騰訊看點CTO徐羽：推動AI技術落地，永遠保持“技術身，產品心，用戶眼”

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"記得在兩年前的「930組織架構變革」不久，同年11月份，騰訊發佈了信息流內容服務品牌——騰訊看點，把 QQ 看點、快報和 QQ 瀏覽器的資訊內容整合到一起，利用機器學習、算法等技術爲不同年齡層的用戶精準推薦信息流，可以看出騰訊在利用AI技術推動內容分發和推薦這件事上開始發力，除此之外，還有基礎設施，基礎架構的演進。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"恰好在一年之後的10月15日，騰訊與搜狗正式交接，搜狗全員轉換身份入職騰訊。交接日當天上午，騰訊發文在 PCG（平臺與內容事業羣）下成立“信息平臺與服務線”，負責 QQ 瀏覽器、看點、搜索、免費小說、文件等業務，爲用戶提供信息搜索、瀏覽消費、編輯存儲、信息服務等平臺與服務，負責人爲騰訊副總裁殷宇（Mel），而徐羽是新成立的“信息平臺與服務線”技術負責人。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"藉此機會，我們邀請徐羽老師來11月5日 "},{"type":"link","attrs":{"href":"https:\/\/aicon.infoq.cn\/2021\/beijing\/schedule","title":"xxx","type":null},"content":[{"type":"text","text":"AICon 全球人工智能與機器學習技術大會"}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"（北京站）2021 上演講，分享 AI 與推薦技術在騰訊看點的應用。在正式演講前，我們採訪了徐羽老師，聊聊他們在技術上有哪些投入，有哪些技術成果，以下是整理的內容。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"騰訊看點 AI 平臺建設和演進的三個階段"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"徐羽作爲騰訊看點技術負責人，親歷並主導了騰訊看點 AI 平臺建設和演進的三個階段，重點圍繞推薦算法、算力以及 NLP 技術方面。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"第一次改造是從淺層的推薦算法模型向超大規模深度推薦算法模型的演進"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，早期偏 LR 和 DNN 的淺層算法模型在泛化能力和用戶興趣表達方面越來越不足，所以在中臺建設了無量大規模機器訓練推理平臺基礎上，把 QQ 瀏覽器的推薦模型升級爲千億級參數量的排序模型。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"第二次改造是推薦系統的訓練和推理從原來 CPU 爲主，升級爲 CPU+GPU 的混合方案，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"充分結合 CPU 擅長的 IO 網絡吞吐計算和 GPU 擅長的神經網絡計算，在降低服務器成本的同時大幅度提升機器學習的性能。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"第三次改造是 NLP 的大規模預訓練平臺，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"目前 QQ 瀏覽器實驗室發佈的神舟預訓練模型在中文語義理解榜 CLUE 上首次打敗了人類，未來騰訊看點所有 NLP\/多模態模型都會基於這套預訓練模型基礎進行 fine-tune 的模式，大幅度提升 NLP 和多模態的訓練和研發速度。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"關於神舟模型介紹，可以點擊這裏查看："},{"type":"link","attrs":{"href":"https:\/\/m.thepaper.cn\/baijiahao_14735653","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/m.thepaper.cn\/baijiahao_14735653"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"融合——新一代推薦架構、新的推薦分發設計"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"最初的騰訊看點，整合了過去的 QQ 看點、快報和 QQ 瀏覽器的資訊內容，當時的架構層面有做一定的融合，不過由於是涉及存量業務的推薦系統改造，範圍要覆蓋底層樣本、特徵、RPC 協議的大量重構，實際上對一個推薦系統來說還是非常的複雜。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"徐羽說，目前騰訊看點的研發團隊正在 PCG 中臺裏面主導開發新一代推薦架構 TRS（Tencent Recommendation System），後續騰訊看點推薦架構會更多的往這個新架構做遷移。 "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"面對來自 QQ 看點、快報和 QQ 瀏覽器的不同年齡羣體的用戶，如何在不傷害用戶習慣的前提下進行內容分發和推薦？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/d7\/a3\/d7aeb8282f5c803c5f10e792b8a28ea3.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"徐羽介紹，這方面的個性化的推薦本質上就應該解決年齡和興趣點範圍不一致的問題，具體到實際的操作層面，有幾個推薦分發設計上的重點："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"第一是區分人羣單獨建模，快報、瀏覽器和看點的用戶模型肯定是需要獨立建模的，這樣才能保證合適人羣的算法不會被大盤帶偏，保持原來的精準度。