Newshot

The Logic Behind SenseTime’s 20% Resource Investment: Lin Dahua on Speed and Long-Term Competitiveness

AI summary

Lin Dahua discusses whether the Scaling Law has failed in spatial intelligence and why SenseTime still devotes 20% of its resources to pre-research with no product outlet. He recalls Seko’s doubts when compared with Seedance, but says Seko found its place in end-to-end short-drama production, with users growing 7–8x this year and a drama exceeding 100 million views. He also argues training costs cannot stop because competitors force continual retraining, while the parameter-scaling race is meeting pushback as model-size choices become more pragmatic.

Translation(model translation, not the source's own text)

Title: The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Racing Against Time and Long-Term Competitiveness

The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Racing Against Time and Long-Term Competitiveness
The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Racing Against Time and Long-Term Competitiveness

When ByteDance's Seedance 2.0 was released earlier this year, its "cinematic-grade" output quality left many industry practitioners excited and shaken, with new creative blueprints unfolding before them—but for SenseTime's Seko team, that was a period of "confusion."

SenseTime co-founder and chief scientist Lin Dahua recalled in a recent interview that during that time, whenever he met investors or attended industry conferences, he was almost always asked the same question: What's the difference between your Seko and Seedance? Can you beat it?

At one point, voices of self-doubt emerged within the team: Should we keep doing this?

The story did not ultimately lead to SenseTime's Seko single-handedly "confronting" Seedance. Seedance renders 30-second video clips to a top-tier standard; the creators Seko targets ultimately deliver 100-episode short dramas—and in between lie scripts, storyboards, character design, and cross-episode visual consistency, an entire creative workflow.

Having thought this through, Seko actually thrived: according to Lin Dahua in the interview, the product's users grew 7 to 8 times this year, and short dramas produced on the platform surpassed 100 million views.

As for SenseTime's prominence in the large model market, Lin Dahua himself offers a candid assessment: "Probably not the biggest." The company's path has also never fully followed the industry's hottest trends.

But SenseTime has a habit it has maintained for years: to this day, the company still invests 20% of its resources in pre-research work that has no product outlet.

To some extent, this is also a form of "sowing seeds"—when seeds go into the ground, there is no need to promise a harvest. Tang Xiao'ou's oft-repeated saying—life finds a way—may be exactly the logic behind this investment.

In a recent interview on the "Tech Talk" program co-produced by Weibo Tech X and Wei Shijie, Lin Dahua started from the current state of the industry, passed through the validation of commercialization, and finally landed on the boundaries of intelligence. But the thread running throughout was only one: that 20% of resources, and the judgment supporting it—why some paths are worth walking for a long, long time, even if no outlet is yet visible.

The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Racing Against Time and Long-Term Competitiveness
The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Racing Against Time and Long-Term Competitiveness

Training Costs Won't Stop Rising: China's AI Race Against Time

Money in the large model industry is burning in a way that won't stop.

There was once a widespread assumption in the industry: train a good model at great expense, and the cost can be amortized across subsequent scaled services.

Lin Dahua pointed out in the interview that this assumption may not hold—the problem isn't that the model isn't good enough, but that "if you don't keep training the next one, others will surpass you." Competition forces companies to retrain round after round, so training costs won't stop, and they keep snowballing.

Although the industry narrative over the past year has been almost dominated by "bigger is stronger," the path of competing on parameters is now backfiring. He mentioned two signals:

First, model size choices have become more pragmatic. Recently, various vendors have released Flash-level small and medium models: GLM 5.3 Flash, DeepSeek 4.1 Flash, and Google and OpenAI have also launched products of similar specifications.

Lin Dahua specifically mentioned DeepSeek's performance in the interview: "The 4.1 Flash's performance level is better than the 4.0 Pro." A model in the several-hundred-B range with well-tuned training beating the previous generation Pro shows that there is still plenty of room for optimization at medium scale.

Second, the quality of pre-training data has been put back on the table. When a model with several hundred billion parameters beats an opponent with 2 trillion parameters in real-world use, many people start going back to examine their own data pipelines.

