Unique Research · Unique Viewpoint · April 7, 2024
Background
On January 24, 2024, Unique Capital and partners held the 2024 AIGC Application Development Summit in Shanghai, with nearly 1,000 AI practitioners attending. This piece is edited from the guest keynote / roundtable segment.
Guests
- Qiu Yiwu (邱懿武) — CEO, Zaowuyun / FabCloud (造物云)
- Chen Dabo (陈达博) — CEO, Fabrie
- Zhou Ze’an (周泽安) — CEO, Biyou Tech (必优科技)
- Zhao Lu (赵鹭) — China Lead, Collov
- Jiang Yuchen (姜昱辰) — CEO, Waveform Intelligence (波形智能)
- Moderator: Hu Xiaojing (胡小婧) — Partner, Unique Capital
Roundtable
Company intros & 2023 milestones
Q: Please briefly introduce your company, the track you focus on, and a quick look back at key developments in 2023.
Chen Dabo, CEO, Fabrie: Fabrie is a product for the design industry. We are live in nearly 120 countries and regions, with hundreds of thousands of professional design users. We focus on internet-facing design — traditional graphic, industrial, apparel, interior/architecture, and game design — so the product is relatively general-purpose.
At the core is an online whiteboard in a visual-document form. Combined with AI, it helps design and creative teams express ideas more easily in digital space.
Designers and teams typically use whiteboards for long-cycle workflows that generate a lot of data — communication process and design-context. That context is precious for AI. In past design/creative workflows, those data were often never captured. A whiteboard is an ideal carrier: it records the full workflow and helps AI raise designers’ productivity.
Zhou Ze’an, CEO, Biyou Tech: Biyou focuses on AI-generated document services. We use foundation models plus industry models to build industry document agents, offering end-to-end document services from generation and beautification to assisted writing, presentation, and motion effects.
Mention “documents” and people think of Microsoft and Kingsoft. Our team spun out of the core creators of Kingsoft’s AI document project — among the earliest domestic teams exploring pure content generation. In 2017 we started the AI Modou project that became today’s WPS AI, with strong user and commercialization validation. Later, for business-direction reasons, core members founded independently.
Even before the AI boom we had built many vertical document models. Document capabilities behind many large companies are still powered by us — AI résumé writing, AI contracts, and similar niches. Tencent Docs, Kingsoft Docs, iFlytek, and others have integrated.
After 2020, foundation-model generalization jumped. In December 2022 we launched UTalk-Doc as a framework project; its first ship was generating “presentation documents.” From Dec 2022 to Mar 2023 we built about 13 sub-models.
Though people now call this a “document Agent,” the essence is document automation. In March we shipped ChatPPT — the name may sound collision-prone, but we may have been the world’s first ChatOffice product, launching 12 days before Microsoft Copilot. It gave the team new labels and partly redefined how document agents are built.
Zhao Lu, China Lead, Collov: We are a Silicon Valley company founded by several Stanford PhDs in vision and machine learning. Our business combines AI with living/space.
The product has gone through three iterations. First we focused on AR room measurement. In North America many students and DIY users need measurement, but professionals are expensive — so we built a small app that scans a whole house from a phone via the web and returns dimensions. By 2021 we enabled online furniture soft-furnishing and basic renovation AR. We also invented “phone scientist” capabilities with underlying vision-foundation-model tech.
Our data volume is 2 billion images; daily processing about 5–6 million. We cover 60 countries and regions: overseas mainly toC with 1.7 million users; China mainly toB. Many listed and near-listed companies at home and abroad use our underlying tech. Since entering China we have hundreds of thousands of registered users, and nearly a thousand enterprise users have inquired about API access.
Jiang Yuchen, CEO, Waveform Intelligence: Waveform focuses on large models for creative writing. On our proprietary Weaver model we ship four sizes — Mini, Base, Pro, Ultra — with Flash Elastic especially notable.
Most models sit in an agent framework, but many real scenes are simple and only need small models. We provide a full routing-model system, so API cost is very low: tasks are dynamically assigned to differently sized models for the creative-writing job at hand.
On Weaver we offer: (1) API for developers and app builders, priced roughly ¥0.008–0.02 per thousand tokens — an affordable option; in creative writing we are faster and better than GPT-4 and beat all other Chinese large models. (2) Customized enterprise writing agents — e.g. influencer marketing copy. (3) A C-end product, Wawa Writing (蛙蛙写作), for professional creators and writers at about ¥29/month.
