Machine Intelligence and Learning (MIL?): a paradigm shift from AI/AGI/ASI to MIL World Models
A Paradigm Shift from AI/AGI/ASI Models to MIL World Models
The world modeling (simulating, learning, understanding and interaction) power is a necessary condition for any real intelligence, natural, real, artificial or alien.
As much as there is NO modern science and engineering or human minds without world modelling, there is NO real artificial intelligence (AI), artificial narrow intelligence (ANI), artificial general intelligence (AGI) or artificial superhuman intelligence (ASI) without the world modeling and simulation capabilities.
We advance a paradigm shift from the AI/AGI/ASI models of human intelligence into the MIL models of the world of reality, physical, mental, social, digital, virtual.
No autonomous existences could exist without having some world models, instinctively feeling or intelligently knowing how the world is organized and works, its reality and causality, structure, complexity and behavior.
Modelling is an essential and inseparable part of all substantial techno-scientific disciplines and knowledge domains.
The world models are fundamental to "what is what and who is who and why is why" and to the development of real-world, general intelligent systems as the Machine Intelligence and Learning (MIL?) machines.
If your would-be intelligent machines, AI systems and machine learning models or neural network algorithms machines or large language models (LLMs), have no world modeling capabilities, they are unintelligent by the very design,
Or, SOTA AI/ML/LLMs are not Generalists, regardless that GPT-4, Bard, Llama 2, or Claude involve a wide range of topics, execute a wide range of tasks, handle multimodal inputs and outputs, operate in multiple languages, and “learn” from zero-shot or few-shot examples, they lack the real-world learning and intelligence of <world modeling> generality and performance.
Mustafa Suleyman, the CEO of Microsoft AI, the co-founder and former head of applied AI at DeepMind , proposed his version of AGI/ACI (“Artificial Capable Intelligence”) as the “Modern Turing Test,” in which an AI would be given $100,000 of capital and tasked with turning that into $1,000,000 over a period of several months.
Now, we suggest more challenging TT, in which Machine Intelligence and Learning (MIL?) would be given $1T of capital and tasked with turning our decaying dumb world into an all-sustainable habitat over a period of 5-7 years. It makes less than a half of the United States' war spending in Afghanistan.
The World is All We Need
Current AI systems and LLMs are not really intelligent, autonomous and interactive, due to lacking of mental models, an internal representation or simulation of reality within their data structures and algorithms, programming and computation and applications.
Mental models are so basic to understanding the world that we are hardly conscious of them. The animal/human neocortex learns world models interacting with the environments, physically or mentally, to be able to make predictions about the world.
Common sense world knowledge and reasoning informs humans or machines what is it, how is something likely, what is plausible, and what is impossible in the world humans/machines navigate.
Meantime, usual AI/LLMs is only rote memorizing and learning what is likely, and what is probable from a linguistic point of view, basing not on meaningful world models but meaningless statistical probabilities.
World models are all what unite humans and machines and all what machines need, for real-world AI is not about simulating reality rather than human intelligence, thinking and actions.
It is about modeling, simulating and understanding the world of reality, everything that exists, ranging from the physical universe, "[t]he totality of all space and time; all that is, has been, to special contexts, associated, for example, with the Earth and all life on it, with humanity as a whole, or with knowledge domains and universes of discourse, as the human mental reality or the world around machines, including all everyday things, objects and events like tables, cars, trees and humans or concerts, music festivals, athletic events, conventions, online behaviors.
For example, to know our internet identity/persona/personality or measure our digital behaviors, MIL bots have a world modeling of the internet, its social media and networks and users and their digital activities, including selecting and visiting web sites, uploading/downloading content, searching for information and completing tasks, blog posting, commenting, liking and disliking, etc.
The Real World Model of Artificial Intelligence
Such world models are necessary conditions for all intellectual processes, from knowing and learning to reasoning, understanding, planning, decision-making and interacting.
The transition from AI/ML/DL models to LLMs to LWMs to the MIL World Models represents a paradigm shift in AI, moving from understanding the world through text, videos or images to knowing it in all its complexity, fundamentally changing how we build a nature-friendly and human-complete general machine intelligence.
