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：做具身，不能一上来什么都干，也不能什么都不干
作者丨齐铖湧
编辑丨岑 峰
董豪是具身行业里一个十足的非典型选手。
大多数人对他的印象，停在 29 岁便是北大博导的年轻学者身份，以及最近任职启元机器人首席科学家、带队攻坚家用具身智能的产业标签。但很少有人知道，他读博之前曾有过一段三年创业最终折戟的经历，住过工厂的八人间宿舍；更少有人知道，智元刚创立时，他就是最早加入、陪着初创团队闯过无数关口的核心成员之一。
当行业还在热议大模型、通用机器人的宏大叙事时，他盯着的是另一个一点都不性感的真问题：靠上万台真机做后训练，成本高到难以落地，这条路要怎么破？
他最近给出的答案是三个关键词：HIL-UMI、Scaling Law 和量产。
而这份答案的底层逻辑，早就写进了他一路 “压缩着走” 的人生里。
硕士毕业，他婉拒了导师的直接读博的建议，回国创办公司做可穿戴设备与脑机接口，一头扎进工厂，住过八人间宿舍。三年后亲眼见证市场被大厂快速挤压，他没有恋战，转身回到帝国理工攻读博士，四年后顺利毕业，期间主导开发的开源框架，还斩获了 ACM MM 最佳开源软件奖。 毕业后，他加入北京大学入职不久就撞上疫情，对着全程不开摄像头的学生，线上讲了一整年的机器人课。2023 年智元机器人创立，他作为核心成员加入，成立北大智元联合实验室；如今他转身出任启元首席科学家，一头扎进了难度最高的 ToC 具身智能赛道。
在行业里，董豪向来朋友众多，创业经历、顶尖学术、头部产业的履历，也带着一种罕见的跨界流动感。一个完整亲历过硬科技公司从起盘到落幕全过程的人，又转身沉潜到学术浓度极高的北京大学，如今再站到产业最前沿，他会怎么看当下的具身智能行业？ 带着这个问题，我们在北大校园里见到了董豪。
以下是访谈内容，《AI 科技评论》在不改变原意的基础上做了编辑。
01
一个"压缩着走"的人
▎AI科技评论：你英本英硕四年读完，29岁北大博导。这个节奏在学术圈压缩了三五年。你自己意识到吗？
董豪：没刻意算过。英本英硕那个体系本来就快，四年硕士毕业是正常操作。
▎AI科技评论：但不到30岁入职北大不正常，你拿到教职第一年还撞上疫情。
董豪：第一次上课是线上，学生不开摄像头，不吱声，我全程自己讲，有时候问问有没有问题，没人回。搞机器人的全线上，机器人摸不到。最惨的是突然解封能用一下，突然又封了，断断续续。很多人没熬过去，我运气还不错。
▎AI科技评论：只归给运气？
董豪：还有就是我做的事聚焦。机器人这个方向，学术和产业都关注，两头都不浪费。
▎AI科技评论：“两头都不浪费”，这个说法很像你。你好像天生不喜欢单维度做事。
董豪：可能吧。比如你提到我以前做的AI框架，当时本来是实验室内部工具，放着可惜就开源了，结果用户越来越多，还拿了奖。一个工作，长出好几个东西，效率高。
02
一位90后北大副教授的创业往事
▎AI科技评论：2012 年硕士毕业，导师郭毅可（中国工程院外籍院士、现港科大首席副校长）劝你读博，你没听，回国创业，具体是什么情况。
董豪：研究生跟着郭老师做科研，郭老师让直接读博，结果我决定回国创业，他支持我，还给了钱，后来 2015 年回去读博，他也没说什么，给了我很大自由，让我自己选课题。
▎AI科技评论：创业做可穿戴设备？
董豪：是的，当时势头很猛。创业的时候住工厂，条件很差，都是八人间，我待遇好一点，一个人住一个八人间，八个床，想睡哪个睡哪个。
▎AI科技评论：那三年应该挺苦的。
董豪：苦倒不至于，就是看得见一个创业团队完整的命。可穿戴设备：便携脑机接口、智能手环，我们先做的，前者是面向未来、后者面向当下。小米华为后来入局可穿戴，但人家一做起来非常快。眼睁睁看着市场没了。
▎AI科技评论：硬科技创业最难的是什么？
董豪：量产。样机半年就能做出来，但你要造一百个产品，成本要控制住，良品率要高，还得耐用，这个要花一两年。一个小改动，改模具、调产线、重新验证，周期按半年算。三年时间，其实没几轮试错的机会。（雷峰网(公众号：雷峰网)）
▎AI科技评论：所以你现在看智元和启元把量产做起来，是什么感受？
董豪：量产很难啃，外行看量产是个数字，做过硬件的人看量产，看到的是几千个细节配合，测试项厚得像一本书。我以前做那么小的东西都觉得难，现在我们要把整机机器人造出来，一个个标准打过去，这里面非常难。
