MaHui Entrepreneur | Sand.ai's Yue Cao: A Decade from Researcher to Entrepreneur Traces the Evolution of Large Models

• This article is republished with authorization from LatePost (ID: postlate); author: Manqi Cheng

Every pivotal choice Yue Cao has made over the past decade serves as a footnote to the evolution of large models: he pivoted to deep learning in 2014, won the ICCV Best Paper Award in 2021, joined BAAI — China's earliest institution focused on large models — in 2022, became a co-founder of Lightyears Away in 2023, and founded Sand.ai in 2024 to work on video generation.

In this interview, Cao shares his complete ten-year journey from researcher to entrepreneur, reflecting the broader transformation of the AI industry from its nascent stages to its rise.

On personal choices and trend judgment, Cao maintains a consistent perspective: paying attention to organization and collaboration. In 2021, Cao won the ICCV Best Paper Marr Prize for Swin Transformer, but after seeing CLIP and DALL·E released by OpenAI that same year, he quickly recognized that the breakthroughs stemmed from fundamentally different organizational and thinking patterns:

At the time, most domestic research groups were still paper-driven, while OpenAI was thinking about the big, fundamental problems in the field — more problem-driven.

When you're very focused on papers, you're influenced by reviewers' tastes, optimizing for so-called novelty... But OpenAI's approach was actually quite simple. The core was designing a scalable system that could maximize compute utilization.

This prompted Cao to join BAAI, which was focused on large language models, a year later. ChatGPT had not yet been released, and large language models were not yet an industry consensus.

From 2024 onward, Cao began building Sand.ai as the founding CEO. Behind the company's two generations of models — Magi-1 to Gaga-1, released just this October — lies the continued evolution of Cao's organizational thinking: from a heavier emphasis on model and technology-driven approaches, to a deep vertical integration of models with product and demand.

Gaga-1 has a concrete goal — making AI-generated video characters that can actually "perform." The biggest pain point in narrative content creation has always been: inconsistent characters, lack of audio-visual synchronization, and performances that feel fake.

How quickly can a research-background founder become a professional CEO? And how much leeway does the environment give them? Cao and Sand.ai will write their answer to this critical question through practice.

From ICCV Best Paper to Studying OpenAI's Organizational Capacity

LatePost: Over the past decade, each of your key choices has been representative. Looking back at your AI experience, what's the first moment that comes to mind?

Cao: It would be the moment I switched to deep learning. That was 2014, my senior year. Deep learning wasn't really hot in China yet — some professors even wrote explicitly on their personal webpages: "We don't do neural networks with more than two layers."

I was relatively lucky. My co-advisor visited UC Berkeley and found that everyone in Silicon Valley was talking about deep learning. Our group was also among the earliest in China to buy GPUs and start working on deep learning. Then in 2017 and 2018, I joined the vision group at MSRA (Microsoft Research Asia).

LatePost: That's a legendary group.

Cao: Yes, they developed ResNet (Residual Neural Network) in 2014-2015. A wave of people left around 2016-2017, some going to companies, some starting their own. (Editor's note: Entrepreneurs from this group include SenseTime founder Tang Xiaoou, Megvii chief scientist Sun Jian, and Momenta founder Xudong Cao.)

This group did have some legacy. What benefited me most was a few things: first, you should work on the most attention-grabbing topics, but ones that still have significant opportunity — not topics that have already converged. Second, once you identify an important opportunity, you need to commit sufficient resources to push it to the extreme, and this is fundamentally an organizational problem.

LatePost: A highly visible direction often already has many people working on it. How do you judge whether it still has substantial room for improvement?

Cao: This really does involve some intuition. When you've actually worked in an industry for a while and thought deeply about many problems, you become sensitive: some subtle foundational signals that seem minor to others register as very strong signals to you.

LatePost: Can we walk through a concrete example — your work with MSRA colleagues that won the 2021 ICCV Best Paper Marr Prize, Swin Transformer, which applied Transformer to vision. How did you see this topic?

Cao: Transformer emerged in 2017. Its adoption across vision and multimodal directions happened in two stages:

• The first stage was embedding Attention modules into convolutional neural networks (CNNs). The main idea was using Attention to compensate for CNN limitations. CNNs are inherently local, with limited receptive fields (the input region that a "neuron" in a deep learning network can see). Attention in Transformer, by contrast, can relatively cheaply expand the receptive field.

There was another path at the time, but fewer people took it: trying to directly replace convolutions with Attention modules. Our group was already working in this direction in 2018-2019, but we were still following ResNet's overall structure, just replacing the convolutional parts.

• The second stage was replacing CNNs with Transformer's overall architecture. Two important works emerged in the second half of 2020: first, OpenAI's Image GPT. After GPT-3, OpenAI brute-forced GPT onto computer vision, but because the results weren't great, most people didn't understand it.

