中國(guó)最大的科技公司之一阿里巴巴在8月25日發(fā)布了兩個(gè)新人工智能模型,大幅提高了人工智能應(yīng)用的可能性。
這兩個(gè)開源模型分別被命名為Qwen-VL和Qwen-VL-Chat,是兩個(gè)視覺語(yǔ)言模型,這意味著它們能夠理解圖像,而不是像ChatGPT和谷歌(Google)的Bard等競(jìng)爭(zhēng)對(duì)手一樣,只可以閱讀文本。Qwen-VL-Chat承諾能夠提供一些復(fù)雜的功能,例如通過(guò)掃描街道標(biāo)志指路,根據(jù)一張照片解決數(shù)學(xué)方程式,根據(jù)多張圖片編寫一段故事等。比如,阿里巴巴表示,它可以掃描醫(yī)院中的普通話標(biāo)志圖像,并翻譯成英文,或者幫助媒體編寫一張圖片的文字說(shuō)明。
8月25日發(fā)布的另外一個(gè)模型Qwen-VL是現(xiàn)有圖像閱讀聊天機(jī)器人的更新版,新版本現(xiàn)在能夠解讀分辨率更高的圖片。
除發(fā)布公告以外,阿里巴巴未回復(fù)《財(cái)富》雜志的置評(píng)請(qǐng)求。
隨著人工智能從噱頭變成了真正的顛覆性技術(shù),開發(fā)者紛紛展開“軍備競(jìng)賽”,推出日益尖端的工具,這些人工智能技術(shù)的迭代正是這場(chǎng)競(jìng)賽的最新進(jìn)展。例如,阿里巴巴表示其最新圖片掃描技術(shù)有許多應(yīng)用機(jī)會(huì),比如幫助視力受損人士購(gòu)物、幫助他們掃描商品,以及由聊天機(jī)器人為他們朗讀商品標(biāo)簽等。
在阿里云的專有模型即服務(wù)平臺(tái)Modelscope和提供人工智能模型數(shù)據(jù)庫(kù)的熱門初創(chuàng)公司Hugging Face,可以獲得這兩個(gè)模型。
在阿里巴巴發(fā)布新模型的前一天,Meta發(fā)布了一款人工智能模型。該模型以7月發(fā)布的開源模型Llama 2為基礎(chǔ),經(jīng)過(guò)微調(diào)能夠用于編寫代碼。過(guò)去幾個(gè)月,阿里巴巴一直在努力追趕Meta在人工智能領(lǐng)域的進(jìn)展。8月早些時(shí)候,阿里巴巴發(fā)布了最早的兩款開源大語(yǔ)言模型Qwen-7B和Qwen-7B-Chat,8月25日發(fā)布的最新版本分別以這兩個(gè)版本作為基礎(chǔ)。7月,阿里巴巴與Meta達(dá)成協(xié)議,將通過(guò)阿里巴巴的云業(yè)務(wù)部門,在中國(guó)市場(chǎng)提供Meta的Llama 2模型。
阿里巴巴希望通過(guò)將新模型開源,可以幫助用戶完善開發(fā)應(yīng)用程序或進(jìn)行研究使用的工具。大多數(shù)人工智能公司希望,用戶能夠把開源模型改造為適合特定使用案例的工具,從而避免從零開始開發(fā)大語(yǔ)言模型的繁瑣過(guò)程。除了提供開源模型以外,人工智能公司還將其專有模型作為服務(wù)提供,希望在這個(gè)新興行業(yè)搶占市場(chǎng)份額。
中國(guó)政府高度重視發(fā)展人工智能
今年7月,中國(guó)政府成為率先頒布綜合性人工智能法規(guī)的國(guó)家之一,專家表示此舉為阿里巴巴和其他中國(guó)科技公司公開發(fā)布自己的產(chǎn)品開了綠燈。
阿里巴巴還準(zhǔn)備進(jìn)行全面重組,把負(fù)責(zé)人工智能研究的云計(jì)算部門阿里云拆分為一個(gè)獨(dú)立部門,此舉受到投資者歡迎。由于人工智能技術(shù)需要大量算力,必須搭配云網(wǎng)絡(luò)才可以正常運(yùn)行,因此將人工智能與云計(jì)算整合到一個(gè)部門,將提高人工智能的效率。阿里云的現(xiàn)任首席執(zhí)行官兼董事長(zhǎng)張勇將在9月卸任,阿里巴巴的兩位聯(lián)合創(chuàng)始人吳泳銘將擔(dān)任首席執(zhí)行官,蔡崇信將出任董事長(zhǎng)。
中國(guó)政府曾經(jīng)不止一次指出,人工智能是中國(guó)科技未來(lái)的關(guān)鍵,并與美國(guó)展開了人工智能領(lǐng)域的競(jìng)賽。即使像阿里巴巴在8月25日發(fā)布的模型這種看似無(wú)害的工具,由于它們所采用的基本技術(shù)并且可能被其他開發(fā)者使用,因此它們也可能被卷入兩國(guó)之間的競(jìng)賽。倫敦國(guó)王學(xué)院(King’s College London)的中國(guó)研究所(Lau China Institute)的所長(zhǎng)凱里·布朗在8月早些時(shí)候?qū)Α敦?cái)富》雜志表示,人工智能“已經(jīng)成為中美兩國(guó)爭(zhēng)奪主導(dǎo)權(quán)的戰(zhàn)場(chǎng)”。
到目前為止,中國(guó)科技公司似乎稍微落后于美國(guó)同行。Meta的Llama 2模型的開源版本基于約700億個(gè)變量(在人工智能領(lǐng)域被稱為參量),比阿里巴巴新發(fā)布的模型多約10倍(阿里巴巴表示其較大的模型不會(huì)開源)。(財(cái)富中文網(wǎng))
譯者:劉進(jìn)龍
審校:汪皓
中國(guó)最大的科技公司之一阿里巴巴在8月25日發(fā)布了兩個(gè)新人工智能模型,大幅提高了人工智能應(yīng)用的可能性。
這兩個(gè)開源模型分別被命名為Qwen-VL和Qwen-VL-Chat,是兩個(gè)視覺語(yǔ)言模型,這意味著它們能夠理解圖像,而不是像ChatGPT和谷歌(Google)的Bard等競(jìng)爭(zhēng)對(duì)手一樣,只可以閱讀文本。Qwen-VL-Chat承諾能夠提供一些復(fù)雜的功能,例如通過(guò)掃描街道標(biāo)志指路,根據(jù)一張照片解決數(shù)學(xué)方程式,根據(jù)多張圖片編寫一段故事等。比如,阿里巴巴表示,它可以掃描醫(yī)院中的普通話標(biāo)志圖像,并翻譯成英文,或者幫助媒體編寫一張圖片的文字說(shuō)明。
8月25日發(fā)布的另外一個(gè)模型Qwen-VL是現(xiàn)有圖像閱讀聊天機(jī)器人的更新版,新版本現(xiàn)在能夠解讀分辨率更高的圖片。
除發(fā)布公告以外,阿里巴巴未回復(fù)《財(cái)富》雜志的置評(píng)請(qǐng)求。
隨著人工智能從噱頭變成了真正的顛覆性技術(shù),開發(fā)者紛紛展開“軍備競(jìng)賽”,推出日益尖端的工具,這些人工智能技術(shù)的迭代正是這場(chǎng)競(jìng)賽的最新進(jìn)展。例如,阿里巴巴表示其最新圖片掃描技術(shù)有許多應(yīng)用機(jī)會(huì),比如幫助視力受損人士購(gòu)物、幫助他們掃描商品,以及由聊天機(jī)器人為他們朗讀商品標(biāo)簽等。
