Wang Guan on Video Generation, Data and Changing Content Platforms

Zhang Xiaojun interviews ONE2X co-founder Wang Guan about AI products, video generation, teams and relationships between creators and platforms.

Wang previously worked on AI products, including at Moonshot. He describes ONE2X as a studio exploring AI products.

He says the studio emphasizes exploration, product quality and individual initiative, hoping people with different backgrounds can take responsibility for tasks aligned with their interests. This describes the team’s approach at the time.

Recalling earlier vision work, he contrasts structured labels with richer unstructured data and says changes in data shape possibilities for models and products.

He sees a continuing role for product managers: understanding a business, translating domain knowledge into model-usable context and workflows, and assessing whether users benefit.

He recalls experimenting with GPT-3 writing assistance before ChatGPT appeared, after which the product environment changed quickly and required reassessing which features foundation services might provide.

The team also explored workflows that organized data sources, prompts and task steps as nodes. He says abstraction should help users rather than merely add configuration complexity.

On “compression is intelligence,” he connects data compression with generalized representations. This is his way of thinking about models, not a settled account of intelligence.

He recalls joining Moonshot in 2023 and leaving to start a venture in early 2024, valuing the team’s shared goals and collaboration with researchers.

Before leaving, he considered where model capabilities come from, how they might develop and how application teams relate to foundation models.

He proposes stages of data: public data, domain-specific data and data generated through product use. This is a personal framework, not a fixed boundary.

In his view, different stages require different mixes of talent, compute, domain knowledge and user relationships, and may favor different organizations.

He discusses the need to clean, understand and validate domain data; possessing data does not guarantee a reliable model or remove privacy and consent responsibilities.

He says early products may look similar, while team goals, data and implementation can eventually differentiate them. Such differentiation still needs market validation.

Wang recalls that video generation was immature when he began his venture in 2024. His team chose the area to seek product opportunities around creators’ workflows.

Video generation involves more than a single image: temporal consistency, control, speed and cost matter. A usable product must connect these capabilities to creative workflows.

He compares video models at that time to an early phase of language-model development, seeing remaining work on data, capability and inference efficiency. This is his assessment of the stage.

He discusses the difference between using a reference video and generating a new one, noting that converting visual information to text and back may lose detail.

Discussing software, the internet and AI, he suggests production, distribution and consumption of information goods may change, and examines how power could shift among them.

He distinguishes what he calls broad and narrow AGI and imagines systems recognizing what they do not know. Conclusions about consciousness or capability remain contested.

He thinks the boundary between model and application companies may blur as foundation providers build products and application companies invest in particular needs and data.

He predicts that some application companies may train models to control cost or improve performance, though whether it pays off depends on resources, markets and technical choices.

Using specialist information as an example, he says organizing reliable sources and understanding workflows can create product value, while expert verification remains necessary.

Wang uses “generation system” for an approach that creates content in response to demand, while recommendation primarily matches users with existing content. Real products can combine both.

He says understanding intent and providing context affect generation, while more personalized content also raises questions of accuracy, sourcing and user-data protection.

From a creator’s perspective, they discuss how changes in production costs, distribution and revenue sharing could affect who creates and how creators are compensated.

Wang anticipates that new content platforms may emerge while acknowledging his team does not yet have a complete consumer-side plan.

He describes his hope for generation systems as removing an intermediary’s markup. This is a view of future industry structure, not proof that intermediaries, costs or creator-rights questions will disappear.

He discusses AI products beginning as capability aggregators, editing tools and other forms, with the key question being whether they help users complete concrete work.

On organization, he says the company then had about 30 people, allowed remote work and retained optional in-person interaction. Headcount and practices reflect the interview’s time.

He values initiative and independent problem solving while saying remote teams need clear communication of goals, responsibilities and feedback.

Looking ahead, he imagines platforms participating in both content production and distribution and users generating new content while consuming. These are scenarios, not settled outcomes.

He uses human-AI research as an example: content produced while solving one person’s problem may help others, though sharing still involves permission and context.

Finally, he mentions language-modeling research that influenced him and admits he does not understand all of its mathematics. His connection between compression and intelligence remains a personal interpretation.