Written by: Techub News Compilation

Introduction

Recently, at the highly anticipated AI summit AI1G, two leading thinkers and practitioners in the field of artificial intelligence — Google DeepMind co-founder and CEO Demis Hassabis and Anthropic co-founder and CEO Dario Amodei — engaged in a crucial dialogue. As the captains of two of the most influential companies in today's AI race, they not only steer the direction of cutting-edge technology but also profoundly influence global thinking about the future of AI. Following their first public discussion in Paris last year, this dialogue focused on "the world after AGI," providing a critical perspective on the speed and limits of AI development, as well as its societal impact, in the context of rapidly accelerating technological breakthroughs and increasingly complex global situations.

Summary

• AGI timeline approaches: Amodei maintains his prediction that AI capable of achieving Nobel-level abilities across multiple domains may emerge in 2026-2027; Hassabis remains relatively cautious, believing there is a 50% probability of achieving human-level general cognitive abilities by the end of this century.
• The self-improvement loop becomes key: Both parties agree that the ability for AI systems to write code and conduct AI research is rapidly forming a "self-improvement loop," which could exponentially accelerate the arrival of AGI and is the most noteworthy variable to watch in the coming year.
• The impact on the workforce is imminent: Both CEOs predict that entry-level white-collar jobs will be the first to be affected, with impacts potentially visible within 1-5 years, expressing concern that society's adaptation speed may not keep pace with the exponential changes in technology.
• Geopolitical and security risks are escalating: They emphasize that the US-China tech competition, chip export controls, and the necessity of establishing international AI safety standards are currently the most pressing global challenges, calling for coordinated actions to avoid catastrophic risks.

AGI Timeline: The Self-Improvement Loop Will Determine Everything

Regarding when AGI will arrive, Dario Amodei and Demis Hassabis provided slightly different but future-oriented predictions.

Amodei reiterated his viewpoint from last year, stating that by 2026 or 2027, we may see AI models that can achieve Nobel-level human performance across multiple domains. He explained the underlying logic as a "self-improvement loop": AI becomes proficient enough in programming and AI research to develop the next generation of more powerful AI, accelerating the entire process. "We might only be 6 to 12 months away from the moment when models can complete most (if not all) of the work of software engineers end-to-end," Amodei pointed out. He acknowledged that not all aspects of the loop can be accelerated by AI (such as chip manufacturing and model training time), but he "finds it hard to imagine this taking longer than a few years," believing developments may "happen faster than people think."

In contrast, Hassabis's prediction is more conservative; he believes there is a 50% chance of achieving systems that display all human cognitive abilities by the end of this century. He acknowledged significant progress in verifiable fields such as coding and mathematics but noted that AI faces greater challenges in areas like the natural sciences. "You may not know if the compounds or physical predictions you build are correct, which may require experimental validation, taking longer," Hassabis emphasized. He highlighted that the "highest level of scientific creativity" involved in posing new questions, theories, or hypotheses is a capability that is currently lacking in AI, potentially requiring one or two unknown breakthrough elements.

However, both agree that the speed at which the self-improvement loop of "AI systems building AI systems" closes will be the most critical factor determining the arrival time of AGI. Hassabis also warned that such a loop, devoid of human involvement, carries risks that need to be taken seriously alongside the pursuit of speed.

Industry Landscape and the Survival Path of Independent Developers

Over the past year, the competitive landscape in the AI field has changed significantly. The host pointed out that at this time last year, OpenAI's ChatGPT was at its peak, while now Google DeepMind has made a strong return with its Gemini series models and products. Hassabis attributed this to the company's deep and extensive research accumulation and the reinvigoration of entrepreneurial spirit and focus throughout the organization. "We are making great progress, but there is still a massive amount of work to do." He mentioned that Google DeepMind is integrating models into various Google products more rapidly.

Regarding whether independent model developers like Anthropic can survive in a capital-intensive competition, Amodei provided an optimistic response based on revenue growth. He revealed Anthropic's astonishing revenue growth curve: from $100 million in 2023 to $1 billion in 2024, and an expected $10 billion in 2025. He believes there is an "exponential relationship" between a model's cognitive capabilities and its ability to generate revenue. "These revenue numbers are starting to approach the scale of the largest companies in the world," Amodei stated. He acknowledged that uncertainty always exists but firmly believes that as long as they can continue to produce the best models, the company will succeed.

Amodei further pointed out that a commonality between Google and Anthropic is that they are both led by researchers focusing on the models themselves and aiming to solve significant hard scientific problems. He suggested that companies driven by research are more likely to emerge victorious in future competitions.

Opportunities and Risks: The Survival Battle from "Loving Machines" to "Technological Adolescence"

Dario Amodei revealed that he is writing a new article as a sister piece to last year's optimistic outlook on AI potential, "Machines of Loving Grace." The new article will focus on the "huge and serious risks" posed by AI. He cited a scene from the film "Contact" as a framework: humanity should ask more advanced civilizations, "How did you avoid destroying yourselves during your technological adolescence?"

