今年是大语言模型爆发式发展的一年，ChatGPT风靡全球，在不到两个月的时间就吸引了上亿规模的用户，影响力空前。
随着ChatGPT的持续火爆，国内市场也热情高涨，陆续发布或宣布研发大语言模型产品，大语言模型赛道的竞争异常激烈。在展望大语言模型未来发展趋势的同时，AI大语言模型爆发又会带来哪些产业变革？
竹间智能总裁兼COO孙彬在腾讯科技Hi Tech Day暨2023数字开物大会上，就以“AI大语言模型爆发的产业变革”为题发表了主题演讲。 数字化转型网（www.szhzxw.cn）
大语言模型从文字到多模态进展非常快，从C端到B端大语言模型给产业带来了突破性变革，同时潜移默化改变用户生活方式，未来甚至会成为人们的生活依赖。关于大语言模型孙彬给出了自己的思考。
孙彬认为，大语言模型突破了人机对话的瓶颈，C端变化是最快的，未来赛道中所有的应用可能都会被对话的模式改变，千千万万的应用会因为大语言模型的方式改变应用范式。
谈到B端的变革，孙彬认为今天大语言模型成为了可以被使用的生产力，生产力的提升也将给产业带来巨大变革。
2023年大语言模型的爆发拉开了数字员工的序幕，同时数字员工是成本和效率完美的结合体。
在企业使用AI大模型方面，孙彬讲到今天大语言模型飞速迭代，企业可以选择外部购买或者寻找伙伴在产业中联合建设的方式，非行业头部企业，不建议去做自研开发。
谈及大语言模型的未来发展，孙彬认为大语言模型会变成能力帮手，成为大家的工作依赖、能力依赖和生活依赖。，时长03:16
孙彬演讲全文：
首先自我介绍一下，我叫孙彬。竹间智能专注在NLP赛道已经八年了，我们团队一直专注于NLP产业的落地。
过去我们自己训练模型，能够在客户场景中寻找数据场景落地，最终达到模型落地产业的效果，涉猎非常多的行业。 数字化转型网（www.szhzxw.cn）
前面几位嘉宾从芯片和数据方面做了很多剖析，下面我和大家分享的是站在自然语义赛道，站在大语言模型爆发的2023年，我们的视角是怎样的。
大家应该都很熟悉这些Logo，2023年是魔幻的一年。一年零两周之前，OpenAI发布ChatGPT引爆整个科技圈，除了AI，跨行业的都很关注这个赛道，很魔幻吧？OpenAI的董事会也很魔幻，前几天有人看到对OpenAI董事会比对自己董事会还了解，因为是产生的一场管理变革。
可以看到不仅是一年之后的今天，Google发布双子星大语言模型，为什么国内所有大模型公司都在奋起直追？就像在光明顶顶端有个叫GPT4的神仙，大家都在围攻光明顶，像不像？
今天大语言模型赛道的竞争非常激烈，效果非常惊艳，但对产业和行业到底有什么影响？
其实今天我们看到的内容是大语言模型从文字、语音、大模态有一个大的趋势，ChatGPT刚开始是文字，现在很多人都在迷恋GPT4的声音，但双子星带来的视觉识别、图像识别和理解才是更深的事情。大语言模型不再是文字的理解，多模态的进展非常快。
实际上科技界还有一件事情是我们都在讨论的，就是闭源和开源的竞争。大家还记得以前苹果和微软在PC开源和闭源的竞争、Android和IOS系统之间的竞争吗？今天大语言模型也有这样的竞争，包括LlAMA这样的开源大模型供大家使用，但依然有很多闭源的模型在专注地保持自己的健康性和安全性。
你们会选择开源模型还是闭源模型做业务？我认为这种竞争只是刚开始，大家会选择不同的赛道完善。
那么各个行业到底有什么变化？相信C端变化最快，大语言模型上来了，效率非常高，效果非常好，可以协作生成。 数字化转型网（www.szhzxw.cn）
去年迅速的有1亿人注册在OpenAI使用GPT，一半的人用来写作业，但今天的智能化才让大家惊艳，有着道德的考虑、生成的能力，相比通用的作者效率更高，写作质量更好，也有很多模板和知识可以帮助大家，交互的方式彻底改变，大家开始习惯对话，现在经常有人拿挑战性的问题去问大语言模型。
我们先不评价道德的判断，但智能化和交互的方式改变了，所以大家看到的是大语言模型的惊艳，从业者看到的是突破。
今天我们看到大语言模型突破了人机对话的瓶颈，可以像人一样理解对话，机器可以理解人的意图和判断，所有应用模式都会因此改变。
过去每次应用的改变都是大趋势，软件从PC端转到移动互联网端也有很大的改变，互联网的应用到手机APP上面重新被改写，我们非常明确地预测接下来的时间内，赛道中所有的这些应用可能都会被对话的模式改变。
你们会用手机点餐，也会用手机订机票，未来可能都是对话的方式进行，我们判断C端的产业变革就是千千万万的应用会因为大语言模型的方式改变应用范式。
谈完C端谈B端，To B一直是拥抱变化的。大家还记得自动化对B端改变了什么？电气时代自动化的方式让产业效率大幅度提升，这是一个巨大的变革。数字化、信息化大家也都了解，所有的生产制造系统可以通过数据流、财务流的方式被管理起来。 数字化转型网（www.szhzxw.cn）
现在到达了智能化时代。还有一个什么化是最近两年最火的？就是国产化。上市的很多公司都是国产替代的佼佼者，但今天大语言模型的爆发，就是国产智能化赛道中间，为什么不火呢？
站在从业者的角度，我们认为火的原因很简单。以前所有科技产品，哪怕计算器、哪怕计算机、哪怕IT设备都是工具类产品，可能叫做生产工具和提效工具，但大语言模型这件事情把生产力的三要素解决了。
生产力是劳动者+生产工具+劳动对象，要是大语言模型有足够的智能，加上私域企业的支持就会变成合格的能力，驱动设备、驱动流程，完成一个生产力的循环。今天大语言模型在To B端真正的变化就是可以被使用的生产力，生产力提升的时候对产业的变革就是巨大的。
这些都是过去经历的，在企业内部和外部，通过机器人对话的方式解决各个岗位和场景。换句话说，这些岗位都可以是数字员工完成。
今天大语言模型的爆发其实是数字员工序幕的拉开，技术拐点的来临，所有的从业者都会关注成本，都会关注效率，但大语言模型带来的数字员工往往是成本和效率非常完美的结合体，也就是数字员工的序幕会在2023年开始拉开。