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"第二是底層的內容特徵會有一定的打通，保證一個新的熱門內容有機會通過多個業務的流量放大，通過多級火箭的方式形成爆款，最終可以把更快更熱門的內容推送給用戶。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"通過 AI 算法大賽促進技術落地"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"今年8月份，騰訊 QQ 瀏覽器組織了 AI 算法大賽，徐羽擔任嘉賓評委。大賽組委會從信息流實際應用場景痛點出發，大賽議題設置了“自動超參數優化”賽道，重點是探索如何用 AI 技術解決模型或策略效果依賴於“超參數”問題的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"徐羽說，傳統的做法在融合多個不同目標模型（例如 CTR、時長、互動等）的時候是需要人工配置一些超參數再去觀察現網的效果，再反饋回來進行調整，這個週期很長而且在超參數空間比較大的情況下通常難以找到最優參數。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"自動超參數優化的目標就是通過 AutoML 中的 HPO 方式，把人工尋參的方式通過網格搜索、非個性化尋參到最終個性化尋參的方式提升效率，在騰訊看點實踐過程中可以減少80%以上的尋參時間，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"大大提升了推薦系統的研發效率。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"另外，由於短視頻消費引來爆發式增長，視頻的語義理解對於提升用戶消費效率至關重要。在“多模態視頻相似度”上，騰訊看點的解決方案是，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"首先採用了圖像、鏡頭、視頻逐級表徵融合方式對視頻內容進行表徵，提高對不同時空間尺度視頻內容的表達能力，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"同時利用業界新的跨模態預訓練模型來抽取能更好關聯視覺和文本的特徵，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"引入了 MLM（詞的掩碼預測）、VTM（視頻、文本的匹配）等多種預訓練任務來提升對比學習的效果，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"並且採用多任務學習範式，將分類、標籤等多種類型監督信息融合到相似度模型中，在純內容的視頻相似度預測準確性上提升了10%以上。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"其次，騰訊看點團隊還利用推薦中用戶點擊、消費視頻的數據來學習視頻之間的高層次語義關聯。另外在此次 QQ 瀏覽器算法大賽中，看點團隊也提出了用 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"Spearman Ranking Coefficient來離線評估視頻相似度，比傳統方法更穩定，能支持更高效的算法迭代。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"以上幾方面的創新也在瀏覽器的視頻推薦中獲得了顯著收益。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"着眼於推薦和搜索場景的大規模預訓練模型探索"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"這兩年各大公司在超大規模預訓練模型和推理 Framework 上都有研究進展，比較有名的模型參數量都已經到萬億級別了，應用場景差異也比較明顯。以 GPT-3 爲例，它依然是偏 NLP 的模型。當然，在AI大模型的探索上，有的機構着眼於提升超大規模AI算力的，有的着眼於突破學術前沿的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"騰訊看點在這方面的思路從實際業務出發，探索如何利用大規模預訓練模型來產出業務價值。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"徐羽進一步解釋說，在實際的推薦和搜索業務場景中，有異常豐富而且持續演化的內容需求，而這些內容需求在開始的時候都是遠不如學術界的數據那樣有着客觀清晰的定義。所以數據標註就成爲內容算法研發的核心瓶頸。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"看點團隊利用騰訊豐富的內容數據，產出了“神舟”預訓練模型，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"基於神舟模型來微調滿足業務的 NLP 需求（例如評論理解、搜索 Query 推薦等），減少了40%以上所需的標註數據量和相應的研發時間。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"另外在學術上，也在權威的中文自然語言理解評測基準 CLUE 上首次超越了人類。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"此外，AI算力的演化，召回的算力不斷提升，Ranking 的算力越來越強，尤其是在雲原生環境下，這也給推薦搜索架構帶來了挑戰。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"目前在單次搜索或者推薦的請求裏面，已經需要數百個模型來進行在線推理和千萬級 Item 的索引檢索。