Author丨Zhou Lei
Editor丨Cen Feng
When Byte's Seedance 2.0 was released at the beginning of this year, many practitioners were excited and shocked by the "movie-level" film quality. A new creative blueprint was slowly unfolding in front of them - but for SenseTime's Seko team, it was a "confused" day.
Lin Dahua, co-founder and chief scientist of SenseTime, recalled in a recent interview that during that time when he went to meet investors and attend industry conferences, he was almost always asked the same question: What is the difference between your Seko and Seedance? Ever done it?
There was a voice of self-doubt in the team: Should we continue doing this?
Later, the story did not end with Seko "fighting" Seedance single-handedly. Seedance renders 30-second video clips to the top level; while Seko is aimed at creators and ultimately delivers 100 episodes of short dramas - which consists of scripts, storyboards, character settings and image consistency across episodes, which is a complete creative process.
After thinking about this matter clearly, Seko's life is better: According to Lin Dahua in the interview, the number of users of this product has increased 7 to 8 times this year, and the platform has produced short dramas with over 100 million views.
Lin Dahua himself has a frank evaluation of SenseTime's popularity in the large model market: "Maybe it's not the biggest." This company's route has never completely followed the hottest direction in the industry.
But SenseTime has a habit that has persisted for many years: to this day, the company still has 20% of its resources invested in pre-research work without any product exports.
To a certain extent, this is also a kind of "sowing". When the seeds are planted in the ground, there is no need to promise a harvest. The phrase that Tang Xiaoou often talks about - life finds a way, "life finds its own way", may be the logic of this investment.
In a recent interview with the "Technology Talk" column jointly produced by Weibo Technology But there is only one clue throughout: the 20% of resources, and the judgment that supports it - why some roads are worth walking for a long, long time, even if there is no exit yet.
01
Training costs cannot be stopped, Chinese AI competes with hand speed
The money in the large model industry is burning in an unstoppable way.
There used to be a common assumption in the industry: spending a huge amount of money to train a good model, the cost can be amortized in subsequent large-scale services.
Lin Dahua pointed out in the interview that this assumption may not be true - the problem is not that the model is not good enough, but that "if you don't continue to train the next one, others will surpass you." Competition forces companies to retrain one round after another, so training costs cannot stop and are getting higher and higher.
Although the industry narrative in the past year has been almost dominated by "bigger is stronger", the path of volume parameters is being counterattacked. He mentioned two signals:
First, the size selection of the model has become more pragmatic. Recently, various manufacturers have released Flash-level small and medium-sized models: GLM 5.3 Flash, DeepSeek 4.1 Flash, and Google and OpenAI have also launched products with similar specifications.
Lin Dahua specifically mentioned the performance of DeepSeek in the interview: "The performance level of Flash 4.1 is better than that of 4.0 Pro." A model with a scale of several hundred B and well-trained training outperformed the previous generation Pro, indicating that the optimization space in the medium-sized model is far from being exhausted.
Second, the quality of pre-training data has been put back on the table. When a model with hundreds of billions of parameters outperformed an opponent with 2 trillion parameters in actual combat, many people began to look back and check their data pipelines.
Lin Dahua has a vivid explanation for this: "There are a lot of impurities (in the data) and a lot of things fighting with each other, so that a lot of efficiency in the middle will be offset." The result is an expensive and unsatisfactory model. The parameters are piled up, but the effect is not improved.
“I don’t think it’s a healthy state when everyone blindly competes in parameters.” In his view, these two signals mean that the industry is returning to rationality: it no longer competes around a single indicator, but returns to the value of the end user.
A similar situation also occurred during the last AI cycle, and SenseTime himself was the one who experienced it. Every project must start with data collection, special training, tuning, and testing. The entire process must be redone, which is difficult to truly replicate. It’s just that the unit of measurement back then was a project, but today it’s been replaced by a parameter.
Fighting hand speed is another form of involution. The cutting-edge laboratories in the United States define the ceiling of SOTA. Once the new model is released, it is equivalent to pointing the way for practitioners in China. A number of products claiming to use similar solutions often emerge within two or three months in China to seize the ecological niche. The recently popular RSI (recursive self-improvement) is also being built by a bunch of companies at the same time.
"To be honest, competing for hand speed is actually quite tiring." Lin Dahua observed that it is difficult to create long-term value if everyone rushes in when they are optimistic about one direction. If ten companies occupy the same ecological niche at the same time, the dividends of the first comer will be quickly diluted.
So what does long-term competitiveness depend on? His answer is: a road that is long enough and deep enough. This question is the starting point for understanding SenseTime’s 20% resources.