LUI after ChatGPT — observations from the product side
Q: ChatGPT opened a new LUI (language user interface) form. Combining with your own products, do you have distinctive observations or discoveries?
Qiu Yiwu, CEO, Zaowuyun: I originally trained as a designer. Designers worldwide face a similar predicament: more than 99% are treated as “a pen” for the client — external or internal. One of NLP’s greatest values for design is freeing designers from that pen role. Some design work need not be done by designers; bosses or clients can express it themselves.
Industrial specialization made many professions highly specialized — design and painting as professional expression. What change will NLP and multimodal bring? With imagination and basic logic, clear natural-language expression may produce work better than a designer’s.
For enterprises, the help is accelerating visualization. Past communication cost caused huge internal friction in new-product R&D. Now visualization plus more natural language can reach consensus — revolutionary for many scenes.
Chen Dabo, CEO, Fabrie: Words do not always convey clearly — a major challenge for design.
Many think LUI is powerful, even the next-gen interface. I look from another angle: it may need to combine with existing interaction modes.
When you need to tweak one detail in a product doc, you usually just edit by hand. In a ChatGPT-like LUI you must describe the change in natural language — sometimes harder and higher cost than direct manipulation.
From product design, LUI is elegant and sharply cuts interaction cost while fitting foundation-model mechanisms — strong in its domains.
Recently Runway shipped a dynamic brush: natural language plus interaction so users can tell the computer which direction an animation should go — pure language would cost more time. I hope in 2024 we see tools that fuse these two interaction modes and raise productivity further.
Zhou Ze’an, CEO, Biyou Tech: I firmly believe in LUI’s potential and insist on 100% transforming toward it. Everyone knows Office — take PPT: people think design and layout are hard, but the bigger challenge is the complex tool system underneath. PPT has 1,765 features; most people know little and Google when stuck. We used to think tools should be as simple as possible, yet many office scenes need seven days of classes to learn a tool. Traditional Office even has MOS certification and global MVPs.
In the GUI era, complex learning was somehow acceptable. As user needs expand, we not only need to generate and layout documents — we want it fast and easy. On top of a huge feature set, LUI offers a more natural input. Sticking with LUI matters: it lowers the barrier and simplifies flow for every industry.
Tool products also hit a common disease: when stuck you can only passively learn the tool’s way — but we really want to solve the problem itself, not explore tool manuals. Learning a tool first to solve a problem is itself a problem; the logic must change.
Another trap in building tools: first insight needs, then define UI and interaction. The biggest issue: needs may not come from our insight — let the product meet users and let users tell you. Unlike interviews or surveys (where public answers come with psychological limits), when there is only a box, users pour in every efficiency-blocking need. As the product iterates, LUI almost never fails to meet demand — user-led input truly expresses good needs.
Our team’s first LUI version took two months to build intent models for all commands, covering needs with 180,000 data points. A week after launch, overall adoption was only 35%. With user permission we used one month of usage data for desensitized referential labeling/training. Sampling rate approached 90% and kept rising. Interaction with users can surface surprise and treasure-like needs — that matters a lot.
Building tools means insight into needs, but slightly changing the traditional insight process can yield new experience. Via LUI our product is closer to customers and breaks terminology limits — less burden and risk for us too.
Zhao Lu, China Lead, Collov: Having moved from landscape/space designer to Alibaba ops to PM, I know how painful design was for a long time before these tools.
In university, 3D or Photoshop was not LUI logic; masters vs beginners closed gaps with tedious action commands — similar to today’s prompts, which save a lot of time.
Entering China with Collov AI, I made key adjustments for the local market. First, simplification. I believe companies that break industry barriers this year will not be those founded only in 2023. Over two and a half years we iterated hundreds of times, including three major UI refreshes. As features grew, how to slim the product to fit LUI logic became a challenge.
In China our first job was stripping non-core features and keeping essential design tools — our idea: “if you can take a photo, you can design.” That also matches AGI pursuit: not only raise professionals’ professionalism, but help novices cross the barrier of professional tools.
Then we introduced a “creative aperture” — turning old university action commands into a creative aperture as a portal for user creativity. Like photography’s countless aperture settings, good photographers adjust aperture for their creative intent. We became among the first Chinese teams to run a data-flywheel strategy — continuous product iteration driving product improvement.
With those two, most users of our Chinese product can quickly find a method and reach expert-level results. At an Alibaba AIGC contest, a judge said our product could reach a 6–8-year professional space designer’s level in a week — actually, if you truly use the tool, you can find that feel in 6–8 minutes. That is LUI’s inspiration for us.
