Fetching from the wire…
01
02
03
04
05
06
07
08
Infra2026-08-06 · source-backed
Published August 4, Exponent Aligned ANS splits BF16 into its 8-bit exponent field and its sign/mantissa, aligns exponents to a block-local center, then entropy-codes only the low-entropy exponent stream while storing sign and mantissa raw. 1.51x lossless, 622 GB/s median decode (peak 628), 391 GB/s encode. Because decode far exceeds typical ~50 GB/s network bandwidth, remote KV cache transfer becomes bandwidth-bound rather than codec-bound, which cuts time-to-first-token. Shipping in OpenLake v0.8 (Apache-2.0, Rust, 2,307 stars) with vLLM integration.
Each link below shares sources, entities, or timing with this story.
A Rust/io_uring storage engine posted to Show HN today moves KV cache off the GPU into a petabyte-scale store co-located on GPU hosts, so a prefix computed on one GPU serves to others (GitHub). Claims million-plus IOPS within 1ms and 66x TTFT speedup on cache hits at 128K cont...
Most coding agents run on JavaScript runtimes and eat 300MB of RAM just sitting idle. Zerostack is built in pure Rust, uses ~8MB on idle, ~12MB while working, and posts 0.0% idle CPU. It hit 1.0 on crates.io this week with 474 points and 252 comments on Hacker News. The archit...
Published August 13 under Apache-2.0 in Rust, nac is the harness Arcee uses internally to supervise training runs, run experiments, and prototype: a central orchestrator with threads and structured episodes keeping long tasks aligned with the original intent (Arcee AI). It shi...
Triple-stream retrieval (BM25 keyword, vector embeddings, knowledge-graph traversal) fused via Reciprocal Rank Fusion on the iii engine, with SQLite for state and an in-memory vector index, no external database. The economic claim: ~170K tokens/year (~$10) versus ~650K tokens...
MemPalace (~53.6K stars) claims best-benchmarked, against mem0 (~57.8K) and claude-mem (~80.8K) (GitHub). Agent memory went from experimental nicety to a competitive subcategory with published benchmarks. Pair this with the local-first angle: Mnemo offers a Rust + SQLite + pet...
Quesma ran the model across GPQA Diamond, IFBench and Terminal-Bench 2.1 (89 agentic coding tasks) on L40S, H100 and H200 via Modal. Q4_K_M at 17 GB matched BF16 at 55 GB within a point on all three. UD-Q2_K_XL at 10.7 GB held instruction-following but dropped Terminal-Bench f...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.