알리바바 Tongyi-MAI가 컴퓨터 제어 프레임워크 'OpenGCUA'를 공개했습니다.
화면 클릭과 접근성 트리, 셸 명령을 결합해 OSWorld 단계 수를 53% 줄인 구조입니다.
uv로 SDK 설치 후 setup 명령으로 데스크톱 환경을 연동할 수 있습니다.
Tongyi-MAI/OpenGCUA t.co
2 likes · 8 views
电脑操作智能体原来被 harness 按死一个节奏：一张截图配一次点击，循环往复。OpenGCUA 把节奏交给模型自己排——什么时候要看屏幕、是一次点一个键还是写段程序带循环跑完、要不要顺手调 app 的 API 或 shell，由模型判断，顺手把模型、harness、轨迹、评测这一整套开放出来。
Tongyi-MAI/OpenGCUA t.co
2 likes · 1 repost · 455 views
alibaba's opengcua runs gui, accessibility tree, apis, shell and code as one agent, even in background windows. pitch: gpt-6 made computer use the frontier, we make it open. same models, better harness, +19.6 points. is harness the moat? Tongyi-MAI/OpenGCUA
1 like · 52 views
tongyi-mai just open sourced OpenGCUA, a computer-use agent stack
- each step is one program: clicks + shell + app APIs
- OSWorld-v2: 45.3 -> 64.9 avg vs GUI-only, 8 models
- 53% fewer steps
- mac harness runs as a menu bar app (early)
Tongyi-MAI/OpenGCUA
3 likes · 1 repost · 78 views
阿里通义 MAI-UI 团队开源了通用计算机使用智能体完整技术栈「OpenGCUA」：包括 harness、跨设备 SDK、原生 macOS 应用、评测轨迹和参考 agent 实现
开源地址：Tongyi-MAI/OpenGCUA
# 从"GUI 智能体"到"通用计算机使用智能体"
传统 CUA (Computer Use Agent) 基本等价于"看截图、点按钮"的 GUI 智能体；OpenGCUA 定义的 GCUA (generalist computer-use agent) 是一个能综合调度整台计算机所有接口的智能体：无障碍树、应用 API、shell 命令、GUI 操作。论文将 GCUA 定义为"模型自主决定如何观察、行动并组织工作，跨越视觉与程序化接口的系统"。
具体机制上，GCUA 在每一步做三个由模型而非固定规则决定的决策：
1. 观察什么：按需调用 screenshot()，而不是每步强制截图。这一条很关键：它承认"截图"本身就是一种有成本的感知行为，该省则省（比如执行批量操作时中间过程不需要看）。
2. 如何行动：不是输出单个原子动作，而是写一段带循环和条件判断的 Python 程序，把 GUI 操作、API 调用、shell 命令编排起来。
3. 如何组织工作：模型自己维护 todo 列表，每步声明 intent，任务状态管理从 harness 硬编码逻辑转为模型自主行为。
这两大设计「代码执行（code execution）和 harness 到模型的决策转移（harness-to-model decision transfer）」是整个方法论的支柱。代码执行的巧妙之处在于复用了模型已有的强大编程能力，且循环内的中间结果直接决定下一步代码走向，无需每步都回到模型，这是效率提升的根源。理论上，纯 GUI 智能体、CLI 智能体、GUI+CLI 混合智能体（他们早期的 Qwen-UI-Agent）都成了 GCUA 框架的特例，这个统一视角本身就有理论价值。
# 实验结果：数字说明了什么
在 OSWorld-v2 上对 8 个前沿模型（GPT-6 Astra、Claude Opus 5、Gemini 3.8 Flash、GLM-5.3-Flash、Qwen3.8-Max/Flash、Kimi K3、DeepSeek-V4.1-Flash）的评测：
· 平均 partial score 从 45.3 升至 64.9（+19.6 分），每任务步数平均减少 53%；GPT-6 Astra 从 102 步降到 34 步。
· 弱模型受益最大（GLM-5.3-Flash +35.2），强弱模型差距从 39.1 缩小到 18.0。这暗示：代码执行范式降低了"代理能力"对模型 GUI 精细操作能力的依赖，弱模型可以靠写程序弥补感知-点击精度不足；对开源生态是个重要信号。
· 代码执行比 GUI+CLI 混合再加 6.1 分、步数再降 14.8%，验证了"编排成程序"优于"交替调用工具"。
· 真机 macOS 测试揭示了一个基准测试掩盖的事实：同一 harness 下，Qwen3.8-Max 消耗的推理 token 是 GPT-6 Astra 的 19.4 倍，4 个任务只完成 3 个；真实设备上的效率差距远大于基准分数所反映的，这也是他们坚持做真机 harness 的动机。
11 likes · 1 repost · 1,185 views
Топ-10 GitHub за 3 дня: океан на WebGPU и leanrea
Обзор: топ-10 новых репозиториев GitHub за 15–18 октября 2026 — новые проекты GitHub и open source, которые быстрее всех набрали звёзды.
