Comment: The original artificial intelligence was the state-of-the-art research in algorithms; it was basically research into anything we couldn't make computers do at the time. Things that were in the nebulous exploratory frontier that required innovation and difficult research. It was considered unrealistic and chasing a pipe dream. After an algorithm became well understood it tended to no longer be considered artificial intelligence. But over time very advanced algorithms like for game playing, search or navigation, as well as knowledge graphs, statistical classifications, and symbolic systems, etc became part of Good old Fashioned Artificial Intelligence or GOFAI. In a parallel fashion the connectionist movement advanced and eventually neural networks showed great capability in areas such as image classification, translation and language processing. These neural networks, which eventually used deep multilayered neural networks, was called deep neural networks or DNNs and was also considered part of ai, despite the algorithms becoming defined and no longer at some unknown frontier for the most part.
While this was happening there was great promise that failed to meet expectations and multiple so called ai winters took place were funding became scarce.
The most recent advancement was with the revolutionary architecture of the transformer, which gave rise to large language models, and has had various modifications and increases in model size as well as quantity of information used during training. Such improvements have led to increased capability. These latest ais, or llms, some call artificial general intelligence, but it can also be called proto agi. In some senses it is subhumans (basic errors almost no human would make, limits in scope and generality) and in some respects it is superhuman (such as encyclopedic knowledge that exceeds any human)
But as to what is artificial general intelligence? it is something that goes from subhuman artificial general intelligence to human level artificial general intelligence or beyond. General intelligence is the master algorithm; it is the algorithm that solves arbitrary problems and comes upon with solutions and algorithms in myriad domains. In animals when tested they can solve arbitrary puzzles presented to them of varying degrees of complexity. The greater the neuron count in cortex, the more complex and difficult the various types of puzzles an animal can solve. In humans there was a qualitative leap due to scaling in the cortex, and this allowed for the development of civilization and technological progress. This master algorithm, you could call it true ai, or true intelligence.
As to how to elaborate on the nature of the master algorithm that can develop or come up with almost limitless solutions and algorithms, I believe it works via pattern completion. Sequence prediction is a type of pattern completion in time. Patterns are structured information which can be interpolated or extrapolated based on the rules that define the relations between the internal elements. In reality there are spatiotemporal patterns, and the brain is designed to handle these, including when there is incomplete data.
Regards a good book on the nature of intelligence, the original book by Jeff Hawkins called On intelligence is a good resource. Though I dont quite agree with the research and path he has taken since then and which he comments on in his other book thousand brains. I believe the cortical algorithm is likely far simpler than he suggests, based on what I've seen from the neuroscience literature.
As for superintelligence or artificial superintelligence, it is believed it might be a qualitative leap on human general intelligence just as human general intelligence is a qualitative leap over the intelligence seen in primates. For example, other primates lack language and have far more limited problem-solving abilities than humans to a significant degree that kept them from developing civilization.
No comments:
Post a Comment