Summary

The strongest investment call from this episode is that the decisive battleground for AI applications has shifted from “whose model is best” to “who captures urgent needs that users will pay for quickly.” Eva uses Final Round’s AI interview product and HIX.AI’s academic writing and AI-detection evasion tools as examples: the former helps users find jobs, while the latter addresses the hard constraint of academic writing and school detection systems. Both are “urgent needs among urgent needs” and more likely than grand narratives to support high LTV. In 2023, the industry was “looking for nails with a hammer”; over time, the scarce capability became choosing the right demand and converting it into revenue.
Paid conversion may matter more than total users in determining whether an AI application is a good business. 卫诗婕’s rough math assumes a $50/month average ticket: a 1% paid rate requires roughly 2M users to reach $10M ARR, while a 10% paid rate requires about 200K users; later in the episode, Eva says 200K users at a 10% paid rate can generate $20M ARR. The discrepancy in ARR and calculation basis is not explained, so the figures should not be treated as one calculation. The conclusion is clear: “I don’t need a Super App, and I don’t need to change the world. 200K users is enough.”
The most dangerous false prosperity in generative AI is rapid user growth alongside revenue that cannot cover recurring compute and infrastructure costs. AI companions, video generation, complex 3D games and enterprise decision-making can all become “sports cars delivering takeout”: multi-turn reasoning, corpus ingestion, API rental, idle servers and peak-capacity expansion can consume margins simultaneously. Eva’s view is blunt: “Most complex generation is a market where the math doesn’t work.” 卫诗婕 challenged that conclusion with Soul’s paid performance, leaving the debate open.
Multimodal products cannot buy growth through KOL exposure alone; virality depends on whether viewers try the product, share it and trigger a second wave of distribution. ElevenLabs represents steady execution against a clear need; AI Mirror and 妙鸭 rely on dramatic before-and-after effects; products such as 海螺 are driven by model quality and emotional content. “Hugging a deceased loved one,” “imagining yourself at 60” and making pets dance translate technical capability into instantly legible use cases. “Drive your photo based on movement” leaves users confused; “make your pet dance” makes them want to try.
The structural benefit of going global is not just higher overseas willingness to pay; it is also the free cold-start network created by the Web, search, external links and cross-border communities. Product Hunt, AI directories, Reddit, Discord and Google SEO can aggregate enough users even around niche needs. The same playbook validated in Southeast Asia and then deployed in the US may produce a “10x” effect. 卫诗婕 relayed Canva’s observation that ARR and MAU in relevant markets are strongly positively correlated with the GDP of the countries where they operate. Tool needs are generally cross-cultural, so “Day 1 monetization” and “Day 1 globalization” can validate product value in parallel.
A startup’s most practical advantage over big tech is not resources but proximity to users, lower ego and concentrated investment behind one clearly defined need. Big tech has data, distribution and large pools of “80-point people,” but organizational friction can turn the output into a 60-point answer—or push it to 90. A startup that concentrates more people on one problem can create a time advantage. The key is not the team’s age but what its members have actually done, and whether a distributed setup—“spend 40 points of money to gain 80 points of experience”—can fill capability gaps.
Eva’s biggest growth correction was moving from running one Campaign after another to building a stable traffic system. Trends always fade; without dashboards, repeatable advocates, app-store keywords, search traffic and vertical distribution partners, teams are trapped in a cycle of “climb, then decay.” The growth flywheel starts with a high-confidence base, followed by 1 or 2 channels selected for the product’s stage. “I don’t want big ups and downs. I want steady happiness.”

Deep dive

1. A marital-argument mediation prompt shows how low the barrier to AI applications has fallen

In 2022 and 2023, when she was still an AI enthusiast, Eva plugged Claude into a Slack group and built an “emotional observer.” During arguments with the partner who later became her husband, she pulled the AI into the group “to help us argue”—more precisely, to act as a neutral adjudicator and mediator rather than taking the man’s or woman’s side.
Her prompt imposed several constraints: the AI was an experienced relationship counselor whose goal was to improve the couple’s relationship; it was to use nonviolent communication to guide both sides toward perspective-taking and participate in the discussion “with highly empathetic language.” It worked not by helping either side win, but because both people, when emotionally activated, were operating inside a “self-consistent logic.” An impartial third party reopened the channel for them to hear each other.
The experience convinced Eva that model companies had already solved the hardest foundational capabilities, while fine-tuning and front-end solutions were readily available. AI applications would therefore proliferate. The real opportunity would go to people who could find needs that had not been served well but that generative AI could address substantially better.

