LLMs often translate math, vision, and engineering problems into text and then reason verbally to solve them. Even multimodal systems reportedly convert images into internal text-like tokens, suggesting a one-way advantage from perception to language rather than from language to pure spatial imagery. This points to verbal abstraction as a general-purpose substrate for high-level thought.
— If language is the central substrate, education, testing, and AI design should emphasize verbal reasoning for transfer and generality.
Alexander Kruel
2025.09.02
72% relevant
It highlights a Nature study where text embeddings from large language models predict fMRI in high‑level visual cortex and even enable caption retrieval from brain activity, implying a shared, language‑like semantic code across vision and language.
Davide Piffer
2025.08.08
100% relevant
The article’s claims about multimodal LLMs internally converting images to text and the domino-tiling proof narrated verbally, plus GRE outperformance by philosophy majors.
← Back to All Ideas