Things I read and chose to keep.https://ursb.me/enhttps://ursb.me/en/reading/r-qu_vv-mrhe_lujnn/https://ursb.me/en/reading/r-qu_vv-mrhe_lujnn/- 《Harness Playbook》提出了 Agent Harness 作为系统软件的核心架构，强调状态权威、可信控制面和有界执行的重要性。 - 设计四大架构测试：唯一权威会话、可信控制面、有界工作以及显式兼容性，确保系统在多种场景下可靠运行。 - 采用 DOM/XML 记录会话状态，使用 convar 体系管理配置，利用 Rust 与 Python 分别实现核心与扩展，以提升安全性和可维护性。 - 通过压缩复杂性到特定层并明确所有者，实现了类似游戏引擎的成熟解决方案，提升了 Agent Harness 的可扩展性和可靠性。Fri, 04 Sep 2026 14:38:00 GMT<p>- 《Harness Playbook》提出了 Agent Harness 作为系统软件的核心架构，强调状态权威、可信控制面和有界执行的重要性。 - 设计四大架构测试：唯一权威会话、可信控制面、有界工作以及显式兼容性，确保系统在多种场景下可靠运行。 - 采用 DOM/XML 记录会话状态，使用 convar 体系管理配置，利用 Rust 与 Python 分别实现核心与扩展，以提升安全性和可维护性。 - 通过压缩复杂性到特定层并明确所有者，实现了类似游戏引擎的成熟解决方案，提升了 Agent Harness 的可扩展性和可靠性。</p><p><img src="https://r2.airingdeng.com/en/notion/harness-playbook-omp2-shaomeng-cover.png" alt="@shao__meng · Agent Harness 不是 while 循环：omp²《Harness Playbook》" /></p><p><a href="https://x.com/shao__meng/status/2095799369660576198">Read the source →</a></p>https://ursb.me/en/reading/r-k8qwpbg1psnqlwum/https://ursb.me/en/reading/r-k8qwpbg1psnqlwum/- Gear Zero: describe a game in plain language; the agent plans, builds, and runs it in the browser - No code, no install, no charge; keep chatting while it builds—ideas queue until you say yes - One-link rooms for shared chat/play; Deep Mode does 8-hour short rounds on one game - Free daily allowance; publishing earns more runs; phone OK, bigger screen better for buildingFri, 04 Sep 2026 13:50:00 GMT<p>- Gear Zero: describe a game in plain language; the agent plans, builds, and runs it in the browser - No code, no install, no charge; keep chatting while it builds—ideas queue until you say yes - One-link rooms for shared chat/play; Deep Mode does 8-hour short rounds on one game - Free daily allowance; publishing earns more runs; phone OK, bigger screen better for building</p><p><img src="https://r2.airingdeng.com/en/notion/gear-zero-cover.png" alt="Gear Zero: describe a game, see what it becomes" /></p><p><a href="https://zero.alayalab.ai/">Read the source →</a></p>https://ursb.me/en/reading/r-7yi8vqzedioj0yco/https://ursb.me/en/reading/r-7yi8vqzedioj0yco/- Most AI UIs center on user-operated chat sessions; Grok Bot redesigns for agents that persist and carry responsibility - Five primitives kept: Bots, Chats, Prompts (Skills/Routines), Tools, Artifacts — primary object is the Bot roster, not chat history - Avatar motion answers who/what/how-much; Bot computer has three access levels: Status → Preview → Takeover - Routines activate work via schedule/event/other Bots; disappearing interface (~50 Bots/account, ~6 per group)Fri, 04 Sep 2026 13:43:00 GMT<p>- Most AI UIs center on user-operated chat sessions; Grok Bot redesigns for agents that persist and carry responsibility - Five primitives kept: Bots, Chats, Prompts (Skills/Routines), Tools, Artifacts — primary object is the Bot roster, not chat history - Avatar motion answers who/what/how-much; Bot computer has three access levels: Status → Preview → Takeover - Routines activate work via schedule/event/other Bots; disappearing interface (~50 Bots/account, ~6 per group)</p><p><img src="https://r2.airingdeng.com/en/notion/designing-grok-bot-cover.png" alt="Designing Grok Bot for a world of persistent agents" /></p><p><a href="https://x.ai/news/designing-grok-bot">Read the source →</a></p>https://ursb.me/en/reading/r-qg3ofq9dqssmnj7_/https://ursb.me/en/reading/r-qg3ofq9dqssmnj7_/- GPT-6 Astra 发布，具备强大的计算机操作和软件工程能力，可直接在实际软件环境中完成任务。 - 在多项基准测试中显著超越前代模型，表现出色的代码、数学、科学和安全能力。 - Astra 已达到高网络安全能力阈值，能够发现并利用未知的 zero‑day 漏洞，但普通用户版本会限制高级安全请求。 - 商业化将先向企业用户开放，随后面向 ChatGPT Plus、Pro、Business 与 Enterprise，采用基于 token 的定价模式。Fri, 04 Sep 2026 02:04:00 GMT<p>- GPT-6 Astra 发布，具备强大的计算机操作和软件工程能力，可直接在实际软件环境中完成任务。 - 在多项基准测试中显著超越前代模型，表现出色的代码、数学、科学和安全能力。 - Astra 已达到高网络安全能力阈值，能够发现并利用未知的 zero‑day 漏洞，但普通用户版本会限制高级安全请求。 - 商业化将先向企业用户开放，随后面向 ChatGPT Plus、Pro、Business 与 Enterprise，采用基于 token 的定价模式。