Lin Dahua has a vivid explanation for this: "(In the data) there is a lot of noise, many things contradicting each other, so much of the efficiency in between gets canceled out." The result is an expensive yet suboptimal model—parameters pile up, but performance doesn't.

"When everyone blindly competes on parameter count, I think that's not a healthy state." In his view, these two signals mean the industry is returning to rationality: no longer fixated on a single metric, but back to end-user value.

Similar situations occurred in the previous AI cycle, and SenseTime itself lived through it. Every project had to start from data collection, with dedicated training, tuning, and testing—the entire pipeline redone each time, hard to truly replicate. Only back then the unit of measurement was projects; today it's parameters.

Racing against time is another form of involution. Frontier labs in the U.S. define the SOTA ceiling; when a new model is released, it essentially points a path for Chinese practitioners—within two or three months, a batch of products claiming to adopt similar approaches tends to emerge domestically, grabbing ecological niches. The recently hot RSI (recursive self-improvement) is likewise being built simultaneously by a bunch of companies.

"To be honest, racing against time is actually quite exhausting." Lin Dahua observed that when everyone rushes into a direction they favor, it's hard to create long-term value. When ten companies compete for the same ecological niche at the same time, the first mover's advantage is quickly diluted.

So what does long-term competitiveness rely on? His answer: walking far enough and digging deep enough on one path. This question is exactly the starting point for understanding SenseTime's 20% of resources.

The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Racing Against Time and Long-Term Competitiveness
The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Racing Against Time and Long-Term Competitiveness

From Anti-Scaling Law to Language and Multimodality

· 20% of Resources and Spatial Intelligence

This 20% of resources is explicitly directed toward work that has "no product as an outlet." Lin Dahua confirmed in the interview that no matter what operational pressure SenseTime faces, this money is still being spent to this day: "At any time, we still retain a certain amount of resources."

In his words, the starting point of such work is asking why: when existing methods fail on a problem, trace back to the most fundamental level. He doesn't seek to explain principles with complex mathematical formulas, but requires that the principle "can effectively solve some problems that other methods currently find very hard to solve."

One of SenseTime's research directions, spatial intelligence, was selected exactly this way.

The selection logic has two steps. First, test the strongest models at the time and find out what they do poorly. Second, ask one level deeper: can these shortcomings be fixed by adding more data? If adding data along the existing path can fix it, then it has application value but cannot support fundamental innovation. Only directions that cannot be fixed no matter how much data you add might harbor genuine fundamental problems—and spatial intelligence happens to be the latter.

The origin of this judgment standard can be traced to his training. When Tang Xiao'ou mentored students, he almost never asked about technical details; he looked at only one thing: can you explain in one or two sentences how your work is fundamentally different from others'? "No two leaves in the world are exactly alike"—technically you can always find differences in the details, but such differences have limited innovative value; if you can't say it in one sentence, it's probably not real innovation either.

Lin Dahua pointed out that the "anomaly" of spatial intelligence is that the Scaling Law fails here.

In industry, publicly questioning the Scaling Law is rare. The industry consensus of the past few years has staked almost all strategy on this path—bigger parameters, more data, stronger capabilities. In such a context, proposing "anti-Scaling Law" at least prompts the industry to re-examine a question: does the approach of simply piling on data and parameters have unnoticed blind spots in certain dimensions of intelligence?

On tasks related to spatial intelligence, the trend observed by the SenseTime team runs exactly counter to the Scaling Law—the more data added, the more unstable the model's performance.

In tests at the time, the model performed unsatisfactorily on perspective-taking tasks. When people move through space, they naturally anticipate how what they see will change: move farther away, and objects before them should appear smaller. But multimodal models at the time lacked this anticipation—even Google Gemini, known for its native multimodal approach and the best performer among them, likewise could not judge what form an object would take from different perspectives and scales.

"It can recognize that it's a black sheep, but it can't answer what that black sheep looks like from another angle. What the model memorizes is the massive correlation between images and semantics: from afar it's a black sheep, up close it's still a black sheep."