徐羽說，隨着模型複雜度進一步提升，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"目前的推薦架構確實有點力不從心了，看點團隊也做了一個比較大的創新，推薦搜索系統的在線推理從原來CPU爲主，升級爲 CPU+GPU 混合的方案，充分結合 CPU 擅長的 IO 網絡吞吐計算和 GPU 擅長的神經網絡計算，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在降低服務器成本的同時大幅度提升推薦和搜索系統的性能。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"永遠保持“技術身，產品心，用戶眼”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"徐羽在技術研究和產品設計上，都有豐富的經驗，最早的時候也參與過數據分析工具 TBI 項目。在產品設計和產品運營上有哪些堅持的原則和方法，按照徐羽的說法，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"因爲他工作以來一直是在前線做業務，所以更多時候會像一個AI產品架構師去思考，對於不同的產品會用不同的理念來區分對待。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"對於偏長線的技術產品，例如對於 PCG 的 AI 中臺，他會在初期就做一個3年的長期規劃，從點到線到面的 AI 路徑圖每年實現一步："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"第一年的點就是全面推動 PCG 多個團隊的 NLP 和 CV 能力的全面 PaaS 化，最終構造成幾百個 AI 技術點；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"第二年的線就是推動所有推薦搜索和 AI 技術上 AI 管線和無量機器學習系統，通過這條 AI 管線串聯上述所有 AI 的點；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"第三年的面，面是最難的，就是要打造一個通用的推薦系統 TRS，把之前所有的線和點全部融合進這個技術面裏。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"對於偏業務層面的技術，例如 QQ 瀏覽器的信息流推薦，徐羽會堅持以用戶體驗驅動來牽引，雖然推薦是一門非常有技術深度還在不斷演進的前沿AI技術，但是推薦產品本質依然是用戶的體驗。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"“對於我們內部每上線一個推薦A\/B實驗，都是要求研發人員除了講清楚技術原理和數據收益，也期望大家能從本質上可以解釋具體解決了用戶的哪些痛點和哪個場景裏面的問題。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"徐羽一直保持着對前沿 AI 新技術的興趣和細節的瞭解，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"但是隨時能跳脫出來以業務產品的角度來 review 是否存在技術的過度追求帶來的用戶體驗或者產品上的負面問題。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"同時，他還認爲，擁有對產品長期技術需求特別是技術難點的預測，是 AI 產品架構師的綜合能力要求和定位。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"嘉賓介紹"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"徐羽，2009 年加入騰訊，現任信息平臺與服務線 CTO 兼總經理、PCG 事業羣推薦 AI 中臺負責人。碩士畢業於加拿大滑鐵盧大學電子與計算機工程系，加入騰訊前在加拿大黑莓公司工作 6 年，參與 BIS 手機郵件研發工作。從 2009 年開始負責手機QQ 瀏覽器從 0 起步到現在億級 DAU 規模的研發工作，在 2018 年建立和負責 PCG 的推薦 AI 中臺，在機器學習平臺、NLP、CV 視頻理解、推薦算法和推薦架構等方面帶領團隊支持 QQ 瀏覽器和 PCG 業務的 AI 落地應用。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"活動推薦"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/aicon.infoq.cn\/2021\/beijing\/schedule","title":"xxx","type":null},"content":[{"type":"text","text":"AICon 全球人工智能與機器學習技術大會"}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#3e3a39","name":"user"}}],"text":"（北京站）2021 將於11月5-6日在北京國際會議中心舉辦。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#3e3a39","name":"user"}}],"text":"除了4個主題演講，大會設置了人工智能前沿技術、通用機器學習技術、計算機視覺實踐、智能金融技術與業務結合、推薦廣告技術與實踐、AI 工程師團隊建設與管理、認知智能的前沿探索、AI 與產業互聯網結合、大數據計算和分析、大規模機器學習算法及應用、智能語音前沿技術應用、大規模預訓練模型進展、自動駕駛技術等，共 14 個專題。目前已全部上線官網，點擊閱讀原文查看。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#3e3a39","name":"user"}}],"text":"會期臨近，還有少量餘票，購票請聯繫票務小姐姐文柳：13269078023（電話同微信）"}]},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/7a\/75\/7aa91a157105dc69bda6eb30bcf0aa75.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}}]}