docs: add community meeting 05 notes

meeting/2026-10-30-community-meeting-05.md

@@ -0,0 +1,100 @@
+# AReno Community Meeting #5
+
+Date: 2026-10-30
+Time: Not recorded
+Attendance: Not counted
+Format: Online community meeting
+
+## Meeting Summary
+
+The fifth AReno community meeting reviewed the main additions in version 0.10.0, including Mac/MLX QLoRA, direct GGUF/Ollama export after CUDA training, Qwen3.5 and multimodal model support, JEV decision-model training, and Harbor integration with a SWE-bench-style agent-training example. The group also discussed speculative decoding, heterogeneous hardware, Agentic RL, and Recursive Self-Improvement (RSI).
+
+The meeting agreed to continue improving JEV and RSI, with a dedicated RSI discussion planned for next week. Near-term community work will focus on technical conferences, university outreach, internal adoption examples, and a consistent cross-platform publishing workflow.
+
+## Agenda
+
+### 1. AReno 0.10.0 Progress
+
+The release delivered the following capabilities:
+
+- QLoRA support on the MLX backend for Apple Silicon, including validation of quantized training and adapter save-and-reload workflows.
+- Direct export of GGUF and Ollama packages after CUDA training, without requiring an external conversion tool.
+- Additional support for Qwen3.5, Bailing V3 MTP, and Phi-4 Multimodal. The multimodal path covers image and audio inputs.
+- JEV, which trains a model-independent decision head on different base models and supports both CUDA and Mac/MLX workflows.
+- Harbor integration connecting agents, execution environments, and training, together with a SWE-bench-style software-engineering example.
+
+### 2. Speculative Decoding and Model Capability Roadmap
+
+The group discussed MTP-based speculative decoding and the need to evaluate newer speculative-decoding techniques. The immediate priority is to stabilize the existing MTP path, followed by comparisons of training cost, inference gains, and adaptation complexity for other methods.
+
+JEV will remain a major focus for the next phase. The project needs tutorials, examples, and blog posts covering decision-head training, saving, loading, and serving, together with compatibility validation across models, quantization options, and runtime backends.
+
+### 3. Q4 Technical Roadmap
+
+The team identified several priorities for the fourth quarter:
+
+- Broaden heterogeneous hardware support, continue collaboration with domestic accelerator ecosystems, and publish a TPU usage guide.
+- Improve memory use, training efficiency, and model coverage on developer-oriented devices such as Mac and DGX Spark.
+- Strengthen Agentic RL and improve the connection between execution environments such as Harbor and AReno training.
+- Continue integrating deployment ecosystems such as GGUF and Ollama so trained models can move into local inference more easily.
+- Explore an RSI loop that updates model weights and provides reusable workflows and evaluation methods.
+
+### 4. RSI Scenario Discussion
+
+The meeting compared several possible RSI proof-of-concept scenarios:
+
+- Kernel-writing agent: repeatedly generate, test, and optimize GPU kernels, using benchmark results or performance leaderboards as feedback. This scenario has a clear evaluation signal and relatively manageable resource requirements.
+- Mathematics: build a loop around problem generation, solving, and training, though large-scale sampling can be expensive.
+- Embodied AI: explore simulation and benchmark environments such as RoboDojo and RoboCasa, which may require less data, while verifying whether the process updates model weights rather than only optimizing high-level skills or prompts.
+
+The team observed that many current RSI systems optimize programs, skills, or context without directly updating model weights. AReno could differentiate itself by connecting exploration, data generation, evaluation, and weight training in one loop. After an internal kernel-agent presentation next week, the team will hold a focused meeting to select the first scenario and define reusable framework requirements.
+
+### 5. Community Activities and Publishing
+
+Upcoming community work includes:
+
+- Supporting booth activities and community outreach around the World Internet Conference in Wuzhen.
+- Preparing an internal adoption example for publication after the website and discussion area are ready.
+- Planning an open-source campus event at Huazhong University of Science and Technology in November, with the date still to be confirmed.
+- Preparing a talk for the Global System Software Technology Conference in Beijing on November 21; the presentation materials are under review.
+- Preparing a research talk at the Hong Kong University of Science and Technology on November 30.
+- Standardizing the publishing workflow: submit each blog to GitHub as Markdown first, then adapt it for Medium, WeChat, and other channels.
+
+## Decisions
+
+- Build on the 0.10.0 release by adding documentation and reproducible examples for Mac/MLX QLoRA, GGUF/Ollama export, model support, JEV, and Harbor integration.