2. 2023 was a model race; the next race was turning multimodality into real demand

Eva describes the industry mood in 2023 as a mix of excitement and overvaluation. ChatGPT displayed astonishing capabilities, and the market kept comparing Model A with Model B, assuming that “whoever had the best model would own the future.” Capital’s focus on models was not entirely misplaced because “the model is the foundation of the world”; people simply had not yet developed an intuitive sense of the capability frontier.
Her sharper summary was: “2023 was looking for nails with a hammer.” Technology came first, followed by the search for problems it might solve. Later, as image and video products flooded the market, multimodality became a clear direction and the narrative shifted toward whether products actually hit a need.
The rapid growth of image and video products was not driven by technology alone. Mainstream social media was already built around images and video, while video could carry text, sound, images and motion at once—effectively accommodating every major generative modality. The more powerful the distribution format, the easier it was for the corresponding productivity tools to capture demand and distribution upside.

3. Big tech treats AI projects as internal investments; startups win on focused execution

Big tech’s strategy teams operate like internal investors: they first assess the existing market, user growth and competitive landscape, then examine their own data, user or ecosystem advantages and whether a new product can feed back into the existing business. Eva cites Tencent’s 腾讯元宝, which can search articles from official accounts and serve both as a front-door entry point for that content and as a source of hard-to-replicate synergies.
The floor for these projects is “even in the worst case, I won’t do too badly”—at minimum, they can reinforce the existing ecosystem, while the upside may produce a new “surprise.” Whether to add AI to an existing product or launch a separate AI-native brand often depends on the decision-maker: owners of incumbent businesses tend to favor SaaS plus AI, while the need for independent attention and branding favors a new product.
Big tech has mature product and engineering teams, data and distribution, but that does not necessarily make it fastest to capture the market. If an internal effort assigns only 5 to 10 people to one direction while a startup keeps 30 people close to users, the startup can build a time advantage. Globalization also weakens the logic of “the winner of distribution takes all”: a product serving a decent niche need can still build a healthy business.

4. A good market is growing—and keeps rewarding new entrants

Eva’s first filter for a direction is whether the market itself is growing; the second is whether a company can address an urgent need and charge for it. Her initial research looks at existing products, market shares and whether a new product can still emerge quickly. Even a market with 10 competitors remains friendly to newcomers if fresh entrants continue to grow rapidly: competition is strong, but so is demand.
Image and video editing are her case in point. The more social media depends on visual content, the more room there is for the tools that produce it. Video also contains a large number of subcategories, leaving multiple directions open for development.
She views Final Round as an AI interview and job-search product, not just ordinary SaaS, because “helping users find jobs” is an urgent need among urgent needs, producing better PMF and per-user LTV. HIX.AI helps users write papers and evade AI detection; users immediately understand why they need it and can make a paid decision quickly.

5. A 1% versus 10% paid conversion rate cuts the required user base by 10x

卫诗婕 uses a rough calculation to remind founders not to fixate on user counts. At a common price of $50/month, a 1% paid rate requires roughly 2M users to reach $10M ARR, while a 10% paid rate requires about 200K users to reach a similar revenue level. Both cases imply roughly 20K paying users; the difference is the total burden of acquisition and service.
That is the appeal of “business-oriented entrepreneurship,” she says: “I don’t need a Super App, and I don’t need to change the world. 200K users is enough.” Solving one small problem for a group of people who keep paying is more consistent with staying alive first than leading with a story about changing the world.
卫诗婕 adds that large-model usage costs remain high, so teams must decide from Day 1 who will pay, how large the market is and whether the economics work. Eva calls payment “to some extent the best retention,” because willingness to pay proves the need is real, while the obligation to deliver after payment forces the product to improve.
Later in the episode, Eva says a product with a 10% paid rate and 200K users can generate $20M ARR. 卫诗婕’s earlier calculation used $10M ARR, and the later figure comes without an explanation of the change in ARR or the calculation basis. The two statements should not be forcibly reconciled.
Niche positioning does not mean building only one feature. If the same group of sophisticated professionals has other needs, the product can add features and secure another 5% or 10% paid conversion. “That is the next step”; the first step is still making one small need work.

