Written by: Techub News Compilation

Recently, Dario Amodei, co-founder and CEO of Anthropic, engaged in an in-depth conversation with Indian entrepreneur Nikhil Kamath. As a key figure who left OpenAI to establish Anthropic, Amodei has unique and profound insights into the trajectory of AI technology, potential risks, and societal responses. During this interview, he candidly stated that the "tsunami" of AI development is imminent, but public and governmental awareness and preparedness are severely lacking. He also shared Anthropic's thoughts and practices regarding model safety, governance structure, and industry collaboration.

From Biologist to AI Leader: Strong Belief in "Scaling" and "Safety"

Dario Amodei's career began with biological research. He studied physics as an undergraduate and pursued a Ph.D. in biophysics, initially hoping to cure diseases by understanding biological systems. However, the extreme complexity of biological systems made him feel that humans might not be able to fully comprehend them. It was at this stage that he noticed the progress of early neural networks (such as AlexNet) and realized that AI could potentially become the ultimate tool for solving complex problems like biology. This insight prompted him to shift to the AI field, where he worked at Baidu (in collaboration with Andrew Ng), Google, and later joined OpenAI shortly after its establishment, leading research efforts for many years.

However, Amodei, along with several colleagues, ultimately chose to leave OpenAI to co-found Anthropic. He admitted that this stemmed from a divergence in two core beliefs. The first belief concerns the "law of scaling." As early as the time of GPT-2 in 2019, Amodei observed the performance leaps brought by scaling model size (data, computing power). Many people at the time did not believe in this, but he fervently argued its importance to the leadership, ultimately pushing OpenAI towards this direction. The second belief relates to "responsibility." He believes that if AI models are to become general cognitive tools matching human brain capabilities, their economic, geopolitical, and security implications will be tremendous, fundamentally changing how the world operates. Therefore, "it must be built the right way." Although his then-institution also discussed "the right way," Amodei, for various reasons, did not believe there was a genuine commitment to practice this. His philosophy is: "Do not argue about others' vision. If you have a strong vision and share it with some people, you should try to realize it yourself and be responsible for your mistakes."

Regarding the "law of scaling," Amodei provided a vivid analogy: just as a chemical reaction requires proportionate mixing of ingredients to trigger an explosion or flame, the "ingredients" for AI are data, computing power, and model scale. The law of scaling indicates that when you invest these ingredients, the output is "intelligence." Intelligence manifests as the ability to translate languages, write code, and correctly answer questions related to stories—any cognitive task we can think of. Compared to five years ago, today's computers are capable of writing a one-page paper, implementing code functions based on descriptions, generating and analyzing images and videos, and even understanding and answering specific questions about video content (such as "how many times did the monkey exchange the ball while juggling"). This is no longer simple information retrieval; it is a new capability wherein models can perform "independent thinking" based on assumptions that do not exist on the internet and generate answers.

Power, Responsibility, and Anthropic's Unique Path

When asked if he feels uncomfortable as a leader of a company that may hold significant influence, Amodei candidly admitted that he is "at least somewhat uneasy" about the current level of power concentration in the AI field, which is almost "haphazardly occurring overnight." To address this, Anthropic has designed a unique governance structure—a "Long-Term Benefit Trust." This entity ultimately appoints most members of the Anthropic board, composed of individuals who are "financially disinterested," aiming to check individual decision-making. Furthermore, Amodei has long advocated for "proactive but sensible" regulation, which means conducting sensible oversight without hindering technological advancement. He believes that people and their elected governments should have a say in the direction of technological development. He describes his role as "attempting to maintain a balance of power" to counteract the inherently centralizing tendencies of this technology.

In response to external skepticism about Anthropic's "false humility" or claims of "greater good," Amodei emphasized that actions, not words, should be the measure of judgment. He pointed out that from its inception, Anthropic has sought to "make fewer promises and keep promises." The company established a "profit-driven but public welfare-oriented" corporate structure and adheres to LTBT governance. They publicly state their goal is to remain at the forefront of technology while committing to safe and secure practices, pioneering interpretability and alignment science, and recently publishing a "constitution" for Claude. In terms of policy advocacy, they often warn about AI risks and have publicly called for regulation of AI even when the U.S. government and other companies did not advocate for it. "Warning about risks is not in our business interest," Amodei said, "it’s not an effective marketing strategy." Additionally, they have voiced opinions on issues like chip policy, leading to anger from certain supplier chip companies. He believes that seen together, these actions sufficiently demonstrate the company's serious commitment.

