2026年9月27日 技术热点总结

项目简介：Every agent’s model. One place. Codex on DeepSeek, Claude Code on Kimi, from the menu bar.

技术栈：Go、claude-code、codex、deepseek、gemini-cli、llm、macos

核心介绍：One place to pick every agent’s model: Codex on DeepSeek, Claude Code on Kimi, Gemini CLI on GLM, from the menu bar. usemagpie.ai magpie is a single screen that lists each AI agent on your machine and the model it is set to. Click a value, pick a model. That is the whole app. It lives in the menu bar: click the icon and a panel drops down; the same screen also opens as a normal window (magpie, or *Open magpie* in the tray m…

项目简介：Turn any LLM into a Jev-style decision model: typed decisions, real probabilities, no training. (continue updating, welcome any issue and PR request)

技术栈：Python、calibration、decision-model、jev、jev-model、llm、system-one、transformers、vllm

核心介绍：Jiamu Zhang1 Tianze Yang1 Yucheng Shi2 Liang Wu1 1 Nokia, Sunnyvale, CA 2 Tencent Hunyuan Qwen3-8B on a real BANKING77 item. Every number is a model output. > 🆕 vLLM serves every level, L2 included. An embed server’s pooler hands back the hidden > state a closed-form head reads, so a decision endpoint is a pooling …

项目简介：微信聊天记录一键转发到 AI Agent 与 Obsidian 的原生 macOS 工具

技术栈：Swift、ai-agent、chat-history、macos、obsidian、share-extension、swift、wechat

核心介绍：WeChatBridge（微信流） 从微信转发菜单，把聊天记录送进 AI Agent 与本地知识库。 原生、轻量、完全本地的 macOS 微信聊天记录转发与归档工具。 简体中文 · English 官网 · render.qmuse.pub > 已发布经过 Developer ID 签名和 Apple 公证的 WeChatBridge 0.1.14 DMG。 macOS 微信 4.1.13 起，多选聊天记录后可以“合并转发”给第三方应用。微信会生成一份包含 TXT、图片和视频的 ZIP，天然适合交给 AI Agent 处理，也适合沉淀进本地知识库。 系统的“转发到其他应用”列表只展示带有 Share Extension 的 App。微信流补齐这层入口，让一次转发直接抵达 Codex、Claude、豆包、千问办公、WorkBuddy、WeSight、Obsidian、剪贴板或你指定的其他应用。 A[微信多选聊天记录] –> B[合并转发] B –> C[WeChatBridge …

项目简介：Coastal town built with Opus 5.5

技术栈：JavaScript

核心介绍：An island fishing game for the browser. Cast from the pier, the beach or your own boat, fight the fish, sell your catch to Joe at the fish stand, and spend it on better gear at Marta’s chandlery. Around it is a real-time tropical island and ocean: swim the reef, drive the boat out to deep water, and watch a humpback breach. It runs directly on WebGPU and WGSL with its own small rendering engine, no framework. the render down on slower machines. faster because the browser caches them.

关键特性：A spinning rod and reel th…

Recently, Large Language Models (LLMs) have been increasingly able to solve advanced mathematical problems, including many that have been open for decades. This opens the door to expansion of mathematical knowledge at unprecedented scale. Yet, while LLMs may be able to conjecture and prove more and more theorems, it remains open whether this new mathematical knowledge is interesting or useful. We define intrinsic interestingness of a theorem as the ratio between the length of its proof and the length of its statement. We show that this correlates strongly with an extrinsic measure of the do…

In this paper, we propose RGBD20K, a novel dataset for facilitating the development of more robust and general RGB-D semantic segmentation by encompassing abundant categories and high-quality annotations. RGBD20K possesses several attractive properties: (1) Expanded Semantic Space. In particular, it covers 160 fine-grained categories, largely surpassing the category diversity of existing popular RGB-D benchmarks (e.g., NYUv2 with 40 classes and SUN RGB-D with 37 classes). With such enriched semantic coverage, we expect to promote the learning of more generalizable segmentation models. (2) L…

Recent years have witnessed major progress in joint audio-video generation. Existing models still suffer from limited per-modality fidelity, insufficient text-modality alignment and weak cross-modal synchronization. While reinforcement-learning post-training offers a promising remedy, directly adapting it to joint audio-video generation is challenging. Heterogeneous multimodal rewards entangle learning signals and complicate credit assignment. Joint optimization of two modality towers is computationally expensive given their divergent dynamics. Moreover, synchronization evaluation difficult…

While Large Language Models (LLMs) rely on highly non-linear components, in this work we demonstrate that they exhibit fundamental linearity: when inputs from distinct text streams are linearly combined, the model outputs a superposition of the individual next-token distributions. We term this the Superposition Linearity Hypothesis. We provide evidence that superposition is an intrinsic property of the Transformer architecture rather than an emergent consequence of training; in fact, we observe that it tends to diminish as pretraining progresses. However, we demonstrate that linearity can b…

Detectors of alignment failures screen deployed language models and score alignment benchmarks. Most are generative judges that spend a decoding pass on every criterion, and classifiers that read token probabilities, such as Llama Guard, still score one fixed label per call. Jev, a model trained with reinforcement learning for calibrated decisions (RLCD), answers many typed questions about one input with calibrated probabilities in a single call. Whether it detects alignment failures has not been measured. We present RLCDAlignBench, which benchmarks Jev on ten alignment failures: sycophancy…

Sparse AutoEncoders (SAEs) offer a promising way to inspect language model representations, but it is still unclear what kind of linguistic structure their latents expose. We use part-of-speech (PoS) categories as a controlled test case to study whether morpho-syntactic information is encoded by individual latents or by structured groups of features. We find that PoS distinctions are highly recoverable from SAE activations, but do not align with one-to-one latent / category mappings. This recoverability is not reducible to lexical memorisation, and Open and Closed PoS classes differ substan…