谈完行业和产业，我们看一看企业会有什么变化。每次技术变革都会是使用技术甲方的一次军备竞赛，也是一次新的消费。 数字化转型网（www.szhzxw.cn）
因为每次技术革命，哪怕是为了企业提效，为了制造行业领先的样板工程，其实甲方都会拥抱这些技术变革。自动化时代如此，数字化时代如此，今天的智能化时代依然如此。
今天运用AI的时候倡导三种模式：大语言模型包括自我研发、外部购买、联合建设。现在看来，大语言模型的翻倍速度是以每两周的方式迭代，所以如果不是行业头部企业，不是自己有足够的IT开发实力的企业，不建议大家去做自研开发，不会拿着模型自己训练。过去的一年中，无数的团队在做模型训练，训练出来的结果其实都不可用。
我建议大家使用后面两种：要是今天选择最好的团队购买成熟产品，时间可能不是今年，也有可能不是明年年初，可能是2024年底。
通过整个产业团队一年多的对模型的沉淀，一定会有最佳的工程方法给到甲方和产业方使用。按照行业寻找最好的合作伙伴，能够在产业中联合建设也是一种路径。
我想给大家提另外一个醒，站在真正的客户方，其实视角如此。借鉴红杉的产业报告，可以看到200-300位CIO和CEO反馈。
甲方在科技变革的时候拥抱技术变革，但依然把投入产出作为第一原则。今天我们的产业界就算是消费也是要对产出有消费，所以必须在科技爆发的时间冷静，大语言模型是不是能够把ROI算得回来？
要是算不回来不会是一个好的生意，换句话说不是一个长久的生意。作为甲方和行业从业者都应该记住，不能变成大模型的业务不是好业务。
每次技术变革，乙方都要使用，因为会有破局者重新打破原来的产业排序，如果不利用新技术就会落后，市场份额就会丢失。 数字化转型网（www.szhzxw.cn）
所有的乙方都存在竞争焦虑，后方排位的企业希望趁机拿到市场份额。但是这种挑战又很痛苦，刚开始大家都在尝试，都是先投入才会有产出。大语言模型上来的时候所有乙方也是痛苦，但必须拥有新技术，拥有大语言模型能力的技术。
造模型的公司又是什么趋势？为什么说一开始就结束了？一开始的时候模型公司的趋势就已经被确定了，可以看到今天大语言模型的公司成长很快，但大语言模型的从业者可能只有三条成长路径：做好大语言模型的基础模型，做到行业领先，但这一点很挑战，因为通用大语言模型结果就是趋同，趋同就会导致切换成本很低。
过去数据中心切换IT需要搬迁，云端现在也是需要迁移，但在大语言模型切换可能只要改一个地址，大语言模型未来的数量肯定很有限。
不管是百模还是千模，最后真正往头部企业使用可能就是那么几家。其他企业怎么办？可能会在工具类、产品类发展。欧美有很多二三十人的企业利用大语言模型的垂类做得非常优秀，相信国内这样的企业也会如同雨后春笋一样爆发出来。
做好产品做精是必然，但防止大语言模型能力往前迈一步，前面企业全部被覆盖和死亡。C端会有这样的风险，B端就一定会专注行业模型吗？行业模型是解决方案，因为需要解决客户方案，很多企业会专注于这个赛道。 数字化转型网（www.szhzxw.cn）
Copilot现象第一步落地的一定是大家的办公，腾讯科技已经分享了很多前沿的Copilot现象，但有另外一个很大的趋势，今天可以做办公助手，但大语言模型可以做一个很好的培训老师。
大家可以看到今天大语言模型不仅可以作为写作生成，也可以角色扮演，不仅可以作为辅助，也可以作为老师，可以改变很多传统需要人员的习惯。
现阶段，大家都把Copilot看作效率帮手，可以在工作中有很大的提效，但仅仅如此吗？大语言模型的发展应该变成能力帮手。
今天我们看到生成视频的能力、写作的能力，包括对事情分析判断的审查能力都是大语言模型的优势，未来我们的工作不仅是依赖效率，而且会依赖能力。我们可以用能力中等的人获得更高的支持，通过机器人辅助。但是这样就够了吗？最难得的是生活的依赖。
前一段时间我的孩子在做科创实验，用人工智能识别图像，告诉交通情况，作为盲人的第三只眼睛。
大语言模型会变成大家的工作依赖、能力依赖和生活依赖，打破对话的瓶颈以后可以变成老人的依赖，解决老人的忧虑问题，包括情感输出和精神焦虑，未来会变成生活的助手。
上一个依赖是什么？手机花了几年成为各位的依赖？现在多少人能够离得开手机？今天大语言模型Copilot可能会是大家的下一个依赖。 数字化转型网（www.szhzxw.cn）
翻译：
What are the real changes that the big model will bring To businesses in B?
This year has been an explosive year for big language models, with ChatGPT sweeping the world and attracting hundreds of millions of users in less than two months.
With the continued popularity of ChatGPT, the domestic market is also enthusiastic, one after another released or announced the development of large language model products, large language model track competition is extremely fierce. While looking forward to the future development trend of large language models, what industrial changes will be brought about by the outbreak of AI large language models?
