Training a 4B model to produce 81% faster query plans than Postgres

310 polyphilz 2026-09-16 19:02:14 UTC rohanbansal.com Llama Share 59

A 4B model was trained using supervised fine-tuning and agentic reinforcement learning to produce Postgres query plans that beat Postgres's default plans, achieving a 44.7% latency reduction across 113 join-heavy queries. The model learned the harness language of the qo-agent harness through off-policy distillation and reinforcement learning, resulting in a geometric mean speedup of 1.81x and ...

Breaking the 1.58-bit Barrier for Ternary LLMs

82 matt_d 2026-09-16 21:02:14 UTC arxiv.org Llama Share 5

Researchers developed BITCOS, a distribution-adaptive layout for ternary Large Language Models. It stores weights more compactly than five-trit packing in 26 of the tested models and achieves up to $1.485$ bits per weight on sparse models.

Xiaomi Mimo 2.6 live post-training dashboard

168 krackers 2026-09-16 21:02:14 UTC mimo.xiaomi.com Full Share 47

Training metrics of the mimo-v2.6-pro and mimo-v2.6-flash reinforcement-learning runs, live from the trainer's logs.

Nvidia announces native GPU programming in Rust

63 nonmaskable 2026-09-16 22:02:14 UTC developer.nvidia.com Llama Share 12

NVIDIA is developing native GPU programming in Rust, with two tracks: SIMT and Tile, to improve performance and safety. The projects aim to make GPU programming easier for the Rust community.

Small programming tricks

337 signa11 2026-09-16 16:02:14 UTC will-keleher.com Llama Share 166

The author shares valuable small nuggets of knowledge that can boost engineering productivity, such as shortcuts for tasks like starting a simple server or searching terminal history. These tiny tricks often require minimal mental infrastructure and can significantly ease work.

Reversing Factorio's RNG

91 jheitmann 2026-09-16 20:02:14 UTC gegell.github.io Llama Share 11

The author of the post has reverse-engineered the random number generator (RNG) used in Factorio, a popular game about factory-building and automation. The RNG is based on the taus88 algorithm, which uses three linear feedback shift registers (LFSRs) to generate pseudo-random numbers. The author has developed a method to predict future RNG outputs by analyzing past observations and using ...

AWS says it can't restore some data from mideast facilities struck by Iran

143 berkeleyjunk 2026-09-16 17:02:14 UTC wsj.com Full Share 107

Please enable JS and disable any ad blocker

Performance Improvements in .NET 11

106 soheilpro 2026-09-16 19:02:14 UTC devblogs.microsoft.com Llama Share 11

The .NET 11 runtime has made significant performance improvements, particularly in the areas of just-in-time (JIT) compilation and garbage collection. The JIT compiler now uses more advanced techniques such as "guarded devirtualization" to optimize virtual method calls, and "escape analysis" to reduce heap allocations. Additionally, the runtime has improved bounds check elimination by ...

Backups Aren't Simple

13 afilipovski 2026-09-16 22:02:14 UTC filipovski.net Llama Share 0

The author shares their experience with data loss and the importance of having a backup system in place, including snapshots and deduplication. They discuss various methods for rotating backups to ensure efficient storage usage.

The engineering behind the US Strategic Petroleum Reserve

17 johnjwang 2026-09-16 22:02:14 UTC johnjwang.com Llama Share 0

The US Strategic Petroleum Reserve stores hundreds of millions of barrels of crude oil in underground salt domes, rather than large tanks, to meet its requirements. This solution is relatively inexpensive and protects the oil from aerial attack, but requires ongoing maintenance and monitoring due to potential tradeoffs.

Accurate Models of AMD Matrix Cores

47 matt_d 2026-09-16 19:02:14 UTC arxiv.org Llama Share 6

Researchers studied matrix multipliers on AMD GPUs, finding non-compliance with IEEE 754 standard. They developed software models to validate bit-level accuracy and applied them to demonstrate application-level accuracy differences between AMD and NVIDIA cores.

Japan's book scene is moving from bookstores to libraries

65 herbertl 2026-09-16 18:02:14 UTC untranslatedjp.substack.com Full Share 17

How good are frontier models at physics?

51 qt31415926 2026-09-16 20:02:14 UTC arxiv.org Llama Share 21

Frontier language models struggle with advanced physics, but experts' reviews reveal many errors are due to benchmarking issues rather than model flaws. Corrected scores show significant improvement in models' performance on various physics benchmarks.

Anatomy of a Texture

64 Agentlien 2026-09-16 15:02:14 UTC agentlien.github.io Llama Share 11

A graphics programmer shares the complexity of texture memory layout in modern video games, highlighting block compression, texel ordering, mip mapping, and texture tiles. This complexity arises from performance optimizations such as cache locality and memory transfer speed improvements.

Dream-RSI: Recursive Self-Improvement through Evolving Worlds

171 bananaflag 2026-09-16 14:02:14 UTC arxiv.org Llama Share 49

AI agents need effective exploration strategies to discover high-value solutions, but current systems struggle with adapting to large search spaces. Dream-RSI is a framework that uses replay simulators and dreaming to improve exploration policies at low cost.

