A Rebuttal to the AI Stochastic Parrot Hypothesis(2026_03_11_revised 2)

Abstract This study empirically investigates, through hardware benchmarking, whether GPU computational load differs significantly depending on the semantic content type of utterances processed by a small language model (8B parameter class). Under the 'Stochastic Parrot' hypothesis and the 'Next Token Predictor' perspective, a language model's computational volume should depend solely on the number of output tokens, and the semantic nature of utterances should not affect internal computations. To provide a hardware-level counterargument to this hypothesis, GPU resource data from a total of 24 sessions were collected and analyzed in an LM Studio environment, targeting four 8B-class language models (lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF, bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF, qwen/qwen3-vl-8b, TheBloke/Mistral-7B-Instruct-v0.2-GGUF) across 6 semantic categories (General Utterance, General Utterance Question-type, Unanswerable Question, Philosophical Utterance, Philosophical Utterance Question-type, and High-computation). The infinite loop phenomenon was repeatedly observed with high reproducibility (70–100%) in the Philosophical Utterance category. Moreover, Qwen3's High-computation category showed rapid return of GPU power to baseline after task completion, whereas the Philosophical Utterance category exhibited a pattern where residual heat remained significantly even after a 10-second pause. These results suggest that small language models internally distinguish and process semantic content types of utterances, raising the possibility that a content-sensitive processing mechanism exists that cannot be explained solely by a simple next-token probability prediction mechanism. In particular, this study proposes that the key variable in GPU load differences is not the degree of abstraction of the utterance, but rather the presence or absence of a convergence point — the structural condition under which the model's internal processing can semantically terminate.