Sun Bin, president and COO of Takema Intelligence, delivered a keynote speech on the topic of “Industrial Transformation of AI big Language Model Outbreak” at the Tencent Hi Tech Day and 2023 Digital Opening Conference.
Large language model from the text to the multi-modal progress is very fast, from the C-end to the B-end of the large language model has brought breakthrough changes to the industry, while subtly changing the user’s lifestyle, the future will even become people’s life dependence. Sun Bin gives his own thoughts on the large language model. 数字化转型网（www.szhzxw.cn）
Sun Bin believes that the big language model breaks through the bottleneck of human-machine dialogue, the change of the C end is the fastest, all the applications in the future track may be changed by the mode of dialogue, and thousands of applications will change the application paradigm because of the way of the big language model.
Talking about the change of the B end, Sun Bin believes that today’s large language model has become a productivity that can be used, and the improvement of productivity will also bring great changes to the industry.
The explosion of large language models in 2023 is the prelude to the digital workforce, which is the perfect combination of cost and efficiency.
In terms of the use of AI large models in enterprises, Sun Bin said that today’s rapid iteration of large language models, enterprises can choose external purchase or find partners in the industry joint construction, non-industry head enterprises, do not recommend to do self-research and development.
Talking about the future development of large language models, Sun Bin believes that large language models will become ability helpers, and become everyone’s work dependence, ability dependence and life dependence. , duration 03:16 数字化转型网（www.szhzxw.cn）
Sun Bin speech full text:
First of all, I’d like to introduce myself. My name is Sun Bin. Takema Intelligent has been focusing on the NLP track for eight years, and our team has been focusing on the landing of the NLP industry.
In the past, we trained the model ourselves and were able to find the data scene landing in the customer scene, and finally achieved the effect of the model landing industry, covering a lot of industries.
The previous several guests have done a lot of analysis from the chip and data, and the following I share with you is standing in the natural semantic track, standing in the 2023 outbreak of large language models, what our perspective is.