"We are knocking on the door of incredible capabilities," Amodei said, "but how we control it is not predetermined." He listed several major challenges that must be addressed in the coming years: how to control highly autonomous systems that are smarter than any human; how to prevent individuals from misusing it (such as bio-terrorism); how to prevent misuse by nation-states (which is why he is extremely concerned about countries like China); and the economic impact, particularly on labor displacement.

Demis Hassabis also expressed concerns about risks, emphasizing that his lifelong commitment to AI is due to its enormous potential as "the ultimate tool of science." He believes that if the best minds can be gathered and there is sufficient time and collaboration, the technical risks can be "handled quite easily." However, if the world becomes fragmented in competition, lacking coordination, the risks will significantly increase. "This technology will be cross-border and will affect all of humanity," Hassabis urged for international collaboration and minimum safety standards.

Economic Impact: The Labor Market Faces the Test of "Adaptation Speed"

Regarding the impact of AI on employment, the two CEOs' predictions are more urgent than current economic data shows. Amodei predicted that half of entry-level white-collar jobs might disappear within 1-5 years. He admitted that there is not yet a noticeable effect in the labor market, but a turning point may come soon. "I can even foresee that within Anthropic, we will need fewer rather than more people in entry-level and mid-level positions in the future." He is concerned that the exponential speed of technological advancement may "overwhelm our ability to adapt."

Hassabis also believes that this year we may start to see impacts on entry-level positions and internship opportunities, but he also noted that powerful AI creative tools are available almost for free, providing people (like university students) with leapfrog capabilities beyond traditional internships. He predicts that in the next five years, some jobs will be disrupted, but more valuable and meaningful new jobs will be created.

However, both believe that once AGI arrives, we will enter a completely unknown territory. Hassabis posed a deeper question: when economic issues can potentially be solved through fairer wealth distribution (even in a post-scarcity world), what will happen to the "sense of meaning and purpose" that humans derive from work? He remains optimistic, believing that humans will find new forms of meaning in activities like art, extreme sports, and interstellar exploration.

Geopolitics: The Most Dangerous Variables and the Controversy of "Selling Chips to China"

Geopolitics, especially US-China competition, is viewed by both as the current most complex and high-risk variable. Amodei candidly stated that his policy recommendations have not changed: "Not selling (high-end) chips is one of the most important measures we can take to buy time to address this issue." He sharply criticized the current logic of the US government — binding China to the US supply chain by selling chips. "If this were telecommunications technology, then discussions about diffusing the US tech stack and ensuring global data centers use Nvidia chips instead of Huawei chips might make sense. But I would compare it to: Should we sell nuclear weapons to North Korea because it makes Boeing money?" Amodei believes that such trade-offs make no sense for a technology with fundamentally strategic significance like AI.

Amodei explained the dilemma of "slowing down": if his radical timeline (1-2 years) is correct, why not proactively slow down to Hassabis's conservative timeline (5-10 years)? The answer is geopolitical rivals. "It's hard to reach a workable agreement for both sides to slow down simultaneously. But if we don't sell chips, it's not a matter of US-China competition; it becomes a competition between Demis and me — and I have great confidence we can solve this problem."

Hassabis also emphasized the necessity of international cooperation but acknowledged that in the current international environment, establishing a global cooperative organization similar to CERN (the European Organization for Nuclear Research) seems far-fetched. He is concerned that a lack of coordination could lead to inconsistent safety standards, increasing overall risk.

AI Safety: From "Doomsday Theory" to the Pursuit of Scientific Controllability

When asked whether he is more concerned about the risk of "malicious super AI," Amodei stated that since its founding, Anthropic has addressed such risks through research like "mechanical interpretability" (attempting to understand the inner workings of models). He remains concerned about this but opposes "doomsday theory" — the belief that disaster is inevitable or the most likely outcome. "I believe it's a risk, but if we work together, we can learn through science to appropriately control and guide these creations." He warned that if parties engage in unchecked racing, there indeed exists the risk of serious errors.

Hassabis again emphasized the potential of "human wisdom," believing that as long as there is time and focus, technical safety issues are solvable. His lifelong work is driven by AI's potential as the ultimate tool for understanding the universe, but he is also keenly aware of its dual-use nature.

In the final rapid Q&A, in response to whether the "Fermi Paradox" (if technological civilizations exist, why don't we see aliens?) supports the doomsday theory of AI, Hassabis answered negatively. He believes that if extraterrestrial civilizations have been destroyed by their technology, we should see signs created by AI, like "paper clips" or Dyson spheres, scattered throughout the galaxy, but we do not. He speculated that humanity may have already passed the "Great Filter" (such as the evolution of multicellular life), and the future chapters will be written by humanity itself.

Looking ahead to next year, both CEOs agree that the most important change to monitor is the progress of "AI systems building AI systems." This will determine whether we have a few years of buffer or whether we are about to face an urgent need for dramatic responses. Amodei concluded, "I would prefer a longer (timeline); I think that would be better for the world." This statement leaves a profound note on the deep dialogue about the fate of humanity's future.

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。