▎AI科技评论：刚提到智能穿戴市场被小米华为吃掉，被大厂碾过去是什么感觉？
董豪：就是你知道再努力也没用。人家供应链、渠道、品牌全都有。我后来总结，前沿技术和市场之间有时间间隔，这是我花三年买的认知。
▎AI科技评论：很多创业者十年才走完的周期，你三年走完了。
董豪：你这么一说……好像是。
▎AI科技评论：那叫止损。
董豪：但其实我不后悔那三年，虽然做研究和下场做，是两回事，但是必须下场做，不做不行。当年造车，融了钱不造的全死了；造了，死掉四分之三。你还是要做。
▎AI科技评论：所以你认为现在很多具身创业公司不花钱不量产，用等待策略是不对的？
董豪：怎么说呢... 说白了，你就算把钱存着也存不了多少，反正你肯定要搏一搏。你的钱多，比得过巨头么，造车不就是例子吗？造了还好一点，还有一丝的胜率。
▎AI科技评论：那你认为具身公司该怎么花钱呢？
董豪：具身智能有个好处，因为它的市场空间太大了，可以突破的点就很多。比如说很多 toB 的场景，它不一定需要一个模型可以做很多不同的任务，所以做这种，短期收益是比较好获得的，比如做工厂里可复制化的工位，比如说分拣。至于怎么花钱，就要看各家主攻的方向了。
▎AI科技评论：但这个思路似乎和启元的在做的不太一致？
董豪：启元的想法，是要一下子能往终局做。
▎AI科技评论：什么是终局？
董豪：面向家庭、消费场景，是必须实现通用作业智能的，总之是要一个通才具身智能大模型，而其他场景不一定需要。
03
从智元到启元：最难的那个选项
▎AI科技评论：很多人告诉我们，你是一个社交小天才，大家都很喜欢你。
董豪：可能因为我一直觉得行业要一起努力，都是朋友，这个行业能做的很多，我还是喜欢一群人共同往前拱的过程。
▎AI科技评论：2023 年智元刚创立你就加入了，志辉和邓总（智元邓泰华）叫的你。现在智元上千人，量产交付跑在头部，为什么要出来做启元？
董豪：分工不一样。智元以 toB 场景落地为主，但启元 toC 不一样，面向家庭的机器人，得真的有大脑，得通用导航，得跟人互动，也要有通用作业智能，最难也最泛化。做具身前沿研究，toC 是最合适的战场。
▎AI科技评论：举个例子？
董豪：比如同样造发动机，民航的和军用的，要求肯定不一样。工业具身不一定需要一个模型做一百件事情、但 toC 真需要。
▎AI科技评论：从智元到启元，很多都得自己做，这不是把难度又往上加了一档？
董豪：该踩的坑都得踩。总之，从之前创业我明白了，等待策略是不对的，你等别人烧钱烧出方向，你在旁边苟着，苟住了胜率更大？不行的。你就算等到了，那又不是你的东西。
▎AI科技评论：从之前的学术界到现在跨界产业，感受是什么？
董豪：节奏和风格都不同，特别是工程化方面。举个例子，我刚进北大，杨耀东（北大博导，灵初智能首席科学家）拉我和王鹤（北大博导，银河通用CTO）做了一个北京科协的项目，关于灵巧操作的。
虽然是立项很顶级，但经费也不多，我记得那时候要买手（灵巧手），只能买Shadow Hand （英国灵巧手公司，一只手10万美元），价格贵得离谱，我和杨老师只能一人买左手、一人买右手，拼着做科研。现在大环境不一样了。
纯学术研究获得的经费少，支撑不了现在大模型的研究，现在不仅仅需要学术界的前沿研究，还需要大规模算力、以及很好的工程infra，学术也需要和产业结合。
▎AI科技评论：大环境会对年轻人有什么影响？你的学生王远非、张继耀现在也在做具身创业，你怎么看？
董豪：他们这一代比我那时候条件好太多，技术视野、动手速度、融资环境都不是我们当年能比的。具身智能空间足够大，年轻人闯是好事，我也很看好他们的创业方向。
▎AI科技评论：你会一味鼓励创业吗？
董豪：不一定，时机很重要。有的学生，我会劝他们先借助大平台把技术搞扎实，积累够了再出来。供应链、生态、技术团队这些东西要时间，花钱买不到。快不是问题，没积累的快才是问题。
▎AI科技评论：启元给自己多长时间？
董豪：长线按三年看。作业智能先做失败了也没风险的场景，比如衣物类的，每天都要做，叠没叠好不会出事，安全边界高，失败了也能重试。我们已经做了全自主洗衣流程：机器人自主规划路径、开合洗衣机门，抓取失败会主动重试，整套流程无人介入。你不能一上来什么都干，也不能什么都不干在那等。先优先打穿一些能力、再横向拓展。
04
