compression

A detailed walkthrough explains how Claude Shannon's 1948 information theory underlies LLMs and shows that the 'next-token prediction' story is misleading, linking compression and prediction mathematically.

A new runtime for the K and Q programming languages that reimagines execution using SIMD, parallelism, fusion, and compression to better leverage modern hardware.

SelKV is a training-free framework for KV cache compression that uses a soft cosine gate for selective merging and an attention-ratio compensation mechanism to correct softmax imbalance, achieving near-lossless generation at 25% cache size and 3.3x decoding speedup on LongBench.

PostgreSQL 19 plans to change the default TOAST compression algorithm from pglz to LZ4, offering better performance and efficiency. The article explains the history of Postgres compression and the reasons for the switch.

A community of tinkerers successfully runs the 753B-parameter GLM-5.2 mixture-of-experts model at near lossless quality on a $15,000 budget using REAP pruning and 2-bit GGUF quantization, sacrificing throughput for affordability.

Introduces VarRate, a training-free method for KV cache compression that allocates variable low-rank budget per token based on query salience, avoiding irreversible token eviction and outperforming uniform-rank methods at matched memory budgets on LongBench.

Introduces Looped Latent Attention (LLA), a post-training codec that compresses KV cache in looped transformers by exploiting low-rank structure across recurrence steps, achieving significant compression ratios while maintaining performance.

This paper presents a video codec that stores video and audio as weights of a sinusoidal representation network, using knowledge distillation and quantization for compression. Experiments show a 2.61x compression ratio compared to the original network, but quality lags behind standard codecs like H.264 and HEVC.

A practitioner recommends a free 33-minute lecture on cross-entropy that reframes language models as compression rather than next-word prediction, likening it to a Stanford ML PhD qualifier.

Researchers propose Latent Thought Flows, compressing 256 text tokens into 8 continuous latents for one-step generation, achieving a better Pareto frontier than autoregressive baselines.

Recommends 6 free and open-source Mac utilities: IINA video player, Stats system monitor, MonitorControl external display control, Ice menu bar manager, AltTab window switcher, Keka compression tool.

ShortOPD proposes a short-to-long on-policy distillation schedule that recovers pruned LLMs for free-form generation by focusing training on effective prefixes, achieving up to 9x improvement over unrecovered models and matching long-horizon distillation with a quarter of the training time.

A tool compresses 60 million text chunks from 201GB to 6GB for RAG without accuracy loss, enabling powerful local retrieval-augmented generation on a laptop.

misa77 is a new LZ-based codec that achieves decompression throughput up to 2x faster than LZ4 while also offering better compression ratios. It targets write-once read-many workloads and has constant memory usage.

Google open-sourced a vector index that compresses 31GB of AI memory to 4GB, fitting 10 million documents and searching faster than FAISS, with zero training or GPU required.

This paper proposes a theoretical framework for intelligence as atomic compression and compositional reuse, introducing the Compression Calculus and the Compounding Cascade thesis.

This paper audits six KV-cache compression methods under query-agnostic protocols, finding that rankings change dramatically compared to query-aware evaluations, with implications for cache reuse in long-context inference.

This paper presents a systematic comparative study of KV-cache compression schemes (TurboQuant and SpectralQuant), introduces a statistical validation methodology, and offers regime-specific recommendations for efficient transformer inference.

uv's lockfile size reduced by 40-50% through compression of environment markers.

A technical blog post exploring methods to reduce Lua string literal sizes by choosing optimal representations (quoted vs raw strings) in the context of ComputerCraft programs, achieving a 400-byte saving.