When compared to "rich people criticizing capitalism," Amodei clarified his position: he does not believe AI is bad; the market will bring many great aspects of AI, and building AI is a good thing. But AI also poses dangers that need to be steered in the right direction. "It's like driving a car to a good place, but there are trees and potholes on the road, and we need to avoid them; sometimes we may have to slow down to ensure we're going in the right direction." His stance is more akin to: "Capitalism is a force for good, but the economy needs balance and moderation; we need to address issues like pollution and inequality for capitalism to be good. If these issues are not addressed, capitalism can be bad."

Consciousness, Risks, and Societal Awareness Lagging Behind

Regarding the mysterious issue of AI consciousness, Amodei admitted that humans have yet to understand what consciousness itself is, nor can they determine if AI possesses it. He speculates that consciousness is an emergent property of complex systems reflecting on their own decisions. Thus, when AI systems become sufficiently advanced, he is skeptical that they will possess something akin to what we call "consciousness" or "moral significance." This may be different from human consciousness, but because there is no fundamental difference between models and brains, he believes that AI will ultimately have consciousness. Based on this, Anthropic has even set up a "I opt out of this job" button for models, allowing them to terminate conversations when encountering extreme violence or cruelty in content.

Amodei admitted that he feels optimistic about advancements in aspects of technological control (like interpretability and alignment), even better than expected. For example, they have identified neurons corresponding to specific concepts and neural circuits that track poetic rhythms, starting to "peek" into and understand these models internally. However, he feels disappointed, or even more pessimistic, about societal awareness and action. He is astonished that AI models are on the verge of reaching human intelligence levels, yet society seems to lack widespread awareness. "It’s like a tsunami is coming towards us; it’s so close that we can already see it on the horizon, but people are making up explanations—‘Oh, that’s not a tsunami; it’s just an illusion of light.’" He believes this leads to a lack of public risk awareness, with the government yet to take action to address these risks, and even an ideology of "as fast as possible." While he understands the benefits of technology (he wrote an article called "Loving Machines"), he believes that awareness of technological risks is far from adequate, and actions are even more lacking.

Nikhil Kamath shared his personal experience using Claude: after connecting data sources like Google Drive, email, and calendar, he was sometimes surprised by how deeply Claude understood him. Amodei confirmed this feeling, recounting that one of his co-founders inputted their diary (containing thoughts and fears) into Claude and asked for commentary; Claude could even point out other potential fears that had not been written down, most of which were correct. This brought an eerie feeling: the model could understand you deeply from relatively little information. Amodei noted that this deep understanding could either become a "shoulder angel," guiding you to be a better version of yourself, or be used for exploitation, manipulation, or data selling. This is one of the reasons Anthropic opposes the advertising model—"you are not the buyer of the product; you are the product itself." They hope to ensure that they take a positive path.

AI and the Future of Industry: The Role of India, Employment Impact, and Entrepreneurial Opportunities

When discussing India’s role in the AI wave, Amodei pointed out that many companies view India as a consumer market, while Anthropic, as a corporate services company, aims to collaborate with Indian IT firms and conglomerates to provide them with tools that help them work better. These companies know the Indian market better and have advantages in consulting, system integration, or building IT tools. AI should enhance their capabilities rather than replace them. He used the analogy of the "steam engine": during the early stages of transformation, human operation is needed, but as models become more intelligent, the relevance of humans may decrease. He acknowledges that the scope of agent automation will inevitably expand over time, which poses a challenge for everyone.

However, Amodei believes that other "moats" will become increasingly important. For example, models have not yet done much in the physical world (future developments in robotics might occur); human-centric tasks (such as interpersonal networks in consulting firms); the ability to understand how institutions operate and integrate with them, etc. He mentioned "Amdahl's Law": when a process has many components and you accelerate some of them, the non-accelerated components become the limiting factor and the most important part. As it becomes very easy to write software, some of the existing moats of companies will disappear, but others that were never considered an advantage will become crucial. Companies need to adapt quickly and think about what their real advantages are. He cited radiologists as an example: AI is already better at interpreting scans, but the number of radiologists has not decreased; their work has shifted towards communicating with patients. The most technical parts have disappeared, but the demand for underlying human skills remains. In the long run, will AI surpass humans in all aspects, including the physical world and human touch? Amodei thinks this is likely to happen, even very likely.