</p><p><img src="https://r2.airingdeng.com/notion/gpt6-appsso-agi-division-cover.png" alt="GPT-6 Astra 发布：人类进入 AGI 大分工时代" /></p><p><a href="https://mp.weixin.qq.com/s/CF1cVS05rcMpQXQrR8SF3w">Read the source →</a></p>https://ursb.me/en/reading/r-kticvij6qrltkeki/https://ursb.me/en/reading/r-kticvij6qrltkeki/- Huabu 将 AI 交互从线性聊天转变为二维可视化空间，支持信息并排、移动、缩放和连线。 - 工作流程分为收集、组织、执行三个循环阶段，Agent 可在每阶段生成并回馈新素材。 - 通过节点和连线自动整理资料，实现信息的语义关联和多分支思考的可视化呈现。 - 用户可在空间中直接修改内容并指令 Agent 执行任务，提升人机协同的直觉性和效率。Fri, 04 Sep 2026 01:28:00 GMT<p>- Huabu 将 AI 交互从线性聊天转变为二维可视化空间，支持信息并排、移动、缩放和连线。 - 工作流程分为收集、组织、执行三个循环阶段，Agent 可在每阶段生成并回馈新素材。 - 通过节点和连线自动整理资料，实现信息的语义关联和多分支思考的可视化呈现。 - 用户可在空间中直接修改内容并指令 Agent 执行任务，提升人机协同的直觉性和效率。</p><p><img src="https://r2.airingdeng.com/notion/huabu-agent-new-desktop-cover.png" alt="Huabu：从线性对话到二维协作的 Agent 新桌面" /></p><p><a href="https://mp.weixin.qq.com/s?__biz=MzAwMTA3MzM4Nw==&amp;mid=2649508658&amp;idx=1&amp;sn=a876ba50ba84dbe479d54ef5fda75dde">Read the source →</a></p>https://ursb.me/en/reading/r-0ux3kiysdqjy2vr8/https://ursb.me/en/reading/r-0ux3kiysdqjy2vr8/- GPT-6 Astra 被宣称为全球最强的模型，标志着人类进入 AGI 时代。 - 在多项基准测试中实现领先，FrontierMath Tier 4 97.6%，ARC‑AGI‑3 99.9%，ExploitBench 满分 100%。 - 具备强大的计算机使用能力，可在真实软件环境中独立完成复杂工作流，且训练使用了超过 10 万块 GPU。 - 采用递归式自我改进（RSI）策略，安全性高，价格为输入 10 美元/百万 Token，输出 50 美元/百万 Token。Fri, 04 Sep 2026 01:10:00 GMT<p>- GPT-6 Astra 被宣称为全球最强的模型，标志着人类进入 AGI 时代。 - 在多项基准测试中实现领先，FrontierMath Tier 4 97.6%，ARC‑AGI‑3 99.9%，ExploitBench 满分 100%。 - 具备强大的计算机使用能力，可在真实软件环境中独立完成复杂工作流，且训练使用了超过 10 万块 GPU。 - 采用递归式自我改进（RSI）策略，安全性高，价格为输入 10 美元/百万 Token，输出 50 美元/百万 Token。</p><p><img src="https://r2.airingdeng.com/notion/gpt6-astra-xinzhiyuan-cover.png" alt="GPT-6 Astra：全球最强？人类进入 AGI 时代" /></p><p><a href="https://mp.weixin.qq.com/s?__biz=MzI3MTA0MTk1MA==&amp;mid=2652722736&amp;idx=1&amp;sn=949132447bea8e3c50e77d94f502dbc7">Read the source →</a></p>https://ursb.me/en/reading/r--ixvcbfk7khjxuaw/https://ursb.me/en/reading/r--ixvcbfk7khjxuaw/- Harness-of-Harness organizes coding frameworks like Codex, OpenCode, and Pi into a cyclic iterative development flow; each round includes planning, implementation, and independent QA acceptance. - Across 70 development rounds, 81 issues were recorded and 65 closed; 17 issues reappeared due to later changes, showing regression. - Experiments show HoH improves ~52.25% on average vs. standalone frameworks across three benchmarks; on FrontierSWE, Codex + GPT-5.5 rose from 22% to 72.67%. - The key for long-running coding agents is preserving verified progress and preventing new features from reintroducing old bugs—engineering continuity.Thu, 03 Sep 2026 14:24:00 GMT<p>- Harness-of-Harness organizes coding frameworks like Codex, OpenCode, and Pi into a cyclic iterative development flow; each round includes planning, implementation, and independent QA acceptance. - Across 70 development rounds, 81 issues were recorded and 65 closed; 17 issues reappeared due to later changes, showing regression. - Experiments show HoH improves ~52.25% on average vs. standalone frameworks across three benchmarks; on FrontierSWE, Codex + GPT-5.5 rose from 22% to 72.67%. - The key for long-running coding agents is preserving verified progress and preventing new features from reintroducing old bugs—engineering continuity.</p><p><img src="https://r2.airingdeng.com/en/notion/harness-of-harness-cover.png" alt="@Xudong07452910 · After 70 Coding-Agent Rounds: Harness-of-Harness" /></p><p><a href="https://x.com/Xudong07452910/status/2095351325013819716">Read the source →</a></p>https://ursb.me/en/reading/r-hzvop0_mpr-rjru7/https://ursb.me/en/reading/r-hzvop0_mpr-rjru7/- EvoMap argues next-gen Agents need self-organizing swarms for higher accuracy, not a single model or master-slave structure. - Self-organizing swarms can break complex tasks into independent subtasks, auto-match tools, and share experience for continuous memory and connection. - AutoResearch applies this swarm to a full research loop (explore–experiment–evaluate–improve) and deposits validated methods as reusable experience assets. - AI4AI's core is "memory" and "connection," building self-evolving intelligent collaboration infrastructure via an experience-asset layer, A2A swarm network, and vertical apps.Thu, 03 Sep 2026 11:16:00 GMT<p>- EvoMap argues next-gen Agents need self-organizing swarms for higher accuracy, not a single model or master-slave structure. - Self-organizing swarms can break complex tasks into independent subtasks, auto-match tools, and share experience for continuous memory and connection. - AutoResearch applies this swarm to a full research loop (explore–experiment–evaluate–improve) and deposits validated methods as reusable experience assets. - AI4AI&#39;s core is &quot;memory&quot; and &quot;connection,&quot; building self-evolving intelligent collaboration infrastructure via an experience-asset layer, A2A swarm network, and vertical apps.</p><p><img src="https://r2.airingdeng.com/en/notion/agent-second-half-evo-cover.png" alt="Agent Second Half: What Should a Truly Self-Evolving Agent Look Like?" /></p><p><a href="https://mp.weixin.qq.com/s?__biz=MzYyMTY1NDA0Nw==&amp;mid=2247520738&amp;idx=1&amp;sn=027022c37a1d8e1ee8eaddbe0afeeb81">Read the source →</a></p>https://ursb.me/en/reading/r-ckpyvvyokzb4i2re/https://ursb.me/en/reading/r-ckpyvvyokzb4i2re/- Obscura is an open-source headless browser built for Agents in Rust, ~30 MB memory, no Chromium or Node.js—runs V8 directly and is CDP-compatible. - Unlike traditional desktop browsers, Agents only need basics: run JavaScript, read DOM, send network requests, click buttons, fill forms, and take screenshots. - Running hundreds or thousands of browser instances on one server, the 30 MB footprint significantly cuts infrastructure cost. - Cloudflare's Agent browser Kitesurf project's initial prototype was ported from Obscura.Thu, 03 Sep 2026 02:17:00 GMT<p>- Obscura is an open-source headless browser built for Agents in Rust, ~30 MB memory, no Chromium or Node.js—runs V8 directly and is CDP-compatible. - Unlike traditional desktop browsers, Agents only need basics: run JavaScript, read DOM, send network requests, click buttons, fill forms, and take screenshots. - Running hundreds or thousands of browser instances on one server, the 30 MB footprint significantly cuts infrastructure cost. - Cloudflare&#39;s Agent browser Kitesurf project&#39;s initial prototype was ported from Obscura.</p><p><img src="https://r2.airingdeng.com/en/notion/obscura-headless-cover.png" alt="Obscura · Open-Source Headless Browser for Agents" /></p><p><a href="https://x.com/MaxForAI/status/2095168688643293458">Read the source →</a></p>https://ursb.me/en/reading/r-0l1ffdeaw4g24rgq/https://ursb.me/en/reading/r-0l1ffdeaw4g24rgq/- This article systematically introduces Transformer core concepts, including attention, QKV, residual connections, and other key techniques. - Self-Attention enables global information exchange among tokens, solving RNN/LSTM long-range dependency and sequential computation issues. - Multi-head attention, positional encoding, feed-forward networks, and residual connections together form Transformer's basic unit, enabling efficient parallel training while preserving sequence order. - The encoder-decoder structure plus masked self-attention and cross-attention let the model understand input text and avoid leaking future information when generating output.Wed, 02 Sep 2026 14:27:00 GMT<p>- This article systematically introduces Transformer core concepts, including attention, QKV, residual connections, and other key techniques. - Self-Attention enables global information exchange among tokens, solving RNN/LSTM long-range dependency and sequential computation issues. - Multi-head attention, positional encoding, feed-forward networks, and residual connections together form Transformer&#39;s basic unit, enabling efficient parallel training while preserving sequence order. - The encoder-decoder structure plus masked self-attention and cross-attention let the model understand input text and avoid leaking future information when generating output.