Solving the Issue of Interpretability of AI

I want to propose a new approach to the problem of AI opacity.

## The Core Problem

Modern AI systems work as "black boxes" - we can't see how they think. Recently, leading researchers warned that we might soon lose even the small transparency we currently have. Here's the difficulty: if we force AI to "think aloud" in human language, it reduces efficiency, but if we allow it to use efficient mathematical representations, we don't understand what's happening.

## Proposed Solution: A Modular System with Translators

I propose dividing the system into four parts:

*1. Free Internal Thinking* Let AI use any mathematical representations that are most efficient for solving tasks. We don't limit its thinking methods.

*2. Multiple Specialized Translator Models* We use several separate models trained to translate AI's internal representations into human-understandable language. Each translator can: - explain the logical structure of reasoning - highlight the main concepts the model is working with - explain how confident the model is in its conclusions Each function is performed by several different translators so results can be cross-checked.

*3. Contradiction Resolution Mechanisms* When translators give different explanations, we: - Highlight areas where they agree (high reliability) - Emphasize discrepancies (likely complex or ambiguous reasoning) - Explain why different interpretations arose If translator results don't contradict each other, we combine non-contradictory aspects into a unified explanation.

*4. Ethics Verification* We use "constitutional AI" (a special rule system, like in Claude.ai) to check: - Compliance with ethical standards - Logical consistency - Alignment with human values

## Main Advantages

- *No delays*: The model can think and produce results without delays (especially important in verbal dialogue), while explanations can be generated in parallel for quality control and, if necessary, future corrections. - *Moderation*: For critically important decisions requiring human moderation, we can wait for the translation and for the human moderator's decision - *Different perspectives*: Different translators show different aspects of thinking - *Transparency of complexities*: When translators disagree, we know the reasoning is complex - *Ethical safety*: An additional verification layer ensures alignment with values

## Open Questions

1. How do we train translators without "correct answers" from humans? 2. How many translators is optimal to use? 3. What to do if all translators cannot clearly explain the reasoning? 4. How to prove that translators accurately reflect internal thinking?

## Next Steps

I would like to: - Create a simple example of such a system working - Develop methods to verify translation accuracy - Combine this approach with existing tools

I would appreciate community feedback, especially regarding potential problems and practical challenges.