Second was ViT, Vision Transformer. Image GPT applied Self-Attention directly on pixels, which was computationally inefficient; ViT turned pixels into patches, such as 16×16 pixel blocks for processing, and achieved good results on ImageNet's image classification task. (Note: Self-Attention is a special attention mechanism where instead of having the output sequence attend to the input sequence, each element in the input sequence attends to itself and all other elements, capturing dependencies between different positions within the sequence.)

The core insight this brought to vision was: stop being fixated on CNNs, and try introducing Transformer's overall architecture. ViT proved this could work for image classification, so we judged: if we could build on Transformer's macro-structure to create a network that adapted to most vision tasks with excellent results, there was an opportunity to replace the entire CNN-based network structure.

LatePost: So Swin Transformer's starting point was not just using Transformer for image classification, but making it more generally adaptable to other common vision tasks?

Cao: Right, including object detection, semantic segmentation, and so on. Before this, computer vision tasks were relatively complex. Different tasks generally used CNNs (ResNet) as their base network architecture, but the overall systems for each task had some differences. So the problem became: how to use the next-generation Transformer-based network architecture to replace the previous CNN-based network architecture.

Because our group had been studying Transformer very early, and we were very familiar with key directions in computer vision — image classification, object detection, and so on — we were among the earliest after ViT to see this opportunity. And we could clearly feel it was a major opportunity, for three reasons:

• Network architecture itself is the most attention-grabbing topic in deep learning;

• If it could be achieved, it would significantly improve results across all vision tasks;

• It had the potential to further break the mental barrier that "computer vision must use convolution."

Another key factor was that after identifying this opportunity, we really mobilized significant forces, pulling in everyone in the group who could participate, and together pushed Swin Transformer to the extreme across every dimension. The final results were markedly better than comparable contemporary work.

LatePost: This ties directly to the other thing you benefited most from — seeing an opportunity, and having the organizational capacity to support how far you could take it?

Cao: Exactly. You need to smell the opportunity; you also need organizational capacity to actually seize it — both are indispensable.

Though these two do sometimes conflict. You can't just charge into whatever topic is hot every time; then every piece of work ends up superficial and won't stand the test of time. I also know a type of person who, whenever they work on a topic, goes extremely deep and stays with it for a very long time — but often by the time they're done, it's too late. If relatively lucky, this type can still produce excellent work.

LatePost: This sounds like it has common ground with R&D and business decisions in companies: needing to think about direction and priorities, how many resources to invest, how teams collaborate.

Cao: But after seeing the series of works from OpenAI and DeepMind, I realized the ceiling of the original methodology was quite obvious. AlphaFold's breakthrough, and OpenAI's DALL·E and CLIP in early 2021, all illustrated this.

But at the time, very few people around me were studying why OpenAI could produce such work, or giving these works evaluations commensurate with their impact. Most people's first reaction was that this required a lot of compute, they had compute and we didn't, so it had nothing to do with me. Actually, if you could put aside ego and carefully study these two works, their methodology, mindset, and organizational form were very different from how we were doing papers at the time.

So in 2021-2022, I went through a fairly significant mindset shift, which prompted me to leave Microsoft Research Asia and join BAAI.

LatePost: What did you see in OpenAI's approach in 2021?

Cao: First, at the time most domestic research groups were still paper-driven. This created many problems — papers have author lists, with first author, second author, which fundamentally doesn't encourage collaboration. OpenAI also published papers, but didn't treat publication as the core goal. Instead, they thought about what big, fundamental problems existed in the field — more problem-driven.

Second, when you're obsessed with papers, you end up catering to reviewers' tastes, optimizing for so-called novelty, trying to come up with methodological tricks. But OpenAI's approach was remarkably simple at its core: design a scalable system that maximizes compute utilization.

The third difference is organizational. A paper-driven organization tends to stay small, with students as your main collaborators. But if you're building a system, you need many different kinds of people: someone to crawl data, someone to clean data, someone to train models, someone to do evaluation, and finally someone to handle PR.

This requires a group of people with shared goals and shared thinking — all wanting to build the system well, rather than worrying about their author order. At the time I didn't know what to call this organizational form. Later I realized: it's a startup.

LatePost: So your starting point was wanting to build more powerful AI technology, which led you to study the methods and organizational approaches behind it?

Cao: Yes. There was another obstacle then — the pandemic. From 2020 through the two-plus years until ChatGPT emerged, communication between China and Silicon Valley dropped significantly. Many academic conferences went remote. This caused the domestic scene to miss OpenAI's transformation. In fact, starting from 2020, they were already producing very different kinds of results.

This is what I thought about most from 2011 through 2022, before joining BAAI: How did they manage to build something so impressive? What kind of organization should we build? Who around me could do this together?

LatePost: Did you consider starting a company directly at that point?

