, truly do research, and truly bring some light and insight to our predicament,” Amodei summarized, “I am working hard to do this. If others also try to do the same, that would be very helpful.”

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support@aicoin.com，本平台相关工作人员将会进行核查。