您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：在商汤没做成的技术，登上了华为发布会。
在商汤没做成的技术，登上了华为发布会。
作者：李秋悦 编辑：吴彤
2022年6月，刘睿港中文博士毕业，赶上的却是一个“坏时候”。
深度学习的高光已持续近十年，新突破越来越难；疫情之下，经济低迷，同届不少同学选择延期，再守一年实验室。
刘睿没有等，他直接去了商汤。
两个月后，Stable Diffusion发布；又过了三个月，ChatGPT引爆全球。刘睿毕业时以为正在退潮的世界，忽然换了一片海——商汤赖以生存的生意，为手机厂商定制算法、出售授权，开始松动；原本需要一个团队单独研发的算法，正被一个个基础模型收编。
然后是2023年12月，汤晓鸥去世。
“定海神针”突然离开，公司内部人心惶惶。刘睿也第一次认真想：汤老师不在了，商汤怎么办，自己怎么办。
回头看，这个“坏时候”里埋着他此前人生的所有伏笔。2014年考入电子科技大学，他最初学光电子，很快觉得天地太窄；转入英才实验学院后，吴恩达的CS229让他迷上机器学习，李飞飞的CS231n把他领进深度学习和计算机视觉。2018年，他进入港中文MMLab，在王晓刚、李鸿升门下研究生成式模型——那正是几年后掀起巨浪的方向。从传统机器学习到深度学习，他几乎踩着AI最热闹的鼓点走完了求学路，却在毕业这年，听见鼓点停了。
2024年，刘睿加入华为。这一次，他想换一种做技术的方式：不再把算法卖给客户，而是直接把它交到用户手里。
此前在商汤未能立项的对话式修图，被他推上了2025年华为Mate 80的发布会；年底，小艺大模型应用实验室成立，刘睿出任首任主任，研究范围涵盖大语言模型、多模态智能、Agent与具身智能。做研究与做产品，在他手里第一次合流。
也正是在这个时候，2022年那个“坏时候”显出了另一副面孔：如果他像同学一样延期一年、守在实验室里等风停，等来的不会是避风港，只会是又一场追赶。所谓坏时机，不过是旧浪退去、新浪未至的间隙——有人站在岸边等，有人先下了水。
现在，刘睿关心的不再是追不追得上技术，而是另一个问题：模型能做的，和用户要的，之间还差什么。
以下是雷峰网左林右狸频道与刘睿（华为小艺大型模型应用实验室创始主任）的对话，经编辑。
01 商汤往事，大模型颠覆算法生意
左林右狸频道：2018年，你在本科大四下学期去商汤科技研究院做实习生，是什么契机？
左林右狸频道：当时面试你的人是闫俊杰。
左林右狸频道：在商汤科技研究院实习那小半年的时间，你在做什么工作？
左林右狸频道：2022年6月，你从港中文MMLab博士毕业，去商汤工作了两年？
左林右狸频道：那两年你在干嘛？
左林右狸频道：2022年底，Stable Diffusion和ChatGPT相继出现，商汤内部什么反应？
左林右狸频道：2022年底，你在商汤内部做了两个重要的talk。
左林右狸频道：到了2023年，你的工作有什么变化？
左林右狸频道：但大模型时代，还可以卖API、卖Token。
左林右狸频道：2024年，你为什么选择离开商汤？
左林右狸频道：还有其他原因吗？
02 下一代智能助手要能主动服务
左林右狸频道：2024年8月你从商汤出来后，为什么选择加入华为？
左林右狸频道：你前面提到在商汤时，有一些自己觉得很好的想法，但是作为乙方实现不了，具体指什么？可以举一个例子吗？
左林右狸频道：加入华为之后，你是怎么一步步建立小艺大型模型应用实验室的？
左林右狸频道：今年上半年，实验室有哪些代表性的研究？
左林右狸频道：Aura要解决什么问题？相比过去的多模态模型，最大的不同是什么？
左林右狸频道：Aura是怎么同时处理视频、语音和文字，实现这种持续交互的？
左林右狸频道：如果想让助手持续地看、听、说，现在最大的瓶颈是什么？
左林右狸频道：另一项工作Boogu，是一个低成本训练的图像生成基础模型，它有什么特别之处？
左林右狸频道：参数更小、训练成本更低，具体是怎么做到的？
左林右狸频道：Math Canvas又在解决什么问题？
左林右狸频道：你们还有一项工作叫Omni-Interact，它主要研究什么？
左林右狸频道：前面这些研究主要涉及多模态。你们在大语言模型和智能体方面有哪些工作？
左林右狸频道：实验室也在探索世界模型和具身智能吗？
03 基础模型重构助手价值
左林右狸频道：你们研究图像生成、视频生成、多模态理解，也研究智能体。这些不同的技术，最终要把小艺变成一个什么样的助手？
左林右狸频道：助手能干活很好理解，但为什么一个终端智能助手，还需要提供情感陪伴？
左林右狸频道：主动服务也意味着可能打扰用户，假如我正在看一幅画，助手怎么判断什么时候应该开口介绍，什么时候保持安静？
左林右狸频道：现阶段，小艺主要通过什么方式理解用户？
左林右狸频道：如果助手想长期理解用户，哪些信息需要一直记住，哪些只在一次对话中有效？
左林右狸频道：这些信息都很私人，小艺如何保护用户隐私安全？
左林右狸频道：小艺和其他助手相比，核心优势是什么？
左林右狸频道：刚刚谈到小艺的智能体开放平台。对于小艺这样的平台、基础模型和第三方应用，这三者各自在创造什么价值？
左林右狸频道：你之前还有一个判断，基础模型会不断吸收应用原本提供的能力。在实际工作中，你观察到了什么？
左林右狸频道：除了角色扮演，还有哪些能力正在被基础模型吸收？
左林右狸频道：基础模型的能力不断扩张，今天很有价值的应用功能，可能下一代模型就直接具备了。小艺怎么判断什么值得做，避免投入很快被覆盖？
左林右狸频道：你刚刚谈了很多智能体能够替用户完成的事情。但有没有一些事情，即使它能做，用户也不一定愿意交给它？
写在最后
（欢迎添加作者李秋悦的微信lqy4605，多多交流～）雷峰网(公众号：雷峰网) 雷峰网
我们欢迎所有能为左林右狸提供新观点新角度新信息的群友，加入我们的实名制社群，与左林右狸一路同行。管理员微信号：leiphonelinli，备注公司-职位-称呼通过更快。
相关视频
往期推荐
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: The technology that didn't make it to the publish conference.
The technology that didn't work out in the soup was put on the publish conference.
By Li Choo Hsiu, ed.