Cao: No. My understanding of entrepreneurship wasn't that deep at that time.

LatePost: What was BAAI's situation then? What made you think it could realize your vision?

Cao: BAAI was among the earliest domestic organizations to embrace large models. It was also a new type of research institution that didn't treat paper publication as its core metric. BAAI had also begun deploying compute clusters early — by mid-2022 it had 1,500 connected A100s, when clusters exceeding 1,000 cards were rare domestically. It also had a goal of open-source openness, releasing models for more people to use.

So it had a relatively relaxed research environment. You could screen for people with a similar mindset to do more frontier exploration together; everyone's goal was also to build a system.

In my understanding at the time, BAAI was the organization most like OpenAI domestically, and also a good place where I could do some exploration myself.

Why Didn't China Produce an OpenAI?

"Looking back in 2025, things are indeed changing: Nezha, Black Myth: Wukong, DeepSeek, Unitree — more cases are emerging. So what about five or ten years from now?"

LatePost: How did Lightyears Away come about?

Cao: After joining BAAI for a while, ChatGPT came out, and the domestic landscape changed instantly. Before that, people probably didn't think my joining BAAI was the right choice.

Then Huiwen Wang posted his "hero recruitment" on WeChat Moments, which truly ignited this wave of large-model enthusiasm domestically.

LatePost: You and Huiwen Wang are both Tsinghua alumni. Did you know each other before?

Cao: No. Between ChatGPT's release and meeting Wang, I had already clearly recognized this as a massive opportunity with significant room for improvement. I also realized that whether you had worked on computer vision or NLP before, you could participate.

Because I had been studying OpenAI for some time, their methodology was different from the past. It wasn't about bringing prior human experience to look at specific tasks, but rather — as mentioned earlier — thinking about how to design a scalable system that maximizes compute utilization.

I was also thinking: how could I participate? This first required model-building capabilities — not just reproduction, but frontier exploration. You also had to consider where the money would come from, how to productize later, and how to build the organization.

LatePost: So you were already considering starting your own company at this point?

Cao: I thought about it. But the conclusion was that this was highly complex. My capability at the time lay in building an organization that could train models; I was lacking in many other areas. So after talking with Wang, it felt like a perfect match.

LatePost: He must have talked to many people. Why did he choose you?

Cao: Yes, with every person he talked to, he would ask: who in this field do you recommend I talk to next? He pretty much talked to all the better researchers domestically at the time. Why me? I don't think he specifically mentioned it. One possible reason: I had already moved from MSRA to BAAI, demonstrating more conviction in this general direction.

LatePost: So Huiwen Wang also talked to Wenfeng Liang then?

Cao: I believe so. In March 2023, I actually only talked to Wang and Liang.

LatePost: How did your conversation with Wenfeng Liang come about?

Cao: He also reached out. It was during DeepSeek's preparation period. Wenfeng Liang had also talked to many domestic researchers. But when I met him, I had already committed to Wang.

LatePost: You mentioned earlier that you and Huiwen Wang hit it off — specifically, how?

Cao: I gained a lot from my first meeting with Wang. I could clearly feel this person was exceptionally strong, with deep insight and substantial practical experience.

A question I had long been thinking about was: why hasn't China produced an organization like OpenAI? I also asked Wang this. He quickly gave an answer: because domestic internet companies, or rather our stage of development, made us "not rich enough."

That is, when you're in catch-up mode with a clear target ahead, you just need to close the gap faster. At this stage, the highest ROI comes from efficiency innovation and model innovation, not original innovation.

But as you get closer to the frontier, there's almost no one ahead of you anymore, and your sense of direction weakens. At this point, transformation is needed across multiple dimensions:

• For entrepreneurs: catch-up and original innovation require different capabilities.

• For investors: previously people sought to invest in things they "understood," or Copy-to-China projects. Now such projects are increasingly scarce; when more original, riskier ideas come before you, do you dare to invest?

• Also society's attitude toward failure: original exploration will certainly have more failures compared to more certain catch-up, but success emerges from failure. Can failure be tolerated more? How do failed companies exit?

The entire chain and ecosystem need to change. Wang's understanding at the time was quite deep. Looking at 2025, things are indeed changing: Nezha, Black Myth: Wukong, DeepSeek, Unitree — more cases are emerging. So what about five or ten years from now? I'm personally very optimistic about the future. I can clearly feel we're in a transition phase.

LatePost: You asked Huiwen Wang "why hasn't China produced an OpenAI?" He said "not rich enough"; while Wenfeng Liang was in a relatively "rich" state when building DeepSeek.

Cao: If I recall correctly, in our early 2023 conversation, Wenfeng Liang's thinking at the time was hoping to build an OpenAI-like organization domestically, not targeting commercialization for a considerable period. He also believed China needed more and more original innovation.

