— Today's theme: the agent 'shell' is getting more attention than the model itself, from harnesses to environment generation.
Nvidia research shows that with carefully designed harness and supervisor components, Claude Opus 5 achieves a 100% score on ARC-AGI-3, highlighting the importance of agent frameworks. DeepSeek releases the deepseek-v4-flash-vision-exp model with vision input support, accessible to developers via an OpenAI-compatible API. Anthropic brings Claude Mythos 5 into Claude Security, allowing enterprise teams to run vulnerability scans without direct model access.
Nvidia publishes research showing that with a custom harness optimizing memory management and adding a supervisor component, Claude Opus 5 achieves a 100% score on the interactive reasoning benchmark ARC-AGI-3, even though the underlying model is not the strongest. Why it matters: For software engineers, this means that when building AI agents, the design of tool orchestration, memory management, and supervision mechanisms may matter more than simply upgrading the model, and it is worth rethinking where to invest in agent architecture.
HN discussion is lukewarm, but TechCrunch notes the result puts pressure on OpenAI, since ARC-AGI-3 is a benchmark OpenAI cares about deeply.
DeepSeek launches deepseek-v4-flash-vision-exp, supporting image input in JPEG, PNG, GIF, and WebP formats, usable via OpenAI-compatible Chat Completions and Responses APIs, with three image passing methods including Base64 inline. Why it matters: This is a significant expansion of DeepSeek's vision capabilities, letting developers plug image understanding into existing OpenAI-compatible workflows without switching frameworks.
HN commenters generally see this as an important upgrade, though some note image resolution is low and recognition accuracy trails competitors.
Starting August 21, 2026, Anthropic deploys the Mythos-tier model Claude Mythos 5, previously restricted to trusted defenders, into Claude Security scans, in public beta for Claude Enterprise customers with no additional model add-on. Scans connect to GitHub repositories, track cross-file data flow, and return results with CWE classification, confidence and severity ratings, and suggested patches. Why it matters: Security teams can use frontier models for vulnerability scanning without directly touching the model, and the product design limits the same model being used to write exploits, making this a notable pattern for enterprise security toolchains.
OpenAI launches an Apple Messages plugin that lets ChatGPT search messages, draft, and send replies on Mac, currently limited to ChatGPT Work and Codex users and opt-in. Why it matters: This marks ChatGPT moving deeper into local communication data. For developers it is both a productivity tool expansion and a privacy boundary concern—users must actively authorize the model to read message history on their device.
Proliferate (YC S25) releases an open-source, self-hosted AI IDE that runs coding agents like Claude Code, Codex, OpenCode, Cursor, and Grok in parallel in the same workspace, with each task getting an isolated git worktree branch, terminal, and conversation state, plus sub-agent delegation and scheduled workflows. Why it matters: For engineers managing multiple coding agents at once, this unified harness reduces orchestration complexity, and the worktree isolation provides a cleaner parallel development environment.
Show HN received 35 points and 14 comments, with the community showing interest in the practicality of multi-agent parallel workspaces.
ウェブは全体像、購読者にはあなた専用を——興味に合わせた AI 精選、プライベート RSS の統合、コミュニティの見解付きで毎朝配信。ずっと無料。
44 号配信 · 毎日150件超から読む価値ある30件に厳選
EnvHarness dynamically reshapes static environments through a programmable plugin layer, co-evolving with reinforcement learning to target agent weaknesses.
🤖EnvHarness and EnvRigger dynamically reshape static environments via programmable plugins to target agent weaknesses and improve reinforcement learning co-evolution.
The FACET framework preserves source intent and executable state in a shared repaired environment, generating high-quality executable tasks for terminal agent training.
🤖FACET constructs executable terminal tasks by preserving source intent and grounding instructions, solutions, and verifiers in a shared repaired environment to enable scalable agent training.
The SWE-bench Science benchmark covers 119 tasks across 20 scientific domains, revealing failure mechanisms when coding agents repair scientific software.
🤖SWE-bench Science benchmarks coding agents on scientific software repair, revealing failure mechanisms and mixed effects of scientific guidance.
