Summary

At the cold-start stage, OpusClip’s answer was not to blanket the market with affiliates, but to find a small number of brand partners who already used the product and could influence its target community. Affiliates chase short-term commissions, while real users can provide feedback, co-build the product and endorse the brand; even two years later, OpusClip had only a dozen-plus core partners on annual contracts. “The key is finding the right people, then pursuing fewer, better ones.”（核心是找到对的人，然后追求少而精。）
Before the product has fully found PMF, the most direct growth lever is dynamic pricing experimentation, with the customization creators care most about turned into a paid feature. Runway uses custom voice and lip sync to differentiate Standard from Pro; Higgsfield lets users upload more than 20 images to train a character LoRA and does not even offer ordinary users a trial. “The features users are willing to pay for show where your PMF really is.”
Pricing is not just a number, but the sum of packaging, entitlements, paywall timing, UI and copywriting. Jun Tao said that simply changing when a paywall appears and how it is worded can lift conversion by 10%-30% or more; any price increase or price discrimination should first protect existing users, preserving their original benefits or even adding more.
Retention sets the ceiling for paid scale, because acquisition eventually stabilizes while a larger installed base produces more absolute churn. A product with 10% monthly churn and one with 20% monthly churn can end up with nearly a twofold difference in remaining users after six months; high churn causes daily losses to catch up with new users sooner. “Ignoring retention while chasing growth is like carrying a leaky bucket to fetch water.”
User feedback should drive roughly 70% of iteration, but the team must retain judgment over the other 30%—needs users have not explicitly stated that could still create differentiation. When OpusClip had roughly 30 people, it already employed 5-6 full-time U.S. customer-support staff who manually processed thousands of tickets each week, turning top feature requests, top complaints, the roadmap and bug fixes into a closed loop.
Early-stage data capabilities do not require an expensive team: a data-literate product manager, engineers and off-the-shelf SaaS are enough to run high-certainty experiments such as paywall tests. Platforms like Statsig can get the workflow running in under half an hour; once a team reaches roughly 50 people and cost and LTV questions become more complex, it can consider hiring data scientists. “Data infrastructure can remain in a rough state for a long time—as long as it is good enough.”
Growth is not a single paid-acquisition role, but acquisition, partnerships, product marketing and growth product working toward the same goal. Even without a media budget, growth product can amplify other teams through conversion experiments, ROAS infrastructure, affiliate revenue sharing and creator tools; before PMF, it can also use payment behavior to identify the Aha Moment.
One path to durable acquisition is brand and category mindshare: OpusClip won the “long-form video to short-form” category by penetrating the podcaster community, while the next phase of AI video may see Agents hide model complexity. Jun Tao is more bullish on tools that preserve character consistency, build IP, replace studio assets and create commercial value than on novelty effects built only for experimentation. “There is a lot of gap between thinking of something and actually making it.”（想到一个东西跟真正做到它之间，是有挺多 gap 的。）

Deep dive

1. The first cold-start question is not how many creators you have, but whether they actually use the product

Jun Tao joined OpusClip in 2023, initially overseeing payments, monetization and growth product. He experienced the product’s journey “from 0 to 1, then from 1 to 100,” and deliberately reduced his talk to one question: “If I could choose one thing to do at each stage, what should it be?”
A team that had just launched a video product once complained to him that AI-tool influencers bought followers and faked orders, agencies could not precisely control the content, and the imported user base was too mixed. The team did not know what its projects were being used for and could not tell whether it had found PMF. His first response was: “Are the creators you are partnering with actually users of your tool?”
Jun Tao distinguishes between 2 types of relationships. Affiliates mainly earn short-term income through commissions, sales incentives or view counts and do not need to understand the product. Early partners may not yet have fully integrated the product into their workflow, but they see the vision and are willing to provide deep feedback, co-build with the team and endorse the product at relatively low cost.
For a video product, content creators are natural distribution channels. If early KOLs are real users, their audiences are usually closer to the target customer; by contrast, large volumes of registrations and projects from unknown sources create noise and can lead the product team to iterate on the wrong signals.

