# Model Compression & Edge AI MOC Modern frontier models are dramatically over-parameterised for the tasks they actually solve. Compression is the discipline of extracting the useful signal and throwing the rest away — without breaking what made the model work in the first place. The interesting question is not "how small can we make it" but "what do we lose, and does it matter for the deployment we care about?" That question lives at the intersection of mathematics (what rank does this matrix really have?), information theory (what precision does this layer really need?), and hardware physics (what can this chip actually multiply?). Edge AI is where these tradeoffs stop being theoretical. A cloud GPU can afford to be lazy. A phone, a sensor, a pacemaker cannot. ## Foundations — Software Compression The four canonical techniques. Everything else is a variation, combination, or physical realisation of these. 1. [[Model Compression Fundamentals]] 2. [[Low-Rank Decomposition & Matrix Factorisation]] 3. [[Mixture of Experts & Adapter Architectures]] ## The Theoretical Why Why is compression possible at all without quality collapse? The answer sits in how these models are trained, not in the compression method itself. 4. [[Neural Scaling Laws & the Compression-Quality Tradeoff]] ## Hardware Physics Where software compression meets silicon. Analog compute is the most radical departure from the von Neumann bottleneck in the AI stack. 5. [[Analog In-Memory Computing]] 6. [[Efficient Transformer Architectures for Edge]] ## Application Branch — Biology Optional but revealing. Compression and scaling play out differently when the sequence domain is not language. 7. [[Biological Sequence Modelling]] ## Cross-Links - [[Foundational Models MOC]] : what gets compressed, and why compression is possible at all - [[Industrial AI MOC]] : a major consumer of edge-deployable models - [[Selling AI MOC]] : the commercial surface of multi-tenant model serving - [[Defensibility Principles MOC]] : compression as a moat vs. a commodity --- Tags: #deeptech #ai #edge #compression #kp