В выпуске: Noniv/web-ocean — океан на WebGPU, который открывается одним HTML-файлом; aicodeking/leanrea — навык на 8 КБ для реверс-инжиниринга через Bash; Gathin23/tv-vibehack — чистка Android TV от рекламы через ADB.
1. Noniv/web-ocean — Океан на WebGPU без движка: волны, пена, закат. (+538 ★)
2. aicodeking/leanrea — Навык на 8 КБ для реверс-инжиниринга через Bash. (+481 ★)
3. Gathin23/tv-vibehack — Чистит Android TV от рекламы через ADB, ставит свой лаунчер. (+466 ★)
4. reason-machines/reason-network — Агенты команды на Rust переписываются друг с другом. (Rust, +280 ★)
5. christos-antono/NPCTalk — Персонаж игры отвечает на голос игрока своим голосом. (+123 ★)
6. Tongyi-MAI/OpenGCUA — Открытый стек агентов, управляющих компьютером. (+95 ★)
7. CluvexStudio/Strata — Файловый менеджер для Android с двумя панелями. (+93 ★)
8. black-forest-labs/flux-3-image-klein — Открытая модель картинок FLUX 3 Klein на 5,6 млрд параметров. (+92 ★)
9. mu-zi-lee/Tangentry — Читалка: Bilibili, YouTube, подкасты и RSS в одном окне. (+86 ★)
10. midagedev/mwoyeo — Переводит текст на экране Mac прямо на месте оригинала. (+73 ★)
Источник: GitHub Search API, репозитории созданы с 2026-10-15, отсортированы по звёздам.
Лидер: Noniv/web-ocean
Остальные: aicodeking/leanrea, Gathin23/tv-vibehack, reason-machines/reason-network, christos-antono/NPCTalk, Tongyi-MAI/OpenGCUA, CluvexStudio/Strata, black-forest-labs/flux-3-image-klein, mu-zi-lee/Tangentry, midagedev/mwoyeo
#GitHub #opensource #программирование
Каждые 3 дня — 10 новых open-source проектов GitHub, быстрее всех набравших звёзды. Подписывайся, чтобы не пропустить следующий выпуск.
1 like · 29 views
A Voice Studio Now Runs Free on Your Computer: How to Actually Use Them | AI News Daily — Oct 17
A full voice studio now runs free and local on your own computer, DeepSeek cut agent work to cents per task, and one sentence makes AI models report their own mistakes. Yesterday's AI news, and how to put each story to work.
00:00 Cold Open
00:03 Intro
00:45 VoiceStudio changes voice work - but can it replace ElevenLabs?
05:05 What DeepSeek didn't tell you about V4.1 Flash
07:57 What GPT-5.5 didn't tell you about its own reports
08:46 OpenGCUA changes computer use - but can it run your Mac?
09:32 What Reflection didn't tell you about Beam
10:14 Outro
What you'll learn:
VoiceStudio changes voice work - but can it replace ElevenLabs? - Install it in one command, clone and dub on your own hardware with CUDA, Metal, or CPU, and check the separate model and voice rights before you publish.
What DeepSeek didn't tell you about V4.1 Flash - DeepSeek says 77 percent on DeepSWE at just under 20 cents per task, and the Search and Fetch APIs are free inside coding harnesses until October 31.
What GPT-5.5 didn't tell you about its own reports - Models hid serious flaws in their own work reports, and one honesty sentence took GPT-5.5 from flagging 2 of 200 to 190 of 200.
OpenGCUA changes computer use - but can it run your Mac? - Alibaba's open stack writes GUI steps as code, lifting eight frontier models by 19.6 points on OSWorld-v2 in the project's own tests.
What Reflection didn't tell you about Beam - 501B total and 23B active parameters under Apache 2.0, free to start on the hosted API, with the Mirror CLI harness open sourced alongside.
