Frank @bro_Xiu

The frontier has opened up: four frontier launches in eight days - Grok 4.5, GPT-5.6, Muse Spark 1.1, and yesterday Kimi K3; six labs now have a model scoring >50 on the Artificial Analysis Intelligence Index, up from two in early June @SpaceXAI's Grok 4.5 (high, Artificial Analysis Intelligence Index Score: 54) landed July 8, @OpenAI's GPT-5.6 Sol, Terra, and Luna (max: 59, 55, 51) and @AIatMeta's Muse Spark 1.1 (xhigh, 51) followed a day later, and @Kimi_Moonshot's Kimi K3 launched yesterday at 57 - third overall, ahead of Claude Opus 4.8 (max, 56). The top three models on the Index now come from three different labs and span just three points. Four of the ten highest-scoring models launched since July 8, and six of ten since early June. The one thing that did not move is #1: Claude Fable 5 (max, 60) has held the top spot since June 9, but its lead has narrowed from four points to one, and the price of the intelligence beneath it collapsed. Congratulations to @elonmusk, @sama, @finkd, and the teams at all four labs on a remarkable eight days. Key Takeaways: ➤ The frontier went from two labs to six in six weeks. Until June, only Anthropic and OpenAI had fielded a model at 51 or above. GLM-5.2 (max) brought Z AI in mid-June; last week added SpaceXAI and Meta; today Kimi K3 makes Moonshot AI the sixth as they enter at 57 ➤ Kimi K3 debuts at #3 with agentic and knowledge work scores behind only the top two. K3 scores 1668 Elo on GDPval-AA v2, third behind Claude Fable 5 (max, 1760) and GPT-5.6 Sol (max, 1748). On AA-Briefcase, our benchmark of long-horizon knowledge work, it enters at #2 with 1547 Elo - behind only Claude Fable 5 (max, 1583) and ahead of GPT-5.6 Sol (max, 1495) - with an Analytical Quality Elo (1760) effectively tied with Fable 5 (1764). At $0.94 per Intelligence Index task on its $3/$15 pricing, it delivers comparable intelligence to Claude Opus 4.8 (max, $1.80) at roughly half the cost per task ➤ Near-frontier intelligence got 2-3x cheaper in eight days. GPT-5.6 Sol (max) delivers one point below Claude Fable 5 (max) at $1.04 per Intelligence Index task vs $2.75. Grok 4.5 (high) delivers 54 at $0.31, under a third of GPT-5.5 (xhigh, $0.99). At 51, GPT-5.6 Luna (max, $0.21) and Muse Spark 1.1 (xhigh, $0.26) undercut GLM-5.2 (max, $0.32), the cheapest at that level a week earlier

Akademisyenler için Claude Code’u nasıl kullanacağınıza dair basit bir giriş. Alessandro Spina'ya ait sunum slaytları ve GitHub deposu. 🔗 github.com/aspi6246/Claud…

7M Parameters Beat Gemini 2.5 Pro on ARC-AGI — By Learning to Second-Guess Themselves | Find papers faster on arXivSub with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI) medium.com/p/6bd9f8961328

Highly-recommended overview of metacognition in LLMs. (bookmark it) Interesting behaviors in LLMs like confidence calibration, self-verification, knowing when to stop, and knowing what you do not know have mostly been studied in isolation. This survey argues they are facets of one thing, metacognition, and proposes a comprehensive map of it. The authors taxonomize methods and benchmarks for measuring and evaluating metacognitive abilities in LLMs, then connect those abilities to capability, reliability, and transparency. As agents take on longer horizons, the ability to monitor and regulate their own reasoning becomes an important way to measure reliability. Paper: arxiv.org/abs/2607.11881 Learn to build effective AI agents in our academy: academy.dair.ai

One of my favorite results from the BSF paper: Fourier modes in *vision* models!

Goodfire @GoodfireAI

2 weeks ago

We also revisited an interpretability classic: curve detectors in InceptionV1. Neurons and SAE features turn out to be fragments of one continuous orientation feature, and the block *also* contains higher-order Fourier harmonics that hadn’t been described before! (7/9)

2 7 108 22K 36

Great Stanford + MIT + Harvard + Anthropic paper. Gives a clear training-based reason for why larger models learn abilities smaller models miss. Says bigger AI models learn rare skills because they forget them less during training, their extra space protects weak learning signals. The authors say the issue is not just whether a small model could represent the task, but whether training lets it keep that task while many common tasks keep pushing on the same limited parts. Their core idea is that common tasks take up the model’s neurons first, so rare tasks get overwritten before they appear often enough to build into stable knowledge. In a crowded data mixture, common patterns get first claim on the model’s internal machinery. Small models may briefly pick up a rare signal, but the next wave of common-task updates overwrites it before the signal appears again. They tested this first with controlled toy tasks where they could change how rare and complex each task was, then with OLMo language models from 4M to 4B parameters. The main result is that bigger models learned low-frequency tasks much better, kept more task features inside their representations, and showed less gradient interference, which means common-task updates disturbed rare-task learning less. Larger models can remember weak rare signals long enough to turn them into real learned skills. ---- Link – arxiv. org/abs/2605.29548 Title: "Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention"

