In this video, I break down the DeepSeek OCR paper and why it’s way bigger than “just OCR.” The core idea is Optical Context Compression: instead of feeding a long document as thousands of text tokens, DeepSeek compresses a whole page into a tiny image (think: 10,000 words into ~100 pixels) and still recovers the content with ~97% accuracy. I explain why long-context LLMs struggle in the first place (cost, latency, and attention scaling), and then walk through how DeepSeek turns pixels into usable context with a vision pipeline that behaves like memory engineering for LLMs. I go step-by-step through the system: DeepEncoder converts the compressed image into vision tokens, and then a Mixture-of-Experts (MoE) decoder reconstructs the text. I also cover the “Tiny Mode → Gundam Mode” intuition, how they compare against GOT-OCR 2.0 and MinerU while using fewer tokens, and why their multilingual data engine matters for robustness. The part that really blew my mind is the “AI forgetting mechanism” framing—this paper is basically showing a new knob for controlling what the model keeps vs. discards, which is exactly what we need if we ever want million-token context to be practical and cheap.