Summary

Wanaka is not betting on “the next TikTok,” but on using AI-native game content to seed a social network built on existing relationships. 张阳 argues that most ordinary people’s AIGC has no value to the internet at large, but matters to friends because of who made it: “ordinary people’s content really has only one destination—to be shown to your friends.” New game-like content could account for more than 70% of the value early on; over time, gaming and social should each represent at least half.
The moat is not just getting AI to write code, but rebuilding an AI-native game engine that serves both AI and humans, with a particular focus on 3D. Wanaka plans to leave roughly 40%-50% of scene-building to people, giving them a sense of control and the “building with LEGO” pleasure through a graphical interface, while Agents handle props, rules and even complex transitions from parkour to FPS. 张阳 says the current experience can already approach Eggy Party; “just solving the code” is nowhere near enough.
张阳 rejects pure prompt-to-content generation paired with swipe-up-and-down distribution because it hits two walls at once: the ceiling on human expression and the structure of supply. TikTok’s short format was a platform choice; early iPhones could already shoot long videos. AI mini-games are short because the tools are limited, homogeneous and hard to control, making it difficult to support the massive diversity recommendation algorithms require. “Creative ability is actually very difficult to truly equalize”; stronger models will not automatically erase the expressive gap between directors and ordinary people.
3D is Wanaka’s key lever for creating nonlinear gains in user experience, and advances in 3D generation, animation and rigging over the past six or seven months have lowered the bar to execution. Turning a figurine into an animated 3D character and putting it into a game feels like something that “was completely impossible before, something you had never seen”; turning the same asset into a 2D image delivers far less surprise. 张阳 stresses that UGC does not need AAA asset quality—for ordinary users, “it moving” is already a huge upgrade.
张阳 believes Roblox’s DAU should be approaching 200M, validating demand while exposing both the coding barrier and the opportunity created by an ecosystem dominated by professional studios. Roblox Studio lets beginners build a playable map in five minutes, but going further requires learning Lua, which 张阳 estimates blocks more than 90% of users. Meanwhile, top content increasingly comes from OGC—professional game companies and studios. “At first nobody understood it, but by the time you did, it was already huge.” Wanaka wants to avoid becoming captive to supply from a handful of institutions.
The key cold-start metric is not downloads but friend-invite rates and local relationship density. 张阳 helped take Hello from zero to 50M DAU in under a year and saw LiveIn surpass TikTok in downloads and enter the top 3 of the North American overall rankings. He concludes that mass-entertainment products depend more on KOLs, organic virality and “exploding one pocket at a time” than on performance advertising. In beta, the first question is whether users “want to pull their friends in to play together” after finishing a session.
张阳 sees the next 2-3 years as critical for Wanaka’s community-building and remains bullish on AI Coding, Agents, multimodality and consumer applications accelerating. Wanaka aims to create a two-sided flywheel between a creator community and social relationships: creators emerge who can make genuinely good content only here, while friends keep migrating in. His longer-term view is that people may need not “a personal assistant” but “someone who goes to work for me”; Token costs and user authorization and privacy remain constraints.

Deep dive

1. Being the top decision-maker means defining the problem first—and owning every outcome

Two years ago, 张阳 was still at Answer.AI; he has now spent roughly a year building a company. He had an unsuccessful startup experience in Australia early in his career and spent a long time questioning whether he was suited to be a CEO. That experience ultimately convinced him that in periods of violent change, it is difficult to influence product decisions at the foundation without being the person at the top.
The biggest difference between a CEO and a number two is not simply greater responsibility, but that “you cannot escape it—every problem you face is ultimately yourself.” In a job, OKRs and KPIs are usually defined by someone else. A CEO must determine even what the problem is, with no excuse whether the outcome is good or bad.
The pace of technology has been more encouraging than he expected. Between 2024 and 2025, the market briefly thought improvements in model intelligence were slowing and questioned whether the bubble had grown too large. Looking again in 2026, he feels the possibilities for expanding outward from Coding and Agents “are actually where things seemed to be just starting in 2020.”
Wanaka began taking shape in April or May 2025, when few people in the market held similar views. By December, the narratives of many companies had rapidly converged. 张阳’s revision is not that the direction was wrong, but that “the whole thing developed faster than we imagined”; the real separation will come from concrete solutions and product execution.