Jiang Yuchen, CEO, Waveform Intelligence: LUI itself simplifies complex workflows. We LLM builders once faced a big problem: models can hit 80 in many scenes, but hitting 100 precisely in a vertical is hard. Over years of accumulation, models can now hit 100 per scene — but users often don’t know which scene or feature to use.
What LLMs essentially do is routing — route user intent to the most suitable tool or model; in trendier terms, an “agent.” We try to recognize novice intent and assign it to very specialized tools or models — a huge help not only for consumers but for whole enterprise workflows. So LUI and agents both play a routing role: simplifying complex workflows.
Who wins as competition intensifies — productivity-tool trends
Q: In your industries, as competition intensifies, what kind of company will win? What are the development trends for productivity-type tools?
Jiang Yuchen, CEO, Waveform Intelligence: In the foundation-model era, as AI advances, general models get stronger. Whether traditional or AI, every company should stick to “stay focused, stay vertical.” In the AI wave, tools will multiply and ecosystems will grow larger in front of everyone.
In the past internet era, Chinese companies tended toward monopoly; in AI, monopoly is impossible. We will see a larger ecosystem where each firm operates with win-win cooperation. Tool products are highly specialized and fragmented; each has advantages in different scenes. Only when those advantages link up do you get better synergy.
Zhao Lu, China Lead, Collov: Last year, starting from the tech front line and the bottom layer, I found: first, you don’t need to obsess over surface things — the spare time becomes real competitiveness. Second, on product/ops: last year’s creative ideas or promo films could attract a batch of users; this year that clearly doesn’t work — the game is more complex. This year is definitely one that demands more ops focus.
Also, though people may think AI already enables easy “super individuals,” it isn’t so. To build a platform at million or ten-million users still needs fine division of labor. Overseas you often hear two people shipping a product for $1B revenue; domestically, better have two people serve one big customer well, or ten people first earn ¥1M — more pragmatic.
Zhou Ze’an, CEO, Biyou Tech: My career has stayed in documents; with the AI era my mindset has stayed steady. As early as 2020 we started vertical small models — e.g. our gallery framework UTalk and ChatPPT. Those weren’t sudden pies from the sky; we didn’t have super insight — just heads-down extreme focus.
As AI foundation models keep rising, only continuous Focus turns quantity into quality; help from large models is not instant — you accumulate base and keep focusing.
Also focus on yourself — find what you’re good at; for me that’s documents. Only by focusing on yourself, not being swayed by competitors, can product quality stay stable. Focus on yourself and there will be opportunity.
Chen Dabo, CEO, Fabrie: Application-layer companies that dig deep into engineering can build tech moats. Take the whiteboard: users see a simple white board, not the engineering years behind it. Last year many AI-app, text-to-image, image-to-text companies sought to partner with us to show generation results on the board — an ideal space to view and iterate. Some considered building their own whiteboard and quickly found matching our UX in a short time nearly impossible. For companies like ours, focusing on our specialty — the accumulated result — is the real value of focus.
But tool/productivity companies easily fall into a trap: only satisfying Maslow’s most basic physical needs. To build higher moats you must further meet emotional and spiritual needs — higher thresholds, longer runway. Everyone is too focused on cutting-edge tech iteration; as product people we can look further — explore higher-level needs and which product forms meet them.
Qiu Yiwu, CEO, Zaowuyun: AI tool development seems more and more like FMCG — assemble and ship. Looking a few years out, besides middleware and infrastructure products with deep engineering value, other products will face fierce domestic competition. For problems that need little code or just engineer stacking, giants like Alibaba and Baidu — with huge advantages and high KPI pressure — often win. Most domestic app tools may eventually lose to those giants.
After deep thought, I believe business essence returns to Buffett’s core moats — IP, brand, scale effects, network effects, and customer switching costs. On that recognition I cut our middle tools and focused on two keys: (1) ensure model quality is excellent — a strong model is itself the best tool; (2) use lightweight tools, leverage AI, apply to new customer scenes.
For example, many Fortune 500 CEOs want to understand AI’s real-application potential. They may take training courses, then try lightweight apps and seek new business via exchange — similar to when SAP launched ERP: enter new application scenes with new cognition and tool/AI integration. Success hinges on two points: unique asset capability; and embedding into the customer’s procurement system to build solid partnership. Only then do firms gain sustained commercial opportunity.
— END —