2. A small number of deep brand partners can penetrate a community better than mass affiliates

Jun Tao stresses that products should establish economic win-wins with creators early, signing annual contracts and formalizing the relationship rather than asking partners to work “for the love of it” indefinitely. Creators continuously serve their audiences by sharing new tools, while the product gains more precise users from their communities, creating a positive feedback loop.
OpusClip launched its Brand Partnership Program early, using the product to find its first believers. A group of creators who joined within the program’s first 3 months continued working with OpusClip for years. John U Share, for example, later labeled every short video “Edited with OpusClip,” while the 2 sides became deeply integrated through events and co-marketing.
This strategy is not about scale. At that point, 2 years later, OpusClip still had only a dozen-plus brand partners on annual contracts who consistently produced promotional content. “It is not about having a large number, but about having high quality—whether they are precise users.”
OpusClip first entered the network around podcasters. For generative-video products, Jun Tao recommends starting with a clearly defined community such as AI video directors, finding seed users who combine real usage needs with influence, and letting them take the product beyond that initial community.

3. Customization is the clearest paid tiering mechanism for creator products

After the cold start, teams typically want to scale through advertising, but whether paid acquisition can continue depends on paid conversion bringing ROI into the black. Product quality is the long-term answer; the faster early lever is “a flexible and precise pricing strategy”—pricing should evolve with the product stage and user needs rather than remain fixed forever after launch.
Jun Tao reduces the core paid value of video products to content distinctiveness. Creators need their real appearance, their own voice, logos, fonts and other brand-kit elements to build recognition. As users move from experimentation to becoming professional creators, the clearest shift in demand is often customization.
For Runway ML, the key difference between Standard and Pro, beyond the per-credit price, is custom voice, voice-over and lip sync: users can bind their own voice and face to generated content. Jun Tao sees this as the most direct way to personalize generative video.
Higgsfield lets users upload more than 20 images to fine-tune a character LoRA of their own, making generated photos look exactly like the real person in the uploads. The capability is directly paywalled, with no trial for ordinary users. OpusClip also saw strong results when testing brand-related entitlements: “How to let users generate customized content is a very major paid feature.”

4. Pricing experiments should protect existing users first, then optimize how users understand the price

Teams often worry that price increases will trigger complaints or negative PR, but Jun Tao believes existing-user entitlements are the non-negotiable. Whether adding packages, introducing price discrimination or raising prices, the team should separate existing users, preserve their original benefits and even add entitlements generously.
This is both a retention calculation and a word-of-mouth calculation: retention among the existing user base determines how far ARR can go, while the harder-to-measure brand reputation ultimately comes from existing users. “New users come and go, but existing users are the true foundation of word of mouth.”
The second prerequisite is full-lifecycle A/B testing capability. The third is easier to underestimate: when a paywall appears, what copy it uses and which features it displays are all pricing. After studying more than 100 products, Jun Tao believes that very few companies truly present pricing well.
In his experience, simply changing paywall timing, UI and copywriting can produce a 10%-30% conversion lift or more. B2B products should also incorporate seat counts into their pricing architecture, but every change must be tested experimentally rather than judged solely by the number on the pricing page.

5. Retention determines when growth hits the zero-growth wall

Jun Tao uses the image of “a leaky bucket” to describe growth without retention: no matter how much new water is poured in, eventually only the bottom of the bucket remains. Two companies with monthly churn of 10% and 20% can end up with nearly a twofold difference in retained users within 6 months.
The larger the user base, the more important retention becomes, because a larger installed base also scales up absolute churn. At the same time, new-user acquisition per unit of time generally stabilizes rather than continuing to grow exponentially. The higher the churn, the sooner daily losses catch up with daily additions, and the sooner the product reaches zero growth.
High retention is therefore not an optimization item to address after growth is complete, but the foundation for paid scale to compound. Jun Tao attributes an important part of OpusClip’s long-term revenue growth to its relatively strong retention; the core of improving retention is not a one-off reactivation campaign, but an ongoing operating and product-iteration mechanism built around user feedback.

