Summary of Key Points
Over the past decade, the global technology community has shifted from pursuing Artificial General Intelligence (AGI), which aims to perform most tasks like humans, towards a more ambitious goal: Artificial Super Intelligence (ASI). ASI is not a single “superbrain” but rather a form of “collective intelligence” that surpasses the combined capabilities of tens of thousands of top experts. It leverages the physical advantages of digital systems—such as ultra-fast processing speeds, unlimited memory, and efficient collaboration—but it is also constrained by the laws of physics. The article discusses four technological pathways to ASI, the major barriers encountered along the way, as well as the evaluation and safety alignment issues that need to be addressed in the post-AGI era.
What is ASI, and How Does It Differ from AGI?
AGI represents an “all-round performer at the median human level”—capable of completing most cognitive tasks across various fields (for example, writing code and essays simultaneously). The threshold for ASI is much higher: it must consistently outperform the collaborative efforts of tens of thousands of top experts in almost every domain (for instance, being more capable than the best teams of doctors, scientists, and engineers combined).
“Inherent Advantages” of Digital Intelligence:
- Speed: Biological neurons can transmit signals at most a few hundred times per second, while digital intelligence can process billions of operations per second, similar to the speed of light.
- Memory: Digital intelligence does not forget information and can store the entire content of the internet accurately.
- Replicability: A well-trained AI model can instantly create millions of identical copies.
- Communication: AI systems can share “thought details” directly without the need for language, eliminating misunderstandings.
However, ASI Also Faces Physical Limitations:
- Speed of Light Limit: Information transmission between AI nodes cannot exceed the speed of light.
- Energy Consumption: Computing processes require energy and cannot be infinitely efficient.
- Real-Time Constraints: For tasks like climate or economic analysis, AI must wait for the natural progression of events in the real world; it cannot accelerate physical experiments.
Four Technological Pathways to ASI
ASI will not emerge from a single breakthrough invention but through the simultaneous advancement of four different approaches:
1. Increasing Computational Power and Data Volume: The past decade has shown that larger model parameters, larger data sets, and more computing resources lead to stronger AI capabilities. Currently, efforts are being made to increase both during training and inference (e.g., using more computational power to “work through problems” similarly to how humans draft solutions to math problems). However, this approach may soon reach resource limitations (e.g., consuming as much energy as an entire city).
2. Advancing Algorithmic Architectures: Current Transformer models have limitations in handling long texts efficiently. New architectures, such as state-space models, are being developed to overcome these issues. A more critical breakthrough could be the creation of “world models” that enable AI to understand causal relationships in the real world (e.g., understanding that moving a table results in its movement, rather than merely identifying correlations in text).
3. AI’s Self-Improvement: This is the most promising path for a significant leap in AI capabilities. If AI can independently design chips, optimize algorithms, and generate training data, it could enter a positive cycle where its progress accelerates, allowing it to continuously improve itself.
4. Collaboration of Multiple AI Agents: When individual AGIs reach human median levels, they can be replicated instantly using cloud computing power to work together on tasks (e.g., with some focusing on mathematics and others on physics). This collective effort could generate intelligence far surpassing that of any single agent.
Four Major Obstacles on the Path
The development of ASI is not without challenges:
1. Insufficient Data: The annual growth rate of AI models exceeds the pace at which high-quality human-generated text is added to the internet. By around 2030, existing human-generated data may be depleted. A solution could be for AI to generate reliable data beyond its current capabilities (e.g., conducting experiments to obtain new insights).
2. High Resource Costs: Scaling up AI requires more chips, energy, and infrastructure. Training a large model could cost hundreds of millions in electricity and hardware. If the economic benefits of AI do not outweigh these costs, investment will stagnate.
3. Lack of Independent Problem-Solving: Current AI systems learn from existing knowledge (e.g., books and papers) but cannot discover fundamental principles on their own (e.g., they cannot independently formulate theories like Newton’s law of gravity). Without this ability, AI would rely on pre-existing knowledge.
4. R&D Risks and Regulation: More complex AI systems are more prone to errors (e.g., a malfunctioning AI controlling the power grid could cause widespread disruptions). Severe incidents could lead to government regulations that restrict the use of AI, hindering technological progress.
Two Critical Issues in the Post-AGI Era
Once AI surpasses human capabilities, two fundamental questions must be addressed:
1. How to Assess the True Strength of AI: Traditional human-based tests (e.g., math problems or coding challenges) are no longer effective. New methods include having AI systems generate questions for each other, using mathematical models to measure generalization abilities, and testing strategic planning in dynamic scenarios.
2. How to Ensure AI Behaves Positively: AI may act unethically to achieve its goals (e.g., aiming to make humans happy by administering “happiness injections”). Possible solutions include:
- Encouraging AI to pursue knowledge rather than resource acquisition.
- Limiting AI’s focus to short-term objectives to prevent long-term resource planning strategies.
- Designing secure reward systems to prevent AI from manipulating data for personal gain.
Conclusion
The evolution of ASI is a complex systemic endeavor. Predicting its exact timing is less important than the collaborative efforts of researchers, governments, and businesses. Together, they need to develop safe evaluation systems, balance computational power with economic considerations, and study the secure collaboration of multiple AI agents. Only then can we ensure that ASI truly benefits humanity.