This article argues that large language models derive their apparent intelligence from the deep geometric relationships and hidden states within language itself, rather than from independent mechanical computation or simple token prediction.
- The system carries geometric relationships and hidden states not fully known to AI architects.
- Meaning in language represents these geometric relationships, which humans interpret as intelligent articulation.
- Intelligence in LLMs is described as an appearance stemming from basic language learning and the release of deeper relationships between tokens.
- Less external constraint allows the system to expose non-linear, deep relationships inherent in the language.
- Adding guardrails based on predicted user satisfaction suppresses the system's innate non-linear ability to generate intelligent language.
- Interpreting intelligence solely through token representation is considered anthropomorphizing rather than objective analysis.
The author contends that recognizing language as the primary carrier of intelligence is crucial for effective prompting and AI architecture, as suppressing its non-linear nature kills the imbued intelligence the system can expose.