In June 2022, Dr. Liu Wei Hong Chinese graduated and caught up with a “bad time”.
After almost a decade of in-depth study, new breakthroughs have become increasingly difficult, and under the epidemic, the economy has been depressed and many of the students of the same class have chosen to stay in the laboratory for another year.
Liu Wei didn't wait, he went straight to merchant soup.
Two months later, Stable Diffusionpublish; three months later, ChatGPT detonated the world. Liu had graduated thinking that he was retreating and suddenly changed a sea-based business — the business of merchants and twigs to survive — and began to relax by agreeing on algorithms for mobile phone factories and authorizing them to sell; an algorithm that needed to be developed by a team alone was being compiled in a basic model.
Then, in December 2023, Tong Xiaogu died.
The "Standing the Sea" suddenly left and there was panic within the company. Liu Wei also thought for the first time: Mr. Tong is gone, what to do with the soup, and what to do with it.
Turning back, this “bad time” covered all the pens in his previous life. In 2014, he was admitted to the University of Electronics, where he first studied optical electrons and soon felt too narrow; when he was transferred to the Institute of Experiments of Excellence, the CS229 of U. N. Enda made him become obsessed with machine learning, and the CS231n of Lee Flying led him into deep learning and computer vision. In 2018, he entered Hong Kong's MMLab, where he studied the generation model under Wang Xiaogang and Li Hung's door — in the direction of a wave a few years later. From his traditional machine to his in-depth studies, he almost completed his studies on the most popular drums in AI, but in the year of his graduation he heard the drums stop.
In 2024, Liu Wei joined China. This time, he wanted to change the way he did it: instead of selling the algorithm to the client, he would hand it over to the user.
He had been pushed to the publish conference in 2025, where he had been unable to set up a dialogue drawing in the trade soup; at the end of the year, the Microtech LLM Application Laboratory was set up and Liu was the first director, covering large language models, multimodular intelligence, Agent and physical intelligence. For the first time in his hands, do research and produce.
And that's when the "bad times" of 2022 showed another face: If he waits for a year in the lab like his classmates, he will not wait for a safe haven, but for another chase. The so-called bad timing is nothing but the retreat of the old waves, the absence of a new wave — people standing on the shore, and others coming down first.
Now, it is no longer about catching up with technology, but about another question: what models can do, what users want, is missing.
