开篇不谈ChatGPT,科技圈内无谈资。
去年底,ChatGPT刚刚出来的时候,有不少业界人士认为这是一款昙花一现的产品。“这种产品火的快,谢的也快”“微软2014年就推出小冰了,ChatGPT不是拾人牙慧吗?”“ChatGPT注定是一款噱头大于应用的软件”。
没想到,短短2个月,ChatGPT成为了全球最受关注的互联网服务,没有之一。它在2023年1月末的月活用户已经突破了1亿,成为史上用户增长速度最快的消费级应用程序,远快于TikTok、微信、Facebook等现象级产品同期增速。
ChatGPT到底有多火?
我列举几个片段:以色列总统使用ChatGPT生成的内容,在一场技术会议上公开发表了一段致辞,骗过现场2万名听众,引发热议;大量美国大学生用ChatGPT做作业,斯坦福研究人员推出了DetectGPT,反制ChatGPT侵入校园;比尔盖茨认为ChatGPT的历史意义重大,不亚于PC或互联网诞生;一家和ChatGPT没有半毛钱的奢侈品平台——寺库,也表示将对ChatGPT相关技术进行深入研究和拓展,以改变现有奢侈品电商的运营模式。
在我的朋友圈,能和ChatGPT比热度的,只有《狂飙》。有必要说明,这种惊人热度是在大部分中国人还无法使用ChatGPT的前提下达成的。
那么,问题就来了。如果说ChatGPT是互联网未来的重要发展方向,中国版的ChatGPT何时面市?
中国版ChatGPT下月推出 百度率先抢跑
其实,关于中国版ChatGPT的相关话题,此前我就接受《环球时报》《证券日报》等媒体采访,谈了一些预判。
在我看来,中国科技企业距离ChatGPT是有差距,但差距没有我们以为的那么大。中国有海量的数据积累,有越来越多的经过实践考验的训练模型,咱们的服务器遍布祖国各地,综合算力不比美国差多少。
中国科技企业的AI能力正在加速落地,百度、华为、腾讯、阿里、商汤、京东、字节的AI落地服务,不见得比微软、谷歌、亚马逊差多少,甚至在一些领域的应用更务实,比如无人驾驶、数字工业、智慧物流等。
至于中国的这批科技企业,哪家最有可能率先搞出来媲美ChatGPT的服务,我认为百度的可能性最大。
果不其然,2023年2月7日下午,百度对外宣布将于3月份推出中国版ChatGPT,项目名为“文心一言”。
据悉,目前文心一言团队正在加班加点,全力赶工做上线前的冲刺,不排除有提前开放内测的可能。对于那些因为众所周知原因无法使用ChatGPT的网友,这是一大福音。
多方看好文心一言 百度市值暴涨超15%
文心一言最终成色如何,我们不得而知,但从各方表现、尤其资本市场的表现来看,文心一言的期待值早已经拉满,各方给予了积极期待。
百度在当天宣布“文心一言”推出,并将在3月份完成内部测试后,公司港股股价大涨15.33%,收盘价162.5港元,创近期新高。
百度文心一言被看好的逻辑在哪?很简单,一句话:虽然文心一言是新项目,但百度的技术储备已久!
可能很多人不知道,关于人工智能领域,百度经过多年积累布局,打造了四层技术架构,包括底层的芯片(比如昆仑)、深度学习框架(比如飞桨)、大模型(比如文心)以及最上层的应用(比如搜索),进行了全链条布局。
文心一言位于其中的模型层,在这个模型之上,又可以基于各个使用场景,融入其他的落地应用服务。
值得一提的是,早在人工智能时代之前,百度在搜索引擎时代就积累了海量的中文语言相关数据、算法、方法论,若论整个中国互联网,哪一家对“文字”“语言”“索引”“语义”的理解更深刻,无疑是百度。这些历经20多年积累的技术和数据,将助力百度不仅仅能推出中国版ChatGPT文心一言,还有较大可能做好文心一言。
这也是为什么百度的度晓晓、希加加推出即出圈的原因所在。尤其百度度晓晓,去年因写作高考作文得高分而引发广泛热议,引发了各界关于“AI写作能否取代人类”的担忧。此后,百度的文心一格推出后,也引发了各界关于“AI绘画能否取代人类”的担忧。
我倒是认为,有这种担忧,恰恰说明了百度在人工智能领域的技术领先。这也说明,关于AIGC涉及到的技术储备,百度其实比ChatGPT背后的OpenAI更早,而且早了很多年。
百家争鸣百花齐放 中美两国更多ChatGPT已经在路上
有人认为ChatGPT是颠覆一切的大杀器,它可以极大程度地提升我们生活、工作效率,可以应用到方方面面。也有人认为ChatGPT是即将打开的潘多拉魔盒,将对现有互联网格局下的诸多主流产品带来毁灭性打击,比如搜索引擎就有可能被ChatGPT彻底取代。
无论如何,不管哪种认知占据主流,关于ChatGPT的发展一刻也不停歇,无人阻挡。中美两国主流互联网企业均在进行类似ChatGPT的产品布局,更多ChatGPT已经在路上。
腾讯申请的“人机对话方法、装置、设备及计算机可读存储介质”专利获授权,以腾讯的产品能力,相信年内有望推出落地产品;谷歌公司母公司宣布将推出聊天机器人“巴德(Bard)”,正面与ChatGPT竞争;微软与OpenAI达成了100亿美元的交易,计划将后者的技术纳入其必应搜索引擎和其他产品;360公司对外宣布,也计划尽快推出类ChatGPT技术的demo版产品;科大讯飞表示有ChatGPT相关技术积累,科大讯飞最有希望把ChatGPT不仅在中文语言模型中落实,同时也最有希望把它应用在教育、医疗这些非常重要的行业赛道。
至于汉王科技、寺库等蹭热度的企业,更是不胜枚举。
不出意外的话,在2023年年内,中美两国会出来最少几十款类ChatGPT或者蹭ChatGPT热度的相关产品,这些产品水平参差不齐,短时间内就会跑出几个梯队。综合各方要素分析,我更看好顶级的科技巨头来做这项工作,至于广大创业企业,应该积极拥抱巨头,利用巨头开发的大模型,服务于自身发展。
附录:笔者接受《环球时报》关于ChatGPT相关采访的报道
《Chinese tech firms eye ChatGPT-style bot as move kicks off new round of AI race in the world》
ChatGPT, OpenAI's highly-regarded chatbot, is making a splash worldwide and has been stirring up excitement within the Chinese AI community recently. Discussing its "impressive yet imperfect performance", some industry players also raised a question: How far are we from a ChatGPT designed by a Chinese company?
