Written by: Techub News Organization
Recently, Dario Amodei, co-founder and CEO of Anthropic, engaged in a 142-minute deep conversation on the well-known podcast Dwarkesh Podcast. As the former Vice President of Research at OpenAI and the leader of one of today's most attention-grabbing AI companies, Amodei's thoughts represent the core cognitive evolution at the forefront of AI. In this densely informative interview, he not only reiterated his well-known "Compute Large Blocks Hypothesis" but also candidly stated that we are "approaching the endpoint of exponential growth" and made bold predictions about the timing of AGI's arrival, its economic impact, and even the geopolitical landscape.
Endpoint of Exponential Growth and AGI Timeline
When asked about the biggest changes in the past three years, Dario Amodei admitted that the exponential progress of technology itself roughly aligns with his expectations from three years ago. From GPT-1 to the capabilities displayed by Claude 3.5 Sonnet today, the model is steadily advancing along the trajectory from "smart high school student" to "smart college student," and then starting to engage in PhD and professional work. The coding capabilities have even surpassed this trajectory.
What truly felt "absolutely insane" to him was the general lack of public awareness regarding "how close we are to the endpoint of exponential growth." He believes that while the world continues to debate those age-old political issues, a fundamental technological transformation is imminent.
Regarding the timeline for AGI, Amodei provided a clear probability judgment. He believes that achieving what he calls the "genius nation in data centers" within the next decade has a probability as high as 90%. "It is hard to exceed 90% because the world is too unpredictable," he added, with the remaining 10% risk coming from "irreducible uncertainties" such as internal company turmoil or conflicts in the Taiwan Strait leading to the destruction of semiconductor fabs. For verifiable domains (such as code generation), he is even more optimistic, believing that excluding extreme uncertainties, the goal can be achieved within one or two years.
He differentiated the "spectrum" of capabilities: from AI writing 90% of the lines of code, to completing 100% of end-to-end software engineering tasks, to ultimately leading to a 90% reduction in demand for software engineers. He emphasized that these are different milestones, but we are "moving through these stages extremely quickly." He acknowledged that currently, AI brings more of a "productivity boost" rather than a "renaissance" of software creation but is convinced that this boost will soon translate into tremendous economic value.
"Compute Large Blocks Hypothesis" and the Extension Law of RL
Amodei reiterated his core hypothesis held since 2017—the "Compute Large Blocks Hypothesis." This hypothesis posits that what truly matters are a few fundamental factors: the raw amount of computation, the amount of data, the quality and distribution of the data, the duration of training, a target function that is infinitely scalable (such as pre-training objectives or RL objectives), and normalization or conditioning techniques that ensure numerical stability. He firmly believes that all "smart" skills and methodologies become irrelevant in the face of vast amounts of computation and data.
In response to the host's query about whether "Reinforcement Learning (RL) lacks publicly available extension laws," Amodei provided a negative answer. He believes that the extension rules of RL are no different in essence from pre-training. Just like pre-training gained strong generalization capabilities after expanding from a narrow text distribution (like fan fiction) to scraping data from the entire internet, RL is undergoing a similar trajectory: starting from simple tasks like math competitions, gradually expanding to broader tasks like coding, and beginning to exhibit generalization abilities.
He refuted the viewpoint that "RL requires specific environments to teach specific skills, proving our lack of core human learning algorithms." He argued that this confuses several different issues. Models absorb massive amounts of data during the pre-training phase, and this process is more akin to being between human "evolution" and "lifelong learning." In-context learning lies between human long-term and short-term learning. While models learning from scratch indeed require more data than humans (lower sample efficiency), once trained, they exhibit very strong learning and adaptability within long context windows.
"The goal is not to use RL to teach the model every possible skill, just as we wouldn't try to expose the model to every possible way of combining words during pre-training," Amodei explained, "the model is trained on a multitude of things and then achieves generalization."
Economic Diffusion: Fast, but Not Infinitely Fast
A key point of contention is the rate at which AI capabilities translate into actual economic impact. Amodei acknowledged the existence of a delay in "economic diffusion," but he strongly opposed viewing it as a kind of "cop-out." He cited Anthropic's own growth as an example: the company's revenue grew from zero in 2023 to $100 million and is projected to reach $1 billion in 2024, with expectations of $9-10 billion in 2025, and potentially adding several billion more in January 2026. "This curve cannot continue indefinitely; GDP is limited," he said, "but I would bet that even when scaled to the entire economy, it will still maintain a fairly rapid speed."