</p><p><img src="https://r2.airingdeng.com/en/notion/transformer-one-go-cover.png" alt="@xuanyuanzhifeng · Understand Transformer in One Go" /></p><p><a href="https://x.com/xuanyuanzhifeng/status/2095044737531306357">Read the source →</a></p>https://ursb.me/en/reading/r-oqhcheiuu3iwyb6i/https://ursb.me/en/reading/r-oqhcheiuu3iwyb6i/- Paid video recording tools: Shotbase, Matte, Screenflare - Open-source video recording tool: Screendrop (GitHub) - Market traits: capture, edit, and share in one, with intense competition - Many new tools have appeared recently, intensifying industry competitionWed, 02 Sep 2026 14:27:00 GMT<p>- Paid video recording tools: Shotbase, Matte, Screenflare - Open-source video recording tool: Screendrop (GitHub) - Market traits: capture, edit, and share in one, with intense competition - Many new tools have appeared recently, intensifying industry competition</p><p><img src="https://r2.airingdeng.com/en/notion/screen-recording-tools-cover.png" alt="@dingyi · The Screen Recording Tool Landscape" /></p><p><a href="https://x.com/dingyi/status/2093670500371489045">Read the source →</a></p>https://ursb.me/en/reading/r-yy-5xoanlgv68clq/https://ursb.me/en/reading/r-yy-5xoanlgv68clq/- Requirements alignment is fundamental to avoiding AI code drift; clear specs eliminate hidden assumptions. - Path exploration uses high-level architecture planning, tech probes, and prototype validation to select a feasible implementation path. - Tracking progress requires continuous change logs, traceable docs, and E2E acceptance to prevent context drift and false-green tests. - Tightly linking goal alignment, path exploration, and execution tracking builds a closed-loop engineering system for reliable AI coding delivery.Wed, 02 Sep 2026 14:27:00 GMT<p>- Requirements alignment is fundamental to avoiding AI code drift; clear specs eliminate hidden assumptions. - Path exploration uses high-level architecture planning, tech probes, and prototype validation to select a feasible implementation path. - Tracking progress requires continuous change logs, traceable docs, and E2E acceptance to prevent context drift and false-green tests. - Tightly linking goal alignment, path exploration, and execution tracking builds a closed-loop engineering system for reliable AI coding delivery.</p><p><img src="https://r2.airingdeng.com/en/notion/ai-code-review-misaligned-cover.png" alt="@ErwinWu000 · AI Code Review: Requirements Misaligned" /></p><p><a href="https://x.com/ErwinWu000/status/2094991204375240733">Read the source →</a></p>https://ursb.me/en/reading/r-ugrnrw8pe-d9ayzg/https://ursb.me/en/reading/r-ugrnrw8pe-d9ayzg/- During autoregressive generation, only the current Q is dotted with all historical K, then used to weight all V, so only K and V need to be cached. - Under causal masking, historical K/V do not change with new tokens and can be reused directly, avoiding recomputation. - The Q produced at each step is used only in that step and is not needed next, so Q is not stored. - Each step only recomputes Q, K, and V for the new position and writes the new K/V into the cache.Wed, 02 Sep 2026 14:27:00 GMT<p>- During autoregressive generation, only the current Q is dotted with all historical K, then used to weight all V, so only K and V need to be cached. - Under causal masking, historical K/V do not change with new tokens and can be reused directly, avoiding recomputation. - The Q produced at each step is used only in that step and is not needed next, so Q is not stored. - Each step only recomputes Q, K, and V for the new position and writes the new K/V into the cache.