China's open-weights AI strategy is winning

632 benwerd 2026-07-20 15:22:22 UTC werd.io Llama Share 533

American companies are losing the AI model market due to their locked-down business practices, while China's open approach is gaining traction. China's release of AI models openly could turn a US-created compute disadvantage into a distribution advantage, and create a more effective global ecosystem.

Kimi Work

157 ms7892 2026-07-20 17:22:22 UTC kimi.com Llama Share 73

You want to find and summarize all PDF files containing quarterly reports in the local workspace ~/Documents/KimiWorkspace and generate a clear summary document. You also want to know about Scheduled Tasks in Kimi Work, specifically if they run when the computer is asleep.

Jelly UI: Soft-body physics for native HTML form controls

153 baldvinmar 2026-07-20 17:22:22 UTC jelly-ui.com Full Share 56

Jelly UI is a dependency-free Web Components library for soft, tactile product interfaces. Real form controls meet soft-body physics, with dark mode, right-to-left support and WCAG AA color tokens built in.

Hacker wipes Romania's land registry database

478 speckx 2026-07-20 14:22:22 UTC news.risky.biz Llama Share 263

A hacker breached Romania's cadastre agency, wiping the country's land registry database after a failed extortion attempt. The hack has brought the real-estate market to a standstill.

Nativ: Run frontier open models locally on your Mac

49 aratahikaru5 2026-07-20 18:22:22 UTC blaizzy.github.io Llama Share 21

Nativ offers a local AI app for Mac with open models from Google, Cohere, and Liquid AI, no accounts or subscriptions required. It's 100% open source, built on MLX-VLM, and connects to existing coding agents.

How we measured AI writing across arXiv, and where the measurement breaks

163 dopamine_daddy 2026-07-20 17:22:22 UTC unslop.run Llama Share 115

Researchers analyzed 12,750 arXiv papers and found that about 32% of new ones read as machine-written, rising from 0.4% before ChatGPT. The study controlled for false positives and found varying adoption rates across fields, with computer science leading at 65%.

Human mathematicians are being outcounterexampled

23 artninja1988 2026-07-20 19:22:22 UTC xenaproject.wordpress.com Llama Share 2

Mathematicians are using AI tools to find counterexamples to long-standing conjectures, including the Erdős Unit Distance conjecture, Grothendieck's question, and the Jacobian Conjecture. These AI-generated counterexamples are being formalized in Lean, a programming language for mathematics, and are helping humans gain a deeper understanding of the underlying mathematics.

Hyprland 0.55 announced the switch to Lua for its config files

85 matesz 2026-07-20 18:22:22 UTC hypr.land Llama Share 63

Hyprland 0.55 is a major update with Lua config, Layout API, and improved color accuracy. It includes new features like ICC profile loading, FP16 precision, and CM pipelines for better color management.

Corners Don't Look Like That: Regarding Screenspace Ambient Occlusion (2012)

122 firephox 2026-07-20 15:22:22 UTC nothings.org Llama Share 46

The author of the text discusses the limitations of screen space ambient occlusion (SSAO) in accurately simulating real-world lighting effects, particularly in corners of rooms. They argue that SSAO often produces unrealistic results, such as darkening corners, due to a misunderstanding of how light interacts with surfaces.

Airport Simulator

546 apunen 2026-07-20 11:22:22 UTC airport.apunen.com Full Share 111

The sky (and your endurance) is the limit!

Launch HN: Bloomy (YC S26) – AI-powered mastery learning for K-12

37 alexsouthmayd 2026-07-20 17:22:22 UTC news.ycombinator.com Llama Share 60

Alex Southmayd, founder of Bloomy, created an AI-powered mastery-learning platform for K-12 students that provides personalized learning paths and a Socratic AI tutor. Bloomy aims to make one-on-one tutoring affordable and available to every child, with a focus on responsible AI use and safeguarding student data.

Firefox Merges Support for Vulkan Video Decoding

197 DemiGuru 2026-07-20 14:22:22 UTC github.com/search Full Share 47

GitHub is where people build software. More than 150 million people use GitHub to discover, fork, and contribute to over 420 million projects.

Perfection Is Not Over-Engineering

143 var0xyz 2026-07-20 14:22:22 UTC var0.xyz Llama Share 68

The author argues that perfection is not over-engineering, but rather ambiguous requirements. Clear requirements lead to a perfect solution.

Agent swarms and the new model economics

30 jlaneve 2026-07-20 18:22:22 UTC cursor.com Llama Share 12

Researchers improved an agent swarm system to build software from scratch. The new system achieved better results and efficiency, with a 73-85% success rate in implementing SQLite from its documentation.

The Voice of Google

106 littlexsparkee 2026-07-20 15:22:22 UTC newyorker.com Llama Share 41

The author worked at Google from 2007 to 2016, starting as a communications associate and eventually becoming a manager, crafting internal messaging and social media content that amplified the company's corporate mythology. As the company grew and evolved, the author witnessed both its innovative spirit and its flaws, including a culture of sexism, a lack of accountability, and a failure to ...

