Relevance 10/10Importance 8/10
A researcher fine-tuned a 4-billion-parameter model to emit PostgreSQL query hints, cutting latency 44.7% across 113 join-heavy queries and hitting a 1.41× geometric-mean speedup over Postgres's own optimizer. The pipeline pairs off-policy distillation from frontier-model trajectories with a custom RL variant built to tolerate the measurement noise that plagues database benchmarking. The model produced valid candidate plans for 89% of queries — a real result from a small model on a decades-old optimization problem.
Relevance 9/10Importance 8/10
Physics experts audited the major physics benchmarks and found the reference solutions themselves were riddled with errors. After corrections, GPT-5.6-Sol's measured mean@4 jumped from 47.3% to 78.7% — meaning the models were being graded against wrong answers. The paper argues frontier models now nearly saturate corrected physics evals, and the field urgently needs expert-validated benchmarks instead of the flawed ones everyone's been citing.
Relevance 9/10Importance 7/10
The authors noticed that in ternary LLMs, zeros aren't evenly distributed — they account for up to 51.5% of all weights. Their BITCOS layout exploits that skew with a bitmap plus sign vector, pushing storage down to 2 minus z bits per weight where z is zero density. Real-world payoff: up to 1.28× gains in matrix ops and 1.27× faster GPU inference throughput across multiple platforms.
Relevance 9/10Importance 7/10
Xiaomi has put its MiMo 2.6 reinforcement-learning post-training run on a public live dashboard with overview, metrics, and about views. It's an unusual move — most labs treat RL training curves as trade secrets, and this one streams them as the run progresses. HN caught it mid-reconnect, which is about as live as live gets.
Relevance 6/10Importance 9/10
Six months after March drone strikes damaged AWS data centers in Bahrain and the UAE, Amazon confirmed it cannot restore access or recover customer data from the affected availability zones. The Bahrain damage spanned multiple availability zones and exceeded the resilience thresholds the infrastructure was designed for; the IRGC claimed responsibility. Most affected customers have migrated to alternate regions using their own backups.
Relevance 7/10Importance 7/10
GPU matrix multipliers aren't IEEE 754 compliant, so this team characterized the actual numerical behavior of AMD's matrix cores across three GPU architectures. They built MATLAB software models validated against hardware with millions of test cases, achieving bit-level reproducibility. They then used those models to directly compare numerical accuracy between AMD matrix cores and NVIDIA tensor cores in real applications.
Relevance 5/10Importance 6/10
Ars Technica's review of Golden Gate covers Apple Intelligence's first significant upgrade in two years alongside a new Siri — and notes the toggle to disable Apple Intelligence and reclaim those gigabytes of on-disk models is now gone. Beyond the AI push, the release delivers fit-and-finish wins: a Liquid Glass opacity slider replacing the old binary toggle, native 5K ultrawide display support, and better multi-monitor window position memory.
Relevance 3/10Importance 6/10
A resurfaced Google post on a SIMD Quicksort that hits roughly 10× speedups over standard library sorts, reaching 1.1 GB/s on AVX-512. It uses compress-store instructions where available and emulates them with permutes on older chips, so it stays portable across architectures. Apache 2.0 licensed and built on Google's Highway library.
Relevance 3/10Importance 4/10
Will Keleher argues that engineering productivity compounds from small, low-context tricks you can absorb in seconds — fzf for command history, SELECT without FROM, git pickaxe, Array.flatMap. His point is that these don't require deep background to adopt, which is exactly why they spread so well. He suggests senior engineers deliberately share one trick a day with teammates.
Relevance 2/10Importance 3/10
By disassembling the game binary, the author found Factorio uses taus88 — three linear feedback shift registers XORed together. Because LFSRs are linear, Gaussian elimination over GF(2) recovers the full internal state from just 88 observations. The author then built the predictor entirely from in-game combinators, burning unwanted RNG rolls through scrap recycling to auto-craft legendary items.
HN Briefing PM
