Ladder of increasingly intelligent systems
Introduction
The aim for many artificial intelligence (AI) researchers is to build intelligent machines that achieve parity with human intelligence, this is human-level AI (HLAI). Once machines can be built with this level of intelligence, if they are given agency it is conceivable such intelligent machines could go on to build ever more intelligent machines, with or without human assistance. With machine intelligence reaching levels beyond human capability the term transcendence or singularity is applied.
The term artificial general intelligence (AGI)?or general AI is widely used to refer to what is called here HLAI. While AGI is commonly used, it suffers from a lack of finer distinctions, whereas the label HLAI offers a rich intelligent system hierarchy, which helps focus research on the different types of AI systems.
This article presents a ladder of intelligent systems. It begins with our current state of intelligent machines, such as generative AI (genAI) and large language models (LLMs), all based on deep learning neural networks (DLNNs), and essentially narrow AI – see Figure 1. The cleverness of LLM and genAI is its ability to reflect, in a filtered manner, the intelligence buried in its training data. For example, LLM calculates probabilities to many decimal places of a most suitable next word – its excellent capability demonstrates the power of statistical analysis, and is the driver behind this narrow AI.
In Figure 1, between narrow AI and transcendence there exists a range of possible levels of intelligent machines, to be explored next.
Figure 1: From current levels of AI to transcendent AI
HLAI
There are three approaches to building HLAI. The first is to understand how the human brain works and with this knowledge build a human-like human-level AI, or (HL)2AI. This approach takes its inspiration from neuroscience. Our brain is the ultimate example to which we aspire to build an intelligent machine: humans have the highest form of intelligence we know and motivates the study of neuroscience to guide building the (HL)2AI machine.
The second approach is to build an intelligent machine on ideas rooted in engineering (knowledge engineering, computer science, mathematics, etc.), let’s call this engHLAI. Modern electronic computing including memory technology, long-term storage, and rapid numerical computation is vastly superior in capability to the equivalent features in the human brain, so the modern information technology stack can be leveraged in building capabilities in engHLAI.
Many aspects of the human brain at the molecular level are a means to achieve some functional step from A to B. If electronics can short cut the molecular complexity that occurs in a living body then using it makes sense. The current research direction of generative AI and LLM may yet succeed in building HLAI and if so then it would be an example of engHLAI – resembling nothing like the human brain, rather a direction taken in a purely engineering one.??
领英推荐
There is another aspect to engHLAI: exploiting the concepts of AI cognitive architectures. Progress with cognitive models has been modest in comparison with recent successes of DLNNs, however, cognitive architectures could play a role in engHLAI models. Adding pre-built information processing structures in the form of cognitive architectures to engHLAI, combined with neural networks that learn about their environment and learn how to solve challenges - this can be compared to how the human genome includes a blueprint for the design of the brain, and once that design has been instantiated experience takes over in adapting it for use in the real world. Similarly, one can conceive of an engHLAI that uses cognitive architecture to kick-start the AI with designed components that can work in combination with neural networks.
The third approach is a hybrid blend of the first and second approaches, hybridHLAI or H2LAI. A hybrid approach exploits the best of computer engineered HLAI combined with (HL)2AI. This may be the ideal approach, exploiting the patterns and concepts we discover in the human brain and accelerated using engineering algorithms that have no correspondence in the brain but can accelerate computation and shortcut the complexity that exists in wetware (i.e., the human brain) through hardware and software.
Animal-level AI
Attempting to build HLAI may be a challenge too complex for our current neuroscientific knowledge. In which case it may be worthwhile to tackle a smaller scale challenge: animal-level AI (ALAI), where by animal is meant a creature lesser than humans. The idea is to crack the hidden neuronal code in creatures with a more manageable number of neurons in their brains, identifying principles that can then be scaled up to emulate HLAI. Evolution reuses successful building blocks throughout the evolutionary tree of life, so solving how neuron communication relates to thinking in a creature with a brain containing millions rather than billions of neurons, is likely to apply to the human brain.?
The same labelling style used above may be applied to ALAI to describe the different possible approaches:
Machine-level AI
It may be useful to consider an intermediate state, between HLAI and transcendent AI: the machine-level AI (MLAI). HLAI is based on the senses we have as humans. Once achieved we could build on HLAI and augment it with new sensors that input a fuller picture of our environment. For example, vision based on the full spectrum of the electromagnetic spectrum. Or a neurorobot (an intelligent robot) may be built with wheels and limbs that bear no relation to the human body. Where ALAI and HLAI have three varieties: pure engineering, animal/human-like, and a hybrid mix of the two, with MLAI it is by definition an augmented engineered HLAI.
Further reading
The article is based on extracts from my forthcoming book:
Towards Human-Level Artificial Intelligence: How neuroscience can inform the pursuit of artificial general intelligence, E M Azoff, to appear, CRC Press, 2024.
The book has a web site with occasional articles: www.hmnlvl.ai