LatePost: During the Lightyears Away experience, which judgments and approaches do you still maintain in your current startup? Which would you adjust?

Cao: My judgment about people hasn't changed: recruit PhDs who have graduated 3-5 years ago or are about to graduate — they're immersed in the front lines, capable, at their peak learning state. Don't worry too much about whether they previously worked on NLP, vision, or speech. This may be consensus now, but each company realized it at different times. In 2023, Lightyears Away and DeepSeek were already recruiting this way. So these two companies had the most talent competition then.

Another lesson: the CEO's pressure is still enormous, so take care of your health and manage anxiety.

LatePost: Any methods for this?

Cao: For me, the most effective method is thinking about big questions at the life level — it actually pushes you to focus on the process itself.

LatePost: We discussed earlier that after DeepSeek exploded in March this year, you met with Wenfeng Liang again. This relieved your anxiety — why?

Cao: The outside world was very excited about DeepSeek, but he himself was quite calm. Actually, in that state, maintaining normalcy is extremely difficult. After meeting him, I also intentionally reduced attention to channels that might input noise to me.

Third, for core model companies, I believe they'll focus more on general Agent and coding capabilities in the near term.

"After Starting a Company, Suddenly Everything Felt Right"

"The other side of this is the extreme pursuit of personal growth. As Charlie Munger said: the most reliable way to get what you want is to deserve it."

LatePost: When Lightyears Away ended, at least to the outside world it seemed very sudden. How did you feel?

Cao: Not being able to continue forward, there was certainly disappointment. But then Teacher Yuan (Jinhui Yuan) and I had to handle many things, including how to smoothly transition people in the organization. So it took some time before things quieted down.

LatePost: That would be after summer 2023, when you began planning to found Sand.ai. What was that process like?

Cao: I was broadly looking at whether to start my own company or join another one. Eventually decided to start my own, looked at some directions, and decided on AI video generation.

This was a process of finding intersections: what excites you? Is there a suitable team to do it together? Is the space large enough? And does it still have massive room for improvement going forward?

LatePost: This resembles your research topic selection approach: important, with massive room for improvement?

Cao: Yes, AI video generation was still relatively early then; Sora (first generation) hadn't appeared yet. Its potential was enormous — on the technology side, and also commercially: it's a direction where unlocking different capabilities unlocks different scenarios and demands, sustainable for a very long time.

LatePost: Why not join another company?

Cao: After the Lightyears Away experience, it was hard to join another company. Wang initially found very right people; his people-finding method was right. The completeness from infrastructure to models to product to fundraising to commercialization was strong. We weren't inferior to other companies in any dimension. I don't know if this mindset was right, but my genuine feeling at the time was that I found it hard to decide to join another company.

LatePost: What is the appeal of entrepreneurship itself to you?

Cao: I found entrepreneurship particularly suits me — suddenly everything felt right. You can probably tell I'm not a very typical researcher. Rather than drilling extremely deep into one thing, I pay more attention to the big direction of a field, to people and organization. Though my research was decent, I didn't have a strong desire to pursue a faculty position.

During that time I was also constantly thinking: what kind of person am I? The core self-awareness was: ambitious. This is my underlying drive.

LatePost: What does "ambitious" specifically mean to you? Is it about accomplishing something, or about winning?

Yue Cao: It's about hoping to do something that has massive impact on the world. The flip side of that is the relentless pursuit of personal growth. As Charlie Munger said: "The best way to get what you want is to deserve what you want." Only when you can develop deep understanding of things and fully train your capabilities can you possibly do something with enormous world-changing impact.

Looking back at my decisions over the past 5 to 10 years, I found they were all driven by ambition. For example, when I saw OpenAI's CLIP and DALL-E, some people's reaction was "I can't do this, so I'll say it's no good." My first reaction was: Why can't we do it? Are we dumber than them, less capable than them? I don't think so. It's about how you organize. When I see people with similar skills to mine achieving extraordinary things, I want to be able to do something with comparable impact too.

So when I truly understood what entrepreneurship entailed, suddenly everything clicked: it requires someone with relatively well-rounded capabilities, its ceiling is extremely high, the range of what you can build is incredibly diverse, and the trials it puts a person through are hell-mode — it forces extreme personal growth.

LatePost: Did you choose AI video because it wasn't on the main axis of the core large model companies?

Yue Cao: Competition is one dimension, but it wasn't the main axis of my decision. Fundamentally, it's because the technology and commercial ceiling of AI video are both very high. At that stage, there weren't even good usable models. So it was very well-suited to a founder with my model background. Doing large language models directly — that timing was definitely too late.

LatePost: Sand.ai was formally established and began operations in January 2024, and officially released its first model Magi-1 in April 2025. Why did it take more than a year?

Yue Cao: Because we chose the autoregressive route. We still underestimated how difficult this path would be.