02
The essence of intelligence:
From Anti-Scaling Law to Language and Multimodality
· 20% resources and spatial intelligence
These 20% of resources are clearly invested in work that “does not have any products for export”. Lin Dahua confirmed in the interview that no matter what operating pressure SenseTime faces, the money continues to this day. "At any time, we still retain certain resources."
In his words, the starting point of this type of work is to ask why: when a problem cannot be solved by existing methods, go back to the basics. He does not pursue the use of complex mathematical formulas to explain the principle, but requires that this principle "can effectively solve some problems that are currently difficult to solve by other methods."
One of SenseTime's research directions, spatial intelligence, was selected in this way.
The selection logic is divided into two steps. The first step is to test the strongest models at the time and find out what they don’t do well. The second step is to ask one more layer down, can these shortcomings be made up for by adding some data? If it can be done well by adding data along the original path, then it has application value, but it cannot support principled innovation. Only in directions that cannot be done well no matter how much data is added, there may be real principle problems hidden - and spatial intelligence happens to be the latter.
The origin of this set of judgment standards can be traced back to his inheritance. When Tang Xiaoou teaches students, he almost never asks about technical details. He only looks at one thing: whether he can explain in one or two sentences the fundamental difference between his work and others. "No two leaves in the world are exactly the same." Technically, you can always find differences in the details, but the innovation value of that kind of difference is limited; if you can't say it in one sentence, it's probably not a real innovation.
Lin Dahua pointed out that the "abnormality" of spatial intelligence is that Scaling Law fails here.
In the industry, it is rare to publicly question Scaling Law. The industry consensus in the past few years has almost all strategies bet on this route - the larger the parameters, the more data, and the stronger the capabilities. Proposing the "anti-Scaling Law" in this context at least prompts the industry to re-examine one issue: whether the practice of simply stacking data and parameters has unnoticed blind spots in certain dimensions of intelligence.
In tasks related to spatial intelligence, the trend observed by the SenseTime team is exactly contrary to the Scaling Law - the more data is added, the more unstable the performance of the model becomes.
In the tests at that time, the model performed unsatisfactory on the perspective-taking task. When people move in space, they will naturally predict the changes they see: as you move farther away, the objects in front of you should become smaller. But the multi-modal models at the time did not have this kind of prediction - even if Google Gemini is known for its native multi-modal route and is one of the best performers, it is also unable to judge what forms objects will appear in different viewing angles and scales.
"It can recognize that it is a black sheep, but it cannot answer what the black sheep looks like when viewed from another angle. What the model remembers is the massive association between the image and the semantics: it looks like a black sheep from a distance, but it still looks like a black sheep when seen up close."
The team continued to add relevant data, but the model's performance did not get better, but became "increasingly confusing": the semantic connections were continuously strengthened, but there was a lack of spatial correlation between the data, and the model relied more and more on semantics to explain everything it saw. The more data, the deeper the solidification and the more stable the error.
An example that Lin Dahua often cited before can illustrate the consequences of this solidification. Give a photo of six fingers to the model at that time, and the answers given by GPT, Gemini, and Grok are all five fingers. The model's judgment path is: first identify "this is a palm", and then directly answer based on the priori "palms usually have five fingers" in the training corpus.
He also reminded that this problem has most likely been fixed in the latest generation of models, but the mechanism behind it is still valid - "It only relies on semantic translation, and then other reasoning is done based on this translation."
When SenseTime released its spatial intelligence model at the beginning of this year, this direction was still a technical reserve in the eyes of the outside world. More than half a year later, the industry began to catch up. Several points that have been widely discussed by the community after the release of GPT-6 - the adoption of the Recurrent Transformer architecture and the robot-driven capabilities demonstrated after pouring massive embodied data into the language model - all point to the same thing: multi-modal and spatial capabilities are entering the core of cutting-edge models.
In Lin Dahua's view, this means that the route of SenseTime has been verified: "GPT-6 actually proves that the path we take is definitely valuable." SenseTime's resources are orders of magnitude inferior to OpenAI, but he believes that continuous accumulation of one or two years will allow SenseTime to develop something different from other teams in the industry.
The particularity of spatial intelligence shows to some extent that a direction with more data and deeper solidification may be the entrance to principled innovation. Finding principles by following "abnormalities" is one of SenseTime's ideas for making technical judgments. Taking one step further along this line of thinking, you will encounter an unavoidable question: what exactly is intelligence, and what is the position of language in it.
· Language and multimodality: The essence of intelligence is interaction with the world
Since the outbreak of big models, "whether language is the key to intelligence" has been a focus of debate.