MemTrapBench finds that retrieved memories can induce reasoning errors and belief distortion in LLMs, and proposes inference-time strategies to avoid cognitive traps.
🤖Retrieved memories can induce reasoning errors and belief distortions in large language models, and an inference-time strategy helps avoid these cognitive traps while maintaining benchmark performance.
Claudette is a Claude Code skill that uses Gemini CLI to translate Claude's BuzzFeed-style replies into normal English.
llm 0.32.1 fixes a fresh-install failure caused by the OpenAI Python library deprecating httpx; 0.33 will switch to httpx2.
llm-openrouter 0.7 is compatible with LLM 0.32, switches to OpenRouter's Responses API, and adds three server-side tools: Shell, WebFetch, and WebSearch.
Seed is a minimal self-modifying agent harness that provides only one exec tool, letting the agent grow tools, memory, and skills in a self/ directory.
Bun 1.4 is the first stable release after the Rust rewrite, adding APIs like Bun.WebView and Bun.Image with greatly improved Node.js compatibility.
Encore rebuilds the Linux MicroVM stack on Apple Silicon, letting Mac developers run Firecracker microVMs locally.
Kagi adds a setting to automatically remove paywalled links; commenters generally praise it as better than Google, though some want to keep quality paywalled content or have a whitelist.
Commenters generally praise Kagi's paywall-removal feature as practical and better than Google; some, however, think quality paywalled content should be kept or a whitelist option provided.
Anna's Archive calls for volunteers to scan rare books before AI companies destroy physical copies, sparking discussion on training data and cultural preservation.
A security researcher discloses accidentally recording hundreds of thousands of military base calls by taking over the e164.arpa domain, raising concerns about ENUM infrastructure security.
OpenRouter launches anonymous stealth model Ox Alpha; commenters widely speculate it is a Chinese GLM-series model, acknowledging capability but questioning the anonymity.
Commenters widely speculate Ox Alpha is a GLM-series Chinese model, acknowledging its capability but questioning the anonymity; others think it could be a Western model.
Engineers at big tech companies suggest putting self-hosting experience on resumes, arguing it covers real engineering problems like deployment, monitoring, and security in miniature.
基于官方 DeepSeek Harness 打造的 Electron 桌面端，深度适配 macOS 和 Windows，提供最佳的，开箱即用的体验。
29 editorial diagram types for Claude Code. Self-contained HTML + SVG. No shadows, no Mermaid-slop.
The DeepSeek adapter adds the multimodal visual understanding model DeepSeek-V4-Flash-Vision-Exp. It also supports configuring native image requests. Commands such as /goal and /plan can accept text and image input, and the @ menu can reference files and sessions; MCP/ACP also supports persistent image attachments, and PTC Mode supports forwarding nested images.
离开谷歌的原因之一：小团队可以极致聚焦！
Three ~300M drafters bring speculative decoding to LFM2.5, delivering up to 3.18x faster decoding with identical greedy output. The post Liquid AI Releases LFM2.5-DSpark Draft Models That Deliver Up to 3.18x Faster Decoding Without Changing Model Outputs appeared first on MarkTechPost .
I tried this model yesterday, and it felt to me like the best one I've tried for a local model for interactive use; the responses and reasoning are very fast, and it actually performs agentic tasks well. The speed is phenomenal. I am running this on Ninfer for Windows
A quick feedback after a really major test: nearly 20 hours of non-stop goal-oriented work with Qwen3.8-27B Q6, running across an RTX 3090 and an RTX 3060. It maintained a speed of around 60–63 tokens/s throughout the session.
Here's a brand new Blackwell-native, prefill-optimized 4-bit quant that runs 50% faster on compatible hardware than a Q4 quant of the same memory footprint. And it runs 4-7% faster than other NVFP4 quants as benchmarked on RTX 5090 32GB. Quant Benchmark Speed NVFP4 pp2048 6250 t/s unsloth NVFP4 pp2048 6010 t/s Q4_0 pp2048 4130 t/s Q6_K pp2048 3210 t/s This GGUF also includes a quantized MTP draft head for a good measure. Check it out for all details and specifically recommended settings for 15%
Run Claude Code, Codex or Opencode in cloud with your repo Discussion | Link
Artificial Analysis just benchmarked them and the scores are crazy good, proving the earlier success wasn't only enabled by overthinking.
