In this video, I break down a really practical research result called LongCodeZip: a code-specific way to compress huge code context (up to 5.6×) without sacrificing the downstream coding performance of your LLM. The whole motivation is simple: as your input context grows, transformer attention gets expensive (latency goes up, token cost goes up), and the model also gets worse at finding the relevant parts inside a massive prompt. And even with “big context” models, real repos can still blow past your budget—plus code isn’t like natural language text because it’s structured and full of cross-file dependencies. Then I show how LongCodeZip actually does the “magic” with a dual-stage pipeline. Stage 1 is coarse-grained function-level filtering: instead of embeddings/similarity (which can fail badly for code), it ranks functions using an AMI-style score based on how much the function reduces the model’s surprise/perplexity when answering your query. Stage 2 is fine-grained block selection inside those functions using perplexity spikes to decide block boundaries and keep only the helpful sub-blocks. The takeaway: you can feed a much smaller prompt (hundreds of tokens instead of thousands), keep accuracy, and cut latency (they show examples like ~15s down to ~6.6s) by doing smart code-aware compression plus dynamic retrieval.