The team continued adding relevant data, and the model's performance didn't improve—instead it became "increasingly confused": semantic associations were constantly reinforced, but these data lacked spatial correlations, so the model relied more and more on semantics to explain everything it saw. The more data, the deeper the ossification, and the more consistently wrong it became.

An example Lin Dahua often cited illustrates the consequences of this ossification. Give a photo of a hand with six fingers to the models of that time, and GPT, Gemini, and Grok all answered five fingers. The model's judgment path is: first recognize "this is a hand," then answer directly based on the prior in the training corpus that "a hand usually has five fingers."

He also noted that this problem has most likely been fixed in the latest generation of models, but the underlying mechanism still holds—"it relies only on semantic translation, and then does other reasoning based on that translation."

When SenseTime released its spatial intelligence model earlier this year, this direction was still seen by outsiders as a technical reserve. More than half a year later, the industry began to catch up. Several points widely discussed by the community after GPT-6's release—the adoption of the Recurrent Transformer architecture, and the robot-driving capabilities demonstrated after massive embodied data was fed into language models—all point to the same thing: multimodal and spatial capabilities are entering the core territory of frontier models.

In Lin Dahua's view, this means the path has been validated for SenseTime: "What GPT-6 actually proves is that the path we're on definitely has value." SenseTime's resources are orders of magnitude behind OpenAI's, but he believes that one or two years of sustained accumulation will let SenseTime grow something different from other teams in the industry.

The peculiarity of spatial intelligence, to some extent, shows that a direction where more data leads to deeper ossification may instead be the entry point for fundamental innovation. Following the "anomalies" to find principles is one of SenseTime's approaches to technical judgment. Taking one more step along this line of thinking leads to an unavoidable question: what exactly is intelligence, and where does language sit within it?

· Language and Multimodality: The Essence of Intelligence Is Interacting with the World

Since the explosion of large models, "whether language is the key to intelligence" has been a focus of debate.

Lin Dahua believes that language structurally injects humanity's most important knowledge into a symbolic system and is "a very high-level representation of human knowledge and intelligence," possibly the carrier with the highest intelligence density and efficiency. Language models have the most mature engineering experience, progress the fastest, and receive greater industry attention—this is also the reality.

But relying on language alone, he believes, is incomplete. Just as being in a bustling downtown and a quiet mountain forest gives a person utterly different experiences, one may not be able to recount the surrounding environment in detail with words. Suppose you give a portrait photo to the best multimodal model and have it generate as detailed a description as possible, over a thousand words, then take that text and have an image generation model draw it—it would still be hard to produce an identical photo.

Article title: The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Moving Fast and Long-Term Competitiveness

Main text: He summarizes this phenomenon as complementarity across the information spectrum: language carries concentrated, high-density information, while images capture large amounts of information that has not been concentrated and cannot be concentrated.

Each modality has its own high-efficiency territory—for writing programs and doing logical reasoning, language is more efficient; for perceiving aesthetics, making quick judgments about spatial structure, and conveying visual details, images are more efficient. As for full-modality input and output, the ceiling remains very high, and bringing them together without losing intelligence "requires paying a price and spending time."

In Lin Dahua's view, the essence of intelligence is the ability of humans to interact with the world—perception, memory of knowledge, multi-step reasoning, issuing instructions, a composite result of multiple elements combined. Just like human IQ test questions, which are also divided into different types such as spatial, memory, and reasoning, reflecting precisely different facets of intelligence.

Looking back at the development of large models over the past three-plus years along this definition, his judgment is "unbalanced":

In mathematics, programming, and long-horizon tasks that can be written as text and program sequences, models have already surpassed the level of ordinary people and are even approaching the top tier; spatial and physical intelligence are still at an early stage—robots have developed much faster than before, but they still only perform simple actions.

And moving from digital intelligence to physical space depends not only on technological judgment, but even more on whether the technology can run a closed loop in real user scenarios and deliver results that stand up to market testing.

The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Moving Fast and Long-Term Competitiveness
The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Moving Fast and Long-Term Competitiveness

Recalling the process by which SenseTime turned AI technological judgment into products and revenue, Lin Dahua put it this way:

SenseTime's path was walked out, not thought up.

In the large-model era, the first "tuition fee" SenseTime paid was a product called "SenseChat." When it was launched in early 2023, it was no different from other chat products. Lin Dahua once gave the product to his wife to try, and "SenseChat" directly made things up when it came to recommending nearby restaurants.

The failure was not entirely due to model quality; the foundation of a conversational product is information infrastructure such as reviews and maps, but that was not where SenseTime's strengths lay.

At the end of 2023, SenseTime narrowed its large-model products to content and interaction, focusing on two directions: work content and entertainment content. The former has now grown into the office agent Xiaohuanxiong—according to this year's interim results announcement, this business grew six or seven times; the latter has grown into Seko.

The approach behind Seko is actually the difference between "creation" and "generation." Generating a polished 30-second clip and creating a 100-episode short drama are separated by scripts, storyboards, character settings, and cross-episode consistency. Seko applies technologies such as unified understanding and generation and interleaved image-text thinking to the creation workflow, while leveraging other models such as Seedance for underlying rendering—if it cannot win on generation quality alone, it turns that into one link in its own workflow. According to Lin Dahua, creators have already used Seko to make short dramas with over 100 million views.

Supporting these product choices is SenseTime's methodology around closed loops. Lin Dahua pointed out bluntly, "For everything, you need a closed loop to iterate effectively; otherwise there is only leaderboard chasing, no iteration." A champion on a Benchmark today may not be one next month, and such results do not persist; what truly persists is the value created for users. He transposed the principle in reinforcement learning that small, fast steps are better than one long leap onto R&D planning: iteration speed and the right direction matter more than how high one single push goes.

Lin Dahua experienced this shift himself. In papers from his early first-author work, twenty or thirty formulas were the norm; but after 2013, that number dropped to no more than five. His criterion changed from "whether the mathematical proof is beautiful" to "whether there is differentiated value in use."

In this year's SenseTime interim report, we saw the official wording of "first profit." However, Lin Dahua believes what SenseTime truly focuses on has always been whether it has become healthier and more vital, and closed-loop business iteration is one manifestation of that vitality.

The 1+X structure is precisely the implementation of this idea: new businesses grow out of the parent body, then independently raise financing and survive on their own in the market, equivalent to extending the "closed loop" from the R&D level to the organizational level. Market competition is harsh enough, but it is growth for both the founders and the businesses.

From SenseTime's perspective, products and businesses are also like living organisms: when immature, they are not forcibly pushed; natural growth is used as a signal before deciding to increase investment—letting products first enter the market's closed loop for testing, rather than iterating internally against preset metrics.

However, to some extent, the commercial closed loop only answers the question of "being able to survive," while for future-oriented competition, Chinese AI companies need to find their own positioning and long-term competitiveness in the post-distillation era, and clearly see where the boundaries of intelligence lie.

The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Moving Fast and Long-Term Competitiveness
The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Moving Fast and Long-Term Competitiveness

The Long-Term Competitiveness of Chinese AI and the Boundaries of Intelligence

The first piece of the long-term competitiveness puzzle, in Lin Dahua's view, is open source.

As early as 2018, Lin Dahua promoted the open-sourcing of OpenMMLab at MMLab: organizing representative algorithms in the field into a unified benchmark system, saving researchers the dirty work of reproduction and comparative experiments. The project later became the most influential open-source algorithm library of the AI 1.0 era, with over 100,000 stars and millions of developers using it.

Years later, the same logic was reproduced unchanged in the large-model era. The intrinsic driving force of open source is influence: influence can raise funds and attract talent, and DeepSeek's demonstration made this pathway clear to everyone. Objectively, this also spread the cost of trial and error across the entire community—model training accounts for only about 10% of R&D costs, with the remaining 90% going to various experiments; opening weights allows every company to pick frontier technical recipes faster.

