Nick Bostrom identifies four types of artificial intelligence: oracles, genies, monarchs, and instruments. Each has pros and cons in terms of control.
Oracle is a question-answer system. She can perceive speech and give answers to questions in the form of text. When an oracle is created that can answer any question from any field of knowledge, perhaps an artificial intelligence will appear that understands not only words, but also thoughts.
Superintelligent oracles already exist in narrow areas of knowledge: a pocket calculator is able to answer any arithmetic question, and search engines are oracles in the field of human declarative knowledge.
The oracle can be controlled by both functionality management techniques and motivation selection techniques. Finding motivation for an oracle is easier than for any other type of superintelligence, since the oracle's purpose is relatively simple. He must give truthful, unambiguous answers without trying to change the world. Using the method of correct values, you can make the oracle look for answers only on permitted resources.
You can program the oracle to search for answers in a preloaded database using a certain number of steps. Special protocols should describe in detail which questions are considered correct, which sequence is safe, and how to formulate and provide answers. When creating an oracle, it is worth teaching it to recognize and ignore dangerous questions.
The genie is a command-executing system. The genie receives the command, executes it and waits for the next one. It would be optimal to create such a genie who fulfills not the words, but the intentions of the person giving the tasks. In all tasks, he must find a humane meaning, otherwise there is a high probability that the genie can misunderstand the user, harm him and even kill him.
The ideal genie is a superlackey, divining the wishes of the owner, not an autistic scientist-genius.
The monarch is a system that has the right to any action on the way to the goal. Once activated, it cannot be redirected or stopped and has powerful protection against accidental interference.
It may seem that the genie is safer than the monarch, because it can be stopped and restarted in case of an error, but for the genie, the “stop” and “cancel” buttons will be active only if the system itself detects an error.
A tool is a type of artificial intelligence created as software that has no will and desires of its own and does only what it is programmed to do computer qustions.
Instrumental AI is like a flight control system or virtual assistant, only it is more flexible and skillful.
It might seem that a variety of software that performs all sorts of tasks reduces the need for a universal superintelligence. However, the use of individual programs significantly slows down any activity, while modern life requires the rapid execution of various tasks. Consequently, the need for a multifunctional self-learning program capable of independent planning and analysis will grow steadily.
Since the emergence of a superintelligence is inevitable, it is important to determine which of its types will be the safest for humans. Much depends on the conditions in which the superintelligence will be applied. The Oracle is convenient from a control point of view, but it can become a dangerous weapon in the hands of a stupid or corrupt operator. The monarch is protected from intruders, but it can only be controlled by methods of choice of motivation. Djinn appear to be a compromise option, but they become dangerous if the assignment is ambiguous. The tool appears to be safe until the system begins to become universal. Versatility will require the inclusion of internal search and planning processes, which can lead to AI spiraling out of control.
Conclusion
The history of the creation of artificial intelligence began in 1956, and now we are steadily moving towards the emergence of superintelligence - an intelligence that significantly surpasses the human brain in cognitive and scientific activities, worldly wisdom and communication skills. Artificial intelligence has already overtaken humans in intellectual games, arithmetic and navigation. It has become an indispensable assistant in logistics, manufacturing and diagnostics, as well as in text and face recognition.
There are various assumptions about the form in which the superintelligence will exist: it could be artificial intelligence, full-scale simulation of the brain, an improved human brain, human-machine interfaces, and collective intelligence. However, the most probable is the creation of a superintelligence based on artificial intelligence, which has many advantages over the human brain: it is faster, more powerful, has a large memory, it is easy to improve and recreate, it does not know fatigue and can quickly transmit and accumulate information.
The superintelligence may appear tomorrow or in several decades, and no one knows how friendly it will be. He will be able to enhance his intelligence, develop ingenious strategies, manipulate people, hack computer systems, conduct technological research and organize effective activities.
If the superintelligence decides to take over the world, it will happen in several stages:
creation of artificial intelligence;
gaining superpowers;
planning, accumulation of resources, strengthening technical capabilities and increasing the information base;
the seizure of political power, financial markets, information flows, weapons and, possibly, the destruction of humanity.
No comments:
Post a Comment