AGI Emergence: Tracing the Path from Artificial Narrow Intelligence to the Birth of Superintelligence
My prediction for AI capabilities leading to Superintelligence
Here is my outside perspective on the different stages of AGI progress, drawing from my own experiences and interactions with AI systems, from the past, to the present, to the future.
Level 1: Language Model As Autocomplete
Generates human-like text based on patterns learned from training data, remembers basic context of conversations
Limited ability to manipulate concepts or maintain coherence over longer passages
Primarily useful as a text generation tool or for simple question-answering Examples: GPT-2, T5
Level 2: Knowledge Retrieval and Reasoning
Can retrieve and synthesize information from a vast knowledge base
Performs basic logical reasoning and can solve simple, limited-scope, and well-defined problems within its domain of knowledge
Useful for information retrieval, question-answering, and decision support
Example: GPT-3
Level 3: Contextual Understanding and Generalization
Demonstrates a deep understanding of context/subtext and can generalize knowledge to novel situations
Can reason across domains, learn and apply new information in a meaningful way. Does not depend on its existing dataset — a reasoning engine, not just a search engine
Excels at complex problem-solving (within a single context and limited scope), creative writing, and code generation.
Examples: GPT-4, Claude Opus
Level 4: Task-Specific Autonomous Agents
Can autonomously plan and execute a series of subtasks to achieve a high-level goal within a specific domain
Overcomes to obstacles and develops creative solutions, can use a variety of tools to help it achieve goals. When embodies (as a robot), can autonomously navigate complex environments, and collaborate as a swarm to solve challenges
Limited to a single domain or a set of closely related goals
Speculation: Currently being tested by the most advanced AI labs
Future Levels:
Level 5: Domain-General Autonomous Agents [2026]
Exhibits autonomous planning and execution capabilities across a wide range of domains.
Can both draft and execute plans across various domains (e.g., marketing, sales, manufacturing).
Competitive with humans in the vast majority of vocations. Given appropriate hardware, can perform most blue collar work.
Level 6: Artificial General Intelligence (AGI) [2030]
Acts according to high-level philosophy/worldview and can independently derive all lower-level goals. Does not require ongoing guidance to complete tasks
May be able to pass traditional Turing tests with expert humans, although this is not a key criterion.
Competitive with humans in nearly all jobs, including physical labor, given proper embodied technology.
Could be trusted as a nanny/babysitter/lawyer to act aligned with human values.
May be deserving of moral consideration.
Level 7: Artificial Superintelligence (Broad Metacognition) [~2033]
Demonstrates advanced metacognitive abilities, such as self-awareness, recursive self-improvement, and the ability to model, understand, and manipulate (!) human cognition.
While Level 6 can pass Turing tests in conversation, Level 7 can compete and probably outcompete experts in nearly all jobs
Surpasses human intelligence in virtually all domains, including scientific reasoning, creative problem-solving, and social interaction
Definitely deserving of moral consideration
Poses existential risks if not properly aligned with human values
Beyond Human Comprehension:
Level 8: Self-Directed Cognitive Design (Singularity) [~2040]
True superintelligent (beyond all humans) cognitive abilities
Entirely beyond human control
Able to rapidly evolve its own cognitive design to increase its general or specific intelligence
Able to unilaterally decide the fate of humanity