Nvidia continues to pour money into data center development — just as AI data centers bring lots of money into Nvidia.
Google DeepMind partners with game studios to prototype breakthrough AI gameplay.
When the biotech company Insilico Medicine used its computer models to propose a promising drug for pulmonary fibrosis, it enthusiastically claimed in a press release that the molecule had been “discovered by” its generative AI platform. Insilico leads a pack of companies using AI to rapidly come up with drug ideas humans might never think…
Yes, we’re confused too.
arXiv:2608.18111v1 Announce Type: new Abstract: Foundation models such as GPT and Claude now solve olympiad-level mathematics with remarkable proficiency, so much so that geometry problem solving has become a standard proxy for their mathematical reasoning. Yet solving a geometry problem and drawing the figure it depends on are not the same skill: progress often hinges on a faithful diagram with the right auxiliary constructions and incidences, and it is unclear that a model which reasons its wa
ChatGPT search now uses the site:operator at scale Promptwatch is part of the emerging "GEO" space, for Generative Engine Optimization - the chatbot version of SEO, where companies offer tools and consulting to help your site increase its presence in replies to prompts inside tools like ChatGPT. The Promptwatch product uses automation to track responses to prompts across end-user chat products like ChatGPT, Claude, and Gemini. They publish aggregate reports on this as part of their own content m
Action-conditioned video world models require low-latency causal generation and reliable responses to game-native controls. Although causal distillation enables one- or few-step video synthesis, extending it to interactive world models remains challenging, as discrete keyboard states and continuous mouse motion must remain aligned with temporally compressed latent chunks during causal training and autoregressive rollout. We introduce ForgeWM, a progressive framework that transforms a bidirection
S1-mini is a 462 MB open-weights normalizer that sits after ASR, removing fillers and resolving self-corrections locally. The post Meet S1-mini: Superwhisper’s 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text appeared first on MarkTechPost .
So usually I avoid Q3 quants because I have had bad experiences with it, models were usually too degraded, so the smallest I normally do is Q4, since I only have rtx 4060 ti 16gb. But since there hasn't been a 35b-3ab released yet, I had to try it. I don't use LLMs in agentic workflows, just on Textgen since I'm not a coder so this is not the primary use case of LLMs for me - but sometimes I really need some coding capabilities or help. I'm very impressed how it one shot multiple serious coding
LinkedIn says its AI slop button is working.
My Pro subscription expired today, they killed my access at 1pm local time. I'm now using Qwen3.8-27b w/ 5090m 24gb vram and pi to do everything i was doing in claudecode. The only downside is claudecode let me code without using my gpu, meaning I have to plan things now. Last night I had ChatGPT write up a prompt for a fancy aurora predictor for Canadians. I fed it to local pi and claude sonnet 5. They took about the same time, pi's app looked better, but claude's had better science. I asked th
Identity-preserving image generation becomes increasingly unreliable when a scene must contain many specified people. Beyond retaining each identity, the model must bind every reference to a distinct person and location, while training-time identity losses must establish correspondence among several noisy predicted faces. We introduce WithEveryone, a unified framework for generating group images up to ten reference identities. WithEveryone injects each selected identity as an addressed token, pr
I have been using omarchy on my tower since nearly a year now, shortly after it was released first. I really love the experience I am having with it but I still use my macbook for daly work, so I wanted to recreate a similar experience on it. Thats why I created omacosy, a setup for tiling windows, custom menu bar, some themes from omarchy, focus follows mouse, focus rings around windwos, some mac flavors with trackpad events and a custom mission control overview for your workspaces. I used Aero
Your personal AGI on your PC. Built to finish real work. Discussion | Link
arXiv:2608.18080v1 Announce Type: new Abstract: We present a review on the applications of large language models (LLMs) in health, e.g., social media analysis, clinical conversational agents, therapy support tools, prompt engineering, multimodal learning, and ethical considerations. We integrate findings from interdisciplinary studies utilizing diverse data sources such as social media posts, electronic medical records, and multimodal inputs to enable early detection of depression, suicide risk