6. Day 1 monetization and Day 1 globalization are the cleanest value tests for AI tools

Eva acknowledges that charging early is sometimes “a necessity” for AI companies. Capital is less active than during the mobile-internet era, while AI tools carry heavier server and model costs, so teams must prove they can make money. Tool value is easiest to validate through payment, although AI social and entertainment products pursuing a larger trend may reasonably monetize later.
卫诗婕 recalls a conversation with a Canva executive: ARR and MAU in different countries are strongly positively correlated with local GDP, reflecting how much ordinary users can pay for ordinary needs. The playbook is to refine the product in markets with purchasing power, then replicate it across other high-potential markets. Big tech can likewise use domestic resources to polish a product before releasing its revenue potential overseas.
Eva’s experience is that basic human needs are broadly shared, with tool products showing particularly little national differentiation. Features with strong demand in China typically have demand overseas as well. Teams can build domestic and overseas versions in parallel from Day 1, although AI social products may first need to satisfy a set of needs and produce a body of content before a business model can “grow slowly.”

7. Startups’ real edge is lower ego and faster iteration against a defined need

Eva believes the critical capabilities in zero-to-one are market sensitivity and insight into user needs. Decision-makers farther from users are more likely to construct grand strategies from first principles. The safer entrepreneurial path is to “lower your ego” and start with a clearly existing need.
Once the need is identified, the startup’s advantages emerge: faster launch, faster iteration, concentrated resources and flexible organizational and staffing decisions. If a product needs only an 80/100 product manager and a sharp user researcher, there is no reason to copy a big-tech staffing model. If algorithms are the key, resources should instead be concentrated on finding outstanding algorithm talent.
Big tech often gathers large numbers of “80-point people,” but headcount makes organization a critical variable. Poor information flow, execution, culture or decision-making can produce a 60-point answer; strong organizational execution can push collective capability to 90. A startup has fewer people, but if it captures a few critical points and maintains speed, it has a chance to break out.

8. Social media, search terms and new-product growth are 3 early demand signals

Eva’s first way of finding demand is to look directly at what users say on social media. She cites the “kitten fill light” as an example: users on Xiaohongshu were already openly discussing their need for better lighting in photos, giving founders zero-to-one evidence of demand.
The second source of evidence is search terms on Google, Baidu or Xiaohongshu, because active search is closer to real intent. The third is user interviews: a team building an anime-oriented product has to speak with anime users rather than imagine their needs on their behalf.
Another practical signal is the growth rate of comparable new products. If a group of products offers similar functionality and one is growing faster than the others, the market is already validating demand for the founder. Looking first for “a new product that is growing quickly” is more reliable than becoming absorbed in whether the technology is sufficiently flashy.
Asked whether big-tech data could create a first-mover advantage for models, Eva drew a clear boundary: “Because I’m not on the technical side, I may not be that authoritative.” In her limited view, high-quality materials and data sources clearly matter, but she did not inflate data advantages into a guarantee of victory.

9. Going global from zero is easier with open links, web search and cross-border communities

Overseas markets have established communities such as Product Hunt and AI directories where AI enthusiasts can try new products and provide feedback. Social platforms also generally support external links, allowing viewers to convert directly after seeing content. Overseas is not a single country: even a niche need can become a sufficiently large initial audience when aggregated across countries.
Product formats also differ. China entered the mobile internet earlier, making Apps generally easier to build than Web products, while Baidu’s SEO ecosystem is less developed than Google’s. Overseas markets still have many desktop-work settings, allowing Web products to capture free traffic from active search. Eva distinguishes SEO, which brings organic traffic, from SEM, which relies on paid acquisition.
Xiaohongshu is one of the few domestic channels suited to Building in Public: new accounts can still receive distribution, the audience is broad, the algorithm is relatively decentralized and the platform is increasingly used for search. That allows niche products to find similar users. The same playbook overseas can cover multiple countries; after gaining traction in Southeast Asia and then moving into the US, the effect may be not “1x” but “10x.”

10. The first zero-to-one trap is mistaking a technical benchmark for user value

Eva’s standard for a startup direction is to “bring users help in a crisis,” not to build a self-congratulatory product simply because a technology is impressive. Digital humans are her counterexample: B2B marketing, SMBs or large enterprises may need them, and the user base may be small while revenue is strong. For ordinary consumer users recording daily life or dancing, it is difficult to explain why a digital human is necessary.
卫诗婕 uses a new ChatGPT version to ask whether ordinary people can perceive a technical upgrade. Eva says the difference between o1 and older versions may be felt mainly by enterprises doing heavy reasoning or long-form work; for mass-market users, it is more a technical benchmark. A stronger capability frontier does not automatically create an existing use case that users will pay for.
Teams therefore need to define whether they serve B2B or B2C and identify who needs the product most. The product will ultimately flow toward its most-needful users, but whether those users will pay the first dollar is the commercial question.