You should all be familiar with these logos, and 2023 is a magical year. A year and two weeks ago, OpenAI released ChatGPT to detonate the entire technology circle, in addition to AI, cross-industry are very concerned about this track, is it magic? OpenAI’s board of directors is also very magical, a few days ago, someone saw that the OpenAI board of directors is more aware of their own board of directors, because it is a management change.
It can be seen that not only a year later today, Google released the Gemini large language model, why are all large model companies in China catching up? It’s like there’s a fairy called GPT4 at the top of the Bright Roof, and everyone’s besieging the bright roof, isn’t it?
Today’s big language model circuit is very competitive, and the results are amazing, but what is the impact on the industry and industry? 数字化转型网（www.szhzxw.cn）
In fact, what we see today is that there is a big trend in large language models from text, speech, and large modes, ChatGPT started as text, and now many people are obsessed with GPT4 sound, but the visual recognition, image recognition and understanding brought by Gemini is a deeper thing. Large language models are no longer about text understanding, and multimodality is progressing very fast.
There’s actually another thing in the tech world that we’re all talking about, which is the closed source versus open source competition. Do you remember the old rivalry between Apple and Microsoft over PC open source and closed source, Android and IOS? There is also competition for large language models today, including open source models like LlAMA, but there are still many closed-source models that are focused on maintaining their health and security.
Will you choose the open source model or the closed source model for your business? I think this competition is just the beginning, people will choose different tracks to improve.
So what has changed in each industry? I believe that the C end changes the fastest, the large language model has come up, the efficiency is very high, the effect is very good, and it can be generated collaboratively.
Last year, 100 million people quickly registered in OpenAI to use GPT, half of them used to write homework, but today’s intelligence is amazing, with ethical considerations, the ability to generate, more efficient than common authors, writing quality is better, there are many templates and knowledge can help you, the way of interaction has completely changed, people start to get used to dialogue, Large language models are now often asked challenging questions.
Let’s not judge the moral judgment, but the intelligent and interactive way has changed, so everyone sees the surprise of the large language model, and practitioners see the breakthrough.
Today, we see large language models breaking through the bottleneck of human-machine dialogue, can understand conversations like humans, machines can understand human intentions and judgments, and all application patterns will change. 数字化转型网（www.szhzxw.cn）
In the past, every application change is a general trend, and the software has also changed greatly from the PC side to the mobile Internet side, and the Internet application has been rewritten to the mobile APP. We very clearly predict that in the next time, all these applications in the track may be changed by the mode of dialogue.
You will use mobile phones to order food, you will also use mobile phones to book air tickets, the future may be carried out in the way of dialogue, we judge that the C-end of the industrial change is that thousands of applications will change the application paradigm because of the way of large language models.
After talking about end C and end B, To B has always been to embrace change. Do you remember what automation did to end B? The way automation in the electrical age has greatly improved the efficiency of the industry is a huge change. We all know that all manufacturing systems can be managed by means of data flow and financial flow. 数字化转型网（www.szhzxw.cn）
Now comes the age of intelligence. And what is the most popular in the last two years? It’s localization. Many listed companies are the leaders of domestic alternatives, but the outbreak of today’s large language model is the middle of the domestic intelligent circuit, why not fire?
From the perspective of practitioners, we believe that the reason for fire is very simple. In the past, all scientific and technological products, even calculators, even computers, even IT equipment are tool products, which may be called production tools and efficiency tools, but the big language model has solved the three elements of productivity.
Productivity is workers + production tools + labor objects, if the large language model has enough intelligence, coupled with the support of private enterprises will become qualified ability, drive equipment, drive process, complete a productivity cycle. The real change in the To B side of today’s big language model is the productivity that can be used, and the change to the industry when productivity increases is huge.
These have been experienced in the past, both inside and outside the enterprise, through the way of robot dialogue to solve various positions and scenarios. In other words, these jobs can all be done by digital employees. 数字化转型网（www.szhzxw.cn）
Today’s outbreak of the large language model is actually the beginning of the digital employees, the advent of the technology inflection point, all practitioners will pay attention to cost, will pay attention to efficiency, but the digital employees brought by the large language model are often a very perfect combination of cost and efficiency, that is, the prelude of digital employees will begin to open in 2023.
After talking about industries and industries, let’s take a look at what changes will happen to enterprises. Each technological change will be an arms race for the use of technology, but also a new consumption.
Because every technological revolution, even for the sake of enterprise efficiency, in order to manufacture industry-leading model projects, in fact, Party A will embrace these technological changes. It was true in the age of automation, it was true in the age of digitization, and it is still true today in the age of intelligence.
