Hello! A lot has changed since the previous article. This time I want to introduce several new mechanisms in AION — mechanisms whose value I did not fully see before.
Have you ever wondered what a human being is? Why do we think, feel, and act in particular ways? Why can emotion become stronger than logic, and why can the same situation produce completely different reactions? What is life experience, how is personality formed, and how does every lived situation influence later behavior?
I ask myself these questions almost every day while working on AION. They are philosophical and psychological questions, and they are difficult to reduce to a simple formula. But they also point toward a practical problem: why do neural systems feel so dead? AION already separates language, logic, and knowledge into modular blocks, but that may not be enough. I want the system to develop dynamically — not only to be a useful assistant, but to have a changing personality, preferences, and internal history.
That led me to look more closely at my own behavior: what I feel, why I react differently in different situations, and what makes me myself. The most interesting mechanism that came out of this process is Experience Space (ES). My current hypothesis is that an analogue of experience could improve both emotional closeness and intellectual development.
How could ES work?
This is not a final solution. Like most AION mechanisms, it is an intermediate stage. My view of experience may change later, and the system may have to be rebuilt.
The central question is simple to ask and extremely difficult to answer: what is experience?
At first I considered representing experience as a particle formed inside ES, working somewhat like a particle in Concept Space (CS), but carrying a richer history. I abandoned that idea because it immediately raised difficult questions: would AION need another CS AI just to find experience? How would experience be updated? Would the whole mechanism become a giant collection of hidden rules?
After reading more and studying my own behavior and the behavior of other people, I arrived at another hypothesis: experience may not be a particle at all. It may be a latent state based on the system's emotional signals — something that changes personality over time.
I experimented conceptually with a CS-like ontology containing Core Experience, User Experience, Between-User Experience, and Science Experience. This division could be useful, but it could also be too rigid: one event may belong to several kinds of experience at once.
The idea is based on a simple observation: experience is what changes a person throughout life. We may remember an event or forget it, but it can still affect our future behavior. The difficult part is determining whether experience was actually absorbed in the intended way. A person can notice that they used to be shy and became more open after finding a supportive group. I do not yet know how to represent that kind of self-understood change.
This may require another space: Self Space, with emotional traces of messages and actions, an internal mood mechanism, and many other components that I have only started to investigate. That makes ES a problem somewhere between philosophy, psychology, and programming. I do not know how well the foundation will work yet; time and experiments will decide that.
What is thinking?
Now we move to another frontier problem. I have to ask the same question I asked about a month ago: how can an AI learn to think?
Experience can at least be studied through a human reference. Thinking is more fundamental and much harder to observe directly. I have been testing many hypotheses for the Cognitive Space (CoS) module: perhaps thinking is a set of cells whose operations are normalized into one state, or perhaps new CS AI signals should be sent into CoS to form a more complete picture of the question.
The current direction uses latent states produced by Librarian while it builds a graph. CoS focuses on the relevant part, runs another competition between candidates, and performs several additional transformations. The details are mostly tensor-level changes, so they are difficult to explain to a casual reader.
During this research I kept encountering the same problem: modules lost too much information while passing data between one another. Imagine that module 1 has A, B, C, D, E, F, and G; module 2 receives only B, C, and E; module 3 receives only E; and module 4 needs D, F, and G. That is not only a model problem — it is a problem with the contract for transferring information between modules.
Inspired by Kafka
I once watched a video explaining how Kafka works and why it can become a nightmare for programmers. I had never used Kafka in my almost five years of programming, but the idea was useful.
The problem with the current contracts is straightforward: module X needs data N, while module Y needs more data, but module X cannot pass information it no longer has.
That led to a new idea: what if AION had one shared information store, similar to a specialized SQL database, where modules could preserve their data and later decide what they needed to read?
This is how Global Module Information (GMI) appeared. It is a deliberately limited, specialized information substrate that resembles a much smaller Kafka. Other modules can publish data, create versions, update or delete records, and preserve a trace of which information was delivered to which module. The purpose is not to make GMI think or select answers. Its purpose is to keep information available and make the handoff between modules more transparent.
I also added a small abbreviation glossary to the article system: hover over an unfamiliar term on a computer, or tap it on a phone, and a short explanation appears.
That is everything I have discovered over the last few weeks. Follow the next articles — it will get more interesting from here.