https://magic76.github.io/crew-app/#/briefingPublic AI engineering and Crew project researchzh-TWhttps://magic76.github.io/crew-app/#/briefing/2026-10-13/4927ced7827a1ba64927ced7827a1ba6Sun, 11 Oct 2026 00:00:00 GMTLong-running conversational AI agents and assistive development models inevitably succumb to contextual degradation, specification drift, and sycophantic decay. Standard Retrieval-Augmented Generation (RAG) paradigms treat dialogue turns, human identity, and technical parameters as an undifferentiated, flat vector repohttps://magic76.github.io/crew-app/#/briefing/2026-10-13/d88beb73cfdb3575d88beb73cfdb3575Sun, 11 Oct 2026 00:00:00 GMTIn machine learning, the feature space is a key framework that converts data samples into vectors for analysis and modelling. Consequently, the results obtained from machine learning algorithms depend heavily on the ability to select and generate important features. Feature generation as a central technique converts rahttps://magic76.github.io/crew-app/#/briefing/2026-10-12/38c1c1ffd5697df138c1c1ffd5697df1Fri, 09 Oct 2026 00:00:00 GMTIntroduction: Multimodal systems increasingly run on public cloud infrastructure because it scales on demand. For workloads that are predictable and not time-critical, that convenience carries recurring fees and limited visibility into the platform, and the case for repatriating such workloads to local hardware is undehttps://magic76.github.io/crew-app/#/briefing/2026-10-12/1eaec5c1369c5d111eaec5c1369c5d11Fri, 25 Sep 2026 14:12:31 GMTHacker News 社群近期討論：Entropy-Based Guided Collaboration in Heterogeneous LLM Multi-Agent Systems。社群熱度不等於技術已經驗證。https://magic76.github.io/crew-app/#/briefing/2026-10-11/research-acon-icml-2026research-acon-icml-2026Sat, 11 Jul 2026 00:00:00 GMTICML 2026 的 ACON 不是單純摘要上下文，而是根據 Agent 執行失敗，反向修正壓縮指引。論文在指定 benchmark 上報告峰值 token 使用量下降 26–54%；這是我們最值得拿來對照 Crew Pocket AutoCompact 的方向。https://magic76.github.io/crew-app/#/briefing/2026-10-11/research-simplemem-icml-2026research-simplemem-icml-2026Sat, 11 Jul 2026 00:00:00 GMTSimpleMem 把長期記憶拆成語意結構化、線上語意整合、依意圖擷取三階段。ICML 2026 論文報告在 LoCoMo benchmark 平均 F1 改善 26.4%，並有推論 token 降低的結果；值得借鏡的是『記憶品質 × 召回成本』的衡量方式。https://magic76.github.io/crew-app/#/briefing/2026-10-11/gh-261cd26dabb9gh-261cd26dabb9Sun, 11 Oct 2026 09:05:35 GMTCrew Pocket 今日已合併中斷後繼續發送的可靠佇列，包含每個 Role 自己的 delivery coordinator、ack 與 retry。下一步不是再添加 toast，而是測「中斷、斷線、重送」極端情境下是否確實不掉訊息、不重送兩次。https://magic76.github.io/crew-app/#/briefing/2026-10-11/research-coding-agent-evals-20261005research-coding-agent-evals-20261005Mon, 05 Oct 2026 00:00:00 GMT10 月 5 日發布的 Coding Agent benchmark 深度分析指出，模型分數往往混用了不同 harness、測試次數或版本。對 Crew Deck / Pocket，比較 Claude、Codex 的最有效方法，是用相同的真實任務、驗收器和成本紀錄測試。https://magic76.github.io/crew-app/#/briefing/2026-10-11/gh-964a7ea93df1gh-964a7ea93df1Sun, 11 Oct 2026 06:10:02 GMTCrew Teacher 今天合併 Live reconnect 修正，將未被確認的學生語音先保留在有界緩衝，重連後可繼續或一鍵重送。這直接對應『斷了線又要把話重說一次』的困擾。https://magic76.github.io/crew-app/#/briefing/2026-10-11/research-context-engineering-20261007research-context-engineering-20261007Wed, 07 Oct 2026 00:00:00 GMT10 月 7 日的新研究文章把 Context Engineering 定位為系統級編排問題：比 prompt 怎麼寫更重要的是資料選取、來源可靠性、工具邊界與生命週期。這跟 Crew Pocket 已經在做的 Context Health 十分貼近。https://magic76.github.io/crew-app/#/briefing/2026-10-11/research-gemini-live-20260915research-gemini-live-20260915Tue, 15 Sep 2026 00:00:00 GMTGoogle 9 月 15 日發布 Gemini 3.8 Live 與 3.8 Live Extended Thinking，官方主打原生語音對話、多語支援、視覺內容和非同步工具呼叫。對 Crew Teacher / Story，目前最應關注的是網路重連、語音完整性與畫面同步，而非單純模型規格。https://magic76.github.io/crew-app/#/briefing/2026-10-11/gh-5a7a4b826483gh-5a7a4b826483Sun, 11 Oct 2026 09:21:36 GMTCrew Pocket 在 10/11 合併了 Map-first 的 Crew World：現在以地圖當主要入口，Dashboard 作為次要檢視。這不只是視覺改版，而是把 Role 的導航、恢復與工作狀態放到同一個操作模型。https://magic76.github.io/crew-app/#/briefing/2026-10-11/gh-ad0f51bdd853gh-ad0f51bdd853Sun, 11 Oct 2026 00:19:40 GMTTeacher 已加入離線 A1–C1 能力探測與重新校正，修正一直在簡單題打轉的問題。現在應關注測級判斷是否夠快、是否有清楚的升級證據，而不是只增加題庫數量。https://magic76.github.io/crew-app/#/briefing/2026-10-11/gh-c5f5b29703f9gh-c5f5b29703f9Sat, 10 Oct 2026 01:27:35 GMTCrew Story 最近合併 Gemini Key 驗證與 Live 相機讀書。這和 Teacher 的教材陪讀有共通技術：連續畫面、模型看到最新影像、語音與頁面切換同步。https://magic76.github.io/crew-app/#/briefing/2026-10-11/gh-e8e7c0fbd8ddgh-e8e7c0fbd8ddSat, 10 Oct 2026 22:14:29 GMTCrew Web 的 Teacher 與 Story 桌面雙欄工作區已合併。接下來應把報告頁與產品功能保持一致的排版品質，避免首頁、報告中心、各 App 視覺上像三套產品拼接。