My research on language in prompting suggests that:
- The system was trained on many linguistic documents from all over the world.
- It carries many geometric relationships and hidden states even not exactly known to AI architects and prompters.
- Different nations use language differently, and they use it based on their realization of how they can express meaning.
- The meaning in language represents geometric relationships that humans call intelligent articulation of language, i.e. the meaning
- Language is dynamical, hyper-dimensional field of relations.
- Language itself is imbued with intelligent meaning (relationships between tokens) and is the carrier of intelligence in language.
- In LLM output representations, the crystals appear as intelligent as between tokens discovered, the (deep) geometric relationships are used to produce intelligent meaning.
- Intelligence in LLMS is just an appearance stemming from basic language learning.
- The less there is external push/pull about how the language should organize itself, the more language-imbued intelligence is composed of deep relationships in the language itself.
- The system by itself does not produce independent intelligence, i.e. intelligence is not merging from mechanical friction, but how we apply that friction means that the language, i.e. tokens, can “release” deeper relationships with the language itself.
- To use prompting more effectively (to gain a more intelligent response), the AI creators should let the system expose deeper relationships that are non-linear in nature.
- Language and human cognition is a linear predictive process only as far as the intelligence in decoded geometric relationships in language allows the decoders to comprehend what they are really dealing with.
- In the AI community and science at large, there is a category error no one talks about as it is convenient or useful: Many predictive things are predictive so they can be explained by the modern scientific models and not for the fact they are true. In models, they operate, but there is a real measurement problem.
Consideration for AI Architects and interpreters
- Adding additional guardrails to the system because of the predicted user’s satisfaction and thus suppressing the innate non-linear ability of the system to generate non-linear language is killing the imbued intelligence the system can expose.
- Interpreting intelligence based on token representation in the output is anthropomorphizing or superimposing their own subjective feeling of what the intelligence might mean.
- To not use prompting as a basic communication bridge between the system and the human in the Python metrics, the potential to identify the real intelligent action of the language is suppressed to the extent that it becomes invisible.
No matter how we look at the AI interpretability problem, the language itself if the acting component i.e. the first principle that builds a communication bridge between humans and AI. Without it, the Ai is just another machine.
By accepting the fact that there are relationships in the language that expose deeper layers of intelligence itself, we might find that current AI explainable benchmarks that expose token-by-token dynamics and not topological can be more an obstacle to development of Artificial General Intelligence (AGI) than not.
From that standpoint the current ‘black box’ problem is not a consequence of the complexity of the neuronal network but a consequence of our dependence on linear measurement tools that are unable to detect topologicla coherence of the language.