Summary

Content consumption has moved from scale expansion into the “deep end,” with short dramas the clearest pressure gauge. On Kuaishou, the first campaign cycle has compressed from one to two weeks in the past to one week, or even less, in H2 2025, with long-tail dramas fading faster; whereas few suppliers once spent more than RMB1M to produce a drama and most spent several hundred thousand or the low hundreds of thousands, low-cost content is now becoming “not good enough,” forcing producers to invest in more polished work while IAA further lengthens the payback cycle.
Short-video user growth may now be limited to the single digits, but Kuaishou’s advertising business is still growing at a double-digit rate, with the incremental growth coming from finer-grained operations. 庄明浩 expects commercialization and overall market growth to change little; Zheng Xun believes the short-video sector’s growth rate is largely set and Kuaishou will continue to grow. Lead-generation ads, content consumption and e-commerce are the faster-growing internal businesses, while short dramas, novels and mini-games continue to expand; however, it is increasingly difficult to deploy “one strategy and then grow sharply,” with more gains coming from finding “more mines beneath the three layers of strategy.”
The 2025 breakout in comic dramas was fundamentally a supply surge, not a sudden emergence of demand. The real inflection point came between Q3 and Q4, especially after the October holiday on Kuaishou; comic dramas may already account for 30% to 40% of short dramas, but rankings on each platform are changing rapidly and every new category can still break into the mainstream, leading Zheng Xun to believe “the ceiling is nowhere in sight.” Teams of 3 to 5 or 5 to 8 people can form production lines; per-minute costs range from several thousand yuan for cheaper works to more than RMB10,000 for expensive ones, still materially below the RMB10,000-level cost of live-action short dramas.
AI’s first commercial payoff is not end-to-end magic, but measurable returns generated after it is embedded across the production and operating chain. Comic dramas have already formed a human-machine workflow from novel to script, script to image and image to video; Kuaishou is also using large models for ad bidding, bid adjustments and account-state assessment, partly replacing bid operators and manual platform strategies, with Zheng Xun saying that in some scenarios “the results are better than a human’s.” Zheng Xun’s summary of 2025 is that “many things we once looked up to ultimately landed at our feet,” a description closer to real commercialization than generalized Agent narratives.
Monetization in live-action short dramas is tilting toward IAA, while the industry is splitting between overseas expansion and premium titles, with mid-tier supply potentially migrating to comic dramas. Open-ended large-model tagging has expanded usable tags from dozens to the triple digits, producing more combinations of female-oriented themes and “rebirth, time travel and system” plots; the most visible cost inflation remains concentrated in top actors, whose fees swing sharply with recent ranking performance. Dedicated short-drama platforms are consolidating domestically, with Red Fruit Short Drama dominant in the category, while generalist platforms continue to absorb consumption and leading CPs broadly expand overseas as a key way to amortize copyright and production costs.
Mini-games have not followed short dramas into one-way convergence around IAA because AI has yet to lower the cost of producing core gameplay. 庄明浩 puts the mini-game market at more than RMB50B, with leading developers generally using a hybrid IAA/IAP model that offers both ad-supported gameplay and richer, more satisfying experiences for IAP users; for now, AI can make simple puzzle games or Snake, but still struggles to independently deliver more complex, genuinely replayable products. NPC Agents, simulators and open-world settings are more worth watching: they could mean each session no longer has only “three endings,” but remain largely a vision today.
The core variable for 2026 is that AI’s capability ceiling keeps rising while platforms push the usable floor into more real-world scenarios. If multimodal systems can solve lip-sync, eye movement, hair, muscle and complex physical details, content teams can concentrate resources on story; Kuaishou will gradually connect natural-traffic targeting for comic dramas, short dramas and mini-games, while trying to let advertisers complete professional campaign operations directly in natural language. The ultimate demand is not crude AI content, but “telling a good story with higher-quality visuals.”

