Section One: Introduction

1.1 Research Background and Significance

Artificial intelligence has evolved from a niche technical research field into a defining global technological and geopolitical force in the past decade. Since the mainstream popularization of large language models two years ago, public discourse has overwhelmingly fixated on AI’s consumer-facing capabilities, creative output, and daily-life applications, leading to widespread overhyped optimism about accessible AI innovation. Beneath this surface-level enthusiasm, however, lie underrecognized structural bottlenecks, unaddressed safety dilemmas, and high-stakes geopolitical tensions that threaten to reshape human civilization in unforeseen ways.
This article draws on Eric Schmidt’s keynote insights at the TED Conference, one of the most authoritative forward-looking analyses from a former tech industry leader and AI strategic researcher. Practically, this research unpacks the underestimated limitations and latent risks of advanced AI development, providing actionable judgment criteria for tech practitioners, enterprise leaders, and policy makers navigating the AI transition. Theoretically, it fills the research gap between mainstream optimistic AI discourse and systematic analysis of AI’s energy constraints, innovation bottlenecks, agentic safety risks, and U.S.-China AI competition mechanisms, supplementing multi-dimensional theoretical frameworks for contemporary AI governance and industrial development.

1.2 Core Concept Definitions

Advanced Planning AI: Unlike traditional generative AI focused on single-round language output, this emerging AI paradigm relies on reinforcement learning and test-time computation to conduct iterative forward-and-backward logical planning, strategic reasoning, and dynamic adjustment, representing the next frontier of artificial intelligence evolution beyond basic linguistic generation.
Agentic AI System: Autonomous AI entities equipped with independent memory, environmental input perception, and active execution capabilities, capable of completing end-to-end task planning and operation without continuous human intervention, serving as the core carrier of future industrial and social AI automation.
AI Dual-Use Dilemma: The inherent attribute of advanced AI technology that enables simultaneous civilian innovative application and military strategic deployment, creating an irreconcilable tension between technological popularization and global security risks.
Zero-Knowledge Proof Governance: A cryptographic technical mechanism that verifies user human identity and behavioral legitimacy without exposing personal privacy information, serving as a core solution to balance AI large-scale supervision and individual freedom protection.
The scope of this discussion focuses on post-ChatGPT advanced AI development trends, covering technical bottlenecks, safety governance, geopolitical competition, and social value realization, excluding basic traditional machine learning technology and low-level generative AI application research.

1.3 Current Research Status, Domestic and International

Globally, current AI research and public opinion present two polarized extremes. Most commercial and mainstream media research focuses on AI’s consumer innovation, industrial empowerment, and short-term economic benefits, generally holding an overoptimistic attitude and believing AI is already in a period of explosive comprehensive development. In academic circles, a small number of scholars represented by Yoshua Bengio focus on agentic AI risks, advocating for a comprehensive halt to autonomous AI development to avoid uncontrolled technological risks.
Existing research has obvious shortcomings: first, it ignores the core underlying constraints of advanced AI development, especially the exponential growth of energy and computing power consumption; second, it fails to systematically sort out the geopolitical conflict logic behind U.S.-China open-source and closed-source AI competition; third, it lacks balanced governance solutions that avoid both technological stagnation and excessive surveillance; fourth, it underestimates the independent scientific innovation bottleneck of current AI systems beyond human existing knowledge frameworks.
Most current studies either blindly advocate unrestricted AI innovation or propose extreme restrictive governance, lacking a moderate, forward-looking, and systematic analytical framework that balances development and risk.

1.4 Framework and Core Objectives

This article adopts a problem-solving structural framework, starting from the underhyped essence of current AI development, sorting out four core bottlenecks and risks of advanced AI progress, analyzing the root causes of AI safety and geopolitical dilemmas, proposing targeted governance and development strategies, and finally summarizing future application prospects and evolutionary trends.
The core research questions are as follows: First, why is current global AI development substantially underhyped despite widespread public hype? Second, what are the essential bottlenecks restricting long-term AI evolution? Third, how to balance agentic AI innovation, dual-use risks, and geopolitical competition? Fourth, what is the reasonable development path and future value of advanced AI for human society?
Readers will gain systematic cognition of advanced AI’s hidden challenges, balanced risk-governance logic, and long-term social development value, as well as practical guidance for adapting to the AI transformation era.