11. The second trap is using a sports car for deliveries; AI companions expose the sharpest divide

Eva describes bad unit economics as “using a sports car to deliver takeout”: the product is extremely expensive to operate, the paid rate is low and user numbers look good enough to create “false prosperity.” She cites AI companions as a typical case because multi-turn interaction and personalization are costly, while the target users may be young and have limited purchasing power.
Her explanation is that after Midjourney generates an image that meets expectations, users can quickly decide whether to pay. An AI companion may require many rounds of interaction and refinement before it approaches an individual’s emotional needs. “If you exclude porn, its paid rate really is quite low—low across the world,” is her assessment of global AI-companion monetization.
卫诗婕’s rebuttal is worth preserving: Soul serves people who need online companionship and lack social assets in the physical world, yet has achieved solid monetization. She also notes that Soul users can talk with real people, write diaries or talk to themselves. Eva restates the disagreement as whether users who have to pay would rather pay a real person than an AI, and argues that these users tend to be younger with weaker purchasing power.
The disagreement was not resolved. 卫诗婕 said she would discuss the issue further with AI-companion founders, while Eva accepted the mechanism 卫诗婕 proposed: companionship carries a heavy delivery burden upfront, and user satisfaction varies by person. Unlike search or image generation, companionship has no clear expected boundary, making “satisfaction” hard to define. Chat screenshots can generate rapid distribution, but long-term retention and high willingness to pay are harder to achieve.

12. Whether complex generation can be a business comes down to compute economics, not just model quality

Eva believes “generation really burns GPUs,” so complex video, enterprise decision-making and 3D game creation require caution. A lightweight image generation request has a different cost structure from sustained multi-turn interaction. AI companions may also require source material such as novels, alongside server, operating, labor and compute costs.
If the underlying capability comes from a long-term API lease, idle capacity may still generate costs. If 10 servers are enough under normal conditions but 50 are needed at peak, the team must decide whether the additional 40 remain idle over the long term or whether to risk service failures during peaks. Cloud services can scale elastically, while private deployments often mean “however many servers you rent is how many you have.”
Eva reduces the calculation to 4 layers: operations, labor, servers and compute. Games are similarly expensive to run. If user growth is not accompanied by validation of paid conversion and costs, it will only expose infrastructure problems faster.

13. The third trap is perfectionism; brand advantage may go to whoever enters first

Eva rejects the idea of fully polishing a product before launch. Startups usually grow incrementally: the first need may be only half right, but once users arrive, in-depth interviews can reveal the other half of what they actually need and a modest product adjustment may be enough. If a team develops behind closed doors for too long, it becomes harder to change direction after drifting off course.
Her principle is to tolerate imperfection, release early and iterate quickly; if the first attempt is accurate, the team can also capture a brand advantage. 卫诗婕 cites MiniMax and ByteDance as examples of companies with strong product capabilities that can run multiple domestic and overseas products in parallel. Eva argues that mature big-tech companies usually complete their analysis before coordinating compute and resources, making them less likely to fall into these 3 foundational traps.

14. Multimodal growth falls into 3 buckets: must-have functions, replicable effects and tech-led generation

The first bucket is steady execution against a clear need. ElevenLabs has gone deep on voice, covering text-to-speech, voice cloning, video translation and voice adjustment. It serves consumers while offering an API to businesses. Generative AI compresses professional voiceover work that once took hours or even a day into cloning a voice from “3 sentences,” while also allowing the emotional state to shift toward anger, cuteness or sadness.
Early on, ElevenLabs would quickly add voiceovers to videos released alongside Sora, helping users understand its capabilities immediately. Eva emphasizes that this type of marketing only pushes the selling point into the market, helping people who care about voice remember and use the product. The underlying demand still comes from real, stable customers such as marketing companies and MCNs.
The second bucket is effect-driven products such as 妙鸭 and AI Mirror, which use obvious before-and-after contrasts—felt, clay and other treatments—to create social distribution. An individual effect can sometimes be copied in 5 to 10 days. Growth is determined not by technical complexity itself but by whether the effect is compelling enough for people to try personally. Glam AI and Remini are also cited as fast-growing examples.
The third bucket is technology-led generation products, where model quality itself may determine “50%” of the outcome and encourage users to create freely. Eva mentions 海螺, Luma, Vidu and VEED, as well as editing products including CapCut, Captions and 美图系 products. Together, they illustrate different growth mechanisms across generation, editing and effects.

