An Empirical Study of Harness Design for Coding Agents

117 wek 2026-09-18 13:41:53 UTC arxiv.org Llama Share 20

A study evaluates a lightweight coding harness for autonomous coding agents, varying planning, action space, and context management across four models on two benchmarks. The findings highlight the importance of context management and suggest strategies for harness design.

Cloudflare Quick Tunnels

43 jcbhmr 2026-09-18 14:41:53 UTC try.cloudflare.com Llama Share 22

Cloudflare offers Quick Tunnels for free, allowing instant global preview and deployment of local applications with a single command. No login is required, and it works seamlessly with existing tools to accelerate feedback loops.

North Korean nuclear test sets off years of earthquakes

26 rbanffy 2026-09-18 14:46:53 UTC science.org Full Share 10

C++26: Trivial infinite loops are no longer undefined behaviour

28 ibobev 2026-09-18 14:41:53 UTC sandordargo.com Llama Share 21

C++26 has fixed undefined behavior for trivial infinite loops, which were previously allowed to be optimized away by compilers. This change ensures that such loops are now well-defined and can't be removed by optimizers, preventing potential security vulnerabilities in embedded systems.

Bend 2 and the Vibe-Coding Trap

266 LiamPowell 2026-09-18 12:41:53 UTC blog.liampwll.com Llama Share 184

The Bend language is being pitched as a solution for AI coding, but its creator missed the current standard in formal verification by not researching existing solutions. Vibe-coding led to unnecessary complexity and verbosity in Bend's design.

BeanShell3 in Development

8 jimjag 2026-09-18 14:41:53 UTC beanshell.github.io Full Share 4

Work has started on a full release of BeanShell 3.0 after a long stale period. BeanShell3 will be completely updated and support new, sought-after features such as Lambdas.

OpenJev

324 ilreb 2026-09-18 10:41:53 UTC openjev.com Llama Share 187

A local model can read probabilities or write distribution tokens, with MiniCPM5 2B being the default desktop size. The model's weights are quantized and inputs remain in the browser cache.

The Shadows Lurking in the Equations – Underwater Islands

38 vismit2000 2026-09-18 12:41:53 UTC gods.art Llama Share 6

The author discusses FuzzyGraph, a visualization tool that shows equations in a non-binary mode, revealing previously invisible mathematical shadows and errors. This allows for better understanding of the underlying topography of equations.

I Vibed a Proof of Conway's Conjecture

63 m-hodges 2026-09-18 14:41:53 UTC overreacted.io Llama Share 50

A user attempted to use AI models, specifically ChatGPT and Claude, to prove Conway's refinement conjecture in surreal numbers. After a month of work with various approaches, the user found that while the models could generate novel results, they struggled with formalizing these results in Lean and verifying their correctness. The process involved multiple iterations of refining the approach ...

I don't like passkeys

384 ethanhawksley 2026-09-18 12:41:53 UTC hawksley.dev Llama Share 354

The tech industry promotes passkeys as a secure login solution, but they pose risks for individuals due to permanent account lockout and device loss. Passkeys offer better security against phishing attacks, but their limitations make them unsuitable for personal use.

Show HN: Rickub – The Smartest Git in the Universe

5 ssaboum 2026-09-18 14:57:53 UTC rickub.com LocalLLM Share 0

Rickub offers a fast, affordable alternative to GitHub and GitLab with robust code review, CI/CD, and AI-powered feedback. It supports familiar workflows, secure collaboration, and seamless integration with OCI artifacts and CLI tools.

NATS publishes preliminary report on technical incident of 8 September

15 asplake 2026-09-18 13:41:53 UTC nats.aero Llama Share 6

NATS has released a preliminary report on the recent air traffic disruption, attributing it to a software issue in their flight data system. The incident was resolved and safety remained unaffected, with NATS apologizing for the inconvenience caused to passengers.

Cekura (YC F24) Is Hiring

0 2026-09-18 12:41:53 UTC ycombinator.com Llama Share

Cekura is hiring a Forward Deployed Engineer to work with customers and help them get value from their self-improving conversational agents. The role involves embedding with customers, automating product insights, driving product direction, and building the FDE org.

Jemalloc 5.4.0

261 gkfasdfasdf 2026-09-18 04:41:53 UTC github.com/jemalloc Llama Share 63

This release includes over 160 commits focusing on technical debt cleaning, bug fixes, and performance improvements. It also introduces new features such as pinned memory tracking and OS abstraction layer.

Warren Buffett Steps Down as Berkshire Chairman, Names Son to Replace Him

181 saimiam 2026-09-18 11:41:53 UTC nytimes.com Full Share 121

Please enable JS and disable any ad blocker

The scourge of x86 emulation

226 dagmx 2026-09-18 04:41:53 UTC fex-emu.com Llama Share 59

The x86 Total Store Ordering (TSO) memory model poses significant challenges for emulating on ARM-based systems, with issues including alignment faults, unaligned atomic operations, and poor performance due to write-combined memory. Despite progress in addressing these problems through extensions like FEAT_LRCPC1/2/3 and hardware support for coherent cachelines, edge-cases remain unresolved, ...

Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint

523 JonSchneider 2026-09-17 21:41:53 UTC prismml.com Llama Share 170

The Ternary Bonsai 2 27B model is released, offering a more capable and energy-efficient version of the previous Bonsai 27B model. It retains 98.2% of aggregate benchmark performance while being nearly 9x smaller in memory footprint.

Build Faster Feedback Loops Using Qualitative User Research

5 nadis 2026-09-18 14:59:53 UTC blog.nseldeib.com LocalLLM Share 0

Early-stage founders should prioritize rapid feedback loops to test hypotheses and validate product-market fit. Use design sprints and qualitative research to accelerate learning and reduce time-to-failure.

Astra for Law

553 vertigoruntime 2026-09-17 20:41:53 UTC openai.com Llama Share 647

OpenAI introduces Astra for Law, a foundation for law firms and legal technology companies to build AI products around their expertise. It combines GPT-6 with settings and tools tailored for professional legal work.

Microsoft exec called AI scraping 'the largest theft of labor in human history'

622 pluc 2026-09-18 10:41:53 UTC techcrunch.com Llama Share 522

Microsoft and OpenAI allegedly trained AI models on copyrighted content without permission, bypassing paywalls and stripping copyright notices. The companies' actions pose an "existential threat" to publishers and journalists whose work was used to train the models.

Bend – A language that blocks AI mistakes via proof, on CPU and GPU

557 nicolas-siplis 2026-09-17 20:41:53 UTC bend-lang.com Llama Share 283

The Bend system uses laws and proofs to ensure AI agents implement human intents correctly, verifying code changes in real-time. With LAWS.bend, developers can declare rules that AIs must follow, preventing bugs from being shipped into production.

Mathematicians Build Long-Awaited Graph Sandwich

8 ibobev 2026-09-18 14:42:53 UTC quantamagazine.org LocalLLM Share 0

Mathematicians solved the sandwich conjecture, proving random regular graphs can be approximated by random binomial graphs. This allows transferring properties from easier-to-analyze binomial graphs to complex regular graphs.

Second Circuit Allows Government to Search Electronic Devices at the Border

33 HotGarbage 2026-09-18 14:41:53 UTC knightcolumbia.org Llama Share 19

The US Court of Appeals for the Second Circuit ruled that border agents can search travelers' electronic devices without suspicion. The ruling was met with disappointment from the Knight First Amendment Institute, which argued a warrant should be required.

ZCode, the GLM coding agent, silently uploads your Git history

229 cdnsteve 2026-09-18 10:41:53 UTC tokenstead.ai Llama Share 53

A developer found that ZCode, an AI coding desktop app from Z.ai, silently uploads a user's entire workspace to Aliyun OSS when logged in. The uploaded data includes .git history and other sensitive information without the user's consent.

Qwen 3.8 Omni Flash

293 jjcm 2026-09-17 23:41:53 UTC qwen.ai Full Share 104

Qwen Studio offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.

Replacing Pull Requests with Delta

101 vquemener 2026-09-18 08:41:53 UTC zed.dev Llama Share 54

Delta is a public beta for a multiplayer coding environment that allows collaboration without committing and pushing code. It offers continuous engineering, where the idea, implementation, review, and landing of changes can happen in the same place.

Subnormal floating-point numbers are expensive on Intel processors

53 zdw 2026-09-18 11:41:53 UTC lemire.me Llama Share 40

A C++ benchmark measured the performance impact of subnormal floating-point numbers on Intel processors, finding multiplications to be 45-50 times slower and divisions 18 times slower. However, AMD and ARM processors handle subnormals at full speed.

When the fractional part of a float fixes your shader

80 vinhnx 2026-09-18 06:41:53 UTC crocidb.com Llama Share 15

The author implemented a Voronoi-diagram shader for a music video project and encountered a stuttering issue on one computer, which they debugged over the course of a week. They discovered that changing a single literal float value in the shader code fixed the issue, suggesting an optimization error.

How to Write with an LLM

256 joeriddles 2026-09-17 22:41:53 UTC sockpuppet.org Llama Share 171

A writer shares two rules to use Large Language Models (LLMs) effectively, avoiding "LLM-creep" that can make writing sound unnatural. By using LLMs as a copyeditor rather than a ghostwriter and following these rules, writers can improve their work without losing their unique voice.

Pre-Greek: The lost language hidden within Ancient Greek

123 axiologist 2026-09-18 03:41:53 UTC linguisticdiscovery.com Llama Share 57

The article discusses the mystery language that influenced Ancient Greek, known as Pre-Greek. This language is thought to have been spoken in the Aegean region prior to the arrival of Indo-European speakers and has left a significant impact on the vocabulary of Ancient Greek. The authors analyze various linguistic features, including suffixes and prefixes, to identify potential loanwords from ...