AGI: WHAT IT MAY LOOK LIKE
FANTASY AND REALITY
The success of any development depends significantly on the technological basis used. More precisely, it depends on the correspondence of the technological basis to the area of potential development use. Since we are discussing the development of AGI systems, the success of the choice of technological basis depends on the adequacy of predicting the scope of AGI applications.
Public attention to LLM systems and the promises of the developers of these tools for the step-by-step evolution of LLM into AGI have formed an implicit idea of the upcoming AGI systems. It briefly boils down to the fact that AGI systems will be some kind of super-systems, the services of which will be used by everyone everywhere utilizing the Internet, just as LLM tools, Google search, and the like are used today. This is one of the reasons for concerns regarding the danger of such systems for society. Are there any grounds for such an opinion?
To answer this question, it is helpful to outline the range of potential applications for which the insufficient level of intelligence of computer systems is an obstacle to improvement.
The first candidates on such a list are autonomous robots, drones, and cars. Insufficiently intelligent control systems for home robots, security robots, and car autopilots are the foremost hindrances. Mechanics and low-level control systems allow complex operations to be implemented - the intellectual component is missing. At the same time, using a remote “superbrain” with access via the Internet is immediately excluded for many reasons. Firstly, decisions need to be made in a time that is less than the time of sending a request via the Internet and receiving a response. Secondly, it is difficult to ensure security in the case of a communication failure. Thirdly, no remote superbrain can cope with a data flow many times greater than the traditional short and infrequent text data for such systems. Finally, each application requires a specialized set of knowledge, and creating AGI systems qualified for different applications does not seem reasonable or feasible. The conclusion is obvious: an AGI system that controls a robot, car, or airplane must be local.
The second candidate is business process management systems. Here, performance problems and the danger of interruption of communication are insignificant - but another consideration exists. If the system is truly intelligent, then its use by competing businesses, states, and armies is no more likely than having the same person as a top manager at the same time at Apple and Microsoft, as the ministers of North and South Korea, as the chief of general staff of Iran, and Israel. And it’s not just about the possible leak of secret information. Having intelligence means accumulating experience through work and using it to make decisions. Creating a system that would not use the knowledge gained from working for one client to make decisions for another client is much more complex than a system that exclusively serves one owner. And it’s even more challenging to prove to potential clients that this is actually guaranteed. This means that an AGI system for business should belong to one client and not be a system for mass use.
The third potential area of application is military systems. All the factors listed above are obviously essential for them, too.
Focusing on developing local intelligent systems guarantees their applicability in the real world. This, accordingly, is a factor in the rational choice of the technological basis of an AGI system, which will be the subject of discussion in the next chapter.