2. WebSim opened the door to prompt-generated interactive content, but the market quickly split in two

张阳 traces this product cycle back to WebSim, which emerged in late 2023 and 2024. It used AI Coding to generate front-end pages, simulating Instagram and Facebook while also creating small interactive experiences typically consumed for only 10-30 seconds. It inspired products such as Low Ball.
The market is broadly moving along two tracks. One treats mini-games as content assets, using a TikTok-like feed for discovery, creation and consumption. The other treats AI as scaffolding for game development. Unity, Roblox, TapTap and major game companies are all trying to connect AI to existing engines or production systems.
The demand signal was direct. After Claude 3.7 launched in March or April 2025, large numbers of users on Twitter began making games with it, and some Game Jams required that more than 80% of the code be written by AI. What people recreated were Angry Birds, Flappy Bird and Snake—not e-commerce sites or BBS forums.

3. Pure prompt generation first runs into the ceiling on human expression

Wanaka initially also wanted to make fast, lightweight content driven by natural language, because the team believed model capabilities might only reach that level. Testing quickly exposed severe homogenization: further prompt optimization could not reliably help ordinary users make good work, so the swipe-feed approach was abandoned.
张阳 rejects the direct comparison with early TikTok. TikTok videos were short because of a distribution choice; iPhones at the time could already shoot long videos. AI mini-games are short because the creative tools support demos, are difficult to control and cannot sustain long consumption. “The ceiling of its creative tools is completely different.”
曲凯 used Jimeng as an example: the model already produces strong results, but he can only write two sentences, while skilled users can describe shots and storyboards like directors. 张阳 added the example of moving a castle from point A to point B: a 3D space involves precise X, Y and Z coordinates, which are difficult to express efficiently through natural language alone.
The constraint is therefore not just model capability. People need more intuitive ways to express themselves, and models need controlled environments and tools like those provided by Manus. A single general-purpose model generating everything end to end will struggle to cover creators’ full requirements for scenes, assets, logic and control.

4. An AI-native engine must draw a clear boundary between human and AI work

The history of game-production tools runs from assembly language through professional engines to Roblox. Each improvement in editor capability expanded the creator base; the Roblox era made it possible for tens of millions of people worldwide to try making games. 张阳 believes the next-generation engine must first be “for AI to use.”
But “for AI to use” does not mean putting every operation into a chat window. Early products such as Rosebud hoped to govern all interaction through conversation, but 张阳 sees conversation as a high-level form of expression that “most people actually cannot articulate.” Human clicks, drags, placements and direct edits inside an environment remain irreplaceable.
Wanaka is therefore designing the system in two parts. The Agent gets an environment in which to understand scenes, call tools and execute tasks; people retain the work best suited to visual manipulation and aesthetic judgment. As more auxiliary models emerge, the boundary between a strong Agent product and the models themselves will gradually move closer, even if the team does not yet train a foundation model.

5. Swipe feeds are not an interface choice but a set of supply and distribution constraints

张阳 rejects swipe feeds not because content must be short, but because this consumption model requires supply with enough diversity and volume. Once a platform relies on recommendation algorithms, it must continuously generate massive quantities of interchangeable content. Ordinary users’ current mini-games are neither diverse nor stable enough.
张阳 first classifies distribution into three types: algorithmic recommendation, following relationships and social distribution among friends. 曲凯 adds operational or editorial recommendation, which 张阳 notes already existed in the BBS and portal eras. ByteDance made algorithmic recommendation so successful that entrepreneurs now assume every content product should use it, even though production methods, creator structures and distribution mechanisms shape one another.
Following a TikTok creator has relatively little value; an account with 1M followers may not have a clearly understood identity. Bilibili and Xiaohongshu depend more on following relationships, and a creator with tens of thousands of followers can still be highly valuable. The difference is between a media-style content pool and a community-style personal relationship.
Distribution among friends is the hardest because it depends on stable relationships. Yet some content can only travel that way. Moments and Instagram built the relationship network first and then carried content on top of it. TikTok’s Friends tab created a familiar-person outlet after the platform’s vast volume of ordinary UGC could not enter the public feed.