LatePost: Why did you choose autoregressive? Why is it harder to build?

Yue Cao: The academic community started working on autoregressive relatively early. After GPT emerged, some teams were already experimenting in 2020 and 2021. At the time, the understanding of autoregressive was using one model architecture to simultaneously process language, images, and video. But precisely because it had to be done end-to-end with language models, the initial results weren't as good as some traditional diffusion models.

When we used autoregressive for video generation in 2024, our thinking was that video data is naturally sequential — it plays continuously in order, just like language, which is also read sequentially. The training approach that maximally compresses language information is predicting the next token. For video, wouldn't the method that maximally compresses video information also be continuous sequential prediction? Even today, I still believe this intuition is correct.

But this approach was very new, with no good reference to follow. During exploration, we had to do enormous amounts of zero-to-one work across data, algorithms, infrastructure, pre-training, post-training, and evaluation — and every link had to be pushed to the extreme for the results to be good. This was a huge challenge for a newly assembled team, and we did face manpower shortages.

LatePost: How many people did you have when building Magi-1 from 2024 to 2025?

Yue Cao: Thirty to forty people. At that stage, on one hand, our sensitivity to the business side was insufficient — business still requires fast delivery and fast iteration. On the other hand, we somewhat underestimated the difficulty of building a novel model type. Ultimately this led to Magi-1's release being relatively late.

LatePost: Did this cause you to miss some business windows?

Yue Cao: Pure visuals, single-asset generation had already become relatively converged and competitive. The new capabilities that emerged this year are audio-visual co-generation and narrative.

The experience of training Magi-1 means the organization still has the capability to train a good autoregressive model. At the same time, in this new phase, we hope to do well on a video generation model centered on character performance capabilities. On the algorithm side, we'll target faster delivery and iteration — this is different from when we were building Magi-1.

The key now is to seize the window of audio-visual co-generation. For professional creators, audio-visual co-generation can produce very realistic characters. For general consumers, it enables low-threshold generation of good-looking, fun videos you can share with friends — consumable, social.

"Sora 2 is an end-to-end model, OpenAI is an end-to-end organization"

"OpenAI has achieved relatively good vertical integration from product R&D to models: after aligning on objectives, the gradient of product needs can propagate back to the model team."

LatePost: Just a few days before our conversation, OpenAI suddenly released Sora 2 and the Sora App. Its characteristics are audio-visual co-generation, plus some social features. I know you tried the product immediately — what was your biggest impression?

Yue Cao: At the model level, Sora 2 has three characteristics: audio-visual co-generation, character ID preservation, but what amazed me most was that it achieved basic narrative within roughly 10 seconds of duration, because it has shot-composition capability. Previously, the more mainstream approach to narrative was using agents — for example, one language model handles the script; one storyboard model handles storyboard-to-image generation, then image-to-video; then there's a corresponding model for soundtrack, and so on.

What's very innovative about OpenAI is that they used an end-to-end model to directly output a 10-second narrative short.

LatePost: Is this related to Sora 2 paying more attention to C-end users? Because more professional video content creators might need the agent approach to more finely control art style, shot composition, and sound.

Yue Cao: It might be the reverse process — with model capability improvements, they then thought of C-end scenarios.

Why did people previously lean toward agents? Because early models couldn't do audio-visual co-generation or shot composition. At that time, a video generation model was just one link in the production pipeline, still requiring humans to design script, storyboard, and soundtrack... Extending from this workflow, the natural thought was: can we use agents to replace humans?

But OpenAI's thinking was: why can't we have the model directly output narrative capability end-to-end? This is a major shift. Of course, its precondition is still model progress — a video with only visuals and no sound, ordinary people don't want to watch. The earliest to achieve audio-visual co-generation would be most likely to first realize that models could directly output a C-end consumable video.

LatePost: How do you think OpenAI achieved end-to-end narrative capability?

Yue Cao: This technical report was very vague. I speculate Sora 2 is still based on diffusion, but I can't determine whether it's bidirectional or unidirectional. However, technically, all these advances were foreseeable — there's no earth-shattering idea: Google's Veo 3 released in May could already do audio-visual co-generation; character ID preservation, the key is visual and audio consistency, which has also been a long-researched direction. Even for narrative capability, before Sora there were already models with storyboard functionality — Sora 2 just does storyboard better.

What truly gave me a cognitive update was OpenAI's organizational capability. I personally believe OpenAI has achieved relatively good vertical integration from product R&D to models: after aligning on objectives, the gradient of product needs can propagate back to the model team. The entire organization is "end-to-end."

When the product wants to achieve some capability, their first reaction isn't to get there through model composition or product-side R&D, but to ask: can we build this directly into the model?