The following is an edited dialogue with Liu Wei (the founding director of the Small Scale Large Model Applications Laboratory) on the LeFung Leftwood Channel.
LLM subversive algorithm business.
Zolin right-wing channel: What was the opportunity in 2018 for you to work as an intern at the Yutsu Institute of Technology in the fourth semester of this class?
Zolin's right-wing channel: You were interviewed by Joo Jie.
Zolin's right-wing channel: What do you do for the first half of your year at the School of Technology and Commerce?
Zolin's right-wing channel: In June 2022, you graduated from Hong Kong Chinese Dr. MMLab and worked for two years in commerce?
Zolin Right Beaver: What were you doing in those two years?
Zolin Right Beaver channel: At the end of 2022, when Stable Diffusion and ChatGPT appeared, what was the reaction within the business soup?
Zolin's right-wing channel: At the end of 2022, you made two important talk inside the shop.
Zolin Right Beaver: By 2023, how did your work change?
Zolin ' s right-wing channel: But in the LLM era, it can also sell API and Token.
Zolin Right Beaver: In 2024, why did you choose to leave merchant soup?
Zolin Right Beaver: Is there any other reason?
The next generation of smart assistants needs to be proactive.
Zolin right-wing channel: Why did you choose to join China when you came out of merchant soup in August 2024?
Zolin right-wing channel: You mentioned earlier that you had some ideas that you thought were good at trading soup, but as a side of B, you couldn't achieve it, what exactly? Can you give an example?
Zolin right-wing channel: How did you build a small-scale, large-scale model application lab after joining Wigand?
Zolin Right Beaver: What are the representative studies in the laboratory in the first half of this year?
Zolin Right Beaver: What's Aura up to? What are the biggest differences compared to previous multi-modular models?
Zolin Right Beaver channel: How does Aura handle video, voice and text at the same time, and achieve this continuous interaction?
Zolin right-wing channel: What is the biggest bottleneck now if you want to keep an assistant watching, listening and talking?
Zolin Right Beaver channel: Another job, Boogu, is a low-cost, trained image generation base model. What's so special about it?
Zolin right-wing channel: how does it work with smaller parameters and lower training costs?
Zolin right-wing channel: Math Canvas solves what?
Zolin ' s right-wing channel: You have another job called Omni-Interact, which focuses on what?
Zolin right-wing channel: These previous studies are mainly related to polymodels. What do you do with Big Language Models and AI agent?
Zolin Right Beaver Channel: Is the lab also exploring world models and intelligence?
03 Base Model Reconstruct Assistant Value
Zolin right-wing channel: You study image generation, video generation, multimodular understanding, and AI agent. What kind of assistant are these different techniques that end up being small?
Zolin ' s right-wing channel: The assistant works well, but why does a terminal smart assistant need emotional support?
Zolin ' s right-wing channel: Active service also means that users may be disturbed. If I'm looking at a painting, how do my assistants judge when to open an introduction and when to keep quiet?
Zolin ' s right-wing channel: At this stage, in what way do small art primarily understand users?
Zolin Right Beaver channel: If the assistant wants to understand the user in the long term, what information needs to be kept in mind and which is effective in only one conversation?
Zolin ' s right-wing channel: This is very personal information. How can small art protect the privacy of users?
Zolin right-wing channel: What are the core advantages of smallness compared to other assistants?
Zolin Right Beaver channel: The AI agent Open Platform just spoke about small art. What are the respective value of platforms such as small art, basic models and third-party applications?
Zolin's right-wing channel: You had a judgment that the underlying model would continue to apply the capabilities that were provided. What do you see in practice?
Zolin Right Beaver channel: What are the capabilities other than role-playing being absorbed by the underlying model?
Zolin ' s right-wing channel: The capacity of the basic models is expanding, and today ' s valuable applications are likely to be available directly to the next generation. How do you judge what is worth doing so that the input is quickly covered?
Zolin's right-wing channel: You just talked about a lot of things AI agent can do for the users. But is there something that even if it can do, the users may not be willing to give to it?
At the end.
WeChatlqy4605, Qatar, WeChat Official Account: Ray Peak Network
We welcome all our friends who can provide new perspectives and new messages to the right of Zolin to join our real-name community and walk with the right of Zolin. Manager WeChat: leiphonelinli, Annotated Company - Position - Name passed faster.
Related Video
Previous recommendations
Original Ray Peak article, unauthorized reproduction. For further information, see the information below.
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.