When using AI today, three models are advocated: large language models include self-development, external purchase, and joint construction. Now IT seems that the doubling speed of the large language model is iterated every two weeks, so if it is not the industry head enterprise, it is not an enterprise that has enough IT development strength, it is not recommended that you do self-research and development, and will not take the model to train yourself. Over the past year, countless teams have been training models, and the results are not actually available. 数字化转型网（www.szhzxw.cn）
I recommend the latter two: If you choose the best team to buy a mature product today, it may not be this year, it may not be early next year, it may be the end of 2024.
Through the precipitation of the model by the entire industry team for more than a year, there will be the best engineering method for Party A and the industry to use. Looking for the best partners according to the industry, being able to jointly build in the industry is also a path.
I want to give you another reminder, standing on the real customer side, in fact, this perspective. Using Sequoia’s industry report, you can see 200-300 CIOs and CEO feedback.
Party A embraces technological change at the time of technological change, but still regards input-output as the first principle. Today, even if our industry consumes, it also needs to consume output, so it must calm down in the time of the outbreak of science and technology. Can the large language model calculate the ROI?
If it doesn’t come back, it won’t be a good business, in other words, it won’t be a long-term business. Both parties and industry practitioners should remember that a business that cannot be turned into a large model is not a good business.
Every time technology changes, Party B must use it, because there will be game breakers to re-break the original industrial order, if you do not use the new technology will fall behind, the market share will be lost. 数字化转型网（www.szhzxw.cn）
All Party B is anxious about competition, and the enterprises in the rear want to take the opportunity to get market share. But this kind of challenge is very painful, at the beginning, everyone is trying, and it is the first input before the output. When the big language model comes up, all Party B is also painful, but must have new technology, with the ability of big language model technology.
What are the trends in model-making companies? Why did you say it was over as soon as it started? At the beginning, the trend of model companies has been identified, and it can be seen that today’s large language model companies are growing rapidly, but practitioners of large language models may only have three growth paths: do a good job of the basic model of large language models, and achieve industry leadership, but this is very challenging, because the result of universal large language models is convergence, and convergence will lead to low switching costs.
In the past, data center switching IT needs to be relocated, and the cloud also needs to be migrated, but in the large language model switching may only change an address, and the number of large language models in the future is certainly limited. 数字化转型网（www.szhzxw.cn）
Whether it is 100 models or thousands of models, the final real use of the head enterprise may be so few. What about other businesses? May be in the tool category, product category development. There are many twenty or thirty enterprises in Europe and the United States that use the large language model to do very well, and I believe that such enterprises in China will also spring up like mushrooms.
It is inevitable to do a good job of products, but to prevent the ability of large language models to take a step forward, the front of the enterprise is all covered and dead. There will be such a risk at the C end, and the B end will definitely focus on the industry model? The industry model is the solution, and because of the need to solve customer solutions, many businesses will focus on this track.
The first step to landing the Copilot phenomenon must be everyone’s office, Tencent technology has shared a lot of cutting-edge Copilot phenomenon, but there is another big trend, today can be an office assistant, but a large language model can be a good training teacher.
As you can see today, large language models can not only be used as writing generation, but also role play, not only as AIDS, but also as teachers, which can change many traditional habits that require personnel. 数字化转型网（www.szhzxw.cn）
At the moment, everyone sees Copilot as a productivity helper that can make a big difference in the workplace, but is that all? The development of large language models should become a capability helper.
Today we see that the ability to generate video, the ability to write, including the ability to review the analysis of things are the advantages of large language models, and in the future our work will depend not only on efficiency, but also on ability. We can get higher support with moderately capable people, assisted by robots. But is that enough? The most rare is the dependence of life.
Some time ago, my children were doing scientific innovation experiments, using artificial intelligence to recognize images, tell traffic conditions, and act as the third eye of the blind.
The large language model will become our work dependence, ability dependence and life dependence, after breaking the bottleneck of dialogue, it can become the dependence of the elderly, solve the worries of the elderly, including emotional output and mental anxiety, and become the assistant of life in the future. 数字化转型网（www.szhzxw.cn）
What was the last dependency? How many years did it take for mobile phones to become your dependence? How many people can turn on their cell phones these days? Today’s big language model Copilot may be the next thing you rely on.
本文由数字化转型网（www.szhzxw.cn）转载而成，来源于腾讯科技；编辑/翻译：数字化转型网宁檬树。
免责声明: 本网站(http://www.szhzxw.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。
本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。