+- Treat JEV and RSI as priorities for the next phase and prepare tutorials, blog posts, and community presentations.
+- Hold a dedicated RSI discussion next week after the kernel-agent presentation and select the first proof-of-concept scenario.
+- Continue Q4 work on heterogeneous hardware, developer-device optimization, Agentic RL, and inference/deployment ecosystem integration.
+- Submit blog posts to GitHub in Markdown before republishing them on other channels.
+
+## Action Items
+
+- Document JEV training, checkpoint saving, loading, and serving, and validate CUDA, Mac/MLX, and quantized-training workflows.
+- Continue validating MTP speculative decoding and compare the benefits and adaptation costs of other techniques.
+- Improve the Harbor integration and SWE-bench-style agent-training example.
+- Prepare a TPU training guide and continue domestic accelerator adaptation work.
+- Improve training experience, resource use, and model coverage on Mac and DGX Spark.
+- Organize the internal kernel-agent presentation and hold the RSI-focused meeting next week.
+- Compare kernel, mathematics, and embodied-AI scenarios, then define the first RSI proof of concept and its evaluation metrics.
+- Prepare materials for the November 21 and November 30 talks and follow up on the university-event date.
+- Submit upcoming blog posts to GitHub before adapting them for Medium, WeChat, and other channels.
+
+## Open Questions
+
+- Should the first RSI proof of concept focus on kernels, mathematics, or embodied AI?
+- How should an RSI loop be designed to update model weights reliably instead of only optimizing skills or prompts?
+- What practical gains can newer speculative-decoding methods provide over the existing MTP path?
+- How should model and feature priorities be balanced across Mac, DGX Spark, TPU, and domestic accelerators?
+- How can upcoming conferences, university events, and internal outreach become reusable examples and sustained community contributions?
+
+## Next Meeting
+
+Date: TBD
+Time: TBD
+Suggested topics:
+
+- Review progress on JEV tutorials, examples, and model compatibility validation.
+- Select the first RSI proof-of-concept scenario, metrics, and reusable framework requirements.
+- Sync progress on heterogeneous hardware, Mac/DGX Spark optimization, and speculative decoding.
+- Follow up on November community events, talks, and publishing plans.

meeting/2026-10-30-community-meeting-05.zh.md

@@ -0,0 +1,100 @@
+# AReno 社区会议 #5
+
+日期：2026-10-30
+时间：未记录
+参会人数：未统计
+形式：线上社区会议
+
+## 会议概要
+
+AReno 第五次社区会议回顾了 0.10.0 版本的主要进展，包括 Mac/MLX QLoRA、CUDA 训练后直接导出 GGUF/Ollama、Qwen3.5 与多模态模型适配、JEV 决策模型训练，以及 Harbor 和 SWE-bench-style Agent 训练示例。团队同时讨论了推测解码、异构硬件、Agentic RL 和递归自我改进（Recursive Self-Improvement，RSI）的后续方向。