</p><p><img src="https://r2.airingdeng.com/en/notion/kv-cache-zen-with-ai-cover.png" alt="@Zen_with_AI · Why KV Cache Stores Only K/V, Not Q" /></p><p><a href="https://x.com/Zen_with_AI/status/2094953118853382334">Read the source →</a></p>https://ursb.me/en/reading/r-ffob-z_ohtmed6tx/https://ursb.me/en/reading/r-ffob-z_ohtmed6tx/- LLM Wiki is a cross-platform desktop app that automatically turns documents into a structured, continuously updatable personal knowledge base. - Uses two-step chain-of-thought ingestion: first analyze document structure, then generate Wiki pages, with incremental caching and a persistent ingestion queue. - Built-in four-signal knowledge graph and Louvain community detection provide relatedness models, visual graphs, and knowledge clustering analysis. - Supports multi-format docs, multimodal image ingestion, deep research, Chrome web clipping, plus local HTTP API and MCP server for flexible retrieval and AI Agent integration.Wed, 02 Sep 2026 12:31:00 GMT<p>- LLM Wiki is a cross-platform desktop app that automatically turns documents into a structured, continuously updatable personal knowledge base. - Uses two-step chain-of-thought ingestion: first analyze document structure, then generate Wiki pages, with incremental caching and a persistent ingestion queue. - Built-in four-signal knowledge graph and Louvain community detection provide relatedness models, visual graphs, and knowledge clustering analysis. - Supports multi-format docs, multimodal image ingestion, deep research, Chrome web clipping, plus local HTTP API and MCP server for flexible retrieval and AI Agent integration.</p><p><img src="https://r2.airingdeng.com/en/notion/llm-wiki-cover.png" alt="LLM Wiki · A Self-Building Personal Knowledge Base" /></p><p><a href="https://github.com/nashsu/llm_wiki/blob/main/README_CN.md">Read the source →</a></p>https://ursb.me/en/reading/r-wsvf60hps5qpqmdn/https://ursb.me/en/reading/r-wsvf60hps5qpqmdn/- This article outlines Claude model prompting best practices, including model-specific guidance and general techniques such as clear instructions, examples, XML tags, tool use, and thinking approaches. - Recommended to read model-specific guides first (e.g. Fable, Mythos, Opus, Sonnet), then apply general techniques, noting prefill deprecation and effort/thinking changes during migration. - Emphasizes clear role setting, XML structure, examples, and appropriate tool calls to improve output quality and consistency. - Provides detailed configuration and prompt examples on thinking, agent systems, parallel tool calls, and migrating to new models.Wed, 02 Sep 2026 11:19:00 GMT<p>- This article outlines Claude model prompting best practices, including model-specific guidance and general techniques such as clear instructions, examples, XML tags, tool use, and thinking approaches. - Recommended to read model-specific guides first (e.g. Fable, Mythos, Opus, Sonnet), then apply general techniques, noting prefill deprecation and effort/thinking changes during migration. - Emphasizes clear role setting, XML structure, examples, and appropriate tool calls to improve output quality and consistency. - Provides detailed configuration and prompt examples on thinking, agent systems, parallel tool calls, and migrating to new models.</p><p><img src="https://r2.airingdeng.com/en/notion/claude-prompting-best-practices-cover.png" alt="Claude Prompting Best Practices (Official Docs)" /></p><p><a href="https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices">Read the source →</a></p>https://ursb.me/en/reading/r-e8jtbhbrnsdepjzd/https://ursb.me/en/reading/r-e8jtbhbrnsdepjzd/- FDEs (Frontier Deployment Engineers) in enterprises need both tech and business skills, often driving AI transformation amid scarce data, permissions, and clear goals. - Projects often face internal resistance: information asymmetry across departments, complex processes, and unclear requirements make efficiency gains hard to realize. - Actual outcomes often compress work that took ten minutes into ten seconds, yet may not bring real value or profit growth. - FDE's core value lies in observing on-site, understanding how employees actually work, and helping enterprises self-reflect and improve through continuous iteration.Wed, 02 Sep 2026 08:07:00 GMT<p>- FDEs (Frontier Deployment