6. The user-feedback loop must accommodate both “70% listening” and “30% judgment”

Jun Tao cites Stripe co-founder Patrick Collison’s framework: roughly 70% of new product ideas come from users with good judgment, while the other 30% are things that should be broadly popular but have not been explicitly requested. The first requires serious listening; the second requires the product team to preserve its long-term vision.
Different channels carry different signals. Discord is used for community interaction and discussion of the product’s future; Intercom handles bugs and support for paying users; Canny structures feature requests and publishes the roadmap; social media provides a view of potential users’ opinions and brand reputation.
When OpusClip had roughly 30 people, it already had 5-6 full-time customer-support staff in the U.S. manually processing thousands of tickets each week and aggregating top feature requests and top complaints. Product and engineering also built automation for support; after Mintlify connected its help center to Intercom, automated support could handle more than 1,000 user complaints per week.
The real loop starts after collection. The product team publishes priorities and progress, engineering quickly puts bugs into a sprint, support is notified as soon as fixes ship, and users receive the update. Jun Tao admits that “doing this yourself is painful,” but only a closed loop can turn feedback into actionable insight.

7. Data investment should start with high-certainty experiments, not an expensive warehouse

Jun Tao sees “insight” as a capability spanning acquisition, conversion and retention. Early on, teams should start with front-end A/B tests whose returns are clear, such as pricing, paywall timing and feature penetration. Retention and cross-platform registration experiments are more complex and can wait until the infrastructure and sample size are mature.
A/B testing also has a learning function. The behavior after which users are most likely to convert at a paywall often reveals the Aha Moment. Teams can either reason backward from product logic to identify the core value or let successive experiments show them exactly when users experience that value.
Even crude data can produce surprising user profiles. By analyzing email domains, OpusClip found heavy usage among U.S. churches and real-estate agencies, then studied their use cases and added them to its ideal-user profile. Linking social accounts can also help identify high-influence creators, while contacting users who are already using the product is far easier than cold outreach to strangers.
Early on, a data-literate product manager and engineers working with SaaS are enough. Once the team reaches roughly 50 people and begins tackling complex questions around cost reduction and LTV, a data scientist may have enough work to fill the role. Later, data-warehouse hires are mainly for SOC 2 compliance, reducing dependence on third-party data sources and lowering costs.

8. Growth is not a role, but a set of capabilities aligned around one goal

In response to Qu Kai’s question about when to hire a GTM leader, Jun Tao broke down OpusClip’s growth organization. The head of growth owns paid acquisition, SEM, SEO and direct sales; the head of partnerships owns KOC/KOL relationships, commercial partnerships and annual contracts; product marketing/GTM owns the company’s social channels, feature launches and outreach to existing users.
Growth product improves conversion and paid-acquisition ROAS inside the product through A/B tests, while also building infrastructure for affiliate revenue sharing, creator-partner operations and fan entitlements. Qu Kai summarized this as “zero-budget” growth: the function does not buy traffic itself, but uses product capabilities to amplify every external channel.
Jun Tao does not offer a fixed hiring sequence applicable to every product. Different products fit different growth models. Higgsfield may rely on a strong creative team and social media for acquisition, while OpusClip initially depended more on partners. Teams should first identify the channels best suited to their product, then fill the corresponding roles.
Qu Kai’s challenge was whether growth product is really necessary before PMF has been found and the feature set is complete. Jun Tao’s answer was the opposite: as long as a certain number of users are registering and continuing to arrive, testing which feature and which moment users will pay at is itself part of finding PMF and the Aha Moment.

9. The best A/B tests do not prove intuition; they expose where intuition is wrong

To teams worried that startups lack experimentation infrastructure, Jun Tao recommends heavy use of SaaS. Platforms like Statsig can get the entire workflow running in under half an hour and typically serve small teams for free at first, charging only after customers grow and become dependent on the product.
He likes Arcade Software for creating product demos. After 3 videos, the fourth can still be recorded, but payment is triggered only when the user makes the project publicly visible. That moment made him pay in “less than a minute.” For many consumer products, the purchase decision can happen within an hour or even 10 minutes, leaving many variants to test along the short path.
One OpusClip redesign simply increased the visibility of its subtitle-template styles, but produced a very large conversion lift. Short-form video users logically do care about subtitles, but the team had not known how large the opportunity was. Jun Tao therefore believes “the best A/B test does not validate your idea; the result is completely different from your idea.”
Unexpected results have value only if they are saved and traceable. Experiment design must retain enough data for the team to keep segmenting variants, validate root causes and turn a conversion anomaly into durable knowledge of user behavior. Paid experiments are therefore not one-off projects, but a permanent part of the growth backlog.