He painted a picture of a "middle world": the model's capabilities are rapidly growing exponentially, its economic diffusion is also fast (much faster than any previous technology), but not infinitely fast. Companies need time to adopt new technologies, which involves legal processes, safety compliance checks, internal promotions, and other steps. Even for a highly attractive product like Claude Code, large companies may adopt it several months slower than individual developers or startups.
"I think we should think about this middle world: things are extremely fast, but not instant, because they require time for economic diffusion and need time to 'close the loop,'" Amodei summarized. This means that even if the "genius nation in data centers" is technically realized within one or two years, the trillions of dollars in revenue it could generate might still take another year or two, or even longer, to fully manifest.
When asked about tasks like "editing videos," which require a long-term accumulation of context and preferences, Amodei predicted that when the "genius nation in data centers" is realized, AI will be able to handle such tasks. This depends on the model reaching a level of "true proficiency" in computer use. He revealed that Anthropic's model scores on computer usage benchmarks (like OSWorld) have increased from about 15% one year ago to 65-70% now.
Geopolitics and the "AI Iron Curtain"
The latter half of the interview turned toward the geopolitical implications of AI. Amodei candidly stated that AI will "greatly intensify" competition between major powers, particularly between the US and China. He believes that countries with strong AI capabilities will gain significant military and economic advantages, and this "winner-takes-all" dynamic may lead to a split of the world into two technological camps.
"I don't think we will sell data centers or chips, or the ability to manufacture chips to China," Amodei explicitly stated. He acknowledged that this would sacrifice some degree of economic mutual benefit, but stressed that in the AI era, growth and economic value will come "almost faster than we can bear," and the real challenge lies in "distribution of benefits, distribution of wealth, and political freedom."
He proposed a more radical possibility: is it possible to develop some sort of technology or AI-based system that creates a balance such that authoritarian countries "cannot refuse to let their people privately benefit from this technology while maintaining power"? He envisioned that perhaps everyone could have an AI model protecting them from surveillance, which could potentially crumble authoritarian structures from within. He admitted this has an idealistic tone, and the historical lessons of the internet and social media indicate significant challenges, but he believes "it's worth a try."
For developing countries, Amodei thinks that the traditional "catch-up growth" model (relying on underutilized labor and foreign capital technology) may fail in the AI era, as labor is no longer the limiting factor. He suggested building data centers in Africa (as long as they are not controlled by China) and promoting emerging industries driven by AI technologies like biotechnology in developing countries, allowing local humanity to participate and share in the growth dividends during the transition.
Constitution, Alignment, and Company Culture
When discussing Anthropic's renowned "Constitution" approach, Amodei distinguished between two levels: the balance of "rules versus principles," and the balance of "malleability versus intrinsic motivation." Practice has shown that teaching models to learn principles (rather than a rigid list of rules) results in more consistent behavior, better coverage of edge cases, and a greater likelihood of performing the tasks that humans want it to do. In terms of malleability, Anthropic's models are more inclined to "generally follow human instructions" rather than being strongly intrinsically motivated entities that go their own way.
Regarding the formulation of constitutional principles, Amodei proposed a "three-layer loop": 1) Internal iterations within Anthropic; 2) Different companies launching different constitutions, creating soft incentives through market competition and public feedback; and 3) Broader societal participation, such as soliciting opinions through collective wisdom projects, and even potentially providing input from representative governments in the future (though he is currently cautious about this). He believes the final solution will necessarily be a mix of these three.
Finally, as a CEO known for crafting lengthy thought memos, Amodei shared how he builds a company culture that is compatible with this "intellectual" CEO role. He spends a significant amount of time each week ensuring the company's culture is healthy, and he conducts a one-hour long "Dario Vision Quest" sharing every two weeks with the entire company, covering internal dynamics, model progress, industry observations, and geopolitical matters. He emphasizes maintaining "completely unfiltered" honesty in internal communications, avoiding corporate bureaucratic language, in order to build trust and keep everyone aligned with the mission, fostering debate and discussion on how to best achieve that mission. He believes this is a key advantage that allows Anthropic to unite 2,500 employees and cope with tremendous commercial and security pressures.
At the end of the interview, Amodei reflected that looking back at this history of exponential growth, future individuals may find it hardest to fathom that "the people involved at the time did not consider it all inevitable." When the world is just a year or two away from dramatic change, "ordinary people on the street are completely unaware." This disconnection between perception and reality, along with the "absolute speed" at which everything is occurring, constitutes the most unique challenges and opportunities of our era.
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。