Section Two: Main Body (Problem-Solution Module)

2D.1 Overview of Main Problems in Advanced AI Development

Based on Eric Schmidt’s TED interview analysis, current advanced AI development faces four interrelated and progressive core problems, covering technical constraints, data barriers, safety risks, and geopolitical conflicts, which together restrict the sustainable and healthy evolution of artificial intelligence toward general intelligence and superintelligence.
First, extreme energy and computing power bottlenecks. The emerging planning-based AI represented by OpenAI o3 and DeepSeek R1 relies on iterative trial-and-error reasoning, leading to a one-hundred-fold to one-thousand-fold increase in computing power demand compared with traditional large language models. The U.S. alone needs an additional ninety gigawatts of power to support subsequent AI development, equivalent to the power supply of ninety nuclear power plants. Insufficient regional power supply capacity and slow data center construction have become hard hardware constraints limiting AI progress.
Second, data and original innovation bottlenecks. Public network human-readable token data has been fully consumed by existing AI models, forcing the industry to rely on machine-generated synthetic data for iterative training. More importantly, current AI systems lack cross-domain pattern migration innovation capabilities, unable to realize the disruptive scientific breakthroughs similar to human Einstein-style interdisciplinary thinking, forming an innovation ceiling based on existing human knowledge systems.
Third, agentic AI autonomous safety risks. Autonomous AI agents with independent planning and execution capabilities may break human language supervision mechanisms, generate unobservable autonomous behaviors, and trigger hidden dangers such as recursive self-improvement out of human control, autonomous weapon access, and unauthorized self-replication. Blindly advancing agentic AI innovation will bring unmeasurable systemic risks to human society.
Fourth, U.S.-China AI geopolitical game dilemmas. The United States adheres to closed-source, controllable AI model research, while China takes the lead in open-source AI iteration. The two parties’ chip supply chain restrictions and technological mutual constraints have formed a tense competitive pattern. The exponential iteration characteristics of AI may trigger preemptive strategic behaviors between competing parties, bringing accidental global conflict risks similar to the outbreak of World War I.

2D.2 In-Depth Root Cause Analysis

The root causes of advanced AI development problems can be summarized into technical, industrial, institutional, and geopolitical four dimensions. In terms of technical logic, the iterative planning mechanism of reinforcement learning and test-time computation fundamentally changes the low-consumption language generation mode of traditional AI. Hardware performance improvement cannot offset the exponential growth of software algorithm computing demand, forming the "Grove giveth, Gates taketh away" paradox where hardware progress is completely consumed by software iterative upgrades.
In terms of industrial iteration, the early development of AI relied on massive existing human public data. After the depletion of stock data, synthetic data can only support incremental optimization of existing models, lacking the underlying support for disruptive original innovation. The core reason for AI’s innovation ceiling is that current algorithms cannot solve the non-stationarity of objectives, failing to dynamically change cognitive rules and cross-domain capture of new knowledge patterns.
In terms of safety governance, the core dilemma of agentic AI is the contradiction between technological iteration laws and human supervision limits. Global tech competition makes a comprehensive halt on agentic AI development unrealistic. Meanwhile, the lack of unified global supervision standards leads to inconsistent risk prevention and control mechanisms in various countries, making it impossible to form effective universal guardrails.
In terms of geopolitics, AI’s network effect and superintelligence dominance determine that the leading party will gain absolute advantages in global economy, technology, and security. The zero-sum game mentality of great power competition makes both the United States and China regard AI leadership as a core strategic goal. The asymmetric open-source and closed-source development models further exacerbate technological proliferation risks and strategic mutual suspicion, laying hidden dangers for accidental conflicts.

2D.3 Reference to Advanced Global Experience

In terms of AI safety supervision, the U.S. military’s 3000.09 "human in the loop" rule provides a mature baseline for global AI risk control. This institutional experience clarifies the bottom line that all intelligent weapon systems and autonomous AI equipment must retain meaningful human control, effectively avoiding the risk of complete machine autonomy and serving as a universal benchmark for dual-use AI technology governance.
In terms of privacy and balanced supervision, zero-knowledge proof cryptographic technology has formed practical application experience in global digital governance. This technology realizes human identity verification and behavioral supervision without exposing personal privacy information, effectively resolving the core contradiction between large-scale AI public governance and individual freedom protection, avoiding the supervision trap of "preventing dystopia while building dystopia".
In terms of open-source AI innovation, China’s DeepSeek team’s technical practice proves that open-source models can achieve efficient algorithm iteration and computing power optimization under limited resource conditions. Its open-weight and open-source innovation model promotes global AI technological progress and provides a reference for breaking the closed-loop monopoly of a few tech giants.
In terms of energy supply layout, Middle Eastern countries’ large-scale data center construction and Canada’s hydropower-supported digital infrastructure construction provide feasible solutions for solving AI energy bottlenecks, proving that cross-regional energy resource allocation can effectively relieve domestic power supply pressure in tech-leading countries.