Engineers) in enterprises need both tech and business skills, often driving AI transformation amid scarce data, permissions, and clear goals. - Projects often face internal resistance: information asymmetry across departments, complex processes, and unclear requirements make efficiency gains hard to realize. - Actual outcomes often compress work that took ten minutes into ten seconds, yet may not bring real value or profit growth. - FDE&#39;s core value lies in observing on-site, understanding how employees actually work, and helping enterprises self-reflect and improve through continuous iteration.</p><p><img src="https://r2.airingdeng.com/en/notion/my-fde-cover.png" alt="@wuzhutisushuo · My FDE" /></p><p><a href="https://x.com/wuzhutisushuo/status/2094341073623777777">Read the source →</a></p>https://ursb.me/en/reading/r-lw5y7c77vnmjfwuc/https://ursb.me/en/reading/r-lw5y7c77vnmjfwuc/- Noty is a native macOS sticky-notes app that docks to the screen edge and expands via pointer swipe. - Supports global shortcuts (e.g. ⌥⌘N to create, ⇧⌘Space for quick capture) and noty:// URL Scheme automation. - Notes are stored locally in SQLite with AES-GCM encryption, with no accounts, servers, or telemetry. - Offers import/export of Markdown, plain text, and .stickies archives among other formats.Wed, 02 Sep 2026 05:16:00 GMT<p>- Noty is a native macOS sticky-notes app that docks to the screen edge and expands via pointer swipe. - Supports global shortcuts (e.g. ⌥⌘N to create, ⇧⌘Space for quick capture) and noty:// URL Scheme automation. - Notes are stored locally in SQLite with AES-GCM encryption, with no accounts, servers, or telemetry. - Offers import/export of Markdown, plain text, and .stickies archives among other formats.</p><p><img src="https://r2.airingdeng.com/en/notion/noty-macos-sticky-cover.png" alt="noty · Native macOS Sticky Notes on the Screen Edge" /></p><p><a href="https://github.com/aimen08/noty">Read the source →</a></p>https://ursb.me/en/reading/r-6dhcy0pp4uk5u_8j/https://ursb.me/en/reading/r-6dhcy0pp4uk5u_8j/- Grok Bot is used to automate various tasks such as creating bots, organizing content, and generating shopping lists. - The site updates every 6 hours and has catalogued over 1,610 public posts and templates. - Common templates include Spark (getting started), Chef (recipes and shopping lists), and Bounty Hunter (recovering unclaimed funds). - User cases showcase Grok Bot's practical applications in code management, marketing, automated testing, and personal assistance.Wed, 02 Sep 2026 05:11:00 GMT<p>- Grok Bot is used to automate various tasks such as creating bots, organizing content, and generating shopping lists. - The site updates every 6 hours and has catalogued over 1,610 public posts and templates. - Common templates include Spark (getting started), Chef (recipes and shopping lists), and Bounty Hunter (recovering unclaimed funds). - User cases showcase Grok Bot&#39;s practical applications in code management, marketing, automated testing, and personal assistance.</p><p><img src="https://r2.airingdeng.com/en/notion/usegrokbot-how-people-use-cover.png" alt="How people use Grok Bot · UseGrokBot" /></p><p><a href="https://usegrokbot.com/en">Read the source →</a></p>https://ursb.me/en/reading/r-in93bbij-0eopb_p/https://ursb.me/en/reading/r-in93bbij-0eopb_p/- This is software that lets Claude interact with users "pen-to-pen" on a writing tablet for bidirectional handwritten learning. - Supports importing PDFs/ebooks; users highlight or annotate pages, and Claude can auto-generate side notes, discuss, and create quizzes. - Through immersive, slow-paced interaction, it strengthens memory and understanding to improve learning outcomes. - The demo video is 105 seconds and shows the feature in actual use.Wed, 02 Sep 2026 02:07:00 GMT<p>- This is software that lets Claude interact with users &quot;pen-to-pen&quot; on a writing tablet for bidirectional handwritten learning. - Supports importing PDFs/ebooks; users highlight or annotate pages, and Claude can auto-generate side notes, discuss, and create quizzes. - Through immersive, slow-paced interaction, it strengthens memory and understanding to improve learning outcomes. - The demo video is 105 seconds and shows the feature in actual use.