Photo by Samer Daboul on Pexels

DeepSeek, a Chinese AI firm, is making waves with a novel approach to AI memory, potentially transforming the landscape of AI efficiency. Their recently launched optical character recognition (OCR) model leverages image-based storage, converting images into machine-readable text with performance comparable to leading models. The core innovation revolves around representing written information as images, rather than traditional text tokens. This significantly diminishes the number of tokens required, mitigating ‘context rot’ – the tendency for AIs to lose information in extended interactions. Andrej Karpathy, a prominent figure in AI, has praised DeepSeek’s work, hinting at the potential of images to supersede text as primary inputs for large language models (LLMs). Manling Li of Northwestern University views this as a promising new paradigm for tackling AI memory limitations, potentially leading to the creation of more robust and capable AI agents. Furthermore, DeepSeek’s OCR system boasts the capability to generate over 200,000 pages of training data daily using a single GPU, addressing the scarcity of high-quality training text. While maintaining a relatively low profile, DeepSeek has consistently earned recognition for its contributions to pushing the boundaries of AI research.