The journey toward the MIL World Models not just an advancement in technology; it's a step closer to creating machines that understand and interact with the world in the most rational, effective and sustainable ways.
Machine intelligence and/or human intelligence
Machine intelligence is all about the world's data processing and real world modelling capacities. while with human intelligence is not so simple.
Today's narrow AI/LLMs follow Gardner's Theory of Multiple Intelligences , the pseudoscientific differentiation of human intelligence into specific intelligences, rather than defining intelligence as a single, general ability with its mental/cognitive models.
The Oxford English Dictionary defines intelligence to be “the ability to acquire and apply knowledge and skills”. Or, if extended, human intelligence is “the ability to acquire and apply knowledge and skills across the range of special intelligences”:
As one could see from Table below, from the paper ‘Levels of AGI: Operationalizing Progress on the Path to AGI’ , with all seeming successes of specialist, narrow AI systems, ChatGPT, Bard and Llama 2, but they completely fail at creating generally intelligent AI systems due to lack of the real-world models.
What are MIL World Models?
As said, no complex creature could exist without having some world models, instinctively feeling or intelligently knowing how the world organized and works, its reality and causality, complexity and behavior.
Domain modelling is an essential and inseparable part of all substantial techno-scientific disciplines and knowledge domains.
MIL Models covers world modeling, as the generation of a ontological, scientific , physical,?conceptual, mathematical or statistical representation of reality and its contents.
MIL models are to explain and predict the behaviur of the world, its entities, objects, systems, phenomena, processes, forces and interactions and are used in a wide variety of contexts, domains, and environments, ranging from?the physical world to mental, social, digital or virtual reality.?
MIL modelling?is an autonomous intelligent activity that produces?models?representing?a particular part or feature or situation of the world to?understand,?define,?quantify,?visualize, or?simulate.
Due to its generality, MIL could operate with different types of models used for different purposes, such as?ontological models for categorization, causal models for cause-effect patterns, conceptual models?for understanding, semantic models for definition, operational models to?operationalize,?mathematical models?to quantify,?computational models?to simulate, and?graphical models?to visualize the data, information or knowledge.
It could be climate modeling of atmospheric and ocean phenomena for weather forecasting or scientific understanding of?global warming or modeling to understand?animal?and?plant?populations and the?dynamics?of interactions between organisms or modeling for?urban planning, construction, and the restoration of?ecosystems or modeling the universe.
MIL models in silico that are rendered in generally intelligent software/hardware allow the human intelligence to leverage computational intelligence power to model, simulate, visualize, manipulate and get deep insights about physical, mental, social, digital or virtual reality, its entities, phenomena, processes, patterns and laws.
Conclusion
Creating AGI as a human-like and human-level intelligence is a way nowhere, as Yann LeCun-like architecture and training paradigm for developing human-brain-mimicking intelligent machines that can learn, reason, and plan like humans and animals. A Path Towards Autonomous Machine Intelligence.’
MIL World Models are a real-world modeling architecture paradigm for developing autonomous intelligent machines that can effectively and rationally interact with the world, humans and other machines.
Resources
We propose a framework for classifying the capabilities and behavior of Artificial General Intelligence (AGI) models and their precursors. This framework introduces levels of AGI performance, generality, and autonomy, providing a common language to compare models, assess risks, and measure progress along the path to AGI. To develop our framework, we analyze existing definitions of AGI, and distill six principles that a useful ontology for AGI should satisfy: Focus on Capabilities, not Processes; Focus on Generality and Performance; Focus on Cognitive and Metacognitive, but not Physical, Tasks; Focus on Potential, not Deployment; Focus on Ecological Validity; Focus on the Path to AGI, not a Single Endpoint;
With these principles in mind, we propose "Levels of AGI" based on depth (performance) and breadth (generality) of capabilities, and reflect on how current systems fit into this ontology. We discuss the challenging requirements for future benchmarks that quantify the behavior and capabilities of AGI models against these levels. Finally, we discuss how these levels of AGI interact with deployment considerations such as autonomy and risk, and emphasize the importance of carefully selecting Human-AI Interaction paradigms for responsible and safe deployment of highly capable AI systems.