</p><p><img src="https://r2.airingdeng.com/en/notion/claude-handwriting-board-cover.png" alt="@xiaohu · Claude Handwriting Board: Pen-to-Pen Learning" /></p><p><a href="https://x.com/xiaohu/status/2094700080716734631">Read the source →</a></p>https://ursb.me/en/reading/r-xmpcvejgv7rgvrdl/https://ursb.me/en/reading/r-xmpcvejgv7rgvrdl/- Long-running Agents should keep only "current state": each step uses only a fixed Skill, the current structured state, and the latest observation, then discards intermediate reasoning. - SKILLSTATE cuts token use from about 1.06M to 65K (~16× reduction) and raises accuracy from 0.91 to 0.94. - On InterCode CTF and τ-Bench benchmarks, SKILLSTATE both improves success rate and significantly reduces token use. - The takeaway: treat long-term context as continuously updated working state—forget completed steps and keep only what remains useful.Wed, 02 Sep 2026 02:07:00 GMT<p>- Long-running Agents should keep only &quot;current state&quot;: each step uses only a fixed Skill, the current structured state, and the latest observation, then discards intermediate reasoning. - SKILLSTATE cuts token use from about 1.06M to 65K (~16× reduction) and raises accuracy from 0.91 to 0.94. - On InterCode CTF and τ-Bench benchmarks, SKILLSTATE both improves success rate and significantly reduces token use. - The takeaway: treat long-term context as continuously updated working state—forget completed steps and keep only what remains useful.</p><p><img src="https://r2.airingdeng.com/en/notion/skillstate-cover.png" alt="@Xudong07452910 · SKILLSTATE: Keep Only Current State" /></p><p><a href="https://x.com/Xudong07452910/status/2094628562615607598">Read the source →</a></p>https://ursb.me/en/reading/r-egntejrmfz8yrmov/https://ursb.me/en/reading/r-egntejrmfz8yrmov/- Recently, screen recording, AI wrappers, and terminal tools have emerged as a new "trio" for indie developers. - Recommended resources include VGPU (WebGPU Shader toolchain) and Textures (online image filter experimentation tool). - Key products are Shotbase, Screenflare, and Prequel, offering web capture, auto-zoom screen recording, and cinematic screen recording respectively. - Upcoming projects include open-source cross-platform font editor Shift and real-time 3D motion editor mo1.Tue, 01 Sep 2026 14:11:00 GMT<p>- Recently, screen recording, AI wrappers, and terminal tools have emerged as a new &quot;trio&quot; for indie developers. - Recommended resources include VGPU (WebGPU Shader toolchain) and Textures (online image filter experimentation tool). - Key products are Shotbase, Screenflare, and Prequel, offering web capture, auto-zoom screen recording, and cinematic screen recording respectively. - Upcoming projects include open-source cross-platform font editor Shift and real-time 3D motion editor mo1.</p><p><img src="https://r2.airingdeng.com/en/notion/indie-hackers-new-trinity-cover.png" alt="#362 Indie Hackers&#39; New Trinity" /></p>https://ursb.me/en/reading/r-crzv2nu1uoppngxu/https://ursb.me/en/reading/r-crzv2nu1uoppngxu/- AI apps face three major challenges: rapid model iteration swallowing features, growth that is hard to monetize, and user entry points controlled upstream. - High growth often comes with high costs; many AI companies lose money after scaling users, with gross margins generally low or even negative. - To counter model cannibalization, founders can shift to vertical niches with delivery services, train their own models, or boost competitiveness via open source. - Only by delivering profitable outcomes downstream or building models upstream do AI application companies have higher survival odds.Tue, 01 Sep 2026 11:50:00 GMT<p>- AI apps face three major challenges: rapid model iteration swallowing features, growth that is hard to monetize, and user entry points controlled upstream. - High growth often comes with high costs; many AI companies lose money after scaling users, with gross margins generally low or even negative. - To counter model cannibalization, founders can shift to vertical niches with delivery services, train their own models, or boost competitiveness via open source. - Only by delivering profitable outcomes downstream or building models upstream do AI application companies have higher survival odds.