This also explains why China's total investment is orders of magnitude behind that of the United States, yet it has been able to catch up so closely. Lin Dahua said, "What I worry about most is not knowing where the path is." Once the path is visible, filling in the details is the strong suit of Chinese engineers.

But open source cannot solve the problem of following the crowd. Lin Dahua said frankly that the mainstream path of Chinese large models over the past three years has been to optimize and innovate on directions pioneered by the United States, which has enormous industrial value, "but if China is to make more lasting contributions, it will certainly have to walk out some different things." He cited DeepSeek as a positive example: it still follows the standard LLM route in the broad paradigm, but has worked very deeply on architectural efficiency.

At the same time, the model market itself cannot accommodate a winner-takes-all player; cost, generality, and specialization form an impossible triangle. A model that can do everything and do everything professionally enough is bound to be huge and expensive. In everyday scenarios, "intelligence surplus" has even become common: an 8B or 32B open-source model can handle most work well, and users have no need to pay tens of dollars for tokens. The more than 3 million models on Hugging Face actually each have their own way of living.

The real watershed may appear in the post-distillation era.

Lin Dahua judges that the topic of distillation has passed: the existing stock of chain-of-thought problems has been constructed, and the next bottleneck is the ability to generate new problems—constructing thousands, tens of thousands, or even hundreds of thousands of real work scenarios and letting models explore and complete tasks on their own.

The bottleneck here is actually not computing power. China happens to hold a good hand: a diversity of life and industry rarely seen anywhere in the world, and the world's deepest reserve of engineers. After the low-hanging fruit has been picked, the fruit higher up depends on people climbing the tree.

But scenarios and manpower only expand the "width" of intelligence; the "depth" of intelligence depends on the speed of intelligence's own growth.

The recently much-watched RSI (recursive self-improvement) is precisely focused on this, but Lin Dahua emphasizes that RSI is only one means of intelligence growth; the bottleneck is not in the automation of training, but in the efficiency of evolution: for a machine to obtain one effective observation point, it still has to run massive training; for a human to master a new skill, a complete training run is not needed. Current AI is still very far from humans in terms of evolution, and once this secret is cracked, "it will reshape the landscape of AI development and competition to a great extent."

He also does not believe the status quo can be extrapolated indefinitely: crude resource consumption will sooner or later hit trajectory explosion, and whether the existing path can reach the endpoint is an unresolved question.

As for expanding boundaries, the first question he cares about is: what role do modalities other than language actually play in intelligence. This understanding is still at an early stage, but the answer will determine where resources should be invested. He places his hopes on the integration of multimodality and space—doing generation alone only improves visual quality; integrating spatial perception into the thinking process will make intelligence more complete.

For intelligence to break through boundaries, the bottleneck lies in the efficiency of learning and evolution, and this is precisely what humans are best at and machines are still far behind in. SenseTime's answer to this question is not all written in models—an organization that can continuously nurture new people is itself a kind of intelligence with the greatest vitality.

Lin Dahua also said that his most important work at present is to create an environment for young people and protect an organization with vitality. This is also the underlying logic of SenseTime's 20% resource investment: actively sowing seeds and protecting growth. This seed can be a product, a business, or the people growing within the organization; what SenseTime has always been doing is quietly waiting for the seed to sprout—life will find its own way.

Leiphone (official account: Leiphone)

Leiphone original article, unauthorized reproduction prohibited. See reprint notice for details.

The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Moving Fast and Long-Term Competitiveness
The Logic Behind SenseTime's 20% Resource Investment: Lin Dahua on Moving Fast and Long-Term Competitiveness
  • Quan Long, Sun Jian, Lin Dahua's Visual Quest | CCF-GAIR 2019
  • SenseTime's revenue in the first three quarters of 2026 grew 30.4% year on year to 3.79 billion, gross margin ...
  • SenseNova 6.8 Flash official version released: office more stable, design more ...
  • SenseTime open-sources SenseNova Loop-9B-Preview: exploring computational depth ...

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by 雷锋网
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