Lin Dahua believes that language injects the most important human knowledge into the symbol system in a structured manner and is "a very high-level representative of human knowledge and intelligence." The density and efficiency of intelligence may be the highest carrier. Language models have the most mature engineering experience and the fastest progress, and have received higher industry attention. This is also the reality.
But language alone, he thought, was incomplete. Just like being in a bustling city and a quiet mountain forest, people's experiences are completely different, and it may not be possible to describe the surrounding environment in detail in words. Suppose you give a portrait photo to the best multi-modal model and let it generate a description as detailed as possible, with thousands of words, and then let the text model draw it, it will be difficult to generate an identical photo.
He summarized this phenomenon as complementarity on the information spectrum: language carries high-density information that has been condensed, while images capture a large amount of information that has not been condensed and cannot be condensed.
Each of the two modes has its own efficient territory - writing programs and doing logical deductions, the language efficiency is higher; the aesthetic feeling, quick judgment of spatial structure, and the transmission of visual details, the image efficiency is higher. As for full-modal input and output, the ceiling is still very high. Putting them together without losing intelligence "requires cost and time."
In Lin Dahua's view, the essence of intelligence is the ability of human beings to interact with the world - perception, memory of knowledge, multi-step reasoning, issuing instructions, and the composite result of a combination of multiple elements. Just like human IQ test questions, they are also divided into different types such as space, memory, reasoning, etc., which reflect different aspects of intelligence.
Following this definition and looking back at the development of the large model over more than three years, his judgment is that it is "unbalanced":
In terms of mathematics, programming, and long-range tasks that can be written as text and program sequences, models have surpassed the level of ordinary people and are even close to the top; spatial and physical intelligence are still in their early stages, and robots are developing much faster than before, but they still can only perform simple actions.
The move from digital intelligence to physical space depends not only on technical judgment, but also on whether the technology can run through the closed loop in real user scenarios and deliver a report card that can withstand market testing.
03
There is iteration only if there is a closed loop:
SenseTime’s business answers
Looking back at the process of SenseTime turning AI technical judgments into products and revenue, Lin Dahua said this:
Shangtang’s path is made by walking, not thinking about it.
In the big model era, the first "tuition" paid by SenseTime was a product called "Discussion". When it launches in early 2023, it will be like any other chat product. Lin Dahua once gave the product to his wife to try out, but the "discussion" ended with a nonsense about recommending nearby restaurants.
The overturn is not entirely due to the quality of the model. The foundation of conversational products is information infrastructure such as reviews and maps, but that is not SenseTime’s talent.
At the end of 2023, SenseTime will focus on content and interaction for its large model products, focusing on two directions: work content and entertainment content. The former has now grown into the office intelligent raccoon—according to this year’s interim results announcement, this part of the business has grown six to seven times; the latter has grown into Seko.
Seko's style of play actually lies behind the difference between "creation" and "generation". Generate beautiful 30-second clips and create 100-episode shorts, all separated by scripts, storyboards, characterization, and cross-episode consistency. Seko uses technologies such as unified understanding and generation, and interlaced thinking of graphics and text in the creative process, and the underlying rendering relies on other models such as Seedance - the quality of generation alone cannot be compared, so it becomes a part of its own workflow. According to Lin Dahua, creators have used Seko to create short plays with over 100 million views.
Supporting these product selections is SenseTime’s closed-loop methodology. Lin Dahua pointed out bluntly, "In everything, you need a closed loop to be able to iterate effectively, otherwise there will be only rankings and no iteration." Being the champion on Benchmark today may not be the next month. This kind of achievement cannot be retained; what really remains is the value created by users. He translated the principle in reinforcement learning that running fast in small steps is better than walking a long distance at once to R&D planning: iteration speed and the right direction are more important than how high a single sprint is.
Lin Dahua has experienced this transformation once in his own life. In his early papers, twenty or thirty formulas were the norm; but after 2013, this number dropped to no more than five. His criterion of judgment changed from "whether the mathematical proof is beautiful" to "whether it has differentiated value in use."
In this year’s SenseTime interim report, we saw the official statement of “first profit”. However, Lin Dahua believes that what SenseTime is really concerned about is always whether it has become healthier and more vital, and the closed-loop iteration of business is one of the manifestations of vitality.
The 1 + The competition in the market is harsh enough, but it brings growth to both the founders and the business.
From SenseTime's perspective, products and businesses are like living things. We do not push them when they are immature. We use natural growth as a signal before deciding to increase investment - let the products enter the market first and be tested in a closed loop, instead of iterating against preset indicators internally.
