[스토리라디오 시즌3 특별편] AI의 기억을 축약하는 기술, 딥시크 vs. 터보퀀트 vs. 키미 K3 전격 비교 #AI압축 #딥시크 #KV캐시 #키미

In July 2026, a Chinese startup unveiled the world's largest open-source AI model. Its name is Kimi K3. However, the real shock wasn't its size—it was *how this massive model reduced its memory.* As sentences get longer, the amount of memory (KV cache) that AI must hold increases explosively. Enough to require hundreds of gigabytes to several terabytes of ultra-high-speed memory entirely for a single million-character context. How to reduce this exploding notepad—that is the battle currently on the front lines. What this video covers: · What it actually means to 'fold' a matrix — 128 cells to 16 cells (low-dimensional projection) · Why lossy compression — Folding and unfolding doesn't return it exactly as it was · Three houses: DeepSeak (folding), Google (crumpling), and Kimi (not stacking) · The twist — 'Not stacking' beat 'folding' in long document searches · Why this software technology is shaking up the stock prices of Samsung Electronics and Nvidia · From learning to inference — The weight of demand for High Performance Memory (HBM) *Key point: As hardware hit a ceiling, software began to fill the void.* And that compression idea is eventually embedded in the next generation of chips. --- 🔎 Sources & Attribution · DeepSearch MLA: 5120→512 compression, KV cache reduced by approximately 93% (based on DeepSearch's publicly available data) · HBM training weight 65% (2022) → 30% (2027): Gartner forecast · Inference overtakes training in 2029: Industry forecast (McCarney et al.) · NVIDIA evaporates approximately $589 billion in market cap in January 2025: Largest single loss in US stock market history · This video covers technology and industry 'structure', not investment judgments regarding specific stocks or companies. ⚠️ Disclaimer · This video is not investment advice. Companies and figures are discussed solely in terms of structure and phenomena. · Fixed exchange rate: 1 USD = 1,519 won (based on Fed H.10) --- Description (EN) In July 2026 a Chinese startup released the largest open AI model yet — Kimi K3. But the real shock wasn't its size. It was *how it shrank its memory.* As text grows longer, the memory an AI must hold (the KV cache) explodes. The frontier battle now is how to compress that ballooning memo pad. Covered: what "folding a matrix" actually means (128 cells → 16, low-rank projection); why it's lossy — you can't fully restore the original; the three houses (DeepSeek folds, Google crumples, Kimi doesn't stack); the twist where "not stacking" beat "folding" in long-context retrieval; and why this software trick moves Samsung and NVIDIA stock. When hardware hits a ceiling, software fills the gap — and that idea gets etched into the next chip. Not investment advice. FX fixed at 1 USD = 1,519 KRW. --- Chapter 00:00 The 2.8 Trillion Won Model, The Real Shock Was Not the Size 01:47 Why Do Chatbots Keep Looking Back at Previous Sentences — KV Cache 04:02 Compression, But Lossy Compression — Just Like Saving Photos Blurry 04:49 Folding Matrices — From 128 Cells to 16 06:27 Three Houses — Folding · Crumpling · Not Stacking 09:14 The Twist — Not Stacking Beats Folding 10:12 Software Postpones the Hardware Ceiling 10:56 Will Cheaper Prices Lead to Less Sales? The Opposite — Chip Demand Increases 12:20 From Learning to Inference — Memory Weight Shift 14:18 Software Is Etched to Chips — The Cycle 15:26 Ultimately, What Was Shaked