技术判断：胜负手在后训练
▎AI科技评论：聊技术。你有个很反共识的判断："预训练和后训练不是比例关系，是阶段性更替。"大部分人还在讨论预训练数据配多少后训练数据，你直接把问题否了？
董豪：不是否了，是阶段不一样。预训练解决起步的世界观构建，路线其实清晰，采数据，训练进去，大力出奇迹，把成功率从 0 拉到有一定成功率，但提升有上限。大规模后训练会比现在说的预训练要大，甚至完全覆盖。
▎AI科技评论：上限在哪？
董豪：模仿学习只能冷启动。训练样本全是正确轨迹，机器人犯了错不知道怎么调整回来。真实世界不按你的演示数据分布。
▎AI科技评论：所以要上强化学习。
董豪：对，当年 AlphaGo 就是这个路径。棋谱模仿学习起步，但真正封神是靠自我博弈的大规模强化学习。后来ChatGPT 也一样，核心还是 RLHF 来做强化。今天的机器人一样，泛化和极致成功率，胜负手全在可规模化的后训练。上半场拼谁先拿模仿学习把机器人点亮，下半场拼谁能用最低边际成本、最快构建机器人的纠错网络。
▎AI科技评论：但后训练的规模化卡在真机上。
董豪：这就是物理约束。你要几万台真机在几万个真实房间里同时做后训练？造一万台出来成本都不得了。智元成立 3 年量产 2 万台，已经是很强了。
▎AI科技评论：所以启元做 HIL-UMI，无本体数据采集的后训练。
董豪：对，这个是我们最近比较火的一个工作。人不需要操作机器人本体，穿戴式、手持式相机记录操作过程，直接转成可训练轨迹。同等预算获得的模型能力提升更高，并且可以支持后训练、有针对性地优化模型缺陷，让成功率逼近100%。
▎AI科技评论：有了万级数采集，后训练就能规模化？
董豪：一半。另一半是数据效率。比如我们现在可以把一条真机轨迹，用世界模型/生成AI，能生成 50 条物体摆放、空间位置都不同的等效样本。真机数据价值放大 50 倍。人类示教、真机锚点、生成式扩增，拼起来才是"可规模化的后训练"。
05
横向 Scaling Law：性格的理论化
▎AI科技评论：你提到的"横向 Scaling Law"，数据量一个轴，任务量一个轴。为什么加任务量这根轴？
董豪：很多公司现在是低效增长曲线：做一个场景，过拟合，交付；下个场景，从零再来。数据和任务只是线性关系，边际收益递减。横向扩展之后，新任务的初始成功率会越来越高，达到高成功率需要的样本量持续下降。越学越省。
▎AI科技评论：我问个可能过度解读的问题。你读书压缩着读，创业三年，现在北大和启元两头同时做。“不在一根轴上死磕、两个维度同时加速”，这个思路是不是你性格的理论化？
董豪：这个问题（笑）
▎AI科技评论：你回答就行。
董豪：有可能。北大的事和启元的事，对我来说不是两份工作，而是同一件事，一边做技术范式创新、一边做规模化验证，两头相互支持。
06
关于“节奏”
▎AI科技评论：聊点别的。你日常喜欢机车和滑雪，滑雪的人都说，恐惧时身体本能往后躲，越躲越摔，你怎么看？
董豪：控制好节奏就行。不是不恐惧，是控制好恐惧。
▎AI科技评论：压缩时间的人，会不会也容易失控，因为速度快。
董豪：所以无论是科研，还是在公司里，节奏才重要。快不是问题，乱才是问题。什么时候投入多少，研发节奏要把控好。
07
后记
董豪和的AI科技评论的对话，在北大校内的一个传统中式院子，叫做静园五院，是北京大学计算机学院前沿计算研究中心的办公教学地点。
北大前沿计算研究中心由图灵奖获得者John Hopcroft和高文院士创建，目的是培养新一代的计算机领域人才，做出有影响力的科研。中心对论文数量、申请经费没有任何的要求，评审方式与国际接轨。
一届又一届的图灵班学子从这里走出去。
在这里，董豪一直在深耕教育、科研，启元的前沿学术研究，也在这里发生。某种程度上，他的说话方式和他的人生、他的理论是同一种东西：压缩。把别人十年的路径压成五年，把一条数据压出五十条的价值，把一维的行业压出一个新的轴。
“横向Scaling Law”也许不只是一个技术趋势，它更像这个人的自我描述。（雷峰网）
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图鉴自动抓取并翻译，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