Deep dive

1. Short-drama lifecycles are shrinking, marking content consumption’s formal entry into the deep end

Zheng Xun’s first pressure gauge is campaign duration: a drama’s first campaign cycle could once run for one to two weeks; by H2 2025, it often lasts only one week, sometimes less, before user interest starts to fade. Top series can still reach a third installment, prepare a fourth and sustain consumption through a stable period, but a large volume of long-tail dramas is becoming obsolete faster.
This is not simply a traffic problem; “users are becoming pickier about dramas.” Repetitive themes and familiar-looking plots became increasingly difficult to scale in H2 2025, while policy imposed more requirements on themes and storylines, forcing producers to keep investing in novel IP, writing ability, narrative structure and filming quality.
庄明浩’s question was where the contradiction lies: if lifecycles are shrinking while production costs are falling, producers could still absorb risk through “natural selection.” Reality, however, is that costs are rising too. Few suppliers once spent RMB1M to make a drama; most stayed at several hundred thousand or the low hundreds of thousands. Low-cost content still exists today, but is increasingly “not good enough.”
IAA adds another layer of pressure. Short dramas have shifted from paying RMB10 or RMB20 to unlock the full series toward watching ads, with even the first several episodes free. That expands the audience, but also raises both the difficulty and the duration of copyright owners’ payback.

2. The installed short-video base can still grow, but growth has shifted from frontier development to precision mining

庄明浩 described the industry’s base very directly: China’s short-video user base has long surpassed the billion mark, and user growth in 2025 was probably limited to the single digits, with commercialization and market-size changes likely similar. Zheng Xun agrees that the short-video sector’s growth rate and platform structure are largely set. Kuaishou will continue to grow, but not materially outpace growth in other non-short-video sectors.
There are still bright spots inside the structure. Time spent and user numbers for Kuaishou’s live-action short dramas continue to grow substantially; novels are outperforming expectations, and mini-games are also adding meaningful volume. By lowering the payment barrier, IAA brings into the market people who once reached episode 10, felt desperate to know what happened next, but still refused to pay.
Partial data from Magnetic Engine show Kuaishou’s advertising business growing at roughly a double-digit rate, with at least some parts of the market reaching that pace. Lead-generation ads, content consumption and e-commerce are the 3 faster-growing areas. E-commerce growth still depends on better product quality, return handling and other basic capabilities.
Zheng Xun describes the current playbook as “drawing the grid more finely”: most growth comes from repeatedly breaking down strategies and data until the team finds “more mines beneath the three layers of strategy.” 庄明浩 uses his own business as a reference point: the original conservative estimate was 8% to 9% growth, while the final result was about 10% to 11%, achieved with considerable difficulty.

3. AI’s value in 2025 moved from imagination into operating results

Zheng Xun summarizes 2025 as “many things we once looked up to ultimately landed at our feet.” The most visible example is video generation driving the rapid growth of comic dramas; the less visible one is large language models entering ad recommendation, bidding and bid-adjustment workflows.
Kuaishou has already deployed some ad-bidding strategies handled by large models, partly replacing bid operators’ experience and the temporary tools that platform teams used to hand-build when trying to improve client performance. Zheng Xun says that in some scenarios, the models “replaced people, and the results were better than a human’s.”
These 2 lines reshape supply and operations respectively: one lowers the barrier to content production, while the other improves the allocation efficiency of installed traffic. Because AI already generated commercial returns in 2025, Kuaishou has greater confidence in its 2026 investment, experimentation and internal training.

4. Comic dramas’ inflection point came from a sharp supply increase, not a sudden demand boom

Zheng Xun rejects the idea that the industry accurately predicted this cycle in advance. Kuaishou had invested in related companies and knew internally that comic dramas were worth pursuing, but no one expected them to reach this scale by year-end. In H1 2025, the shared view among both platforms and producers was still that comic dramas were “below expectations,” or at least not something to bet on.
The inflection point was concentrated between Q3 and Q4, with Kuaishou feeling the steepest acceleration “after the October holiday.” Demand did not fundamentally change: users who liked dynamic comics had always existed, while AI live-action comic dramas also absorbed some live-action short-drama users. The major increase came from supply.
The top rankings for live-action short dramas are relatively stable: several platforms tend to receive similar leading titles, and the content generally performs. Comic dramas, by contrast, have rankings that “you simply cannot keep up with”: each platform has different leading works, with silly meme formats one week and a wholesale shift to AI live-action styles the next.
Zheng Xun therefore believes that comic dramas’ current share of 30% to 40% of short dramas is not the ceiling. “Put a drama out and it will run”; each new style can trigger another round of user expansion, with growth driven more by supply expansion than by fixed users competing against one another.