在阿里云的專有模型即服務(wù)平臺(tái)Modelscope和提供人工智能模型數(shù)據(jù)庫(kù)的熱門初創(chuàng)公司Hugging Face,可以獲得這兩個(gè)模型。
在阿里巴巴發(fā)布新模型的前一天,Meta發(fā)布了一款人工智能模型。該模型以7月發(fā)布的開源模型Llama 2為基礎(chǔ),經(jīng)過(guò)微調(diào)能夠用于編寫代碼。過(guò)去幾個(gè)月,阿里巴巴一直在努力追趕Meta在人工智能領(lǐng)域的進(jìn)展。8月早些時(shí)候,阿里巴巴發(fā)布了最早的兩款開源大語(yǔ)言模型Qwen-7B和Qwen-7B-Chat,8月25日發(fā)布的最新版本分別以這兩個(gè)版本作為基礎(chǔ)。7月,阿里巴巴與Meta達(dá)成協(xié)議,將通過(guò)阿里巴巴的云業(yè)務(wù)部門,在中國(guó)市場(chǎng)提供Meta的Llama 2模型。
阿里巴巴希望通過(guò)將新模型開源,可以幫助用戶完善開發(fā)應(yīng)用程序或進(jìn)行研究使用的工具。大多數(shù)人工智能公司希望,用戶能夠把開源模型改造為適合特定使用案例的工具,從而避免從零開始開發(fā)大語(yǔ)言模型的繁瑣過(guò)程。除了提供開源模型以外,人工智能公司還將其專有模型作為服務(wù)提供,希望在這個(gè)新興行業(yè)搶占市場(chǎng)份額。
中國(guó)政府高度重視發(fā)展人工智能
今年7月,中國(guó)政府成為率先頒布綜合性人工智能法規(guī)的國(guó)家之一,專家表示此舉為阿里巴巴和其他中國(guó)科技公司公開發(fā)布自己的產(chǎn)品開了綠燈。
阿里巴巴還準(zhǔn)備進(jìn)行全面重組,把負(fù)責(zé)人工智能研究的云計(jì)算部門阿里云拆分為一個(gè)獨(dú)立部門,此舉受到投資者歡迎。由于人工智能技術(shù)需要大量算力,必須搭配云網(wǎng)絡(luò)才可以正常運(yùn)行,因此將人工智能與云計(jì)算整合到一個(gè)部門,將提高人工智能的效率。阿里云的現(xiàn)任首席執(zhí)行官兼董事長(zhǎng)張勇將在9月卸任,阿里巴巴的兩位聯(lián)合創(chuàng)始人吳泳銘將擔(dān)任首席執(zhí)行官,蔡崇信將出任董事長(zhǎng)。
中國(guó)政府曾經(jīng)不止一次指出,人工智能是中國(guó)科技未來(lái)的關(guān)鍵,并與美國(guó)展開了人工智能領(lǐng)域的競(jìng)賽。即使像阿里巴巴在8月25日發(fā)布的模型這種看似無(wú)害的工具,由于它們所采用的基本技術(shù)并且可能被其他開發(fā)者使用,因此它們也可能被卷入兩國(guó)之間的競(jìng)賽。倫敦國(guó)王學(xué)院(King’s College London)的中國(guó)研究所(Lau China Institute)的所長(zhǎng)凱里·布朗在8月早些時(shí)候?qū)Α敦?cái)富》雜志表示,人工智能“已經(jīng)成為中美兩國(guó)爭(zhēng)奪主導(dǎo)權(quán)的戰(zhàn)場(chǎng)”。
到目前為止,中國(guó)科技公司似乎稍微落后于美國(guó)同行。Meta的Llama 2模型的開源版本基于約700億個(gè)變量(在人工智能領(lǐng)域被稱為參量),比阿里巴巴新發(fā)布的模型多約10倍(阿里巴巴表示其較大的模型不會(huì)開源)。(財(cái)富中文網(wǎng))
譯者:劉進(jìn)龍
審校:汪皓
Alibaba, one of China’s biggest tech companies, announced the release of two new A.I. models on August 25 that dramatically level up the possibilities of artificial intelligence.
The open source models, called Qwen-VL and Qwen-VL-Chat, are vision language models, meaning they “read” images rather than text, unlike competitors ChatGPT and Google Bard. Qwen-VL-Chat promises complex features like providing directions by scanning street signs, solving math equations based on a photo, and weaving together a narrative based on multiple pictures. For example, it can scan an image of a sign in a hospital written in Mandarin and then translate it into English, or help a news organization write a caption for a photo, the company says.
Qwen-VL, the other release on August 25, is an updated version of its existing image-reading chatbot that can now read pictures in higher resolution.
Alibaba declined to comment to Fortune beyond its public announcement.
These new iterations of A.I. are the latest shots fired in the arms race among developers to create increasingly sophisticated tools, as the technology graduates from gimmick to genuine game-changer. For example, Alibaba says its new image-scanning technology has significant opportunities to help visually impaired people with shopping, allowing them, for instance, to scan an item and have the chatbot recite the label back to them.