+
+会议决定继续完善 JEV 和 RSI 能力，并在下周安排一次 RSI 专题讨论。近期社区工作将围绕技术大会、高校分享、内部推广案例和多平台内容发布展开。
+
+## 会议议题
+
+### 1. AReno 0.10.0 版本进展
+
+本次版本主要完成了以下能力：
+
+- 在 Apple Silicon 的 MLX 后端支持 QLoRA，并验证量化训练、适配器保存与重新加载流程。
+- CUDA 训练完成后可直接导出 GGUF 和 Ollama 所需文件，无需额外转换工具。
+- 补充 Qwen3.5、Bailing V3 MTP 和 Phi-4 Multimodal 等模型能力；多模态路径覆盖图像和音频等输入。
+- 引入 JEV，支持在不同基础模型上训练模型无关的决策头，并覆盖 CUDA 与 Mac/MLX 路径。
+- 集成 Harbor，打通 Agent、执行环境和训练流程，并提供 SWE-bench-style 软件工程任务示例。
+
+### 2. 推测解码与模型能力规划
+
+会议讨论了以 MTP 为基础的推测解码能力，以及后续跟进新型推测解码方法的必要性。团队将优先保证现有 MTP 路径稳定，再评估其他方法的训练成本、推理收益和适配复杂度。
+
+JEV 将作为下一阶段重点能力继续完善。后续需要补充教程、示例和博客，说明决策头的训练、保存、加载和部署流程，并验证不同模型、量化方式和运行后端的兼容性。
+
+### 3. Q4 技术路线
+
+团队明确了第四季度的几个重点方向：
+
+- 扩展异构硬件支持，继续推进与国产加速卡生态的合作，并完善 TPU 使用教程。
+- 针对 Mac 和 DGX Spark 等个人开发者设备，继续优化显存占用、训练效率和模型覆盖。
+- 加强 Agentic RL 能力，完善 Harbor 等执行环境与 AReno 训练链路的集成。
+- 继续优化 GGUF/Ollama 等生态连接，让训练后的模型更容易进入本地推理和部署流程。
+- 探索能够更新模型权重的 RSI 训练闭环，形成可复用的流程和评估方法。
+
+### 4. RSI 场景讨论
+
+会议重点比较了几个 RSI 概念验证场景：
+
+- Kernel 编写 Agent：让 Agent 反复生成、测试和优化 GPU Kernel，以性能榜单或基准测试作为反馈。该场景评价信号清晰、资源需求相对可控。
+- 数学任务：可通过生成题目、求解和训练构建闭环，但大规模采样成本较高。
+- 具身智能：可关注 RoboDojo、RoboCasa 等仿真与评测环境，数据规模可能更小，但需要进一步确认训练是否会真正更新模型权重，而不仅是优化高层技能或提示词。
+
+团队认为，当前许多 RSI 实践主要优化程序、技能或上下文，较少直接更新模型权重。AReno 的潜在差异化方向是把探索、数据生成、评估和权重训练串成完整闭环。下周将在 Kernel Agent 内部分享后安排一次专题会议，进一步选择首个场景并梳理通用框架需求。
+
+### 5. 社区活动与内容发布
+
+近期社区工作包括：
+
+- 配合乌镇世界互联网大会相关展台和社区传播。
+- 准备内部推广案例，待网站和讨论区等基础设施就绪后推进发布。
+- 筹备 11 月华中科技大学开源校园活动，具体日期待确认。
+- 准备 11 月 21 日北京全球系统软件技术大会分享，演讲材料正在评审。
+- 准备 11 月 30 日香港科技大学科研分享。
+- 统一内容发布流程：博客先以 Markdown 形式提交到 GitHub，再根据不同平台调整后发布到 Medium、微信公众号等渠道。
+
+## 会议决策
+
+- 以 0.10.0 已完成的 Mac/MLX QLoRA、GGUF/Ollama 导出、模型适配、JEV 和 Harbor 集成为基础，继续补充文档和可复现示例。
+- 将 JEV 和 RSI 作为下一阶段重点方向，并准备教程、博客和社区分享材料。
+- 下周组织 RSI 专题讨论，在 Kernel Agent 分享后确定首个概念验证场景。
+- Q4 持续推进异构硬件、个人开发者设备优化、Agentic RL 和推理部署生态集成。
+- 博客统一先提交 GitHub Markdown，再同步到其他内容平台。
+
+## 待办事项
+
+- 补充 JEV 的训练、检查点保存、加载和部署教程，并验证 CUDA、Mac/MLX 与量化训练路径。
+- 继续验证 MTP 推测解码，并评估其他推测解码方法的收益和适配成本。
+- 完善 Harbor 集成和 SWE-bench-style Agent 训练示例。
+- 准备 TPU 训练教程，并继续推进国产加速卡适配合作。
+- 优化 Mac 与 DGX Spark 上的训练体验、资源占用和模型覆盖。
+- 组织 Kernel Agent 内部分享，并在下周召开 RSI 专题会议。
+- 对比 Kernel、数学和具身智能场景，确定 RSI 首个概念验证及评估指标。
+- 准备 11 月 21 日和 11 月 30 日的演讲材料，并跟进高校活动日期。
+- 将后续博客先提交至 GitHub，再适配 Medium、微信公众号等平台。
+
+## 开放问题
+
+- RSI 首个概念验证应优先选择 Kernel、数学还是具身智能场景？
+- 如何设计能够稳定更新模型权重，而非只优化技能或提示词的 RSI 闭环？
+- 新型推测解码方法相对现有 MTP 路径能带来多大实际收益？
+- Mac、DGX Spark、TPU 和国产加速卡应如何安排模型与功能支持优先级？
+- 如何将近期大会、高校活动和内部推广转化为可复用案例与持续社区贡献？
+
+## 下次会议
+
+日期：待定
+时间：待定
+建议议题：
+
+- 评审 JEV 教程、示例和模型兼容性验证进展。
+- 确定 RSI 首个概念验证场景、指标和通用框架需求。
+- 同步异构硬件、Mac/DGX Spark 优化和推测解码进展。
+- 跟进 11 月社区活动、演讲和内容发布安排。