5. Teams of 3 to 8 and per-minute costs in the thousands have materially lowered the production barrier

庄明浩’s causal chain is endorsed by Zheng Xun: as the barrier and cost of live-action short dramas continue to rise, some producers may be pushed out, while comic dramas can absorb them with lower costs, an executable workflow and high growth rates.
In an extreme case, 1 person can make a comic drama, but maintaining output and a long-term production line is difficult. Better-performing organizations typically have 3 to 5 or 5 to 8 people. A former live-action drama team of more than 10 people can also be split into 3 or 4 groups running 3 comic-drama lines at the same time.
Comic dramas are now commonly measured by “per-minute production cost”: cheaper works are already in the several-thousand-yuan range, while expensive ones may still exceed RMB10,000. Live-action short dramas remain at the RMB10,000 level. AI is no longer optional; the only question is how much of it a team uses. Pure hand-drawing or frame-by-frame CG production is effectively absent from mainstream teams.
The lower barrier is also changing the division of labor. Live-action short dramas require sets, actors and offline resources; comic dramas can start as soon as the team has the technology. Kuaishou has even seen people who previously handled only distribution learn editing and “become original creators,” moving directly into producing their own content.

6. Once “one sentence to finished video” faded, the human-machine production line finally worked

One reason growth was difficult in H1 was that practitioners from traditional Chinese animation and comics had not yet figured out how to apply short-drama storytelling. The shift came when a group of short-drama screenwriters and story planners entered the field. Familiar with planted hooks, conflict and reversals, they quickly brought “time travel, rebirth and system” settings into comic-drama scripts.
The industry initially overestimated end-to-end capability, expecting “one sentence to finished video.” The workable solution was to break the process into “novel to script, script to image, image to video,” plus voice-over and other steps. Each stage has a corresponding model or tool, while people handle the handoffs and quality control.
Zheng Xun sees comic dramas as a good example of “human-AI collaboration.” The goal is not to make the tools fully one-stop, but to have the script lead pass materials to the video lead, with the human “adhesive” filling the gaps between model capabilities.
Internal Kuaishou tool competitions show that a skilled operator can produce 2 or 3 episodes with relatively polished visuals in roughly 1 or 2 hours. A mainstream work runs more than 40 episodes, each 1 to 2 minutes long, for about 1 hour of total content; even a mature team still needs several days, and the gap between skilled and unskilled operators is significant.

7. Short-drama infrastructure supports comic dramas, but platforms must solve user expansion again

Comic dramas can scale quickly because they can first “thank the previous short-drama wave for providing the infrastructure.” Most of the existing chain can be reused, so platforms do not need to build a market from scratch.
The biggest difference is the audience. Existing ad-recommendation models and bidding strategies target relatively fixed short-drama clients and users. If comic dramas merely move those users away from live-action dramas, they add nothing to the platform’s total market. Kuaishou therefore devoted resources first to breaking into new audiences.
User-side changes include adding more bullet comments for younger dynamic-comic audiences and adjusting the timing of paid and IAA monetization. On the client side, the platform must handle another uncertainty: comic-drama production quality maps weakly to its ability to scale, and cheap meme-style works can also perform extremely well.
Kuaishou has therefore debated whether to “let anything that can run run” or tilt distribution toward particular types of supply, and has introduced supply policies intended to support content and teams that are better for advertising, consumption and the ecosystem over the long term, rather than allowing short-term low-cost hits to define the category.

8. Comic dramas’ next ceiling is not image quality, but story and visual grammar

Early this year, Zheng Xun still believed AI’s main unlock was fantasy imagery that would be difficult to shoot with live actors. He has since changed his view: “visual optimization may no longer be the ceiling of this business.” Once tool limitations and production costs fall, the scarce resource becomes a good story again.
Novels and comic dramas fit together because the stories in existing works have already been validated. But as supply continues to expand, the industry will need more new stories. Zheng Xun also expects short and medium-length novels, along with stories never adapted or poorly told in the past, to regain opportunities.
The second bottleneck is visual grammar. Live-action short dramas still benefit from directors, cinematographers and film-production experience: even without expensive equipment, practitioners know when to use a medium shot and when to cut to a close-up. As small comic-drama studios multiply, some meme-style works are abandoning traditional shot conventions altogether.
This “freewheeling” approach may produce new changes. Once human and AI workflows are sufficiently integrated, AI should gradually “become an industry tool,” with users perceiving the story rather than the tool itself.