Introductory: I can't do anything with my body. I can't do anything with my body.
Author
Edit Peak
Dong Hao is a perfect atypical player in the trade.
Most of them had the impression that at the age of 29, he was a young scholar of the Great North, and that he had recently been a leading scientist in the initial robotics, leading a team to attack the smart industrial label of his family. Few knew, however, that he had had a three-year career before reading the book and had lived in the factory's eight-person dormitories; fewer knew that when Ji-won was first created, he had been one of the first core members to join the team that had broken through countless doors.
When the industry was still discussing the grand narrative of LLM, the universal robot, he was staring at another real problem that was not sexy: How do we break this road?
His recent answer was three keywords: HIL-UMI, Scaling Law and volume production.
And the bottom logic of this answer was written into his life of "compressing walking."
He graduated from a master ' s degree and declined the teacher ' s direct reading proposal, returning to his country to set up a company to make wearable equipment and brain interfaces, placing his head in a factory and living in an eight-person dormitory. Three years later, he witnessed the market being rapidly squeezed by big factories, and he turned back to Imperial Technology for a doctorate, graduated after four years and won the ACM MM Best Open Software Award. Upon graduation, soon after joining Beijing University, he ran into an epidemic and gave a year of robotic lessons online to students who had not opened their cameras. Founded in 2023, he joined as a core member to set up the North Great Jiwon Joint Laboratory; he has now turned to be Chief Scientist of Kaiwon, and has entered one of the most difficult ToC intelligent tracks.