Specifically for Sora 2, they may have aligned on an objective at some stage to make "narrative shorts," then the model team defined internal benchmarks, then achieved the goal through model-side optimization and released the capability.

I feel that a shared problem-solving philosophy at OpenAI is: if a problem can potentially be solved end-to-end, it should be prioritized to be solved end-to-end.

Actually, looking back at large language models, OpenAI had a similar thread: after GPT-3 emerged in June 2020, OpenAI's phased focus was aligning GPT-3 into an interface ordinary people could use. This was essentially vertical integration — looking from the demand side at how to use GPT-3, which led to InstructGPT, the entire post-training process, then SFT (supervised fine-tuning) and RL (reinforcement learning). This is the process of how you make a foundation model usable by people. (Note: InstructGPT, announced in January 2022, introduced RLHF — reinforcement learning from human feedback — on top of GPT-3, enabling the model to generate responses more aligned with human intent, habits, and preferences.)

This shows that OpenAI already possessed a product sense, or rather, an end-to-end optimization capability from model to product.

Gaga-1 Focuses on Solving Character Performance: The Biggest Bottleneck in Narrative Content

"Previously people battled with memes in group chats; when video generation becomes simple enough, they can battle with videos too."

LatePost: Sand.ai also released its new model Gaga-1 after the October holiday. What are its characteristics?

Yue Cao: Our new model can do audio-visual co-generation, focusing mainly on characters speaking and performing. Previously, video generation with only visuals could only satisfy B-roll in narrative films (like establishing shots, transitions, etc.), but not A-roll (parts with characters and performances). For example, generating visuals first then doing AI lip-sync looks very weird, hard to cross the "uncanny valley," and doesn't reach ordinary people's consumption threshold. Yet in the vast majority of videos people consume, more than half the frames are people. Many short drama and commercial filmmakers told us that the biggest bottleneck with past models was "characters look too fake, no performance."

So when we started building this generation of models around mid-year, our goal was to prioritize solving character performance well — this requires doing well on character consistency, emotional expression, audio-visual co-generation, and other capabilities.

LatePost: At that point in time, was this insight industry consensus, or a relatively unique judgment?

Yue Cao: Judging from results, not everyone was making this a priority at the time. Behind this is: are you starting from demand, or from technology? We believe you should start more from demand; at the same time, you need to be able to make technical judgments about what's achievable and what's not — this makes it easier to find the fit between demand and model.

Zooming in to the "creative agent" category, besides us, others are still relatively small. Overall, there are three types of promising companies in this space. First, the 2023 cohort doing multimodal generation, like Krea, Higgsfield, Freepik — they may all transition toward agents. Midjourney probably not, it's more model-focused. Then there are vertical industry giants like Adobe, Canva, but new products need to reach certain revenue scale, like over $100 million annual revenue, to be meaningful to them. Third are the core model companies — I believe they'll focus more on general agents and coding capabilities in the near term.

LatePost: Who are Gaga-1's users? What's the priority ranking?

Yue Cao: There are two categories. The first is professional narrative content creators — people making AI short dramas, advertising assets, and commercial promos. All of this requires vivid character performances.

The second is everyday consumers. When video can generate synchronized audio and visuals, it really hits a tipping point for "for fun" use. Before, people traded memes in group chats; when video generation becomes simple enough, they can "trade videos" instead. Emotions and feelings that couldn't be fully conveyed through text alone can now be expressed through video — it's a natural social distribution scenario.

Right now, professional users like short drama makers are the highest priority. Our Gaga online product, launching October 11, will let people fully experience the model's capabilities.

LatePost: When might you build a consumer-facing product?

Yue Cao: It's in the plans. We believe the right time is after the model has narrative capabilities. It will likely be an app, but we need to spend more time refining the positioning. It'll also be an end-to-end design from model to product. OpenAI has certainly provided a lot of inspiration.

LatePost: Your current priority users are narrative content creators, and they care deeply about cost. Where does Gaga-1 stand?

Yue Cao: Our use cases are relatively focused, so we've controlled costs well with this generation — less than one-tenth of Sora's cheapest tier. Before this, even when the performance wasn't great, the cost gap between making an AI short drama and shooting one wasn't actually that large. (Note: Sora 2's cheapest standard tier charges $0.10 per second for 720p video.)

LatePost: What did short dramas cost to shoot before? How much can your model help reduce that?

Yue Cao: Around 2021, shooting a 60-to-100-minute short drama cost roughly 100,000 RMB total. Then the industry got competitive, and by 2023 it was around 300,000–400,000 RMB, sometimes even higher. It's come down slightly now.

Using AI, the mainstream price was previously 2,000–5,000 RMB per minute, so a full drama ran roughly 200,000–400,000 RMB. But the problem was, even spending that much, AI short drama quality still couldn't match real production. Now we've gotten below $1 per minute with room for margin — the cost drop is massive.

