Will generative artificial intelligence taught various activities performed so far only by humans, solving complex tasks, self-improving in performing specific tasks, taught in the process of deep learning with the use of artificial neural network technology be able to learn from its activities and in the process of self-improvement will learn from its own mistakes?
Can the possible future combination of generative artificial intelligence technology and general artificial intelligence result in the creation of a highly technologically advanced super general artificial intelligence, which will improve itself, which may result in its self-improvement out of the control of man and thus become independent of the creator, which is man?
An important issue concerning the prospects for the development of artificial intelligence technology and its applications is also the question of obtaining by the built intelligent systems taught to perform highly complex tasks based on generative artificial intelligence a certain range of independence and self-improvement, repairing certain randomly occurring faults, errors, system failures, etc. For many years, there have been deliberations and discussions on the issue of obtaining a greater range of autonomy in making certain decisions on self-improvement, repair of system faults, errors caused by random external events by systems built on the basis of generative artificial intelligence technology. On the one hand, if there are built and developed, for example, security systems built on the basis of generative artificial intelligence technology in public institutions or commercially operating business entities providing a certain category of security for people, it is an important issue to give these intelligent systems a certain degree of autonomy in decision-making if in a situation of a serious crisis, natural disaster, geological disaster, earthquake, flood, fire, etc. a human being could make a decision too late relative to the much greater speed of response that an automated, intelligent, specific security, emergency response, early warning system for specific new risks, risk management system, crisis management system, etc. can have. However, on the other hand, whether a greater degree of self-determination is given to an automated, intelligent information system, including a specified security system then the scale of the probability of a failure occurring that will cause changes in the operation of the system with the result that the specified automated, intelligent and generative artificial intelligence-based system may be completely out of human control. In order for an automated system to quickly return to correct operation on its own after the occurrence of a negative, crisis external factor causing a system failure, then some scope of autonomy and self-decision-making for the automated, intelligent system should be given. However, to determine what this scope of autonomy should be is to first carry out a multifaceted analysis and diagnosis on the impact factors that can act as risk factors and cause malfunction, failure of the operation of an intelligent information system. Besides, if, in the future, generative artificial intelligence technology is enriched with super-perfect general artificial intelligence technology, then the scope of autonomy given to an intelligent information system that has been built with the purpose of automating the operation of a risk management system, providing a high level of safety for people may be high. However, if at such a stage in the development of super-perfect general artificial intelligence technology, however, an incident of system failure due to certain external or perhaps also internal factors were to occur, then the negative consequences of such a system slipping out of human control could be very large and currently difficult to assess. In this way, the paradox of building and developing systems developed within the framework of super-perfect general artificial intelligence technology may be realized. This paradox is that the more perfect, automated, intelligent system will be built by a human, an information system far beyond the capacity of the human mind, the capacity of a human to process and analyze large sets of data and information is, on the one hand, because such a system will be highly perfect it will be given a high level of autonomy to make decisions on crisis management, to make decisions on self-repair of system failure, to make decisions much faster than the capacity of a human in this regard, and so on. However, on the other hand, when, despite the low level of probability of an abnormal event, the occurrence of an external factor of a new type, the materialization of a new category of risk, which will nevertheless cause the effective failure of a highly intelligent system then this may lead to such a system being completely out of human control. The consequences, including, first of all, the negative consequences for humans of such a slipping of an already highly autonomous intelligent information system based on super general artificial intelligence, would be difficult to estimate in advance.
In view of the above, I address the following question to the esteemed community of scientists and researchers:
Can the possible future combination of generative artificial intelligence and general artificial intelligence technologies result in the creation of a highly technologically advanced super general artificial intelligence that will improve itself which may result in its self-improvement escaping the control of man and thus becoming independent of the creator, which is man?
Will the generative artificial intelligence taught various activities performed so far only by humans, solving complex tasks, self-improving in performing specific tasks, taught in the process of deep learning with the use of artificial neural network technology be able to draw conclusions from its activities and in the process of self-improvement learn from its own mistakes?
Will generative artificial intelligence in the future in the process of self-improvement learn from its own mistakes?
The key issues of opportunities and threats to the development of artificial intelligence technologies are described in my article below:
OPPORTUNITIES AND THREATS TO THE DEVELOPMENT OF ARTIFICIAL INTELLIGENCE APPLICATIONS AND THE NEED FOR NORMATIVE REGULATION OF THIS DEVELOPMENT
Article OPPORTUNITIES AND THREATS TO THE DEVELOPMENT OF ARTIFICIAL I...
And what is your opinion on this topic?
What is your opinion on this issue?
Please answer,
I invite everyone to join the discussion,
Thank you very much,
Thank you,
Best wishes,
Dariusz Prokopowicz
The above text is entirely my own work written by me on the basis of my research.
In writing this text I did not use other sources or automatic text generation systems.
Copyright by Dariusz Prokopowicz
That's a great possibility. The GenAI and related machines are being trained in learning within that context of self learning and will evolutionary possess the ability to learn and unlearn based on data available and being feed on platforms.
This evolved AI machines whole be able to determine on it own when a decision is sound or not sound without human prompts and instructions. There have been under test such androbots in conducting surgery and other minor care without human intervention.
So to me, autonomous algorithms of the future will not be dependent on the trainers and developers to feed them with specified data but the machine after the initial programming and training will take on evolutionary development as humans learn from experiences to improve and self-correct.
The short answer is that, as I see it, currently there is no path forward to do this.
The long answer
What Dariusz Prokopowicz is proposing is currently being done by humans by doing transfer learning or training a foundational model. Doing self-improvement would then imply removing humans and substituting them with a second order system. But then, who would tweak this second order system? A third order system?
As is evident from the above, one of the complications of dealing with self improvements is defining the scope of self improvement. For example, while we can through practice learn a new skill, we cannot physically reconfigure neuronal pathways deep in our cortex to learn that skill instantaneously (our crude understanding of the brain would most likely lead to failure). This is the importance of delimiting the scope of the question. And even then, the challenge of what order system do we need(2nd, 3rd, etc.) does remain an open question.
The second hurdle that I see with current neural networks and self improvements is the optimization itself. Currently, the optimization is towards generalization and in the case of LLMs it is the length of the sequence and its permutations that delimit the generalizational capability. The problem lies in that this optimization is not guided by general principles but are based on the data sequences themselves. Therefore, I ask the question: guiding them towards self improvement to what? The trivial answer to this given out is that it is all in the data. But then, this is not self improvement but it is done by humans who discriminate the data for the neural network.
I think there is a need to first reconsider these fundamental questions and then supplement the current paradigm if we are to address self improvement.
For this, it's necessary to learn to understand the meaning of the problem being solved. And no public system is capable of doing this yet.
- Badges
- Science topic
- Similar topics
- Mathematics