Community reviews next-generation coding benchmarks like DeepSWE, Terminal-Bench, and Program-Bench, believing they better reflect deep software engineering capabilities.
— Today's main thread: open-source models are running wild locally, while OpenAI is disclosing its own research acceleration data.
TL;DR du jour
OpenAI announced reaching its 'automated research intern' goal and released internal data on coding agents accelerating research; GPT-6 Astra impressed in robotic manipulation but cost and stability remain questionable. The local LLM community launched large-scale comparative evaluations of abliterated variants of Qwen 3.8 27B, alongside multiple local agent and inference optimization practices. Nitter, after receiving a cease-and-desist letter from X Corp, decided to continue operations following legal consultation, sparking discussions in the open-source community about platform monopoly.
À la une
1
OpenAI achieves 'automated research intern' goal, releases internal data on agent-accelerated research
OpenAI announced it has achieved an 'automated research intern'—a system that can complete work under human guidance that would take a skilled researcher days—and stated it is moving toward building an 'automated AI researcher' by March 2028. Meanwhile, OpenAI released internal data showing how coding agents improve experiment speed, task complexity, and research acceleration. Why it matters: This marks frontier labs beginning to use agents to accelerate their own research. For software engineers, it means the boundary of AI-assisted development is expanding from code completion to full research tasks, and its engineering practices and reliability are worth attention.
Hacker News commenters acknowledged the potential of agents in research acceleration, but some remained cautious about the credibility and safety of 'automated researchers'.
2
GPT-6 Astra robotic manipulation evaluation: 19/20 success on bowl task, only 2/20 on puzzle task
GPT-6 Astra, when performing the 'pick up red block and place in bowl' task on a YAM robotic arm, succeeded 19 out of 20 trials, far exceeding Claude Fable 5.1's 8/20, at a cost of about $0.94 per trial; however, it succeeded only 2/20 on the puzzle task, matching Fable 5.1, with the final insertion step stalling. Why it matters: This shows frontier models are approaching practicality on simple manipulation tasks, but fine manipulation remains a bottleneck; for engineers focused on embodied intelligence and agents, this is a typical sample of the trade-off between model capability and cost.
Most acknowledged Astra's performance in coding and game building, but some felt its robotic control is costly and ineffective, and that the evaluation methodology has limitations.
3
Comparison of 8 Qwen 3.8 27B abliterated variants: 167 GPU hours, 11-day evaluation
The community conducted a comparison of 8 abliterated (de-censored) variants of Qwen 3.8 27B on Hugging Face, taking 11 days and 167 GPU hours, using methods such as weight comparison and KL divergence to verify their claimed de-censoring effects. Why it matters: Local LLM users have strong demand for 'uncensored' models, but quality varies widely; this evaluation provides quantitative evidence for selecting reliable variants and reflects the vitality of the open-source model fine-tuning ecosystem.
The community showed strong interest in the evaluation results, believing some variants are not what they claim to be, but others emphasized that abliteration harms the model's overall capabilities.
4
Nitter and XCancel decide to continue operations after legal consultation following X Corp cease-and-desist letter
X Corp sent a cease-and-desist letter to the Nitter project on August 24, 2026, demanding permanent removal of instances and repositories; after legal consultation, the Nitter project announced it will continue operations, with more details to be released later. Why it matters: Nitter is a privacy- and performance-focused Twitter front-end alternative; its survival concerns the open-source community's practices in resisting platform enclosure and tracking, and tests the resilience of licenses like AGPL under real legal pressure.
Commenters generally supported Nitter resuming service, believing X's monopoly is unfair, but some noted legal risks remain and caution is needed.
5
Asahi Linux officially supports M3-series Macs, GPU and DCP still pending breakthroughs
Asahi Linux announced that M3-series SoC support has been merged into the installer, with almost everything working including camera, microphone, USB 3 at 10 Gb/s, hardware video decoding (including AV1), WiFi, and Bluetooth; the main exceptions remain full DCP support and GPU. Why it matters: This is a major milestone in reverse engineering Apple's closed hardware, providing developers who want to run Linux on M-series Macs with a more complete option, but the missing GPU means 3D acceleration and power efficiency remain limited.
Commenters generally praised the Asahi Linux team's reverse engineering achievements, but some felt support is insufficient, the target user base is unclear, and worried about Apple legal risks.
