This series of posts presents the idea of building a knowledge database that allows for the collection and exchange of reliable information between people, computers, and between people and computers. It also discusses how this database could increase the intelligence of both people and computers, and increase the safety of AGI.
Correct information often lets us make better decisions. AI will also perform better if it has access to a larger amount of credible information, both during learning and when answering questions. Therefore, this database could increase the collective intelligence of people and the intelligence of computers.
The database would be a vast and credible model of a significant part of the physical world we live in, understandable to people, computer programs, and AI. The fact that the model is understandable means that both people and computers can interact with it intelligently.
A child learns about the world through interactions resulting from being "immersed" in the physical environment. One of the primary goals of a living organism is survival, and the absence of pain is an important indicator of whether it is achieving this goal well.
How could AI independently understand the world and function within it? Instead of immersing AI-controlled robots in the physical environment of people, I propose to immerse intelligent agents in an information space that reflects our world. Interactions with this space would involve exchanging information (i.e., reading from and writing into this space) and evaluating its correctness. One of the primary goals of an intelligent agent would be to answer questions correctly, and its reputation would be an indicator of achieving this goal well. With this defined goal, AI could autonomously pursue objectives important for people.
Because the information space would be human-readable, the actions of AI would be more transparent and, therefore, safer. AI and people could interact with the database in the same way, but using unique abilities of their intelligence, complementing and correcting their actions and learning the most effective reasoning techniques from each other.
The idea originated from fundamental considerations rather than by analogy to existing systems.
Initially, I hypothesized that building AGI will be easier if AI has access to a credible, highly detailed, and computer-understandable knowledge database describing the world. I concluded that due to the size of the database, crowdsourcing is the easiest way to build it.
I also thought that to obtain AI that is not only able to read the database but also understands the knowledge it contains, it is good, analogously to living organisms, to allow AI to interact with this database and define a goal it can autonomously pursue within the database, along with a criterion for its successful achievement.
Organisms of various levels of complexity live on Earth, with increasingly advanced forms appearing as a result of evolution. Similarly, in the proposed system, there would be a place for AI of varying complexity and efficiency. Over time, more advanced and efficient forms would emerge, for example, those able to reason better based on existing information or add new raw data more effectively. Initially, human intelligence would play a leading role in the system.
The shared information space provides an opportunity for both intelligences (AI and humans) to understand the surrounding world and immerse themselves in it, just as living organisms immerse themselves in the physical world during learning. Through this database, we could immerse AI in a world it understands and is consistent with our reality. We could also connect robots to the system, allowing computers not only to be immersed in knowledge about the world but also in the actual world itself, just like people.
In the subsequent posts in this series, I describe details of the knowledge base introduced here and outline its application to increase collective human intelligence and intelligence of computers.
The second post describes the structure of the knowledge base and the method of building it.
The third and fourth posts describe some of its possible initial uses that increase intelligence of people and computers, including those that are important according to Effective Altruism.
The eighth post introduces the concept of an information space being an interface for exchanging knowledge between all intelligent agents (including people and AI). It also presents a hypothesis that truth is an attractor in the information space implemented by the proposed knowledge database. If it turns out that truth is actually an attractor, then it seems this will improve cooperation among people to the extent that it can theoretically lead to the emergence of collective superintelligence, with this database being its shared knowledge.
The ninth post lists some of the most important problems of current AI, including LLMs, and proposes to use the described knowledge database to try to address them.
The remaining three short posts describe other aspects of this knowledge database.
The posts were originally published on consensusknowledge.com in the years 2018 and 2023.
This series of posts describes the idea of increasing intelligence of people and computers and improving AI safety by building a knowledge database understandable to people and computers and an interface to exchange information between them. In particular, the database could be used:
The idea originated from fundamental considerations rather than by analogy to existing systems, so it has unique features and capabilities.
I think that a necessary and sufficient condition for the success of this idea is to prove the hypothesis presented in the post Knowledge Base 8: The truth as an attractor in the information space, which states that in the described system, people can effectively discover the truth. I believe that the hypothesis can only be tested by implementing the entire system. The implementation is not very difficult but is not trivial either.
Given the potential significance of the idea, its uniqueness, and its verifiability, I ask for the following: