Shane Legg, Founder and Chief AGI Scientist at Google DeepMind, shares his insights on the future of artificial general intelligence. He predicts we might see AGI by 2028 and discusses the importance of new architectures in achieving this goal. The dialogue dives deep into aligning superhuman models with human values and the balance between AI capabilities and safety. Legg also highlights multimodality as the next frontier for AI, emphasizing its potential to interpret diverse inputs beyond text.
AGI is expected to be achieved around 2028 and will require new architectures and improved alignment with superhuman models.
To ensure the ethical development of AGI, it is crucial to enhance reasoning abilities, develop comprehensive ethics understanding, and establish safety benchmarks.
Deep dives
Measuring progress towards AGI
Progress towards AGI (Artificial General Intelligence) is difficult to measure concretely. AGI aims to achieve the cognitive capabilities of humans, which encompass a wide range of tasks. While specific tasks can be measured more easily, measuring the generality of AGI is more challenging. It requires a comprehensive set of tests that cover various cognitive tasks humans can perform. Human performance serves as a benchmark, and an AI system must match or exceed human performance across different tasks. When an AI system consistently performs at or above human levels across a broad range of cognitive tasks, it can be considered as an AGI candidate.
Shortcomings of language models
Existing language models, such as large-scale transformers, have made significant progress in natural language understanding. However, they have certain limitations. For example, they lack episodic memory, which enables humans to learn specific things rapidly. Language models also struggle with understanding streaming video or having episodic memory, which humans possess. These models primarily mimic and generalize from data they have seen, rather than engaging in creative problem-solving through search processes like humans do. While efforts are made to compensate for these limitations, such as extending context windows, there is a need to improve language models' ability to understand different modalities and aspects of human cognition.
The need for better reasoning and ethical understanding
To achieve AGI, it is crucial to enhance AI systems' reasoning abilities and ethical understanding. Current architectures lack the capability for fine-grained reasoning and deep understanding of the world and ethics. To ensure alignment and ethical behavior, AI systems should possess robust reasoning capabilities and a profound understanding of ethics. Grilling the system and verifying its ethical reasoning processes can help evaluate its understanding and alignment with ethical principles. Developing systems that combine powerful world models, comprehensive ethics understanding, and reliable reasoning is essential for building highly ethical and alignable AGI.
Achieving alignment and addressing safety concerns
Ensuring alignment between human values and AGI involves specifying ethical principles and training the system to consistently apply them during decision-making. AI systems should possess deep understanding of ethics and use their world models and sound reasoning processes to analyze potential actions from ethical perspectives. This requires collaborative dialogue and continuous checks to verify ethical understanding and decision-making processes. Establishing frameworks and safety benchmarks as capabilities advance can provide concrete guidelines for ensuring safety and alignment as AGI development progresses.