llm

A practitioner shares lessons from a keynote Q&A on why Strands agents pick the wrong tools or call them in the wrong order, and writes up a practical guide with links and code snippets on how to fix tool selection and ordering issues.

spinf-serve is an open-source (Apache 2.0) inference server that reads content once and answers many calibrated classification/decision questions in a single GPU forward pass, returning probabilities over user-defined options via an OpenAI-compatible API.

QuixiDesk is a native Rust desktop workspace for AI (macOS, Windows, Linux) with pluggable LLM providers, built-in MCP tools, browser/computer use, and a unified task/chat interface, running without Electron or Node.js.

Qwen4 has been released, with hints of a 120B parameter model surfacing via a GitHub PR in the qwen-code repository.

TryMirai open-sourced uzu, an Apple Silicon inference engine, featuring Weaver + DFlash speculative decoding that makes a 9B LLM (Qwen3.5) 4x faster locally and 2.6x faster than standard speculative decoding.

Artificial Analysis 更新了 HLE Diamond 基准评测，显示开源权重模型在该高难度基准上落后闭源模型超过 30 分。

该推文整理了15位值得关注的AI研究领域X账号，涵盖LLM、扩散模型、量化、NLP和AI评测等方向，并简要说明了每位账号的内容特色。

Claude Opus 5.5 and GPT-6 Astra independently audited 138 economics papers, flagging 2,035 potential problems and 147 potentially headline-changing findings, while a third AI attempted verification, sparking debate over AI systematically re-examining decades of published research.

Meta 分享其 Catching JiTTests 方案在生产环境大规模部署的成果：利用 LLM 自动生成会失败的即时测试，在代码合入前捕获传统测试和 LLM 代码审查难以发现的 bug，并配套发布了详细论文。

Anyscale is hosting a webinar on reinforcement learning for LLM post-training, covering RLVR with GRPO, on-policy distillation (OPD/MOPD), terminal agent post-training, and a deep dive into SkyRL's async RL capabilities.

The author argues that two common coding-agent complaints—burning through weekly usage quotas and ignoring style guides to write defensive TypeScript—stem from one attention bug: pasting 'DON'T' code examples into prompts loads those token sequences into the KV-cache, where they act as semantic attractors and prime the model to emit exactly the banned pattern, triggering long, expensive correction loops. They propose a 4-part positive override structure for prompts and passing call-graph blast radii to subagents via Graphviz strict digraphs instead of prose or Mermaid.

A practitioner shares scores for AI agent retrieval/RAG tools from a rankings site they built, noting LlamaIndex leads at 7.0 while Pinecone and Ragas lag on release freshness, and asks production teams about their retrieval stacks and evaluation practices.

A developer with a free month asks for recommendations on the best AI coding tools, models, or combinations to spend $100 on for full-stack development with strong instruction following and low bug rates.

The user claims that Chinese open-source models will have to strip out cybersecurity capabilities before they can be released in the future, speculates that future models such as K3.1 Pro, GLM 5.5, and Qwen4 will all be affected, and recommends saving existing open-source models with their full capabilities intact.

Martin Fowler discusses the phenomenon of "LLM Voice" in prose — the uncanny-valley feeling readers get from AI-written text — and urges authors to avoid letting LLMs draft their writing, warning that delegating writing also delegates thinking.

A user reports that running the quantized RedHatAI/Qwen3.8-27B-INT4 model with QEngine on dual RTX 3090s is significantly faster than their previous FP8 vLLM setup, sharing links to the config and model repo.

Transformers v5.22.0 发布，新增对 Mistral Large 4 (Le Chonk) 模型的支持。

IEEE Spectrum's Benjamin Skuse ran an experiment with IEEE Access to see if an amateur could use free LLMs (ChatGPT, Gemini) to generate a convincing fake scientific paper and slip it past a journal's defenses, highlighting how AI slop is eroding trust in academic publishing.

The open-source project academic-commercialization-agent uses a team of AI agents to automatically turn a paper PDF or research direction into a citation-backed commercialization assessment scorecard, covering dimensions like technology readiness, patent strength, and market accessibility. All experimental results — including failure cases — are published openly.

作者受到 qwenputer 项目启发，在 Qwen 模型权重内构建的虚拟机上成功运行了基础版 Doom，虽然速度很慢，但实现了「在 LLM 上跑 Doom」的趣味演示，并开源了项目仓库。