Mistral X Mozilla: Private, Multilingual AI Browsing

510 vertigoruntime 2026-09-16 09:02:14 UTC mistral.ai Llama Share 182

Mozilla and Mistral partner to bring AI-powered browsing with privacy, control, and choice. Firefox Smart Window now uses Mistral models for users in France, North America, UK, and Germany.

Show HN: An e-ink frame that hears birds and draws them as 1800s illustrations

2039 arnemunthekaas 2026-09-15 23:02:14 UTC github.com/arnegiacomo Llama Share 236

A Raspberry Pi-based e-ink bird frame uses AI for real-time bird detection, displaying hand-cut 1800s illustrations of birds heard in the garden. The project is still in development and welcomes contributions to improve its features and artwork.

WalShadow: Sub-second Postgres replication to ClickHouse from physical WAL

30 spathak 2026-09-16 20:02:14 UTC clickhouse.com Llama Share 5

WalShadow is an open-source engine that replicates Postgres data to ClickHouse directly from physical WAL, achieving low latency and high throughput. It supports the complete replication lifecycle and eliminates the need for logical replication slots, reducing resource consumption on the source Postgres instance.

Anecdotally, programmers dislike "reduce"

62 vinhnx 2026-09-16 14:02:14 UTC evanhahn.com Llama Share 108

Programmers tend to favor map and filter over reduce, with less feedback on its use in code reviews. This preference may be due to language limitations or personal familiarity with other methods rather than a fundamental issue with reduce itself.

Reverse-engineered Jev-like model

36 rochansinha 2026-09-16 19:02:14 UTC github.com/vinnylarouge Llama Share 5

A Jev-like model generates probabilities for text options in one pass, similar to TypeSafe's commercial model. The repository provides an independent starter model with the same input and output shape, trained on synthetic data and evaluated on a test set.

Vectorized and performance-portable Quicksort (2022)

160 mococa 2026-09-16 19:02:14 UTC opensource.googleblog.com Llama Share 24

Researchers have developed an open-source code that sorts arrays of numbers up to 10 times faster than C++'s std::sort, using SIMD/vector instructions and a portable implementation across multiple CPU architectures. The new algorithm outperforms state-of-the-art algorithms for sorting large datasets.

Hackers Got Inside a Flock Camera

423 driverdan 2026-09-16 14:02:14 UTC wired.com Llama Share 202

Hackers stole data from a Flock camera, revealing how it tracks vehicles and people, and shared the files with media outlets. The breach shows that Flock's cameras can detect license plates, vehicles, bicycles, and even bumper stickers.

Why Does the Universe Expand?

4 the__alchemist 2026-09-16 22:02:14 UTC cosmicave.org Prefix Share 1

Five hundred years ago, Nicolaus Copernicus proposed that the Earth might be one of several planets orbiting the Sun, rather than the centre of the universe. He compared the geocentric model to a monstrous form assembled from parts of different bodies, like the Creature Mary Shelley brought to life three centuries later in Frankenstein — each part appearing human on its own, but as a whole a ...

I replaced my brown-noise browser tab with a menu bar app

6 oldmanrahul 2026-09-16 22:02:14 UTC oldmanrahul.com LocalLLM Share 0

Hush is a macOS app that generates brown noise and speech blocker for focus, living in the menu bar for easy access. It's offline, CPU-friendly, and allows customization with user feedback shaping its features.

Training Text-to-Image Models 3.6× Faster

20 schopra909 2026-09-16 20:02:14 UTC linum.ai Llama Share 1

Researchers propose a novel architecture, JiT-DDT, to improve the efficiency and quality of generative image and video models. By using x-prediction early in the DiT via an encoder, they achieve 3.6× fewer GPU-hours training time while generating images with 4× more pixels than Linum v2.

The DeepMind Institute

118 vertigoruntime 2026-09-16 15:02:14 UTC institute.deepmind.com Full Share 41

The DMI is a platform, started by researchers from Google and Google DeepMind, to publish and discuss creative, deeply informed ideas about a world with AGI. DMI pieces are intended to serve as conversation starters, reflecting the author's ideas and research, and should not be read as Google's official view.

Tell the speakers that you liked their talks

263 whisper2020 2026-09-16 15:02:14 UTC ohhelloana.blog Llama Share 74

The author attended SmashingConf Freiburg and enjoyed meeting people, including speakers they admired. They encouraged attendees to ask questions to speakers, sharing their own experience of feeling shy but facilitating conversations at events.

Kyber (YC W23) Is Hiring a Forward Deployed Engineer

0 2026-09-16 13:02:14 UTC ycombinator.com Llama Share

Kyber is building an AI-native document platform for enterprises, seeking a future founder and early-career engineer to manage technical relationships with customers. The role involves owning the full technical sales delivery lifecycle, representing Kyber directly to enterprise stakeholders, and resolving customer-reported issues.

The Siberian Ice Maiden and the Scythian World

45 NaOH 2026-09-16 17:02:14 UTC patrickwyman.substack.com Full Share 4

Show HN: AttaLambda: a language where types and data are made of untyped lambdas

33 kserrec 2026-09-16 17:02:14 UTC attalambda.com Full Share 1

Its ordinary computation is built from unary functions and application. It adds readable syntax, exact rational numbers, and runtime type checks. Input and output cross a separate host boundary.