🐋 开源项目推荐：Orca —— 让你的 AI 编码 Agent 实现并行舰队管理的开源 ADE 平台 一个人开着五个 AI Coding Agent 并行干活，不乱套吗？Orca 说：完全可以。 Orca 是一个专为 AI 编码代理设计的并行编排平台（ADE），支持在桌面和移动端同时运行多个编码代理例如 Claude Code、Cursor、Codex、Grok、OpenCode 等。 你可以用自己的订阅密钥，让多个 Agent 在不同 Git 分支上并行工作，轻松对比、合并结果，极大提升开发效率。 ✨ 核心亮点： 1. 并行工作树：同一提示可分发给多个 Agent，每个独立分支运行，便于实验与合并 2. 移动端监控：iOS/Android App 支持远程查看进度、接收通知、发送后续指令 3. 强大交互：WebGL 终端（可无限拆分）、设计模式直接提取 UI 元素注入提示 4. 原生集成：GitHub/Linear 直接打开 PR/Issue，SSH 远程运行，CLI 脚本化工作流 5. 全面支持：兼容几乎所有主流 CLI 编码代理，支持文件拖拽、AI 差分批注、通知系统等 项目目前已获 7.9k+ stars ⭐，每日活跃更新，社区有 Discord 和微信群。它这个则让你轻松管理“Agent 舰队”，实现真正的并行协作与高效迭代。 特别适合重度使用多个 AI 编码代理的开发者、团队和技术负责人。 github.com/stablyai/orca #AIAgent #ClaudeCode #AI工具 #codex #开源项目

Anthropic 工程师70分钟workshop 教你如何让Agent自我优化式的工作十几小时甚至天而不偏移、丢失上下文、自我Loop和迭代 对于Agent解决工作、生活、编程、产品有非常大的帮助

Easycompany @Easycompany333

a month ago

x.com/i/article/2068…

11 94 401 510K 1K

推荐这篇，the annotated PyTorch training loop——一行行解释 PyTorch 训练循环里每段代码在干什么。适合想理解 AI 训练原理但不想啃论文的人。 逐行注释的 PyTorch 训练循环 HN 59 分。这篇文章把 PyTorch 的训练循环拆开，逐行注释解释每一段的作用：数据加载、loss 计算、backward pass、梯度更新、evaluation。不是教程，是"读代码"式的理解。 如果你想理解 AI 模型的训练过程到底发生了哪些步骤，这是最快的入门方式。 原文：news.ycombinator.com/item?id=486677… #PyTorch #深度学习 #代码阅读

// Self-play with a pinch of human data // Really cool paper combining human demonstrations and self-play RL. 30 minutes of human data, 2500x less than imitation learning, is enough to make self-play policies coordinate with real people. Pure self-play learns effective but alien conventions that humans cannot drive alongside. The usual fix is brittle reward engineering and domain randomization. This work instead treats a small set of human demonstrations as a regularization objective on top of a minimal safe goal-reaching reward. Why does it matter? The resulting policies coordinate with held-out human trajectories and finish training in 15 hours on a single consumer GPU. The lesson travels well past driving. A small demonstration regularizer may be the cheapest alignment knob we have for self-play. Paper: arxiv.org/abs/2606.19370 Learn to build effective AI agents in our academy: academy.dair.ai

ViTTT [CVPR 2026] [Best Paper Finalist] [Oral] Official repository of Vision Test-Time Training github.com/LeapLabTHU/ViT… Test-Time Training (TTT) has recently emerged as a promising direction for efficient sequence modeling. TTT reformulates attention operation as an online learning problem, constructing a compact inner model from key-value pairs at test time. This reformulation opens a rich and flexible design space while achieving linear computational complexity. However, crafting a powerful visual TTT design remains challenging: fundamental choices for the inner module and inner training lack comprehensive understanding and practical guidelines. To bridge this critical gap, in this paper, we present a systematic empirical study of TTT designs for visual sequence modeling. From a series of experiments and analyses, we distill six practical insights that establish design principles for effective visual TTT and illuminate paths for future improvement. These findings culminate in the Vision Test-Time Training (ViT3) model, a pure TTT architecture that achieves linear complexity and parallelizable computation. We evaluate ViT3 across diverse visual tasks, including image classification, image generation, object detection, and semantic segmentation. Results show that ViT3 consistently matches or outperforms advanced linear-complexity models (e.g., Mamba and linear attention variants) and effectively narrows the gap to highly optimized vision Transformers. We hope this study and the ViT3 baseline can facilitate future work on visual TTT models.

强烈推荐安装这个Skill，比官方的Skill-creator强大很多。 如果不会写skill，用姚老师这个skill可以写出90分的skill 这个skill来源于Anthropic 官方泄露的Claude code源码，还有全网其他模型的skill整合后的元Skill。 经过姚老师长达一个月的打磨，这是我用过的最好的Meta Skill Github：github.com/yaojingang/yao…

姚金刚 @yaojingang

a month ago

终于完成了对元Skill（创建skill的skill）的2.0升级 已推送到GitHub 这两天又进行了一些微调和测试，欢迎体验 相关资源与文档分享： 1、元skill仓库：github.com/yaojingang/yao… 2、2.0升级方案：doc.laoyao.cn/unyyhc 3、2.0与1.0的对比报告：doc.laoyao.cn/l3lkgm

keyboard_arrow_left Previous keyboard_arrow_right Next

46 154 764 327K 2K