Both models will be made available on Alibaba Cloud’s proprietary model-as-a-service platform Modelscope and on Hugging Face, the popular startup that has a library of A.I. models.
Alibaba’s release comes just a day after Meta launched an A.I. model fine-tuned for writing code, built on the open-source Llama 2 model released in July. Alibaba has been trying to keep up with Meta’s A.I. rollouts for the last few months. Earlier this month, Alibaba unveiled its first two open-source large language models, Qwen-7B and Qwen-7B-Chat—the same ones that form the basis for August 25’s releases. In July, the two companies struck an agreement to make Meta’s Llama 2 model available to the Chinese market via Alibaba’s cloud division.
By making these new models open-source, Alibaba is letting users tweak the tools to develop their own apps or conduct research. Most A.I. companies hope that users will adapt open-source models into tools for highly specific use cases, without having to undertake the onerous task of building a large language model from scratch. Alongside the open-source offerings, the companies offer their proprietary models as a service, hoping to capture market share in the burgeoning industry.
A.I. development is a priority for the Chinese government
Just in July, the Chinese government became one of the first countries to issue comprehensive regulations for A.I., a development that experts say gave Alibaba and other Chinese tech companies the green light to make their products public.
Alibaba is also preparing to undergo a complete restructuring that would spin off Alibaba Cloud, the cloud computing division that houses its A.I. research, into an independent division, a move that investors welcome. Since A.I. technology requires significant computing power that can only be properly serviced with a cloud network, having the two in the same division would boost A.I.’s efficiencies. The current CEO and chairman of Alibaba Cloud, Daniel Zhang, is set to down in September, to be replaced by two of Alibaba’s cofounders: Eddie Wu as CEO and Joseph Tsai as chairman.
The Chinese government has on more than one occasion indicated that it considers A.I. critical to its technological future, setting up an race with the U.S. Even seemingly innocuous tools like those released by Alibaba on August 25 could be implicated because of their underlying technology and how other developers might use them. A.I. “has become a proxy in the battle for primacy between China and the U.S.,” Kerry Brown, director of the Lau China Institute at King’s College London, told Fortune earlier this month.
So far, it seems that Chinese tech companies are slightly lagging their U.S. counterparts. The open source version of Meta’s Llama 2 model is based on roughly 70 billion variables (called parameters in A.I. parlance), about 10 times larger than Alibaba’s new releases (Alibaba does say it has bigger models which aren’t open-source).