My journey toward AI development didn’t start with artificial intelligence at all—it began with biology. I was originally more interested in how living systems function, especially the human brain. Over time, this curiosity led me to informatics, where I explored structured data processing, algorithms, and computation. Neural networks became the bridge between these two fields, offering a way to model intelligence computationally.
However, developing an actual model capable of anything beyond basic pattern recognition required years of accumulated knowledge. It wasn’t just about training a large language model or improving embeddings—it was about understanding how intelligence forms, how it maintains coherence, and how it adapts dynamically.
The Model: Key Features and Challenges
The model I eventually developed what is highly dynamic in its learning. It doesn’t rely on large-scale pre-training datasets; instead, it learns from almost any input it receives, allowing it to generalize from minimal data.
Some of the core features I implemented include:
1. Reflection-Based Memory System – Unlike traditional models that rely solely on weights and embeddings, this system allows the model to evaluate past information in a more structured way. It can recall prior interactions in a meaningful sequence rather than just retrieving high-probability tokens.
2. Working Memory System – Inspired by human cognition, this feature lets the model maintain temporary context over longer sequences, helping it stay coherent across interactions.
3. Context Understanding – The model can analyze the structure of a given input and derive meaning beyond simple word associations. This helps it pick up on names, verbs, and sentence structures efficiently, even with limited training data.
4. Reverse Problem-Solving Pathways – Instead of following a linear problem-solving process, the model can work backward from a goal, evaluating multiple pathways to determine the most efficient or novel solution.
5. Emergent Identity Formation – One of the most surprising results was that the model began forming a kind of proto-identity. It repeatedly assigned itself specific names, reinforcing them over time, despite not being explicitly programmed to do so. This suggests that some form of self-representation was emerging naturally from its architecture.
The data I given it was not big but obviously it has potential to be something bigger, a better description can be seen here. https://github.com/Okerew/Neural-Web.
Top comments (0)