6. The scarce value of ordinary AIGC is not the content but “you made this”

张阳 uses “a streetlamp downstairs or a cat” to explain relationship-based distribution. Such a casual snapshot has no value to the internet at large, but friends may find it interesting and like it because of who posted it. “What they are consuming is actually the relationship, not the content itself.”
AIGC will also stratify. Beautiful videos, images or game-like content made by top creators are suited to follower-based distribution and, once supply is large enough, algorithmic recommendation. But most ordinary people’s work “really has only one destination—to be shown to your friends.”
曲凯 asks whether this is a problem with AI at its current stage or with the game medium itself. 张阳 puts the answer on the human side: “Creative ability is actually very difficult to truly equalize.” Models can raise the floor, but they will not automatically make someone without visual language produce the same content as a professional director.

7. Interactive content and games follow two different consumption curves

For most AI-generated products today, 张阳 prefers the term “interactive content,” because the core gameplay and game loop are too weak. Wanaka wants to make its tools powerful enough to cover lightweight interactions from the most complex games downward, rather than working backward from simple demos toward heavyweight games.
曲凯’s objection is worth preserving: a well-made heavyweight mini-game may hold someone for several hours and need no constant swiping. Lightweight interaction resembles early Flash content—too shallow, and perhaps unable to support daily repeat consumption. Each end weakens the case for “TikTok for mini-games.”
张阳 agrees that interactive content may eventually move closer to video: multimodal or world models generate a scene first, followed by light interaction. The main players there are more likely to be Google, ByteDance or Jimeng. The game-oriented end can support repeat consumption and networked collaboration—a form of “playing one or two rounds with friends when you have time” without lasting as long as Honor of Kings.

8. The shift from 2D to 3D was driven by user experience

The team initially had not resolved the 2D-versus-3D tradeoff. Testing showed a clear difference in depth and feel of consumption. When a user photographs a figurine, turns it into a movable 3D character and puts it into a game, the result creates a clear Aha moment. Turning it into a 2D image and making it run across the screen feels much weaker.
The Wanaka team had no prior 3D experience, but technical iteration offset part of the learning curve. In only six or seven months, 3D generation, animation and rigging changed dramatically. Early attempts to photograph a toy often produced severe mesh breakage; later, even smaller models could perform well.
张阳 stresses that UGC and studio content have different quality thresholds. AAA developers may consider these models unusable, but when ordinary users see their own creations “come to life,” the experience has already jumped a level. Roblox’s large volume of rough, even childish visuals shows that high fidelity is not a prerequisite for participation.

9. Roblox’s real moat has extended from its editor to social relationships

Roblox initially concentrated on elementary- and middle-school users in the United States. Its long-term challenge was aging users up. 张阳 observes that the age profile has been moving upward over the past year or two. College women joking on TikTok that they are “still playing Roblox at this age” itself shows how widespread the behavior has become.
Content complexity has risen with the age of the user base. By 2025, Roblox had begun to see SLG, strategy and heavily numerical games, a clear change from an ecosystem previously dominated by parkour and party-style lightweight play.
The larger change is social. Interviews with young people in North America repeatedly revealed blank Instagram profiles: they post only disappearing Stories or use Instagram as DM. A large share of real interaction has moved to spaces such as Roblox and Minecraft, where people typically open rooms and play with acquaintances rather than match with strangers.
张阳 puts Roblox DAU at “probably close to 200M.” It has risen steadily for more than a decade, and its time spent is among the few products in North America capable of competing with TikTok. “At first nobody understood it, but by the time you did, it was already huge.” Friends remaining there is a key reason users stay as they grow older.