LatePost: Gaga-1's release timing came right after Sora 2. How did that affect you? I also spoke with Malvin from Lovart about Sora recently; he believes the giants will accelerate investment in similar AI-to-consumer and social products because the upside is huge and the cost of missing out is too great. Does that squeeze out room for smaller companies to "lay low and grow"?

Yue Cao: Even without Sora, the video space was already taken seriously — it was already hard to "lay low and grow." But frankly, while big companies are investing heavily, they won't move that fast. Also, whether the Sora app is truly a consumer platform opportunity still needs observation. I believe a new consumer platform needs at least two direct conditions: whether it's a new content format, and whether there's a new distribution chain. So far, Sora hasn't shown disruptive performance on either front. It still feels more like a tool — many people will make videos with Sora, then post them to Moments, Xiaohongshu, Douyin, Kuaishou.

LatePost: What metrics will you watch to see if Sora can become a new consumer platform?

Yue Cao: The most critical is retention — can it first satisfy some people's long-term needs? Beyond that, can it become a large consumer platform? Nobody has the answer right now.

The Core of a Vertically Integrated Organization: People from Different Backgrounds Aligning on Context

"Product people understand models better; model people understand product better."

LatePost: From Magi-1 to Gaga-1, Sand.ai went through a shift from being more model-driven to doing "vertical integration" between product needs and model training. How did that happen?

Yue Cao: The core is that for a startup, achieving PMF (product-market fit) matters more. Being too technology-driven might make it hard to hit阶段性 user or commercialization goals; but going purely from demand risks falling behind technically. Balancing these two is the hardest part.

LatePost: Logically that makes sense. But why was it hard to do from the start?

Yue Cao: This stage is still fundamentally technology-driven. People who understand PMF, product managers, or commercialization folks need time to grasp model boundaries and development directions; and model people, especially the top talent, mostly haven't done product or business before. Both sides need to build the other's capabilities. I've personally spent a lot of time learning productization, commercialization, and organization-building. My current belief: you need a more "end-to-end optimized" organization — a vertically integrated one.

LatePost: With this realization, what changed for you and your team?

Yue Cao: The core was mixing model, product, and operations people together. Really just getting everyone to communicate more, to "align on context." For example, having the most critical people on the technical and product sides do one-on-ones with each other; making myself an important distribution hub — as someone from an algorithms background, I've spent the most time on product and operations this year. This way, at least everyone who talks with me has very aligned context, and others in the organization align through them.

I think the core of a vertical organization is this: people from different backgrounds gradually aligning on shared context. Product people understand models better; model people understand product better.

LatePost: How do you view "model as product" now?

Yue Cao: When people first said "model as product," they meant you didn't need to over-"polish" on the product and operations side — there was a sense that product should somewhat follow the model. Early on that might have been true: product was about maximizing model capability display, without adding too much product-side prior.

But vertical integration goes further. If product is designed well enough, it not only showcases model capabilities but can also amplify the model experience when capabilities are similar. Sora 2's Cameo is an example: this first comes from the model now being able to do synchronized audio and video, so you can insert people. But how does that become a consumer product feature? OpenAI did a series of designs: invite-code referrals, @-ing friends for collabs, Cameo authorization tiers. Model and product amplify each other. So at this stage, there's also a lot of room for product iteration.

LatePost: How are you practicing "vertical integration" yourself on Gaga-1 and its corresponding product?

Yue Cao: The new model's core capability is generating realistic character performances. Then, once that goal is achieved, which features matter most? One example is ID preservation — consistency in character appearance, look, and voice.

Behind these surface changes is still organizational vertical integration: from product to model, we first aligned on priorities like "character performance" and "ID preservation." If you were a pure product company only calling model APIs, you'd first struggle to adjust model optimization direction based on product-side feature priorities, and you'd also struggle to use the model side to optimize for unique user experiences that other products don't have.

From Researcher to CEO

"It's not about liking or disliking — more like there's an angle between the two roles that takes time to work through."

LatePost: I know that besides learning through conversations with people, you also enjoy talking with large language models. You've chatted quite a bit with Gemini.

Yue Cao: The world's strongest models are already incredibly capable — most people just haven't realized it yet.

The last time I clearly felt its capability was in May this year, when our team was discussing an algorithms problem. Someone threw the problem to Gemini 2.5 Pro, and it not only filled in the parts we'd missed in our discussion but proposed a pretty solid solution. We joked at the time that this was "Vibe Research."

After that I started discussing things with it regularly and found a few interesting things:

• First, human thinking often relies on analogy, but analogy is dangerous because two things are both similar and dissimilar. Language models are especially good at deconstructing analogies — they can accurately point out the connections, disconnections, and why, letting you see that "seemingly related but actually not" structure.

