A Framework for Thinking About AI Application Startups in the Age of GPT | VOICE
Operating systems are a game for the few, but application startups are open to many.
3,318 words, 8-minute read
Author: Bingjian Wu
Operating systems are for the few, but application entrepreneurship is for the many. Windows's greatest contribution was birthing the browser as a super-app; the browser's greatest contribution was spawning super-apps like Yahoo and Google, kicking off the internet era. The history of information industry evolution is one of continuous value-chain migration and upward float. All consumer-facing products compete for two things: share of user time and share of user wallet. Whoever sits closest to these two wins as the killer app and captures value from the underlying platform.
GPT is, without question, operating-system-level. The first killer app to emerge atop it is chat search. Built on this foundation, growing ten more apps at the scale of Baidu or Douyin is entirely possible — triggering a redistribution of user time and wallet share. As an early-stage investor, many friends around me are considering jumping into AI application entrepreneurship and come to me to discuss which direction to choose and what hand to play. Here I'll share my thinking framework.
Framework One:
Super-apps emerge from new platform capabilities
From PC to mobile, the big opportunities came from new smartphone capabilities. LBS enabled Meituan and DiDi. The address book enabled WeChat. The camera enabled Douyin and Kuaishou. Voice enabled WeSing. The swipe gesture enabled ByteDance.
Compared to the phone itself, WeChat's new capability was an address book expanded a hundredfold — more contacts, point-to-many communication. This enhanced capability made group buying, distribution, and bargaining possible, with Pinduoduo as its largest application.
What new attributes does this generation of AI bring? My preliminary thoughts:
First: language, natural-language interaction. Windows and Android used graphical interfaces because the technology simply couldn't make machines understand human speech. Now it's possible. Based on "listening to human speech" as an interaction mode. From mouse movement on PC to swipe gestures on mobile, my guess is that future natural-language-based interaction will drastically reduce interface pages — most operations completed in a single page, plain-language commands calling up various results, even one page pulling from multiple apps. Extreme simplicity for the user; complexity buried in the backend. AI-native applications may not even take app form. PC's key metric was page views; mobile's is DAU; AI applications' might be "call volume."
Second: Generative, real-time generation. From generating text and code to images, audio, video, and eventually robot movements and preliminary scientific research — outputs that can be generated grow ever more numerous and complex. Abstractly: x-axis is media form (text, audio, image, video), y-axis is presentation form (digital 2D, digital 3D, physical 3D), z-axis is industry scenario (advertising, e-commerce, gaming, film, programming, design, research, robotics...). The z-axis alone has hundreds of categories. The key is finding a valid point within this xyz space — that's true PMF. "Video-digital 2D-advertising" points to AI-generated video ads. "Video-physical 3D-robotics" points to AI-generated robot movements.
This brings two changes. One, tenfold efficiency gains: AI-generated video ads, AI programming are ten times faster than traditional approaches. Two, unlocking the previously impossible: past AI education couldn't be truly AI, resorting instead to exhaustively pre-producing video素材, cutting into fragments, stringing together via logic trees. Now truly AI-native education becomes possible. Looking ahead, video might see a file format beyond MP4 and AVI — based on initial materials and real-time interaction, generating subsequent video on the fly, which in turn unlocks many industries.
Third: multimodal perception. Each information source is a modality; AI can simultaneously perceive language + image + many others — input becomes extraordinarily rich. Humans are the ultimate multimodal perceivers: hearing, sight, touch, smell, taste. So many inputs, real-time and integrated, enable precise situational judgment. PC input was mainly mouse and keyboard. Mobile input was mainly touchscreen, location, gyroscope (movement). In the AI era, large models can real-time perceive our language, images, and video while encompassing all PC and mobile perception forms — simultaneous perception, integrated judgment, beginning to approach human perceptual levels. For example, in a roundtable meeting, through microphone and camera, perception of people, content, and environment could help distill core viewpoints, provide partial answers to questions, and suggest prompts — making the meeting far more efficient.
Mobile compared to PC, due to its portable nature and added sensors, collected data orders of magnitude greater. AI-era data volume may rise several more orders of magnitude, as all language and images become valid input. What energy this contains is hard to say now.
These are my preliminary thoughts on AI's new capabilities, still being refined. I believe next-generation AI applications will grow from AI's new capabilities. And the next-generation smart terminal will likely be unlocked by these AI capabilities, with AR as the most probable candidate. Wearing AR, its input synchronizes with our eyes and ears; its output can generate in real-time; its interaction can be natural-language-based. AI's three new capabilities unlock three previously impossible aspects of AR.
Framework Two:
Playing open cards or hidden cards
The above discusses direction selection. With direction chosen, you still need to assess which card game your background suits.
Today any idea easily thought of is an open card — laying everything on the table to compare experience points, combat power, and health bars. Virtual assistants, virtual teachers, AI customer service, copy generation — these most-discussed directions are obvious. In such directions, masters and veterans abound; merely playing an "early mover" card rarely wins.