15. KOLs buy attention; real virality requires a second wave of sharing

卫诗婕 asks whether, once product quality is good enough, concentrating on top and mid-tier KOLs can manufacture virality from the top down. Eva says no: KOLs can only make more people aware that “this thing exists.” Real popularity depends on whether viewers share and try it, and whether people who see the resulting work continue to distribute it.
Generative video is particularly suited to completing that chain through emotion. Eva cites AI hugs, imagining oneself at 60 from a photo, and hugging a deceased loved one or a favorite celebrity. Content tied to the US presidential election can likewise spread through public figures and IP. Chasing trends and borrowing IP are common, but viral marketing always contains an element of luck.
Product messaging has to translate technical capability into a concrete desire: “If I tell you that the model can drive your photo based on movement, you may be completely confused. But if I say it can make your pet dance, you’ll want to try it.” The capability has not changed, but the user’s comprehension cost and the product’s distribution power have.

16. Consumer-to-business conversion can bring decision-makers in and reduce costly enterprise acquisition

Eva believes HeyGen found a smart consumer-to-business path. Viral videos such as 郭德纲 delivering crosstalk first created awareness among ordinary users. Enterprise decision-makers are ordinary users too, and after seeing the content they may proactively ask about an enterprise solution. That is closer to attraction than push-driven sales teams cultivating leads and knocking on doors one by one.
Kimi used a similar mechanism by focusing on long-form text, with events such as loading 《甄嬛传》 into Kimi taking the capability mainstream. Early attention may also have been amplified by the emergence of “Kimi concept stocks.” Later, when B2B platforms such as 扣子 needed to connect to models, some potential partners may have reached out because they already knew Kimi.
Large-scale consumer marketing events are not cheap and success is probabilistic. But if the goal is to attract a mid-sized or large enterprise customer, the return may still compare favorably with the cost of sustained sales outreach. At a later stage, the product still needs to return to vertical channels such as overseas e-commerce trade shows, LinkedIn content and the relevant target industries.

17. Virality can travel across markets; UGC and search gaps form a pool of product opportunities

Eva cites 通义千问’s “全民舞王” campaign: upload a photo and make a person or 2 golden retrievers dance. After going viral on Douyin, a similar format could catch on again in Southeast Asia and the US. Cuteness, dancing and pets combine broadly shared needs. Users may not want to make themselves dance, but they are very willing to make their pets dance, allowing teams to exploit time differences between markets.
Vidu first accumulated seed-user content on Discord before breaking out through a YouTube video. Generative products such as Suno and Midjourney face an extraordinary variety of prompts, so teams need to watch which results generate the most discussion in their communities, select those outputs as social-media assets and build new creative work around use cases that have already been validated.
China lacks an equally visible “Chinese Discord,” but overseas communities can aggregate niche creators across more than 50 countries. Even an extremely narrow direction such as AI tattoos can become a business. Eva ultimately reduces globalization to organizational execution: can the team capture opportunities globally, identify trends quickly and act in time?
For founders with limited resources, she recommends a demand-backward approach modeled on 万兴科技: start by mining search terms, find low-competition pockets where demand is large but existing products are underserving it, then quickly ship imperfect lightweight tools, add adjacent needs and build a product matrix. This differs from ByteDance’s strength in testing multiple directions simultaneously. It is closer to choosing the site first, then using the first small business to finance the next one.

18. Sustainable growth means defending the base, while team design determines whether the flywheel turns

Eva’s biggest failure lesson was focusing too heavily on social-media performance early on and treating growth as a series of Campaigns. She later realized that trends inevitably “climb, then decay”; what matters is whether dashboards, distribution networks and fixed channels can keep a stable flow of traffic in place.
Stable traffic can come from a group of people who consistently help test and share the product, app-store keywords, fixed channels or organic search. Their daily scale may not be spectacular, but they keep the team from repeatedly betting on a new probability. “You need to think about what your baseline looks like,” she says. “I don’t want big ups and downs. I want steady happiness.”
Channels also have to match the product’s cadence. Product Hunt and AI Tools have limited but precise reach, while Reddit is useful for finding specific user groups. For education products, the users are students but the payers are parents, so partnerships with local schools or training institutions may outperform broad traffic. A startup can only focus on 1 or 2 priorities at a time; it should not copy Discord from Day 1 simply because another team succeeded there.
When talent is scarce, Eva recommends pairing experienced people with an execution team and using a distributed setup to “spend 40 points of money to gain 80 points of experience.” AI-native capability should not be judged by age but by what a person has actually done. She wants to serve globally oriented, PLG-driven rather than SLG-driven teams, filling gaps across 3 layers—strategic direction, product tactics and channel sequencing—while continuing to look for her own opportunities to create.

How I Use RL + LLMs to Build Agents | A Conversation with Pokee AI Founder 朱哲清 Bill Red-pilled Billionaires, LA Fire Update, Newsom's Price Caps, TikTok Ban, Jobless MBAs