• Second, it's an extremely powerful structuring tool. When I'm thinking through a problem, I'll input scattered points to the model, and it can quickly organize fragments into logical structure.

So it's become my thinking partner, "Professor Gemini." I might spend an hour or two daily chatting with it about various issues. It constantly throws out new perspectives, helping me understand things more fundamentally.

LatePost: Do you trust large models more than people? For example, when a person and Gemini tell you the same point, how do your reactions differ?

Yue Cao: The hardest thing between people is aligning context. When I express an idea, there are massive premises behind it, but I can only compress it into a few dozen words; the listener can only receive those few dozen words. To truly align, we'd need to spend half an hour, an hour filling in background — exhausting for both sides.

Language models are extremely strong here. For example, when product and algorithms people discuss, a paragraph might contain three or four concepts requiring background knowledge; the algorithms person can send that paragraph to the model, tell it the speaker's background, and have it fill in the context to aid understanding. I lead by example in doing this.

So it's not about trusting one more than the other — it's that models can dramatically reduce friction in human communication, bridging cognitive gaps between people. This is a foundational transformation that most people haven't realized yet.

LatePost: Could this itself be made into an AI application?

Yue Cao: I know some friends working on similar things, like AI dating assistant apps — a guy screenshots a chat and asks the model: "What does she mean by this?" Women also ask "What's he hinting at?" The essence of this scenario is helping people fill in missing context.

Context gaps relate to two things: one is background differences between people — gender, culture, upbringing; the other is the topic being discussed, like algorithms people hearing product terminology that seems obscure, or vice versa. Language models hold nearly all public knowledge in the world, so they can understand what another person is saying faster than most people can.

LatePost: What have you gained from recent conversations with real people?

Yue Cao: You can probably tell by now that conversations with people aren't as enlightening as with Gemini (laughs).

Actually the most direct gains came from potential customers. I chatted with a veteran from the film industry recently, and he said something that stuck with me: "The endgame of all content is narrative." Narrative is the fundamental structure that makes content work: short videos started as "recording a beautiful life," then evolved to competing on viewing experience, even if just 15 seconds; including being an up主, persona-building is also a form of narrative.

LatePost: Your WeChat signature used to be for a while: "Think about how Ilya thinks?" How have your learning targets changed?

Yue Cao: That was from when I was deep into studying OpenAI. Now I think more broadly — Huiwen Wang, Yiming [Zhang], Xiang Li, Lei Jun... abstracting the angles from which they view problems. At first you might only grasp one layer, but you can keep validating that in practice.

LatePost: What did you learn from Huiwen Wang when you started a company together?

Yue Cao: The first time I talked to him about wanting to do AI video, he gave me one piece of advice: go study Pixar — that company has a great business model. The insight from Pixar is that it started from graphics technology to make films, but box office is only the first step. What's more critical is that the character IP stays with the company, enabling long-term operation and monetization. With live-action films, the character's mindshare gets carried away by the actor — production companies can hardly accumulate IP assets sustainably. So Pixar has both the narrative production attributes of the film industry and the sustainable revenue attributes of the IP industry, and all of it is technology-driven.

LatePost: In the transition from researcher to entrepreneur, what parts do you enjoy? What parts do you dislike but have to adapt to?

Yue Cao: It's not really about liking or disliking — more like there's an angle between two roles that takes time to grind down. From initially focusing more on what technology has greater impact, to what systems have greater impact, to focusing more on product and commercial value.

LatePost: What's your highest-priority work right now?

Yue Cao: Understanding the AI video direction clearly, identifying the biggest opportunities in the coming period; building the Sand.ai organization to capture those opportunities.

LatePost: Are you already seeing some signs?

Yue Cao: The core logic of AI video is the evolution of capabilities across language, image, video, and audio models. In the near term it's still more of a tool-type opportunity. But as models become increasingly stable and output quality keeps improving, consumer-facing opportunities may emerge. This should be consensus among people who've really thought about the industry.

Where there's no consensus is timing — that's also the hardest thing to judge, and I can't say too much about it now. Essentially, you need to prepare in advance for that direction after you've judged the timing. This also tests an organization's sense of rhythm and execution capability. Even if you judge the direction right, if the organization's rhythm can't keep up, you'll miss the opportunity. That's why I'm so focused on organization now.

LatePost: Final question — share something you've been thinking about recently, or a question you want to validate going forward.

Yue Cao: I sometimes think about an open-ended question: when will ASI (Artificial Superintelligence) emerge? When it truly arrives, how will it transform the world? The intelligence level of language models is actually improving continuously — if we roughly analogize to human IQ curves, it's already gone from several tens to one hundred, and may reach one thousand in the future. The gap between it and humans might be as large as that between humans and monkeys. The intelligence difference between an adult and a child is already enough for the former to "handle" the latter. So how would a system with intelligence far exceeding human levels view and influence human society?