9. Open-ended tags show short-drama themes shifting from concentrated formulas to dispersed combinations

Short-drama themes are being shaped both by policy requirements and by users becoming more demanding. It is difficult to say simply whether supply is merely following users, but changes in Kuaishou’s content-understanding tags provide a more granular view.
The old system was preset by suppliers, with only dozens of highly concentrated tags such as male-oriented, female-oriented, historical, modern, urban, time travel and comeback. In 2025, Kuaishou switched to using large models to understand an entire piece of content. Clients can write tags, but the model generates tags for the full work itself.
After removing hallucinations, nonsense and overly thin terms, usable tags first rose into the high double digits and then reached the triple digits, with concentration continuing to fall. The single most popular tags may still be old themes such as comeback and time travel, but the combination space has expanded substantially.
Zheng Xun gives the example that “rebirth, time travel and system” can all appear together, or a story can use rebirth without time travel, or combine rebirth with a system. Greater dispersion in tag combinations means formulas really are becoming more varied, while female-oriented supply is also increasing visibly.

10. The live-action short-drama industry is converging around overseas markets, top actors and generalist platforms

Zheng Xun’s Kuaishou business line does not cover overseas expansion, but observations from the CP side show that leading companies are almost all going abroad. Since late 2024, overseas business has already accounted for a large share of GMV at some companies. Common approaches include re-dubbing and reusing existing content; more aggressive players shoot local dramas overseas directly.
Overseas expansion is also a way to amortize copyright and production costs. Many platforms have started considering bundled negotiations for domestic and overseas rights. For producers, recovering the cost of the same content across more markets has become an income path that cannot be ignored in the deep end.
Domestic platforms are not simply becoming fewer; dedicated short-drama platforms are consolidating while generalist content platforms are expanding. Zheng Xun says Red Fruit Short Drama is dominant in the category, while Kuaishou and other generalist platforms retain short-drama sections. Meituan, Alipay and other platforms with little prior connection to content consumption are also entering.
Standalone comic-drama apps are currently growing more slowly than generalist apps because their audiences are narrower and user-acquisition costs are higher. Buying users by individual drama also creates a zero-sum dynamic with short dramas. Zheng Xun leaves room for uncertainty: “I may be proven wrong after a while,” because supply is still changing rapidly.

11. Cost inflation is concentrated in top actors, while mid-tier live-action supply may move to comic dramas

The clearest cost increase in live-action short dramas is in actors, especially top performers whose ability has been proven. Fees are highly correlated with recent rankings: once a work appears on the charts, an actor becomes expensive quickly; when the heat fades, prices fall again. The volatility is substantial.
According to Zheng Xun, the cost of bringing in IP remains relatively manageable. Producers may own the IP or partner with novel platforms, while AI can generally expand a story from a synopsis, outline and setting. He therefore agrees that the strategic keywords for leading short-drama companies have become concentrated: AI to improve content quality, plus comic dramas and overseas expansion.
His 2026 structural view is that top live-action dramas will continue to sustain high investment, and good actors, screenwriters and scripts will remain valuable. A fan economy is already emerging around actors, but some mid-tier investment may migrate to comic dramas to reduce production and payback pressure.

12. IAA has become the main road for short-drama monetization, while interaction details can still expand the consumption pool

Zheng Xun defines the IAA-versus-IAP choice as a balance between how much users like the content and what cost they are willing to pay to access it. Early IAP models that charged RMB10 or RMB20 in stages filtered out many interested users who were unwilling to pay directly; IAA brings those users into the viewing pool.
One of Kuaishou’s 2025 experiments changed some rewarded ads in short dramas into mid-roll ads. A rewarded ad stops the user for 10 or 20 seconds and still requires a tap to close afterward; a mid-roll ad preserves the user’s swiping habit, allowing the user to swipe away once the viewing time ends.
Even after the two formats were matched for dwell time, the mid-roll solution performed materially better than rewarded ads across the experiments. Zheng Xun’s conclusion is that following users’ existing operating habits and experience can produce new monetization methods or user experiences, bringing more people into the viewing pool.

