從虛擬學生“華智冰”到冬奧手語播報，人工智能大模型究竟還有多少可能性？

{"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":"9月28日，清華大學首位AI虛擬學生“華智冰”首次正臉亮相，併爲大家帶來一首歌曲，亮相視頻中，“華智冰”是一名眉目清秀的女孩，表情動作也十分真實，視頻一經發出，立馬引起網友熱議。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"那麼“華智冰”到底是誰？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"video","data":{"id":"423630","name":"5583c3b39602945a1a55241c742e76c4.mp4","poster":"","url":"https:\/\/media001.geekbang.org\/9122c1ac95df44e68c21e48f5c1c9914\/8d528c48b6464d2892a10382d927bd9c-ca43ce14db99d4c750c41d527902b2ff-sd.m3u8"}}},{"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":"三個多月前的“"},{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/theme\/102","title":null,"type":null},"content":[{"type":"text","text":"2021北京智源大會"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"”上，“華智冰”第一次出現在大衆視野中，她基於超大規模智能模型“"},{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/kE1hGZY6yaouPIKblgYa","title":null,"type":null},"content":[{"type":"text","text":"悟道2.0"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"”誕生，具有豐富知識、與人類有良好交互能力，會創作音樂、詩詞和繪畫作品。這次“她”憑藉一首彈唱成功圈了一波粉絲，也成爲頗受喜歡的“網紅”。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"“華智冰”的走紅同時讓“她”背後的世界最大的超大規模智能模型“悟道2.0”再次受到大家關注。自“悟道2.0”發佈之日起，關於它怎麼用起來就成爲業界關心的熱點話題。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在本月24日，在"},{"type":"link","attrs":{"href":"https:\/\/www.baai.ac.cn\/","title":null,"type":null},"content":[{"type":"text","text":"智源研究院"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"承辦的2021中關村論壇“人工智能與多學科協同創新”平行論壇"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"上，智源研究院學術副院長、清華大學教授唐傑公佈了“悟道2.0”應用的最新進展，並帶來了兩個最新應用案例。大模型時代的到來對於人工智能究竟有何意義？在會上，智源研究院院長黃鐵軍也給出了我們答案。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/84\/84e404c0e3c6058d78ae13e75cd03905.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"“悟道”大模型落地實踐"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"本次創新成果展示主要圍繞"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"“悟道”開放平臺及技術創新"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，包括：“悟道”大模型開放平臺、大模型高效推理工具包BMInf、全新升級的AI文圖繪畫大師CogView和實現大模型羣聊的對話模型EVA。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"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":"AI工業化開發模式的“悟道”大模型開放平臺，使得企業可以規模化、產業化地進行AI應用開發，降低了大模型開發門檻。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"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":"大模型高效推理工具包BMInf，實現了“一卡頂四卡”，支持在低配置的NVIDIA GTX 1060單卡機上運行百億大模型，同時具備能力全面、應用便捷的優勢。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"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":"AI文圖繪畫大師CogView本次又取得重大升級，4倍提速，4倍清晰，實現圖像生成品質與速度的雙重突破。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/5d\/5d5a6fec83aadd1294821a789d23f262.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"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":"對話模型EVA實行開源上線，除特別善解人外，還具有大模型羣聊功能，展望“人與物”交流的現在時代邁進到“物與物”交流的人機共融未來時態。