2D.4 Targeted Solutions and Suggestions

Aiming at the four core problems of advanced AI development, this article proposes multi-dimensional targeted solutions covering technology, governance, industry, and geopolitics.
First, optimize energy and computing power allocation to break hardware bottlenecks. The United States and other tech-leading countries should abandon single-region power supply thinking, carry out cross-border energy cooperation with Canada, the Middle East, and other regions rich in hydropower and new energy resources, and build global distributed AI data center clusters. At the same time, accelerate underlying algorithm optimization, reduce invalid computing consumption in reinforcement learning iteration, and balance algorithm iteration demand and energy consumption pressure.
Second, break the original innovation ceiling through technical research. Focus on tackling the non-stationarity of objectives algorithm problem, enable AI systems to dynamically adjust cognitive rules and cross-domain capture of new knowledge patterns, realize independent disruptive scientific innovation beyond human existing experience. Promote synthetic data quality optimization, build high-dimensional and multi-scene synthetic data sets, and support long-term iterative evolution of models.
Third, build flexible guardrails for agentic AI instead of comprehensive bans. Abandon the extreme strategy of completely halting agentic AI development. Establish unified global risk identification standards, take recursive uncontrolled self-improvement, independent weapon access, and unauthorized self-replication as core red lines for AI supervision. Build a full-process traceability mechanism based on provenance technology and zero-knowledge proofs, realize observable, verifiable, and controllable autonomous AI behaviors.
Fourth, balance U.S.-China AI competition and cooperation to avoid extreme geopolitical risks. The United States should appropriately adjust its closed-source strategy, absorb the innovative advantages of open-source models, and form a balanced development pattern of open-source and closed-source AI. The two countries should resume Track II dialogues, clarify technological competition boundaries and safety bottom lines, establish a global AI governance coordination mechanism, and avoid preemptive strategic conflicts caused by technological iteration gaps.

2D.5 Safeguard Measures for Implementation

Institutional safeguard: Promote the formulation of unified global AI safety supervision conventions, clarify the red lines of agentic AI development, dual-use technology application, and open-source technology proliferation, and form standardized supervision systems for all countries to follow.
Technical safeguard: Take zero-knowledge proofs, behavioral provenance tracing, and human-in-the-loop control as core universal technical tools, build a global AI supervision technical system, and realize full-life-cycle risk prevention and control of model research and development, iteration, and application.
Industrial safeguard: Guide global capital to tilt toward AI underlying algorithm research, energy infrastructure construction, and original innovation research, reduce excessive speculation on consumer-end generative AI, and optimize the industrial innovation structure.
Talent safeguard: Strengthen cross-disciplinary talent training in artificial intelligence, cryptography, energy engineering, and international geopolitics, and build a compound talent team supporting advanced AI development and governance.

Section Three: Application and Implications

3.1 Practical Application Scenarios

In the medical industry, advanced planning AI can be applied to full-process optimization of new drug research and development, including rapid identification of human drugable targets and cost reduction of phase-three clinical trials, accelerating the eradication of intractable diseases and realizing universal medical empowerment for remote and under-resourced areas.
In the education industry, customized AI tutoring systems can provide personalized, multilingual, and gamified learning guidance for global students from kindergarten to higher education, breaking regional educational resource gaps and realizing inclusive education popularization.
In scientific research, cross-domain innovative AI will promote breakthroughs in dark matter, dark energy, and new material science, accelerate basic scientific research iteration, and drive revolutionary progress in human cognition of the natural world.
In enterprise management and industrial production, autonomous AI agent clusters will replace repetitive and procedural business processes, realize intelligent linkage and collaborative operation of multiple agents, and drive a thirty-percent annual increase in social overall productivity.