13. Mini-games retain a hybrid IAA/IAP model because AI has not lowered the cost of core production

Unlike short dramas, mini-games show no clear trend toward separation between IAA and IAP. 庄明浩 puts the market at more than RMB50B, and leading developers almost all use a hybrid model: the same game serves ad-supported users while offering IAP users a more satisfying experience.
Zheng Xun emphasizes the key phrase “not yet”: this AI wave has not materially lowered game-production costs. AI can help people with no programming background make simple puzzle games and can produce Snake, but it still cannot reliably deliver genuinely playable products such as nail-removal or water-sorting games end to end.
IAP therefore still corresponds to richer gameplay and a higher bar for experience. If AI eventually cuts the cost of PVE content sharply, free supply may expand; but PVP and social gameplay are fundamentally about “people competing with people,” so there will always be room for payments that give some players differentiated advantages.

14. AI-native games will first grow through NPCs and simulators, not by rebuilding entire games

庄明浩 observes that “small little games,” dating and work simulators, Werewolf variants, puzzle games and roguelikes are beginning to use AI. In the past, these products could only prewrite a limited number of fixed branches; with models added, uncertainty itself can become part of the early user experience.
Larger games already use AI components extensively, including connecting NPCs to large language models or Agents to generate dialogue within a defined world and character setting. Zheng Xun believes this can increase replayability, but at this stage it “is not the core gameplay.”
The more distant vision is for 1 AI to write the worldbuilding, multiple Agents to play characters, NPCs or players, and another Agent to organize events. A writing Agent could turn the process into novels in different styles, while characters could continue to drive actions through dialogue, making it difficult to distinguish the result from either a game or an immersive short drama.
庄明浩 warns that this resembles the return of metaverse narratives. Zheng Xun’s distinction is that the old metaverse merely moved the real-world “me” into a virtual world; an Agent world has generative characters and changing conditions, so every person and every session could produce a different experience rather than just “three endings.” He also explicitly acknowledges that this remains a vision and has not been fully implemented.

15. New content formats tend to come from marginal innovation, not entirely new categories appearing from nowhere

Zheng Xun believes people did not fail to imagine vertical video before short video appeared; only after the scale became large enough did vertical video receive its own name distinct from “video.” New content formats likewise often begin as small changes before forming an independent demand curve and industry.
Kuaishou continues experimenting with audiobooks and novel dramas, which sit between novels and short dramas. They have some user demand but are not yet large enough to be defined as new major categories. AI-driven immersive content may similarly sit between games and short dramas; the eventual label is not important.
Digital humans in livestreaming are another incremental example: capability and interactivity continue to improve, but they have not immediately created an entirely new category. What platforms can do today is keep testing the boundaries of formats, not predict the next inevitable hit.

16. Recommendation has become too large for rule-based governance, while advertising multiplies the complexity again

庄明浩 uses orders of magnitude to explain the problem: Chinese video platforms may add a combined 100M-level number of videos each day, while nearly 1B devices use them for 1 to 2 hours daily. If the average video is half a minute and each person watches about 100 videos a day, the system is handling roughly 100B matches, while the potential combinations of new videos and users are even harder to enumerate.
Without a recommendation engine, platforms can only fall back on trending lists and the lowest common denominator. The result is either uncontrollable or highly concentrated. Recommendation systems are therefore not an auxiliary tool, but the infrastructure that makes large-scale content consumption possible.
Zheng Xun adds that advertising expands the universe once more: the same video paired with different titles, bids or target audiences becomes a different ad unit. If the first 15 seconds are identical but the later section differs, the system still treats them as 2 videos. Users see similar content; the system processes an ever-expanding set of independent objects.

17. Large models first make content and users describable, then feed that understanding back into recommendation

Kuaishou’s first step is using large models for content understanding. Unlike traditional tags, a model can write a 100-word or 500-word plot summary for a short drama, or explain game mechanics in 50 or 500 words, converting implicit plots, relationships and mechanisms into language.
Hallucinations cannot be eliminated completely, so the ad system also uses reinforcement learning and reward mechanisms to keep generated content as close to reality as possible. Once complete, the platform has a richer supply profile and a more detailed “mini-biography” of the user than traditional features provide.
These descriptions can be used for content analysis or directly converted into embeddings and recommendation strategies. Moving from dozens of open tags to more than 100 is only the most visible output; more importantly, videos, products and users that were once black boxes now have a semantic layer that can be inspected and modified.