On January 30, roughly three months after its official launch, Chinese tech giant Baidu was reportedly to be planning the launch of an AI chatbot service similar to ChatGPT, which may be unveiled in March, with the initial version having been integrated into its search service, and it will allow users to receive conversational search results.
Baidu CEO Robin Li hailed the project as "leading the generational reform in the search experience," saying internally that relevant technology has reached a critical point and that Baidu has a great opportunity in it, according to a report from the domestic news site chinastarmarket.com.
China's explorations
"It's ambitious and exciting. But we believe it's achievable because the technology has reached a critical point. Technologies like ChatGPT service will get popular, but how to apply it to search scenarios is more difficult to explore than the technology itself. How can we make this happen, so that hundreds of millions of people benefit from it every day, and their efficiency is significantly improved? We have a great opportunity in this regard," Li said in the report.
"Baidu does have plans for (making) such a tool, and research was accelerated at the end of 2022," several people close to the matter told the Global Times, declining to disclose further details.
Baidu did not reply to an interview request from the Global Times.
So, what is ChatGPT? "I am ChatGPT, an AI language model developed by OpenAI. I can answer questions on various topics and perform various language tasks such as text generation, summarization and translation," the chatbot told the Global Times while being asked to introduce itself.
The chatbot goes beyond traditional ones that are focused on personal communication and provides a range of responses including code, essays, and scientific reports, and sometimes, proves to outperform human beings.
Baidu, a Chinese internet search leader in China, has its tech strength to do the "cash-burning", Liu Dingding, a veteran internet industry watcher, told the Global Times on Sunday.
Although in terms of semantic understanding and core presentation ability, it may not necessarily be better than the US one, after the first step of the official launch, as the number of users and data collected increase and computing power strengthens, its understanding will improve, and it will definitely become an indispensable tool in answering content creation questions, Liu said.
Apart from Baidu, industry watchers told the Global Times that other Chinese firms have also made forays into the sector and made some early explorations.
"Our team is currently exploring using chatbots to create short videos and other similar content, and enable creators to rev up production efficiency in content creation," a staffer at a domestic short-video company, told the Global Times.
The individual noted that though user experience of ChatGPT is "breath-taking" compared with its predecessors, disadvantages are also obvious - for instance, ChatGPT is prone to giving biased, incorrect answers, or outdated ones since its training data is unfiltered and only goes up to 2021.
"We are trying to avoid that for our training," the person said on condition of anonymity.
Other chatbots in China are also used in more professional tasks, such as programming, essay writing or game development.
Chinese tech giants such as Bytedance, Huawei, Tencent, Alibaba, and SenseTime all have strong AI capabilities, as well as the accumulation of relevant data needed to develop a model, Liu said.
However, regarding to basic AI algorithm structures and the models, Chinese enterprises lag behind by about two years, according to some industry insiders.
AI algorithm is key
ChatGPT's question-and-answer logic and performance as an AI are very good, which proves that AI has become a useful tool. The difference from the previous technology is that ChatGPT is a large model, which has a stronger understanding of contextual semantics than other AI algorithms, and the data accumulation speed is faster than its competitors, with growth and usability, Pan Helin, joint director of the Research Center for Digital Economics and Financial Innovation affiliated with Zhejiang University's International Business School, told the Global Times on Sunday.
But so far, "we have not seen a relatively strong AI algorithm model in China at this stage", Pan said. It may be technically capable, but by combining the three levels of algorithm, computing power, and data, there are currently no domestic companies that can launch ChatGPT-equivalent algorithms, Pan said.
A report released by UBS on Wednesday showed that ChatGPT is estimated to have reached 100 million monthly active users in January, two months after its launch, becoming the fastest-growing consumer app.
This is an unprecedented rate of expansion - TikTok took about 9 months to reach 100 million users after its launch, while Meta's Instagram took about 2.5 years.
"I had very mixed feelings after I tried ChatGPT - I am excited to see that artificial intelligence can develop to this point, but also concerned whether Chinese companies can rapidly catch up, though we can anticipate that," Liu said.
In essence, there exists a competition of AI capabilities between China and the US - to compete with such a strong competitor, China has to go ahead and overcome all challenges it faces, from management ideas, computing power, algorithms, and data circulation, to capital investment needed to achieve the set goal, Pan said.