However, to a certain extent, business closed loop only answers the question of "survival". Facing future competition, Chinese AI companies must find their own positioning and long-term competitiveness in the post-distillation era, and see clearly where the boundaries of intelligence are.
04
In addition to hand speed:
The long-term competitiveness of China’s AI and the boundaries of intelligence
Lin Dahua believes that the first piece of the puzzle for long-term competitiveness is open source.
As early as 2018, Lin Dahua promoted the open source OpenMMLab at MMLab: organizing representative algorithms in the field into a unified benchmark system, saving researchers the dirty work of reproduction and comparative experiments. This project later became the most influential open source algorithm library in the AI 1.0 era, used by hundreds of thousands of stars and millions of developers.
Years later, the same logic reappeared in the era of large models. The endogenous driving force of open source is influence: influence can raise funds and recruit people. DeepSeek's demonstration allows everyone to see this path clearly. Objectively, this also distributes the cost of trial and error to the entire community - model training only accounts for about 10% of R&D costs, and the remaining 90% is various experiments; open weighting allows each company to pick cutting-edge technology recipes faster.
This also explains why China’s total investment is far behind that of the United States, but it can catch up so closely. Lin Dahua said, "The biggest worry is not knowing where the road is." Once the road is visible, filling in the details is the strength of Chinese engineers.
But open source cannot solve the problem of following the trend. Lin Dahua bluntly said that the mainstream path of China's big models in the past three years has been to optimize and innovate in the direction pioneered by the United States. The industrial value is great, "but if China wants to make a more lasting contribution, it must come up with something different." He used DeepSeek as a positive case: it is still the standard LLM route in terms of large paradigm, but it has made great efforts in architectural efficiency.
At the same time, the model market itself cannot accommodate takers. Cost, versatility, and professionalism form an impossible triangle. A model that can do everything, and does everything professionally, is bound to be huge and expensive. In daily scenarios, there is even widespread "intelligence surplus": an 8B or 32B open source model can handle most tasks, and users do not need to pay dozens of dollars in tokens for this. There are more than 3 million models on Hugging Face, but each one has its own way of life.
The real watershed may come in the post-distillation era.
Lin Dahua judged that the topic of distillation was over: the existing thinking chain questions had been constructed, and the next bottleneck was the ability to generate new problems—constructing thousands, tens of thousands, or even hundreds of thousands of real work scenarios, allowing the model to explore and complete tasks on its own.
The bottleneck in this matter is actually not computing power. China happens to hold a good hand: the diversity of life and industry is rare in the world, and the reserve of engineers is the largest in the world. After the fruits at the lower places are picked, the fruits at the higher places have to be climbed up the tree by people.
However, scenarios and manpower only expand the "width" of intelligence; the "depth" of intelligence depends on the growth rate of intelligence itself.
RSI (recursive self-improvement), which has attracted much attention recently, focuses on this, but Lin Dahua emphasized that RSI is only one of the means of intelligent growth. The bottleneck is not the automation of training, but the efficiency of evolution: a machine still has to run a massive amount of training to obtain an effective observation point; a complete training is not required for a person to master a new skill. The current AI is still very far away from humans in terms of evolution. Once this secret is cracked, "it will largely reshape the landscape of AI development and competition."
He also does not believe that the current situation can be extrapolated indefinitely: Extensive resource consumption will sooner or later hit a trajectory explosion. Whether the existing path can reach the end is an open question.
As for expanding the boundaries, the first question he is concerned about is: what role do modalities other than language play in intelligence. This understanding is still in its early stages, but the answer will determine where resources should be directed. He places his hope on the integration of multi-modality and space - simply generating will only improve the visual quality; integrating spatial perception into the thinking process will make intelligence more complete.
For intelligence to break through its boundaries, the bottleneck lies in the efficiency of learning and evolution, which is exactly what humans are best at but machines are still far behind. SenseTime's answer to this question is not all written in the model - an organization that can continue to breed new people is itself the most viable intelligence.
Lin Dahua also said that his most important job at the moment is to create an environment for young people and protect a viable organization. This is also the underlying logic of SenseTime’s 20% resources: actively sow seeds and protect growth. This seed can be a product, a business, or a person who grows up in the organization. What SenseTime has been doing is waiting for the seeds to sprout, and life will have its own way out.
Leifeng.com (public account: Leifeng.com)
Original articles from Leifeng.com, unauthorized reproduction is prohibited. See the reprint instructions for details.
edit
Scan to follow the author on WeChat
Resend Email Modify Email




