Core Viewpoint - Google's recent findings suggest that large language models (LLMs) exhibit a surprising correspondence with the human brain's language processing mechanisms, indicating a linear relationship between brain activity during real conversations and the internal embeddings of speech-to-text models [1][15]. Group 1: Research Methodology - Google introduced a unified computational framework linking acoustic, speech, and word-level language structures to study the neural basis of everyday conversations in the human brain [4]. - The research involved recording neural signals from participants during open-ended conversations for a cumulative total of 100 hours, while simultaneously extracting embeddings from the Whisper model [4]. - A coding model was developed to linearly map these embeddings to brain activity during speech generation and understanding, accurately predicting neural activity in new conversations not used for training [4][7]. Group 2: Key Findings - The study revealed that for each word heard or spoken, two types of embeddings are extracted: speech embeddings from the model's encoder and language embeddings from the decoder [6]. - The neural response sequence in the brain during language understanding and generation was found to be dynamic, with specific brain regions activated at different times [10][12]. - The results indicated that the embeddings from the speech-to-text model provide a coherent framework for understanding the neural basis of language processing in natural conversations [15]. Group 3: Comparison with Human Brain - Despite the parallel processing of words in large models, the human brain processes them serially, reflecting similar statistical patterns [16]. - The research highlights the concept of "soft hierarchy" in neural processing, where lower-level acoustic processing overlaps with higher-level semantic processing in the brain [17]. - Although the computational principles are similar, the underlying neural circuit architectures of LLMs and the human brain differ significantly [23][25]. Group 4: Future Directions - The accumulated research aims to create innovative, biologically inspired artificial neural networks to enhance their ability to process information effectively in real-world applications [26].