Semantic Core and Evolutionary Approach to Artificial Intelligence

Abstract

In this article, I explore an alternative approach to developing artificial intelligence based on semantic core and evolutionary principles. Instead of an endless race for model scaling, I propose focusing on deep understanding of limited knowledge domains through operational definitions. The concept of a "virtual kindergarten" for AI agents and external regulators is described. Practical steps towards implementation are suggested, starting with the use of existing lexicographic databases. The article offers a pragmatic view on creating AI with genuine understanding.

Introduction

Modern large language models demonstrate impressive results, but a fundamental problem remains: they don't understand the content of text, rather they imitate understanding based on statistical patterns. In this article, I present an alternative approach to AI development, based on a semantic core and evolutionary principles.

Fundamental Limitations of Modern AI Systems

Working with various language models, we repeatedly encounter their limitations:

1. Statistical understanding instead of semantic — models operate with probabilities of word sequences rather than their meanings
2. Inability to acknowledge ignorance — tendency toward confabulation in the absence of information
3. Lack of logical consistency — random contradictions in responses
4. Unreliable context retention — loss of important information during lengthy interactions

These limitations are not eliminated by increasing model size and data volume. A fundamentally new approach is needed.

Pragmatic Approach: Lowering Expectations, Increasing Depth

Instead of pursuing comprehensive artificial intelligence, we propose a more realistic path:

1. Domain-specialized understanding instead of general intelligence

   • Creating systems that deeply understand limited domains
   • Focus on quality of understanding in a narrow area rather than breadth of coverage

2. Operational definitions as the foundation of understanding

   • Prioritizing precise definitions of concepts over statistical associations
   • Recognizing that emotional experience is not required for many tasks

3. Built-in uncertainty

   • Principled refusal to express certainty about unverified facts
   • Systems capable of honestly saying "I don't know" or "I need to check"

Semantic Core as the Foundation of Understanding

Our approach is based on the concept of a semantic core — a structured set of clearly defined concepts:

1. Clear definitions as foundation
   Each concept must have an operational definition that allows unambiguous identification of its applicability, without relying on "similarity" to other concepts.

2. Hierarchical organization of concepts
   More complex concepts are defined through simpler ones, forming a natural hierarchy without cyclic dependencies.

3. Pragmatic understanding

   • Focus on the functional use of words and concepts
   • Sufficiency of operational definition without emotional component
   • Emphasis on practical applicability of knowledge

Evolutionary Approach to AI

An alternative model of artificial intelligence development, inspired by natural processes:

1. "Virtual kindergarten"

   • Creating multiple AI agents with different parameters
   • Collective learning through interaction with human "educators"
   • Natural selection of the most successful models

2. Embracing uncertainty

   • Abandoning complete control over each AI agent's development
   • Allowing contradictory instructions as part of learning
   • Developing the ability to independently resolve contradictions

3. External regulators

   • Pragmatic understanding of the consequences of incorrect actions
   • Systems for access restriction and behavior control
   • "Survival" of the most useful and adaptive agents

4. Specialization instead of universality

   • Development of highly specialized AI with deep understanding of their domain
   • Evaluation of AI based on ability to solve specific tasks
   • Abandoning the idea of creating "universal intelligence"

Practical Steps to Implementation

We propose a phased approach to creating AI with real understanding:

1. Utilization of existing lexicographic databases

   • Connecting existing dictionaries with clear definitions
   • Working with polysemous words and their contexts
   • Recognition of parts of speech and their functions in sentences

2. Factual information extraction

   • Finding and presenting facts from structured sources
   • Clear identification of requested entities
   • Transparency of information sources

3. Basic input error correction

   • Recognition of typos and simple errors
   • Finding the nearest semantically correct variant
   • Confirmation or automatic correction of errors

Conclusion

The proposed approach to developing artificial intelligence represents an alternative to the dominant paradigm of model scaling. Instead of a "race for size," we focus on quality of understanding, clarity of definitions, and natural evolution of intelligence through interaction.

The key distinction of our approach is the recognition that operational understanding without the need to recreate human consciousness in all its fullness is sufficient for many practical tasks. This is a more realistic path that can yield useful results in the near future.