After continuous iteration of large models, product competition shifts toward scenarios and experiences

Author: Frank, PANews

As AI gradually shifts from flashy demonstrations to practical applications, the deployment of AI solutions is accelerating to meet the growing consumer demand. Meanwhile, with the continuous enhancement of large model capabilities, AI seems to have entered an era where “everyone can create product prototypes.”

During muShanghai AI Week, the roundtable discussion hosted by PANews, “Innovative Practices and Path Exploration in the AI Consumer Ecosystem,” focused on the real deployment pathways of consumer-grade AI products. Participants included Feng Wen, head of the MiniMax open platform product; Levy, CEO of FateTell; Anita, head of Sentient APAC; and Gao Jiafeng, electronic musician and independent developer, representing different fields such as model open platforms, cultural outbound applications, open-source AI ecosystems, and music creation practices.

In their view, the core issues of consumer-grade AI have not become simpler due to technological iteration. After the leap in model capabilities, the true barriers are shifting toward scene understanding, data organization, user education, emotional value, and the construction of open ecosystems.

AI has not lowered the difficulty of entrepreneurship; the real barriers still lie in application scenarios

A common contradiction in the AI industry is: as models become more powerful, the threshold for entrepreneurship seems to lower, but many products struggle to find long-term viable scenarios. Applications that seem feasible today may quickly lose relevance with the release of the next version of models.

Feng Wen believes that for consumer-grade AI products, idea generation and scene judgment remain more important. As the provider of large models and open platforms, MiniMax emphasizes underlying model capabilities, token-related product design, and end-to-end developer experience. But from an entrepreneur’s perspective, products should be designed based on “the model’s intelligence level after six months.”

His judgment is that, given the ongoing scaling laws of models and continuous improvements in capabilities, entrepreneurs should not be overly constrained by current model speed, cost, or capability limits. Instead, they should think more boldly about target users, specific scenarios, and problems to solve. Model providers will keep offering cheaper, faster, and more cost-effective capabilities, while application layers need to clearly answer “why this scenario.”

Levy adds from the application layer that another source of barriers exists. He believes that while technology changes rapidly, the data and understanding associated with scenarios are not quickly erased. Many previously thought that only fine-tuning models could create data barriers; but with mature context engineering and prompt engineering, the data and structures accumulated in context management can also influence model performance. Especially for highly vertical, culture-related, or personalized data, which may not enter the general model weights, this can become a differentiating foundation for consumer AI products to resist model iteration.

Anita offers a more cautious view on “AI lowering the entrepreneurial threshold.” She believes that AI indeed makes generating demo samples, building prototypes, and quickly launching initial products easier, but the core challenges of entrepreneurship—such as customer acquisition, community engagement, commercial implementation, and establishing human connections beyond programming—remain. She mentions that the concept of super-individuals and “one-person companies” is currently popular, but truly capable individuals often need more diverse skills, not just calling large models.

From Bazi to Music: Better Understanding Users as a Barrier in Consumer AI

As technological capabilities advance, the value of consumer-grade AI products ultimately returns to human needs.

FateTell provides a typical example. Levy explains that FateTell is an AI + Eastern astrology/Bazi consumer app aimed at overseas users, with users in over 90 countries. The team initially avoided pure efficiency tools and instead focused on spiritual consumption and emotional value.

He sees understanding one’s destiny, seeking explanations and comfort as fundamental psychological needs that cross cultures and have persisted over time. AI has historically struggled to build trust in this scenario, but with the improved capabilities of models like DeepSeekR1, objectively, users and investors can understand that “large models can perform complex reasoning and explanations.” The barriers FateTell faces are not just about model capabilities but also about how to translate Chinese cultural concepts like the Heavenly Stems and Earthly Branches, the I Ching, and Bazi for overseas users, and how to use language, visuals, and interactions to help people from different cultural backgrounds appreciate their appeal.

Gao Jiafeng raises a similar question from the perspective of music creators: AI should not only deliver results but also preserve the process. He mentions tools like Suno that make music generation very straightforward, but they skip the creative process, leading to a lack of participation and a sense of belonging for users. For musicians and ordinary users, creation is not just about producing a “finished song”; the process itself is part of the experience.

He uses playing football as a metaphor: even if ordinary people can never surpass Messi or C Ronaldo, they still play out of love. The same applies to music creation. Gao is developing MusicAIGameBoy, a music AI gaming device that uses large or small models to generate music code, combined with gamified interactions, allowing even those who don’t understand music to participate in creation through gameplay. For him, the real scenario isn’t “automatically generating a song,” but returning the interactive process of music creation to users.