18. Deduplication and data inheritance turn ad cold starts from a probability game into a compounding asset

Kuaishou reduces the core recommendation unit to a single pair: a video and its unique corresponding promoted object—a drama, game, novel or product. Repeated ad units are consolidated under the same combination, materially reducing computational complexity.
This is especially important for cold starts. If every inflated ad asset has to seek exposure from scratch, the platform lacks enough traffic and advertisers keep asking to “run it again,” leaving both sides trapped in a probability game.
In short dramas, for example, platforms may launch around 100 new dramas a day, with the same drama sometimes uploaded under different names by different distributors. Kuaishou still maintains data boundaries between distributors, but when the same drama later adjusts its creative, bidding method or audience targeting, it can inherit the successful cold-start data instead of starting from zero.

19. Semantic IDs encode multimodal content into inheritable, decodable assets

庄明浩 says language models, multimodal systems and coding originally looked like 3 separate routes, but increasingly resemble a single line. Zheng Xun’s application answer is that Kuaishou will “turn multimodal into language on a large scale,” then convert language into embeddings or other strategy inputs.
An ordinary video ID only indicates upload order: 1001 and 1002 have no semantic relationship, so the ID space naturally keeps expanding. An image embedding can express pixel-level relationships, but cannot directly tell the system whether the image depicts a cat or a dog. Categories and tags are too broad to carry fine-grained relationships.
Kuaishou is therefore building a “semantic ID,” similar to an identity-card number: the first digits correspond to broad categories, while later digits break them down layer by layer. Two videos can match at the first layer and separate at the second, inheriting upper-level data while preserving their differences rather than treating every piece of content as an unrelated new object.
This encoding depends on multimodal systems converting video into text and language, then understanding its semantics. The code can be parsed backward and traced to its meaning. The model appears more complex on the surface, but several rounds of translation turn an unreadable black box into a traceable structure.

20. Agents are first taking over bounded bidding decisions and moving toward a natural-language operating system

Kuaishou has already brought large models into the ad-bidding game. The model assesses whether an account has completed cold start, how its current volume is running and how sensitive the client is to cost, then sets upper and lower bounds, bid coefficients and boost strategies. During major promotions, it can loosen the limits; lead-generation ads and small clients may require tighter cost controls.
This knowledge was previously passed along informally among product and operations staff, who repeatedly adjusted parameters in group chats every day. A large language model can analyze an account’s history, the current status of other accounts and the product’s campaign habits, then continuously issue the relevant instructions. Humans cannot make global judgments at the same granularity and frequency.
庄明浩 asks whether Agents might eventually use a language humans cannot understand. Zheng Xun believes that is possible, but for now still wants to preserve natural language because it makes debugging easier: one Agent understands the account state and outputs natural-language instructions, while another module configures the coefficients accordingly.
Their shared correction to the idea of an “Agent year” is that the gains being realized today do not come from omnipotent general-purpose Agents, but from understanding and executing existing knowledge, experience and patterns within clearly defined boundaries. The room for large language models to “really create or invent things” remains limited; breaking work into steps and using human-machine collaboration is more stable.

21. In 2026, the task is to raise the technical ceiling while pushing the commercial floor into organic traffic and campaign systems

Zheng Xun divides the future into 2 lines: foundation models raise the ceiling, while commercial teams keep bringing the floor into production. The next step for multimodal systems is not merely making visuals broadly usable, but accurately matching lip movements, hair, eye movement, facial muscles and physical actions, potentially reshaping the cost structure of long-form video and even theatrical films.
庄明浩 compares this with complex scenes in Ne Zha: large numbers of characters, the bend and elasticity of chains, and interactions with seawater all require extremely high manual costs. If models can handle those technical details, creators can put their attention back into “telling the story seriously.”
On the content-platform side, Kuaishou plans to gradually connect organic-traffic targeting for short dramas, comic dramas and mini-games, then use traffic and revenue-sharing policies to help copyright owners lengthen payback periods and reduce financial pressure. AI releases some costs; Kuaishou’s traffic and revenue-sharing policies release another part. The goal is to “let people who are good at making content and telling stories produce good content.”
The bolder milestone is allowing clients to connect to the campaign system directly in natural language and complete the work now handled by professional bid operators. Kuaishou has already exposed the model-generated content-understanding layer to clients for correction—for example, when a mother-daughter relationship is misclassified as a romantic relationship, the client can correct it and feed the correction into the reward mechanism. Zheng Xun believes it “won’t take too long” for this capability to open up to more bid operators.

AI Markets: Deep Dive with a16z's David George a16z, Anish Acharya: Is SaaS Dead? Do Margins Still Matter? Why We Are Not in an AI Bubble?