Chaque matin, un digest tech fait pour vous
Le web montre la vue d’ensemble ; les abonnés reçoivent la leur — sélection IA selon vos intérêts, votre RSS privé intégré, avec les avis de la communauté, livrée chaque matin. Gratuit à vie.
58 numéros publiés · 150+ infos filtrées à 30 chaque jour
Actu IA
Skild AI releases robotic foundation model S1, which can complete unseen long-horizon manipulation tasks from just a single task demonstration video, making ICL a new direction in embodied intelligence.
Users believe Qwen 3.8 Flash Next (Max) excels in casual conversation and factual knowledge, with coding ability overshadowing its general dialogue value.
Dev & open source
Show HN: Mador – Make any DOM reactive with a tiny 80-line Proxy state tupleShow HN1 minOpen sourceOutils dev
Mador is a minimal reactive DOM runtime with only 80 lines of Proxy state tuples, suitable for developers who don't want to use frameworks.
Show HN: Kadō – open-source habit tracker, with non-binary habit score, for iOSShow HN4 minOpen sourceProduits
Kadō is an open-source iOS habit tracker with non-binary habit scores, offline-first, no accounts, no telemetry, and lossless data export to JSON/CSV.
Anubis implements a Proof-of-Work anti-scraping solution using WebAssembly, taking a year to complete WASM integration, to defend against large-scale scraping by AI companies.
GrapheneOS is rewriting its default Messaging app with Android Compose, with plans for long-term RCS and end-to-end encryption support.
Échos de la communauté
Qwen3.8-27B "Unhacked" my PCreddit4 minIASécurité
A user shares how Qwen3.8-27B helped them investigate and clean up a PC intrusion; commenters have mixed views on the reliability of local models for security analysis.
2x R9700, 64 GB DDR5 is an absolute beast machine with vLLM Radiance / R9V and Qwen 3.8 27b and Flash nextreddit3 minIACloud & infra
Dual R9700 + 64GB DDR5 with vLLM Radiance/R9V running Qwen 3.8 27B, which users call a 'beast machine' for local inference.
Block KV cache streaming: bound VRAM at long context via a shared CUDA phase arena by giveen · Pull Request #357 · TheTom/llama-cpp-turboquantreddit1 minIAOpen source
Block KV cache streaming limits VRAM under long contexts via a shared CUDA phase arena, now extended to multiple models with benchmarks completed.
An article about complex feelings toward AI sparks heated discussion: amazement at emergent reasoning in neural networks, fear of superintelligence risks, and disgust at AI companies destroying the open web coexist.
Commenters mainly discuss fear and optimism about AI; most worry about its social impact, but some believe AI will bring progress.
GitHub Trending
affaan-m/ECC★ 251280
Sponsor Star affaan-m / ECC The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
mattpocock/skills★ 254497
Sponsor Star mattpocock / skills Skills for Real Engineers. Straight from my .agents directory.
cathrynlavery/diagram-design★ 32327
Star cathrynlavery / diagram-design 38 editorial diagram types for Claude Code, Codex, and Pi. Self-contained HTML + SVG. No shadows. No Mermaid slop.
NousResearch/hermes-agent★ 242531
Star NousResearch / hermes-agent The agent that grows with you
openai/skills★ 25608
Star openai / skills Skills Catalog for Codex
anomalyco/opencode★ 205246
Sponsor Star anomalyco / opencode The open source coding agent.
blader/humanizer★ 44213
Star blader / humanizer Agent skill that removes signs of AI-generated writing from text
llvm/llvm-project★ 40207
Sponsor Star llvm / llvm-project The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.
DietrichGebert/ponytail★ 129311
Sponsor Star DietrichGebert / ponytail Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
ruvnet/ruflo★ 70973
Star ruvnet / ruflo 🌊 The original agent meta-harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
Aussi à voir（6 de plus）
Feels like AI has been especially taxing on the mental states of developers and the outlook of this industry in general (understandably so). Would anyone else be interested in a weekly thread to share your "wins" or other good things as a brief reprieve?
Reproducibility seems to be headed towards irrelevance in ML research. Is it too late? [D]reddit1 minRechercheIA
I feel that reproducibility is now a lost cause in machine learning research for three reasons: Many research is moving towards the physical AI territory, where you need expensive hardwares or even entire laboratories with high-speed cameras, in order to perform an experiment. You truly have no idea if the experiment can be reproduced and have to trust the demo. But demos are not perfectly reliable. Plus people are incentivized to only show the part of the demo that works. The entire system can