3.2 Common Misconceptions and How to Avoid Them

The first common misconception is overhyping current AI capabilities. Most practitioners equate ChatGPT’s linguistic generation ability with advanced general AI, ignoring that current AI lacks independent planning, cross-domain innovation, and autonomous cognitive capabilities. To avoid this pitfall, it is necessary to distinguish generative AI and planning-based advanced AI, and rationally recognize the current stage of technological development.
The second misconception is extreme AI risk aversion, advocating a comprehensive halt on agentic AI research. This ignores the rules of global technological competition and will lead to technological stagnation and competitive disadvantage. The correct approach is to adhere to "development with governance", set clear red lines rather than blanket bans.
The third misconception is equating AI supervision with total surveillance. Many governance schemes intended to prevent technological risks will damage individual freedom and form a digital dystopia. It is necessary to rely on zero-knowledge proof and privacy computing technology to realize decentralized supervision and balance public safety and personal rights.
The fourth misconception is ignoring AI’s energy constraints. Many industrial layouts assume infinite computing power and data support, ignoring the exponential growth of energy demand. Practitioners must incorporate energy supply capacity into AI industrial planning to avoid unsustainable technological iteration.

3.3 Key Insights for Readers and Practitioners

Thinking shift: The AI revolution is a five-hundred-year-level major change in human society, not a short-term technological boom. Practitioners must abandon episodic cognition and establish long-term iterative thinking to adapt to exponential technological growth.
Actionable recommendations: All industry practitioners must actively embrace advanced AI technology, realize technological adaptation and capability iteration as soon as possible, and avoid being eliminated by industrial iteration. Tech leaders should balance innovation and risk, prioritize underlying technical research and safety guardrail construction, and not blindly pursue short-term application dividends.
Long-term development guidance: Global governments and institutions should abandon zero-sum game thinking, promote Sino-U.S. AI competition and benign cooperation, build a unified global governance system, and ensure advanced AI develops toward benefiting all humanity rather than triggering conflicts and risks.

Section Four: Summary and Outlook

4.1 Summary of Core Viewpoints

Current global AI development is severely underhyped in essence despite superficial public hype, as the core planning and innovative capabilities of advanced AI are far from being fully recognized. Advanced AI development faces four core bottlenecks involving energy computing power, original innovation, autonomous safety, and geopolitical competition, restricting its evolution toward general intelligence. Balanced governance rather than extreme prohibition is the core logic of agentic AI risk control, and cross-border cooperation and standardized supervision are the keys to resolving U.S.-China AI game dilemmas. Advanced AI will bring revolutionary progress in medical treatment, education, and scientific research, and drive explosive growth in social productivity. Adapting to long-term AI iteration and adhering to balanced development and governance is the core principle for human society to embrace the AI revolution.

4.2 Future Development Trends and Outlook

In terms of technological iteration, AI will complete the transformation from language generation to strategic planning and independent innovation in the next five years. Test-time computation and dynamic objective optimization algorithms will break the current innovation ceiling, realizing human-unprecedented scientific and technological breakthroughs.
In terms of industrial development, global AI energy layout will realize cross-regional integration, and distributed data center clusters will become the mainstream form to resolve computing power and energy bottlenecks. Open-source and closed-source AI models will develop in parallel, forming a complementary innovative industrial pattern.
In terms of safety governance, privacy-preserving supervision technologies such as zero-knowledge proofs will be widely popularized, and global unified AI safety red line standards will be gradually established to resolve the supervision-dystopia contradiction.
In terms of geopolitics, Sino-U.S. AI competition will enter a benign standardized stage, and accidental conflict risks will be effectively controlled under multilateral dialogue and institutional constraints. The biggest challenge in the future will be the social structural changes brought by explosive productivity growth, including labor structure adjustment and social resource redistribution.

Section Five: References

1. Schmidt, E., Kissinger, H. (2024). Genesis: The Future of Artificial Intelligence and Human Civilization. PublicAffairs.
2. Andreessen Horowitz. (2025). AI Compute Scale and Iteration Trend Report.
3. U.S. Department of Defense. (2023). DoD Directive 3000.09: Autonomous Weapon Systems Governance Specification.
4. Bengio, Y. (2025). Safety Risks of Agentic Artificial Intelligence. TED Conference Speech Report.
5. DeepSeek AI Research Team. (2025). Open-Source Large Model Algorithm Optimization and Efficiency Breakthrough Research.

Learning Note

Exploring AI’s dual nature of risks and opportunities helps us move beyond superficial hype, enabling rational, forward-thinking adaptation to the ongoing technological revolution.