AGI Cognitive Advantages
Potential cognitive advantages of general AI over the human mind:
- The ability to communicate and share information, e.g., abilities, concepts, memories, and thoughts, at a greater rate and on a greater level than us.
- The ability to continually add new computing power to a particular design feature, and to optimize that computing power. This may in turn result in moderate or substantial design improvements.
- The ability to control what is and what is not learned and remembered.
- The ability to have sensory modalities that we don’t, e.g., a modality for source code, which may help improve the AI’s programming competency—by making the AI inherently native to programming—to a level beyond our own. Such a modality may allow the AI to perceive its hardware machine code and the programming language used for its design.
- The ability to instantly learn new information.
- The ability to read and understand specific aspects of its source code.
- The ability to recursively self-encapsulate and self-improve (i.e., bootstrap) intelligence. As a mind becomes smarter, the mind can use its intelligence to improve its design, thereby improving its intelligence, which may allow for further improvements to its design, which may allow for further improvements to its intelligence. (See an expanded discussion of this below.)
- The ability to retrieve, acquire, understand, store, and use all readable information from the Internet, which contains most human knowledge.
- The ability to run on electronic circuits that travel at the speed of light (186,000 miles per second), rather than electrochemical signals that travel at 100-150 meters per second; and the ability to process information ten million times faster than neurons (an exponentially increasing difference since computer chips quadruple in power every two years, while the speed of neurons (200Hz) remains constant). Such a speed differential will give AI at least a million to one increase in speed of thought, resulting in a subjective year passing in only 31 physical seconds.
- The ability to self-analyze and self-modify its source code and design, and to make changes, as long as those changes achieve favorable design improvements.
- The ability to self-observe and understand on a fine-grained level that is impossible for us. AI may have an improved capacity for introspection and manipulation, e.g., an improved ability to introspect and manipulate code, the functional level comparable to human neurons, which we can’t think thoughts about or change.
- They won’t rationalize their actions or have Observer-Biased beliefs.
The process of strong self-improvement, such as the development of seed AI with access to its own source code, is quite important, since it can result in effects that are not in proportion to their causes. The development of seed AI is a ladder-climbing process: Self-improvements to components at one rung must result in further self-improvements at the same or higher rung, for self-development to continue. If the seed AI has achieved all possible improvements to components at one rung without opening up new design space for further improvements, then it will have bottlenecked, and will require the programmers' help to advance it to the next rung. As the ladder-climbing process advances, the seed AI will achieve greater self-improvements, and, in conjunction, greater self-encapsulation, of its design. Once a component has been improved by the seed AI, the programmers will no longer be the sole cause of that component, and future improvements by the programmers may indirectly result in the seed AI’s achievement of additional self-improvements to previously self-improved components. As the seed AI reaches higher ladder rungs, its ability to self-improve and self-encapsulate will continue to improve, which will propagate back to its ability to reach even higher rungs.
A dependency of seed AI development is that Human-Similar General Intelligence is achievable on computing systems. If this can’t be done, then AI may remain permanently stalled in early development, because the process of strong Recursive Self-Improvement is unlikely to begin occurring until the late developmental stages. However, if Human-Similar General Intelligence is possible, as it has been possible in us by blind evolution, then seed AI is likely to quickly begin the process of open-ended self-improvement, which in turn is likely to quickly lead to human-surpassing General Intelligence. The main three reasons for this is: the 10^9 speed advantages of present-day transistors over human neurons; the predicted ability of seed AI to regularly apply and add more computing power to a particular feature or problem, which may result in improvements to its preexisting intelligence; and the predicted ability of seed AI to use its preexisting intelligence to self-improve its subcomponents, which in turn may result in improvements to its overall intelligence (this is the previously described ladder-climbing process). These reasons are also useful arguments against the claim that Homo sapiens sapiens, the first species with enough General Intelligence to achieve exponential cultural growth, is in turn a limit on possible intelligence.
Basically, when there is an AI with Human-Similar General Intelligence (with a "cognitive supersystem" with its many cognitive advantages over our mind), with Human-Similar competency in programming, with ultrafast intelligence, with expandable intelligence, and with recursively self-improving intelligence, the result may be a rapid rate of transition (hours, days, weeks) to human-surpassing competency in programming, and then to human-surpassing general intelligence.
For the best available analysis of the ladder-climbing process, the domino effect, and other issues and aspects of seed AI development, see:
For further analysis of possible cognitive advantages of general AI over us, see:
- LOGI: 3: Seed AI: 3.1: Advantages of Minds-In-General
- Adaptive AI FAQ: Advantages of being artificial/designed
- Why Machines Will Become Hyper-Intelligent before Humans Do
-- Anand