In the industry, Tung Hao has a long history of friends, entrepreneurial experience, top academic, head industry, and a rare sense of cross-border mobility. What would a man who has experienced a hard technology company from the beginning to the end of the scene, and who has sunk to the high academic concentration of Beijing University, now at the forefront of the industry, think about the current smart industry? With that problem in mind, we saw Dong Ho at the North campus.
Below is an interview that was edited by the AI Science Review without changing its intentions.
01
A man who's "crushed."
@AiTech: You've been reading English for four years, and you're 29 years old. The rhythm has been constricted in academic circles for three or five years. Do you realize that?
Dong Hao: Never calculated. The system is fast. Four years of master's graduation is normal.
In the first year of your teaching career, you hit an epidemic.
Dong Hao: First class is online, students don't turn on cameras, they don't squeak, I say it all myself, and sometimes ask if there's a problem and no one returns. You can't touch a robot on the whole line. The worst part is the sudden unsealing of the seal, and then it's closed and it's intermittent. A lot of people didn't make it. I got lucky.
@AiTech: Only luck?
Tung Haugh: And focus on what I did. The robot's direction, academic and industrial concerns, and neither is wasted.
It's like you. You don't seem to like to do one-dimensional things.
Tung Hao: Maybe. For example, you mentioned the AI framework that I used to do, which was a laboratory in-house tool, which, unfortunately, turned out to be an increasing number of users and received awards. One job, several things, efficient.
02
The business history of a 90-year-old associate professor from North.
▎AI Science and Technology Review: In 2012, Master Guo Yikko, a mentor (a foreign student at the Chinese Engineering Institute and Chief Vice-Chancellor of the University of Hong Kong) advised you to read Bo, you did not listen, you went home to start a business, and what was the situation?
Dong Hao: The graduate students followed Mr. Guo in scientific research, and Mr. Guo allowed me to study directly, and I decided to go home to start a business, and he supported me and paid me back, and then he went back to study in 2015, and he didn't say anything, and he gave me a lot of freedom to choose my own subjects.
@AiTech: Business start-up to wearable equipment?
Tung Hao: Yes, it was strong. I live in a factory in bad conditions, eight. I'm better off living in one room, eight beds, one sleep.
@AiTech: Those three years should have been tough.
Dong Hao: It's hard to see the life of an entrepreneurial team. Clothes: brain interfaces, smart bracelets, which we make first, looking to the future and now. Mihua is dressed for later entry, but he does it very quickly. Watching the market go.
▎AiTech Review: What is the hardest part of hard technology entrepreneurship?
Dong Hao: Quantified. The prototype can be made in six months, but it takes a year or two for you to build a hundred products, keep the cost under control, have a good taste, and have a good taste. A minor change, modified moulds, routing, re-validation, six-monthly cycle. In three years, there's not really a lot of chances. (Ray Peak Network (WeChat Official Account:Ray Peak Network))
@AiTech: So what do you think it's like to see Ji-won and Kai-won make a quantum product?
Dong-ho: It's hard to eat, it's hard to watch, it's a number for amateurs, it's hard for people who have made hardware to read it, it's thousands of details, and it's like a book. I used to make things that small, and now we're going to build the whole robot, and we're going to hit it one by one, and it's going to be really hard.
▎Ai Technology Review: How does it feel to say that the market for smart clothing was eaten by little Mihua and run over by big factories?
CHUNG: You know, it's no use trying. People have supply chains, channels, brands. I later concluded that there was a time lag between frontier technology and the market, which was the perception that I had spent three years buying.
▎AiTech Review: Many entrepreneurs have gone through a cycle of 10 years and you've gone three years.
Dong Hao: You say that like that.
It's called stopping damage.
CHUNG: But I don't regret it for the past three years. It's two things to do, but it's one thing to do. When they built the car, they all died, and they built it, and they died three quarters. You're still gonna do it.