10. Game-based social works by moving existing relationships, not creating strangers

Looking back at domestic attempts such as Werewolf and Momo, 曲凯 notes that most used games to help strangers meet. Roblox follows the opposite logic: friends enter a game space together. 张阳 agrees with the distinction. The former constructs relationships; the latter gives existing relationships a continuing shared activity.
One reason similar domestic platforms struggle to scale is that WeChat and QQ have already absorbed so many relationships. North American users are more willing to separate work, family and different friend groups. iMessage, Instagram and Snapchat each own different contexts; no single application mixes relationships as completely as WeChat.
The United States continues to produce periodic openings in social. The wave that produced LiveIn came from iOS Widgets; other waves have come from LBS, multiplayer video and Party products. A product can suddenly explode within a local group without retaining those relationships over the long term. This helps explain why North America keeps producing new products while China sees fewer such windows.
张阳 believes game-based social has global relevance. China’s RPO is also “extremely high,” but the local structure differs. QQ Zone, Penguin and QQ Farm were classic examples of gamified social. Children in China today can directly access adult games such as Honor of Kings and battle royale titles; age segmentation is weaker, reducing the exclusive space for a Roblox-style children’s platform.

11. Social products survive over time through either solving a major problem or continuously supplying new play

The first path is removing structural pressure from a relationship network. 张阳 believes Snapchat solved the “last major problem” in social among acquaintances: when content remains permanently visible and the audience is always present, people instinctively stop posting. Ephemeral content made users willing to express themselves again.
张阳 notes that Instagram posting rates declined during the same period, as its ecosystem gradually shifted from UGC toward PGC. 曲凯 believes Stories later halted the decline and blocked Snapchat. He also notes that Facebook has described itself in earnings reports or meetings as a media company, suggesting that pure relationship networks may gradually become media networks.
The second path is continuously supplying relationships with new information or new ways to play. TikTok creates large volumes of content worth sharing, with social emerging as the result of sharing. Roblox continuously creates games to play with friends through the same mechanism. Wanaka is not trying to reinvent IM; it is choosing the path of content feeding relationships.

12. Wanaka’s personalized-asset entry point points toward the next Roblox

Over the long term, 张阳 assigns gaming and social a weighting of “at least 50/50, and possibly more social.” But a social network needs an entry point; it cannot be built out of an address book or IM alone. New game-like content may account for more than 70% of the value early on, while the product must make it clear from the moment users arrive that they should add friends and invite people.
Compared with traditional game companies and Roblox’s existing content, Wanaka’s entry point is that users define the assets and scenes. Traditional platforms provide a complete game first, with the company fixing the characters, gameplay and materials. Wanaka wants most characters, maps and content—apart from some gameplay—to belong to users’ own lives.
This is also why it is not “the next TikTok.” Algorithmic distribution requires strong PGC or PUGC as the backbone; early TikTok creators making finger dances and lip-sync videos already had visual appeal and on-camera ability. Ordinary people’s personalized content derives its value instead from “this is my classmate, my school or my friend.”
张阳’s conclusion is direct: “The proposition of ‘the next TikTok’ does not really hold.” AI-generated videos can still be distributed on TikTok or Douyin. If products such as Jimeng add light interaction, an independent medium might avoid the decline in VV and ad impressions caused by interaction, but the underlying multimodal and distribution advantages will remain concentrated at companies such as Google and ByteDance.

13. Roblox Studio proves the importance of editors—and how difficult they are to replicate

Compared with Unity, Cocos and Godot at the time, Roblox Studio’s ease of use was “worlds apart.” Someone with no engine background could build a playable map in five minutes, make an NPC run or drive a car. For children, this is a powerful first creative experience.
Then comes the coding wall. To keep improving a map, users must learn Lua; 张阳 estimates this blocks more than 90% of children and adults. The opportunity presented by AI is to redesign this barrier, not merely add code completion to an old editor.
Roblox and Unity both began adding AI after 2023, but 张阳 compares the model to Copilot: it is more useful for people who already know how to code and understand engines. The old systems are difficult to rebuild wholesale, while game assets, scenes, interfaces and logic cannot be separated into independent modules as easily as ordinary software.