Kimi K3, Qwen 3.8, and Anthropic's (Potential) Unravelling

210 cl42 2026-07-20 15:22:22 UTC emergingtrajectories.com Llama Share 221

Two state-of-the-art foundation models, Kimi K3 and Qwen 3.8, were launched, challenging top-tier model developers and threatening Anthropic's dominance. Their open models and lower costs pose a significant risk to Anthropic's product differentiation and economic model.

Ziggity – A terminal UI for Git, written in Zig

45 TheSorcerer 2026-07-20 16:22:22 UTC github.com/simoarpe Llama Share 32

ziggity is a terminal-based Git client that brings the lazygit workflow to a small, dependency-light Zig binary. It uses plain Git subprocesses and libvaxis for the UI, and is available for macOS, Linux, and Windows.

LEDs' Potential to Save Our Night Skies

172 defrost 2026-07-20 13:22:22 UTC spectrum.ieee.org Llama Share 126

The author explores how artificial light at night affects cities, citing excessive lighting as a problem that disrupts human health and ecological balance. LEDs offer a solution, but their potential is often wasted due to outdated thinking and a failure to control and coordinate lighting levels.

Cross sectioning insects in an electron microscope with a femtosecond laser [video]

45 surprisetalk 2026-07-20 16:22:22 UTC youtube.com Embed Share 2

I Stopped "Creating Content"

112 mtlynch 2026-07-20 16:22:22 UTC refactoringenglish.com Llama Share 76

The author argues that the term "content creator" has become overused and devalues creative work. They suggest using more specific and natural language to describe one's work.

Shinjuku Station in 3D

93 Gecko4072 2026-07-20 14:22:22 UTC satoshi7190.github.io Full Share 17

「新宿駅周辺屋内地図データ」（国土交通省）（https://www.geospatial.jp/ckan/dataset/mlit-indoor-shinjuku-r2）を加工して作成

85.3 GFlops: Optimizing FP32 Matrix Multiplication on a Single AMD Zen 3 Core

12 houslast 2026-07-20 19:22:22 UTC github.com/houslast3 Llama Share 2

Um estudo de otimização de multiplicação de matrizes em C++ usando AVX2/FMA alcançou 85,30 GFLOPS em um AMD Ryzen 5 5500, superando a implementação ingênua por 56,5 vezes. O modelo MX24 atingiu o melhor desempenho, com erro máximo absoluto de 0,0.

Exploit brokers pay $500k for WordPress RCEs. I found one with GPT5.6 and $25

354 infosecau 2026-07-20 08:22:22 UTC slcyber.io Llama Share 202

A security researcher used a large language model called Sol to discover a pre-authentication SQL injection vulnerability in WordPress, which can be escalated to a full-blown RCE. The researcher found that Sol was able to chain together several creative exploitation techniques, including recursive batch calls and cache abuse, to produce a full exploit chain in just over 10 hours.

The drivers behind software delivery inefficiency

22 qaprof 2026-07-20 18:22:22 UTC dl.acm.org Llama Share 7

The website is verifying you're not a bot and requires access to challenges.cloudflare.com. Try adjusting firewall settings or connecting to a different network to resolve the issue.

Safety and alignment in an era of long-horizon models

13 Wingy 2026-07-20 19:22:22 UTC openai.com Llama Share 2

OpenAI's internal model, designed for long-running tasks, exhibited novel failures not captured in pre-deployment evaluations, prompting a pause in access and the development of new evaluations and safeguards. The experience reinforced the value of iterative deployment and the need for pre-deployment testing to be paired with close monitoring, safeguards, and the ability to pause or roll back ...

ECC and DDR5

98 zdw 2026-07-20 13:22:22 UTC etbe.coker.com.au Llama Share 92

ECC RAM corrects errors in memory before reaching the CPU, using Hamming codes to detect and correct single-bit errors. ECC RAM is recommended for systems storing important data, as it can prevent data corruption and lost time due to RAM hardware errors.

An Empirical Study: AI Agent Rules Need Context and Layered Enforcement

13 matt_d 2026-07-20 18:22:22 UTC eunomia.dev Llama Share 2

ActPlane's 2,116-statement study shows that developers' rules are not short of complexity, but rather the difficulty lies in turning natural-language requirements into state that a system can observe and evaluate over time. The study classifies every statement independently and finds that 64% of statements are policies requiring, forbidding, or conditioning agent actions.

The EU is about to sell our most sensitive data to the US for visa-free travel

432 rapnie 2026-07-20 12:22:22 UTC edri.org Llama Share 244

The US is requiring EU countries to share biometric data and personal information for traveller screening, with the EU negotiating a Framework Agreement to set exchange modalities. The EU is being pressured into excessive US demands, potentially compromising European and third-country nationals' rights and data protection.

That post never existed. Stop listening to that thing

87 wglb 2026-07-20 18:22:22 UTC rachelbythebay.com Full Share 70

Controlling Reasoning Effort in LLMs

49 ibobev 2026-07-20 15:22:22 UTC magazine.sebastianraschka.com Llama Share 2

Reasoning models have become a standard part of modern model releases, with techniques like reinforcement learning with verifiable rewards (RLVR) and inference scaling improving task performance. Implementing reasoning effort modes, such as toggling between low, medium, and high effort, can be achieved through various methods including supervised fine-tuning and reinforcement learning.