@AiTech: So you think it's not right to use a wait-and-see strategy for a lot of start-up companies that don't pay and pay?
Tung Hao: Well... you can't save much of the money, you'll have to fight. You have more money than a giant. It's better. There's a little chance.
How do you think the company should spend it?
Dong Hao: Smarty has one advantage because it has too much market space to break through. For example, many toB scenarios do not necessarily require a model that can do a lot of different tasks, so short-term benefits are better in this case, for example, doing replicable jobs in factories, for example. As for spending, it depends on the direction of the masters.
@AiTech: But the idea seems to be different from what Kai-won is doing?
Dong Hao: The idea of Kai-won is to get to the end.
@AiTech: What is the end?
Tung Hao: For households and consumption, it is necessary to achieve universal operational intelligence, in short, a generalist who is smart LLM, while others are not necessarily required.
03
From Ji-won to Kai-won: The hardest option
▎: A lot of people tell us that you're a social genius and everybody loves you.
CHUNG: Maybe because I've always felt that the industry was working together, that it was friends, that the industry could do a lot, and I still liked the process of moving forward together.
Zhihui and General Deng called you when you were created in 2023. Now there are thousands of people in Ji-won, and delivery runs in the head. Why did you come out to do it?
Dong Hao: The division of labour is different. Ji-won is based on the ToB scene, but the original toC is not the same, but the family-oriented robot has a real brain, has a universal navigation, interacts with people, has a common operational intelligence, is the hardest and most general. Toc is the most appropriate battlefield.
@AiTech: For example?
Dong Hao: For example, the same engine, civil aviation and military, the requirements are certainly different. Industry does not need a model to do 100 things, but toc really does.
@AiTech: From Ji-won to Kai-won, much of it has to be done by itself.
Dong Hao: Every pit to step on. Anyway, I understand that it's not right to wait for someone else to burn the money out. No way. Even if you wait, it's not yours.
The AAI Science and Technology Review: What does it feel from the former academic world to the present transboundary industry?
Dong Hao: The rhythm and style are different, especially in engineering. By way of example, I've just entered the North, and Yang Yoon-dong has been doing a Beijing science project with Wang cranes.
It's top-of-the-art, but it's not very expensive, and I remember having to buy a hand at that time, and having to buy a Shadow Hand. It's different now.
Pure academic research is poorly funded to sustain the current LLM research, which now requires not only front-line academic research, but also large-scale calculus and good engineering infra, and academic integration with industry.
@AiTech: What will the big environment affect young people? Your student, Zhang Ji-Yo, is now doing his own business. What do you think?
Tung Hao: Their generation is far better off than I was then, technology horizons, speed of action, financing environment than we could have been. Smart space is big enough for young people to break in, and I'm looking at their entrepreneurial direction.
@AiTech: Will you encourage start-up?
Dong Hao: Not necessarily. Timing is important. Some of the students, I'm going to persuade them to use the big platform to build up the technology and accumulate it. The supply chain, ecology, and technical teams take time and cannot be bought. It's not a problem. It's a problem not to accumulate.
@AiTech: How long do you think you're going to give yourself?
Dong Hao: The long line is three years. Job intelligence fails without risk, such as clothing, every day, with no folding, no trouble, high security borders, and retry if it fails. We've done a fully autonomous laundry process: robots plan their own paths, open the washing doors, capture failure will take the initiative of retry, and the whole process is uninvolved. You can't come up and do anything. You can't wait. Priority is given to cutting through some capacities and later later lateral expansion.
04
Technical judgement: training behind the winner and the loser
I'm talking about technology. You have a counter-consensus judgment: "Pre- and post-training are not proportional relationships, but stage-replacement." Most people are still discussing how much pre- and post-training data are available, and you're going to question it?
Dong Hao: No, it's different. Pre-training to solve the initial worldview, with clear routes, data collection, training, miracles, pulling success rate from 0 to a certain success rate, with ceilings. Large-scale post-training will be larger than, or even more fully covered, what is now said about pre-training.