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"同時針對大家關心的"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"應用落地方面"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，唐傑教授爲我們展示了兩個 “悟道”大模型賦能產業的最新技術應用案例，分別爲：“悟道”大模型與"},{"type":"link","attrs":{"href":"https:\/\/www.oppo.com\/cn\/smartphones\/series-find-x\/find-x3-pro-lensman\/?utm_source=baidu&utm_medium=pcppzq&utm_campaign=pcppzq0368&utm_term=oppo","title":null,"type":null},"content":[{"type":"text","text":"OPPO"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"小布助手開啓“生成式回答系統”以及基於智源“悟道”大模型的冬奧手語播報數字人。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"現階段，從長尾問題來看，OPPO小布助手以及整個AI語音回覆行業普遍存在以下難點："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"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":"單個問題命中率低、體量大且無法窮盡"}]}]},{"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":"自建語料庫週期長，人工數據標註成本高"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"而這些問題也導致海量問題無法通過AI語音助手找到答案。針對這一行業痛點， 智源“悟道”大模型與OPPO小布助手合作，研發“生成式回答系統”，功能上線後將解決行業共通性的長尾問題，單條回答的建設成本降低99%。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/1e\/1e11813a5e1038c5160b2d76133593f4.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"對於此次與OPPO小布助手的合作，智源研究院黃院長也在會後採訪中表示"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，“對於像OPPO這樣的大企業，直接對接大模型並沒有問題，但對於中小型企業，目前來講可能還有難度，我們目標是希望有一天能讓他們也可以簡單、方便地用上大模型，用AI技術解決各行業應用場景中的具體問題。畢竟我們相信它有滿足各類企業需求的潛力，只不過把潛力變成現實的中間還需要更多生態建設，悟道大模型正在朝着這個方向努力。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"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，本次論壇上我們還認識了一位“新朋友”——基於智源“悟道”大模型的冬奧手語播報數字人。“悟道”手語播報數字人將在北京冬奧會期間正式投入應用，它將提供全流程智能化的數字人手語生成服務，方便聽障人士也能收看賽事專題報道，提升他們的社會參與度和幸福感。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/cb\/cbed1fe404f478cd0fe7123cdc8a52ab.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"除了論壇上展示的兩個應用案例，在黃院長看來，大模型的“通用智能”能力在醫療、金融、新聞傳播等行業中同樣具有非常大的應用前景。例如，在醫療健康領域，大模型在醫療數據格式化、病歷自動解讀與分析、自動問診系統等都可以發揮巨大效用；在新聞傳播領域，基於模型可實現智能新聞線索收集、機器寫作、輔助編輯、虛擬主播等應用。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"因此通過以上對“悟道”應用生態的觀察，我們不難發現，對於人工智能的發展來說，從“大煉模型”邁向“煉大模型”已經是大勢所趨。那麼大模型時代究竟有何意義？對於廣大技術人員來說，又該以何種姿態面對其中的機遇與挑戰呢？"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"大模型時代的到來，意味着什麼？"}]},{"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":"駐足於讓人工智能普惠各行各業的願景上，黃院長曾說過：“之前全球形成了‘千村萬戶大煉模型’的熱鬧場面，但這種‘自家鍊鋼自己用’的作坊方式成本較高，顯然不符合現代產業發展規律。而大模型泛化能力強，可以做到‘舉一反三’，同一模型利用少量數據進行微調或不進行微調就能完成多個場景的任務，中小企業不需要招聘很多AI算法專業人員即可以進行應用開發，可以顯著降低中小企業的研發門檻，促進AI技術的落地。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"而今在思考“悟道2.0”該如何用起來以及落地的實踐過程中，黃院長以及科研團隊"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"對於大模型又有了更深刻的體會"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，“大模型就是把社會的各種數據資源、最強的算法以及算力整合在一起，變成公共基礎平臺”， 這也對大模型下一步的努力方向進行構想，即通用化、普及化、標準化。因此黃院長也在會上發出呼籲產希望學研能共同努力發展大模型，使其能更好地服務於各種應用。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/d8\/d8e7c511e483278fbd2491ae43f44765.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"同時，面對大模型的諸多挑戰技術人員更需不斷創新進步。例如針對模態信息的理解和建模的統一問題，需要加入多模態的數據，比如文本、圖片、視頻等類型的輸入，多模態的數據可以使得模型具備更多樣的能力，以更好地應對各類下游任務，提高模型的“智商”。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"從多才多藝的中國首個AI虛擬學生“華智冰”到助力冬奧數字人，可以說，屬於人工智能大模型的時代已經正式開啓。對此，由黃院長領攜的科研團隊也向全社會廣發英雄帖：歡迎各位有志之士加入到智源大模型的生態中來，爲大模型研發的技術路線及應用構建等提供具體指導，爲探索新一代人工智能發展提供助力。"}]}]}