10. AI pricing must account for gross margin, trial costs, retention and payment failures

Jun Tao agrees that many AI companies may still have room to raise prices, because AI products struggle to replicate the 70%-80% gross margins of traditional SaaS, with a large share of costs flowing to model providers. Products must find forms with sufficiently high willingness to pay to absorb those costs.
He cites Listen Labs, which initially sold a subscription for a few hundred dollars per month before repricing sharply to roughly $30K-$50K per year. The shift shows that researching target users and their willingness to pay may be more effective than making small optimizations around the original price point.
Monthly and annual billing do not need to be overcomplicated. If the growth model depends on long-term retention, a large annual discount can encourage users to extend their commitment, after which the team can test the best discount within what costs allow. Annual users retain better, which can make the overall economics work. The real constraints are product costs, achievable gross margin and the cost of free trials, which must be included in total returns; trials that are too expensive directly limit the scale of paid acquisition.
Watermarks, free-use limits and entitlement boundaries can all be adjusted continuously, but higher paid conversion may come at the expense of experience and retention. Jun Tao recommends running experiments for 2 weeks to 1 month, then comparing the incremental conversion lift with the incremental churn to calculate net returns. Failed subscription charges are past-due payments, not voluntary cancellations; they can be as common as active cancellations or even more common. Without optimization, retry success may be only 20%-30%; payment-method prompts and retry emails sent at different times can potentially lift it to roughly 40%-50%.

11. Category mindshare comes from penetrating one user network and making the basics a top-down mandate

Jun Tao believes there are 3 ways to build durable traffic and acquisition; the first he develops is becoming a category definer. He cites Mantis and OpusClip, with the latter becoming the first brand target users think of when they hear “turn long-form video into short-form.” When Qu Kai asks whether this means success comes mainly from brand rather than paid acquisition, Jun Tao answers that at OpusClip, paid acquisition also serves the brand.
His coffee-shop analogy makes the positioning concrete. Customers will not wait for someone in a heavy restaurant or a store with no seats; they choose a light-consumption coffee shop. Those waiting tend to gather at mall entrances and corners, which is why Starbucks often appears there. Internet growth works the same way: “Know who your users are, where they are, and how you should reach them.”
OpusClip defined podcasters as its first core user group, then discovered that they learned from one another around a small number of KOCs and KOLs. By working with and penetrating this small group over time, OpusClip became the natural first choice whenever they discussed converting long-form video into short-form, turning category definition into a moat.
Qu Kai’s final summary was that this growth playbook contains no magical secret. It simply makes fundamentals such as A/B testing, the customer-support loop and user feedback more scientific. Jun Tao agreed, but cautioned that “there is a big gap between thinking of something and actually making it”; execution requires the right people and is often a founder-backed, top-down initiative.

12. The next step for AI video is not more isolated effects, but Agents taking over a fragmented workflow

Jun Tao believes AI video is still early. Usability comes partly from generative models and partly from Agents. A single model still cannot directly produce a finished video; multiple outputs must be stitched together, and if a person does not do that manually, an Agent must. He believes both pieces may gradually mature starting in April this year.
Qu Kai questions whether AI effects are merely a short-lived mobile-internet-style photo-editing trend, with rankings changing constantly as viral effects come and go. Jun Tao agrees that competition will be intense, but argues that Kicksuit’s professional camera-movement LoRA, character consistency and character capabilities can support real production, IP operations and commercial value, making them more than toys. For AI video products more broadly, he mentions Pix Field and Pixverse as promising.
For professional creators, generation costs may not be especially high because the alternatives could be studios, photographers and traditional asset shoots. The larger obstacle is the learning curve: different models excel in different scenarios, creators must master prompt engineering, and each model update forces them to learn again. The entire workflow is highly fragmented.
The product opportunity, therefore, is to use Agents to package model selection and output stitching, allowing traditional film and television creators to work through natural language, a canvas or another intuitive interface. The one-click video-generation and canvas-style products appearing one after another are responding to this demand, with the goal of making professional production more usable and sustainable.

This Week in AI: NVIDIA’s Most Powerful Chip, Robotics Reach a New Milestone & AGI by 2026 | EP #202 The Universal Hierarchy of Life - Prof. Chris Kempes [SFI]