The Rise of Agents and Changing User Education Logic

In consumer-grade AI products, user education often determines whether the product can be truly used.

Feng Wen mentions that among users of the MiniMax open platform, some have development experience but are still hindered by API documentation, parameters, error codes, and token usage. To address this, the platform provides model trial platforms, development guides, demo cases, and video tutorials to help developers move from understanding to calling more quickly.

With the development of Agents, the way user education is conducted is also changing. In the past, users needed to read documentation, understand interfaces, and troubleshoot errors. Now, with upgraded agent performance, many users rely on the agent to read documents, search for solutions, select appropriate models, and automatically correct paths. Model providers need to ensure good model, documentation, and platform experiences, while communities, developers, and various product forms will collectively lower the usage barriers.

For Sentient, the open ecosystem itself is also part of user education and product deployment. Anita explains that Sentient focuses on open-source AI ecosystems and related infrastructure, gathering developers through hackathons, funding programs, and other initiatives. She emphasizes that products must first clearly identify their target users: who they are, where they appear, and how to build trust. For developer tools, hackathons and ecosystem collaborations are effective entry points; for consumer products, KOLs, KOCs, and social media content are equally important.

Against the backdrop of rapidly decreasing AIGC costs, startup teams can produce trailers, visual materials, and promotional content at lower costs, enabling faster acquisition of initial users. Gao Jiafeng also believes that product design should be closer to users, allowing them to learn naturally through interaction and entertainment rather than relying on extensive manuals. This “learning through use” approach may be more suitable for consumer AI than traditional tutorials.

Hardware Enters the Real World, and Personalization & Emotional Value Continue to Amplify

In the next three to five years, the guests generally agree that the consumer AI market is still in early penetration, but product forms will undergo significant changes.

Feng Wen predicts that in three to five years, intelligent hardware, robots, and embodied intelligence will reach important inflection points. As model capabilities improve, AI will no longer be confined to software interfaces but will also enter the physical world, performing more interactions and tasks. Some products will target humans, providing efficiency boosts or emotional value. Others may serve as agents, offering environments, tools, and infrastructure to connect AI with the physical world. Regardless of form, products should ultimately keep humans at the center, allowing more time for human-to-human connection, family, real-world experiences, and richer lifestyles.

Levy believes that predicting three to five years ahead in AI is already very difficult, and even three to five months ahead is uncertain. While advanced users are deeply using tools like ClaudeCode, most ordinary users are still in early AI adoption stages. In the coming years, AI will further satisfy more fragmented and personalized needs. Compared to the relatively “one-size-fits-all” services of the mobile internet era, AI has the potential to offer more specific and niche services for each individual. Additionally, the anxiety over unemployment and uncertainty caused by technological progress may further amplify the demand for psychological companionship and emotional consumption.

Anita summarizes this change as “technological democratization.” She believes that in the future, distinctions between liberal arts, science, arts, and technology will weaken. A small vendor might use AI to create ads and targeted push notifications to improve their business. The value of AI is not necessarily to make everyone a top-tier programmer but to help people in various life scenarios access better tools. At the same time, fears of unemployment and loneliness will drive up emotional value needs, giving more opportunities to hardware, AI pets, companionship devices, and multisensory interactive products.

Gao Jiafeng approaches from the perspective of cultural shifts. He thinks that future content forms like music, movies, and videos will be reorganized, and even whether “songs” remain the smallest unit of music consumption is uncertain. Current concepts like multi-track audio and stems may continue to be broken down into more atomic creation units. But as formats dissolve, emotional connections carried by IP, brands, and specific personalities will become even more important. People no longer always seek perfect works but rather objects with flaws, warmth, and the ability to build emotional bonds.

Although the guests did not provide a unified answer for consumer AI, the discussions across different fields—model platforms, cultural applications, open-source ecosystems, and music creation—point to a common trend: as model capabilities continue to improve, competition in consumer AI will shift from “who calls the more powerful model” to “who understands more specific users, real scenarios, and emotional needs.”

The future consumer AI ecosystem may include stronger open infrastructure, lower development barriers, more personalized services, more companion-like hardware, and new product forms centered around culture and creation processes. Models will keep evolving, but what will truly endure are products that are needed by people, understood by people, and capable of establishing human connections.