Take AI education: education is a long-chain industry, from user acquisition to course operations to R&D to sales, each with established methods. Ultimately it's about "turning screws" efficiency. This industry has many masters — people who've raised over $100 million, managed thousand-person teams, just waiting for a good PMF to restart. If a newcomer enters AI education, just as their PMF clicks and they're learning user acquisition, managing 100 salespeople, upgrading from pistol to rifle — masters pop up with Gatling guns and mow them down.
What are hidden cards? Hidden cards are typically non-consensus, completely 0-to-1 product innovation,挖掘 latent needs, with extremely fuzzy market space — giants and masters can't imagine or disdain them. Snapchat, Airbnb, Kuaishou, Douyin, Bilibili, Pinduoduo, Xiaohongshu — all felt somewhat niche early on, their generalizability unknown. In 2013, if you'd asked whether the entire nation would watch 30-second short videos, there'd be no answer. Such hidden cards buy several years of early survival window, allowing furtive development. By the time cards are revealed and combat power compared, everyone's holding Gatlings — mutual firefight still offers winning chances.
Open cards mean consensus, typically addressing existing markets with new solutions replacing old. Tenfold improvement is the best weapon — perhaps better copywriting tools, better teaching tools, enhancing experience and efficiency.
Hidden cards mean non-consensus, typically corresponding to incremental populations, satisfying unmet needs. 0-to-1 innovation is the best weapon. Before Kuaishou and Douyin, short video was barely validated — a short video paradigm invented from thin air. Before Pinduoduo,下沉 populations rarely shopped online.
If you're experience-rich, skilled at winning through business acumen and efficiency, playing open cards works — at high enough level, you can even intimidate and clear the field.
If you're experience-light, with sharp user insight and innovation skills as an AI product manager, better think hidden cards — entering from a very unique, non-consensus切入点. Starting with open cards is disrespectful to the Gatling gun.
Framework Three:
Every hundred-million-user product generation rests on "unconscious" product forms
With direction chosen and card game determined, you still need to land on specific product form.
What's the classic product form for AI-native apps?
PC era's classic forms were directories and search.
Mobile era's classic forms were feed streams (Douyin, ByteDance), timeline (WeChat), 140 characters (Twitter), QR codes (mobile payment).
The larger the DAU, the simpler the product form, the broader the适配 population, the easier users fall into "unconscious" usage — the ultimate time-killing weapon.
AI-native apps will also develop classic product forms, perhaps chat form, perhaps copilot form. Classic forms need time to emerge. In 2010 when mobile was rising, apps had a strong web-page feel. Not until around 2015, when several dominant apps established themselves, did mobile product aesthetics converge — content input copied Weibo, content display copied ByteDance, chat interface copied WeChat, adding friends and payment used QR codes. All articles copy the same few best forms.
Correct product form often achieves great results with little effort, enabling leapfrog development. One example:
QR code is new payment. One QR code撬动了 mobile payment普及; China leapfrogged credit cards to become a mobile payment nation, turning weakness into strength.
Similarly, AI is new SaaS. Future SaaS forms will change dramatically. AI as every company's foundation will level past costs: first, reducing deployment costs, as structured and unstructured data all connect to large models; second, reducing customer acquisition costs, as large models may enable app store-like platforms, perhaps in plugin form, lowering marketing costs and making product itself matter more; third, eliminating many long-tail pages, as interaction becomes natural-language-based rather than graphical. In the extreme, software will no longer be "as a service" but "result as a service" — might a China-form SaaS emerge here?
From building mobile products in 2010 to mobile investing in 2013, a major retrospective: had I understood then that "big application opportunities come from smartphones' new capabilities," category selection would have been much easier.
I've attended several AI entrepreneurship discussions, brainstorming enough ideas to fill three A4 pages — so many! This abundance easily gives entrepreneurs a gold-rush impulse; grabbing one idea to start equals rolling dice. Better reject five of your own ideas before choosing. We must ask ourselves: what's the thinking framework, how to select valid choices. Hope PMF be with you!
Founded in 2022, Heart Capital is a China-based early-stage venture capital fund focused on technology and digitization, established by Yan Han, formerly a founding partner at Lightspeed. The Heart Capital team comprises founding partners, the CFO, core investors from Lightspeed China, and senior industry investors from Cainiao and Baidu. The team's past investments include Series A investments in Full Truck Alliance (NYSE: YMM), Xpeng Motors (NYSE: XPEV, HK: 09868), as well as FinVolution (NYSE: FINV), 06810.HK, RoboSense, World Logistics, LandSpace, Lanhu, Micro-nano Star, Starfield, and others.
Rooted in China with a global outlook, Heart Capital seeks to identify world leaders who will disrupt the future. Heart Capital champions the value of "people" and "heart," aspiring to accompany more young Chinese entrepreneurs onto the world stage.