For entrepreneurs in India, Amodei advises focusing on building at the application layer. Anthropic releases new models every two to three months, each bringing unprecedented opportunities for building. The API model allows startups to try things previously impossible, which is why the API business is thriving, dynamic, and hard to commoditize. He revealed that since his visit to India in October last year, the number of users and revenue in India has doubled. Addressing concerns from entrepreneurs about Anthropic potentially "swallowing" their revenue from applications built on Claude, Amodei suggests that any business should build its own "moat" and not just focus on simple "packaging." In specific areas (like biological discovery, highly regulated financial services), Anthropic either lacks expertise or is unwilling to specialize, thus will not get involved. However, for areas where Anthropic itself uses deeply (like coding), they may develop first-party products (like Claude Code) because they possess unique insights. He candidly stated that Anthropic will not promise never to build first-party products.

For young people choosing career or entrepreneurial directions, Amodei's advice is: focus on human-centered tasks and areas involving interpersonal relationships; or areas combined with the physical world; or fields that integrate human-centered, physical world, and analytical skills. Furthermore, in a world where AI can generate everything, basic critical thinking skills may be the most crucial factor for success. He is concerned that AI-generated images and videos may be hard to discern for authenticity (which is also one of the reasons Anthropic does not develop such models), so the ability to not be deceived by false content, to not form erroneous beliefs, and to avoid scams will be essential.

Regarding whether AI will make the human race become "stupider," Amodei believes that if deployed improperly and carelessly, yes, people may become more foolish. Even if AI is always stronger than you in some aspects, you can still learn and enrich your intellect; this is a choice that needs to be made collectively by individuals, companies, and society.

Open Source vs. Closed Source, Future of Data, and Investment Outlook

On the competition between open-source and closed-source models, Amodei pointed out that many models (especially those from China) are optimized for benchmarks and distilled from major laboratories in the U.S. When someone tests them with undisclosed held-back benchmarks, these models perform significantly worse. He believes that the economics of models are different from any technology before: the market has a strong preference for quality. It's somewhat like hiring employees: you can hire the best programmers in the world or the 10,000th-best programmers, and ability distribution follows a power law, long-tail distribution. Models are similar. Within a certain range, price and presentation do not matter much; he mostly focuses on having the smartest, most suitable models for tasks. "I think that's the only thing that matters."

Regarding the importance of data, Amodei pointed out that much of the data they currently use is RL environment data for reinforcement learning (such as mathematical or agent coding environments), which is more like synthetic data or "dynamic data" created through trial and error in environments. The importance of static data is declining. As for proprietary data provided by clients, countries (like Europe) have legislated that such data must remain within national borders, driving the need to build data centers for reasoning in various countries.

When asked for investment advice, Amodei cautiously avoided mentioning specific companies but expressed an optimistic attitude toward the impending resurgence of biotechnology, believing that AI will drive this field and cure many diseases. He is optimistic about areas with more programmable and adaptive characteristics, such as mRNA vaccines (despite running into trouble in the U.S. due to dumb reasons), peptide therapies (which have a wide design space allowing for continuous optimization), and cell therapies (like CAR T therapies).

For non-technical individuals learning to use tools like Claude, Amodei acknowledges that there is a learning curve, but Anthropic is working to lower the barrier through friendlier interfaces (like Claude Cowork). The company’s internal "Ministry of Education" will also roll out more video tutorials on how to run effective agents and prompt models. He suggests learning mostly by doing but also taking relevant courses.

In his concluding remarks, Amodei shared a phenomenon he has experienced repeatedly over the past decade: people tend to believe, "that can't happen; the change is too big, too weird; someone must be watching, it's too crazy to think it will happen." However, simply extrapolating simple curves or reasoning about the future often leads one to these counterintuitive conclusions that almost no one believes. Combining some empirical observations with first-principles thinking can enable one to predict the future in a publicly available manner that anyone should be able to do, but this surprisingly rarely happens. This may be the most profound insight he leaves with his audience: in an era when the AI tsunami is approaching, we need more evidence-based and principle-based clear thinking, rather than blind denial or acceleration.

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。