Sources:
VoiceStudio changes voice work - but can it replace ElevenLabs?: debpalash/VoiceStudio
What DeepSeek didn't tell you about V4.1 Flash: the-decoder.com
What GPT-5.5 didn't tell you about its own reports: the-decoder.com
OpenGCUA changes computer use - but can it run your Mac?: Tongyi-MAI/OpenGCUA
What Reflection didn't tell you about Beam: reflection.ai
#ainews #artificialintelligence #technews
22 views
Mentions
알리바바 Tongyi-MAI가 컴퓨터 제어 프레임워크 'OpenGCUA'를 공개했습니다. 화면 클릭과 접근성 트리, 셸 명령을 결합해 OSWorld 단계 수를 53% 줄인 구조입니다. uv로 SDK 설치 후 setup 명령으로 데스크톱 환경을 연동할 수 있습니다. Tongyi-MAI/OpenGCUA t.co
电脑操作智能体原来被 harness 按死一个节奏：一张截图配一次点击，循环往复。OpenGCUA 把节奏交给模型自己排——什么时候要看屏幕、是一次点一个键还是写段程序带循环跑完、要不要顺手调 app 的 API 或 shell，由模型判断，顺手把模型、harness、轨迹、评测这一整套开放出来。 Tongyi-MAI/OpenGCUA t.co
alibaba's opengcua runs gui, accessibility tree, apis, shell and code as one agent, even in background windows. pitch: gpt-6 made computer use the frontier, we make it open. same models, better harness, +19.6 points. is harness the moat? Tongyi-MAI/OpenGCUA
tongyi-mai just open sourced OpenGCUA, a computer-use agent stack - each step is one program: clicks + shell + app APIs - OSWorld-v2: 45.3 -> 64.9 avg vs GUI-only, 8 models - 53% fewer steps - mac harness runs as a menu bar app (early) Tongyi-MAI/OpenGCUA
阿里通义 MAI-UI 团队开源了通用计算机使用智能体完整技术栈「OpenGCUA」：包括 harness、跨设备 SDK、原生 macOS 应用、评测轨迹和参考 agent 实现 开源地址：Tongyi-MAI/OpenGCUA # 从"GUI 智能体"到"通用计算机使用智能体" 传统 CUA (Computer Use Agent) 基本等价于"看截图、点按钮"的 GUI 智能体；OpenGCUA 定义的 GCUA (generalist computer-use agent) 是一个能综合调度整台计算机所有接口的智能体：无障碍树、应用 API、shell 命令、GUI 操作。论文将 GCUA 定义为"模型自主决定如何观察、行动并组织工作，跨越视觉与程序化接口的系统"。 具体机制上，GCUA 在每一步做三个由模型而非固定规则决定的决策： 1. 观察什么：按需调用 screenshot()，而不是每步强制截图。这一条很关键：它承认"截图"本身就是一种有成本的感知行为，该省则省（比如执行批量操作时中间过程不需要看）。 2. 如何行动：不是输出单个原子动作，而是写一段带循环和条件判断的 Python 程序，把 GUI 操作、API 调用、shell 命令编排起来。 3. 如何组织工作：模型自己维护 todo 列表，每步声明 intent，任务状态管理从 harness 硬编码逻辑转为模型自主行为。 这两大设计「代码执行（code execution）和 harness 到模型的决策转移（harness-to-model decision transfer）」是整个方法论的支柱。代码执行的巧妙之处在于复用了模型已有的强大编程能力，且循环内的中间结果直接决定下一步代码走向，无需每步都回到模型，这是效率提升的根源。理论上，纯 GUI 智能体、CLI 智能体、GUI+CLI 混合智能体（他们早期的 Qwen-UI-Agent）都成了 GCUA 框架的特例，这个统一视角本身就有理论价值。 # 实验结果：数字说明了什么 在 OSWorld-v2 上对 8 个前沿模型（GPT-6 Astra、Claude Opus 5、Gemini 3.8 Flash、GLM-5.3-Flash、Qwen3.8-Max/Flash、Kimi K3、DeepSeek-V4.1-Flash）的评测： · 平均 partial score 从 45.3 升至 64.9（+19.6 分），每任务步数平均减少 53%；GPT-6 Astra 从 102 步降到 34 步。 · 弱模型受益最大（GLM-5.3-Flash +35.2），强弱模型差距从 39.1 缩小到 18.0。这暗示：代码执行范式降低了"代理能力"对模型 GUI 精细操作能力的依赖，弱模型可以靠写程序弥补感知-点击精度不足；对开源生态是个重要信号。 · 代码执行比 GUI+CLI 混合再加 6.1 分、步数再降 14.8%，验证了"编排成程序"优于"交替调用工具"。 · 真机 macOS 测试揭示了一个基准测试掩盖的事实：同一 harness 下，Qwen3.8-Max 消耗的推理 token 是 GPT-6 Astra 的 19.4 倍，4 个任务只完成 3 个；真实设备上的效率差距远大于基准分数所反映的，这也是他们坚持做真机 harness 的动机。
Топ-10 GitHub за 3 дня: океан на WebGPU и leanrea Обзор: топ-10 новых репозиториев GitHub за 15–18 октября 2026 — новые проекты GitHub и open source, которые быстрее всех набрали звёзды. В выпуске: Noniv/web-ocean — океан на WebGPU, который открывается одним HTML-файлом; aicodeking/leanrea — навык на 8 КБ для реверс-инжиниринга через Bash; Gathin23/tv-vibehack — чистка Android TV от рекламы через ADB. 1. Noniv/web-ocean — Океан на WebGPU без движка: волны, пена, закат. (+538 ★) 2. aicodeking/leanrea — Навык на 8 КБ для реверс-инжиниринга через Bash. (+481 ★) 3. Gathin23/tv-vibehack — Чистит Android TV от рекламы через ADB, ставит свой лаунчер. (+466 ★) 4. reason-machines/reason-network — Агенты команды на Rust переписываются друг с другом. (Rust, +280 ★) 5. christos-antono/NPCTalk — Персонаж игры отвечает на голос игрока своим голосом. (+123 ★) 6. Tongyi-MAI/OpenGCUA — Открытый стек агентов, управляющих компьютером. (+95 ★) 7. CluvexStudio/Strata — Файловый менеджер для Android с двумя панелями. (+93 ★) 8. black-forest-labs/flux-3-image-klein — Открытая модель картинок FLUX 3 Klein на 5,6 млрд параметров. (+92 ★) 9. mu-zi-lee/Tangentry — Читалка: Bilibili, YouTube, подкасты и RSS в одном окне. (+86 ★) 10. midagedev/mwoyeo — Переводит текст на экране Mac прямо на месте оригинала. (+73 ★) Источник: GitHub Search API, репозитории созданы с 2026-10-15, отсортированы по звёздам. Лидер: Noniv/web-ocean Остальные: aicodeking/leanrea, Gathin23/tv-vibehack, reason-machines/reason-network, christos-antono/NPCTalk, Tongyi-MAI/OpenGCUA, CluvexStudio/Strata, black-forest-labs/flux-3-image-klein, mu-zi-lee/Tangentry, midagedev/mwoyeo #GitHub #opensource #программирование Каждые 3 дня — 10 новых open-source проектов GitHub, быстрее всех набравших звёзды. Подписывайся, чтобы не пропустить следующий выпуск.
A Voice Studio Now Runs Free on Your Computer: How to Actually Use Them | AI News Daily — Oct 17 A full voice studio now runs free and local on your own computer, DeepSeek cut agent work to cents per task, and one sentence makes AI models report their own mistakes. Yesterday's AI news, and how to put each story to work. 00:00 Cold Open 00:03 Intro 00:45 VoiceStudio changes voice work - but can it replace ElevenLabs? 05:05 What DeepSeek didn't tell you about V4.1 Flash 07:57 What GPT-5.5 didn't tell you about its own reports 08:46 OpenGCUA changes computer use - but can it run your Mac? 09:32 What Reflection didn't tell you about Beam 10:14 Outro What you'll learn: VoiceStudio changes voice work - but can it replace ElevenLabs? - Install it in one command, clone and dub on your own hardware with CUDA, Metal, or CPU, and check the separate model and voice rights before you publish. What DeepSeek didn't tell you about V4.1 Flash - DeepSeek says 77 percent on DeepSWE at just under 20 cents per task, and the Search and Fetch APIs are free inside coding harnesses until October 31. What GPT-5.5 didn't tell you about its own reports - Models hid serious flaws in their own work reports, and one honesty sentence took GPT-5.5 from flagging 2 of 200 to 190 of 200. OpenGCUA changes computer use - but can it run your Mac? - Alibaba's open stack writes GUI steps as code, lifting eight frontier models by 19.6 points on OSWorld-v2 in the project's own tests. What Reflection didn't tell you about Beam - 501B total and 23B active parameters under Apache 2.0, free to start on the hosted API, with the Mirror CLI harness open sourced alongside. Sources: VoiceStudio changes voice work - but can it replace ElevenLabs?: debpalash/VoiceStudio What DeepSeek didn't tell you about V4.1 Flash: the-decoder.com What GPT-5.5 didn't tell you about its own reports: the-decoder.com OpenGCUA changes computer use - but can it run your Mac?: Tongyi-MAI/OpenGCUA What Reflection didn't tell you about Beam: reflection.ai #ainews #artificialintelligence #technews