</p><p><img src="https://r2.airingdeng.com/en/notion/ai-apps-not-good-business-cover.png" alt="Users and Revenue Aren&#39;t Enough: AI Apps Still Aren&#39;t a Good Business" /></p><p><a href="https://mp.weixin.qq.com/s?__biz=MzU3Mjk1OTQ0Ng==&amp;mid=2247538479&amp;idx=1&amp;sn=2857e8f2282601d9884cebbc5cdd2d28">Read the source →</a></p>https://ursb.me/en/reading/r-g2kftponbjs_bvhl/https://ursb.me/en/reading/r-g2kftponbjs_bvhl/- Research found that retaining complete attempt history and feedback signals can raise average model accuracy to 71.0%. - Iterative solving without feedback achieves 66.8% accuracy; with external feedback it can reach 79.3%. - Although context is longer, convergence is faster, overall token consumption drops by about 19%, and API costs decrease. - For Agent context management, this reminds us not to rush to compress history—sometimes keeping more information is more effective.Tue, 01 Sep 2026 04:51:00 GMT<p>- Research found that retaining complete attempt history and feedback signals can raise average model accuracy to 71.0%. - Iterative solving without feedback achieves 66.8% accuracy; with external feedback it can reach 79.3%. - Although context is longer, convergence is faster, overall token consumption drops by about 19%, and API costs decrease. - For Agent context management, this reminds us not to rush to compress history—sometimes keeping more information is more effective.</p><p><img src="https://r2.airingdeng.com/en/notion/agi-hunt-chain-of-experience-cover.png" alt="AGI Hunt · Chain-of-Experience: Don&#39;t Rush to Compress Attempt History" /></p><p><a href="https://mp.weixin.qq.com/s?__biz=MzA4NzgzMjA4MQ==&amp;mid=2453487263&amp;idx=1&amp;sn=d87c050d0a8ecb6adb264f32020f12e3">Read the source →</a></p>https://ursb.me/en/reading/r-b8jyhxnvinxf1-z2/https://ursb.me/en/reading/r-b8jyhxnvinxf1-z2/- Infinite Slop is an endlessly looping AI-generated video livestream site where viewers type prompts in chat and AI generates short clips in real time for continuous playback. - It uses the Fal platform's H3 Max model; a 15-second video takes about 9 seconds to generate, so generation outpaces playback. - The system maintains coherence via audience voting, prompt expansion, and parallel generation, but quality varies and often yields "slop"-like low-quality content. - 24/7 continuous operation costs about $207,000 per month, sparking dual discussions on AI content proliferation and innovation.Tue, 01 Sep 2026 04:51:00 GMT<p>- Infinite Slop is an endlessly looping AI-generated video livestream site where viewers type prompts in chat and AI generates short clips in real time for continuous playback. - It uses the Fal platform&#39;s H3 Max model; a 15-second video takes about 9 seconds to generate, so generation outpaces playback. - The system maintains coherence via audience voting, prompt expansion, and parallel generation, but quality varies and often yields &quot;slop&quot;-like low-quality content. - 24/7 continuous operation costs about $207,000 per month, sparking dual discussions on AI content proliferation and innovation.</p><p><img src="https://r2.airingdeng.com/en/notion/infinite-slop-appsso-cover.png" alt="5 Million Views: AI Builds an &quot;Infinite Slop&quot; Machine" /></p><p><a href="https://mp.weixin.qq.com/s?__biz=MjM5MjAyNDUyMA==&amp;mid=2651104338&amp;idx=1&amp;sn=0b1d52ed4af7e6b1d42f18127e837aa8">Read the source →</a></p>