@AiTech: What is the ceiling?
Dong Hao: Imitation can only be started in cold. Trained sample is the right track. The robot made a mistake and didn't know how to adjust it. The real world is not distributed according to your demo data.
▎AiTechnology Review: So we need to step up on learning.
Dong Hao: Yeah, that's the way Alpha Go was. Chess spectrum mimics the beginning of learning, but the true seal of God is a massive and intensive learning of self-playing. And then ChatGPT was the same, and the core was RLHF to strengthen it. Today's robots, like success rate, are trained in scale. The first half of the game will light up the robot with an imitation, the second half with the least marginal cost, and the fastest way to build the bug network of the robot.
@AiTech: But the scale of the training is stuck in real time.
Dong Hao: That's physical constraint. You want tens of thousands of real machines to train in tens of thousands of real rooms? I can't afford to build 10,000. Ji-won was set up to produce 20,000 stations per year, which is already strong.
▎AiTechnology Review: So start HIL-UMI, post-training in hard-core data collection.
Tung Hao: Yeah, this is one of our hot jobs lately. A person does not need to operate a robotic body, wear a uniform, handheld camera to record the operation, and to be transferred directly to a training track. The model capabilities acquired by the same budget are higher and support post-training, targeted optimization of model deficiencies, bringing success rate close to 100 per cent.
@AiTech: With a 10,000-degree collection, training can be scaled up?
Dong Hao: Half. The other half is data efficiency. For example, we can now use a real track, using a world model/generation of AI, to generate 50 objects in different space positions and equivalents of sample. The value of real data is increased 50 times. Human teaching, real anchors, generation build-up is "scalable post-training".
05
Scaling Law: Theory of character
"Scaling Law," you mentioned, data one axis, mission one axis. Why do you add a mission to this axis?
Dong Hao: Many companies now have an inefficient growth curve: make a scene, match and deliver; and the next scene, from zero. Data and tasks are linear and marginal gains are diminishing. After the horizontal expansion, the initial success rate of the new mission will grow higher and the quantity of sample required to meet the high success rate will continue to decline. Over and over.
I asked a question that might be overreaded. You read hard, you start three years, and now you're doing both North and First. "Don't die on an axle and accelerate at the same time at both dimensions," is that a theory of your character?
Dong Hao: This question.
@AiTech: Just answer.
Dong Hao: It's possible. The North thing and the Enlightenment thing are not two jobs for me, but the same thing. They support each other in terms of technological paradigm innovation and scale certification.
06
About the rhythm
@AiTech: Talk about something else. What do you think about when you like bikes and skis and skiers who say that when you're afraid, your body can hide and fall?
Tung Hao: Just control the rhythm. It's not fear, it's fear control.
▎AiTechnology Review: Will people who compress time be easily out of control because of speed?
Dong Hao: So it's the rhythm that matters whether it's scientific or corporate. Hurry up, it's not the problem. When you invest, the R & D rhythm will be controlled.
07
Rear
The dialogue between Dong Hao and the AI Science and Technology Review, a traditional Chinese courtyard at the North University, called the Shizuku College, is the premises of the Centre for Advanced Computing Studies of the University of Beijing.
The Northern Frontier Centre for Computing Research was created by the Turing Prize laureate John Hopcroft and the Maester of the University of Highland, with the aim of developing a new generation of computer talent for influential research. The Centre does not have any requirements for the number of papers and applications for funding, and the manner of evaluation is consistent with international standards.
The Turing class went out from here.
Here, Tung Hao has been working on deep-farming education, scientific research, and forward-looking academic research. In a way, he speaks the same way as his life, his theory: compression. Putting a 10-year path down for five years, crushing one data out of 50 values and pushing a one-dimensional industry out of a new axis.
"Scaling Law" across the horizon may not be just a technological trend; it's more like the person's self-presentation. (Ray Peak Network)
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
For your account security, please verify the mailbox.
Check your mailbox.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.