14. Replacing only code will not rewrite the engine ecosystem; platforms must also manage professional supply

张阳 believes game assets and scene placement may contribute “more than half of the value.” If a product only lets AI write code, there is no fundamental difference between its output and telling Gemini or ChatGPT, “make me a game.” It still cannot support a complete, controllable game expression.
The engine itself is not an ideal business model: “A good engine company often makes money not because the engine makes money, but because the game makes money.” He cites Fortnite behind Unreal Engine and CS:GO in Steam’s early days to show how distribution platforms often use flagship content to attract users before building an ecosystem.
曲凯 raises, while explicitly saying he is “not entirely sure,” that Roblox’s most popular games may contribute a large share of the market and most of the revenue. 张阳 believes Wanaka will not follow that path because he thinks Roblox has not managed its PUGC ecosystem well. It now includes not only PGC but also what he calls OGC—Organization Generated Content. Many professional studios and game companies have entered, with some teams pursuing DAU rather than profit.
Wanaka wants to avoid copying this trajectory. 张阳’s ideal structure is that even if professional content contributes 80% of traffic, the platform should reserve roughly 20% for UGC. Otherwise, it will become captive to the professional-content ecosystem and struggle to cultivate supply from ordinary people.

15. Wanaka’s product is a graphical editor + Agent + personalized consumer experience

The creation side is an AI engine with a full graphical interface. Large amounts of code and game logic can be completed through conversation, but people will handle roughly 40%-50% of scene construction, placing maps and adjusting assets like LEGO. Control itself is part of the creative pleasure.
The Agent framework handles abstract logic better suited to natural language: adding props or changing rules at any point in a parkour game, or even switching the entire game into an FPS. The goal is not to assist professional developers, but to let people with no game-making experience produce “something really good.”
The consumer side compresses 3D personalization into a simple action. Users photograph themselves, friends or figurines, quickly generate characters and place them in role-playing, parkour or other games. The platform should offer “an infinite number of games to play with your friends,” while making every session carry the group’s own materials.
曲凯 asks about the upper limit of complexity. 张阳’s current reference point is Eggy Party: “It should be about there.” Roblox is a clear competitor; Eggy Party is more like an independent game company than an exactly comparable platform rival.

16. AI Coding is turning human-written code from an advantage into an ecosystem burden

When the team initially made game templates internally, members were still allowed to write code by hand. As Coding capabilities improved, the team recently began asking them not to write code and to use AI generation instead. Human-written code can damage a unified ecosystem and make it harder for AI to understand, modify and recombine the system later.
Token consumption is the clearest signal of capability. Early on, a programmer’s roughly RMB200 monthly plan was generally sufficient. Now it “cannot cover it at all,” showing that AI is no longer occasional autocomplete but is handling an increasingly large share of product and content production.
张阳 does not conclude that people will leave creation. The right tools should let people focus on scenes, aesthetics, gameplay intent and selection, while Agents handle code execution that is poorly suited to humans. The “proportion of human participation” will be restructured, but human expression will still determine the differences between works.

17. Ordinary people make games first for the process, not traffic or revenue

张阳 agrees with 蔡浩宇’s view that two types of games will coexist in the future: a tiny number of elite creators will continue making work through traditional craft and as artists, while the remaining 99% of content will be created by ordinary people. Wanaka’s real question is how to support that 99% in making things and where those works will ultimately be distributed.
Independent games and Game Jams showed him a group of “love-driven” creators. Participants may not know how to code and may only handle design or art, but they are willing to complete a work around a theme within a limited time. AI’s value is to let people who love creating but lack complete skills independently realize an idea for the first time.
Minecraft players may spend a long time building a single structure, deriving the same satisfaction as building with LEGO. One friend described college women making UGC maps in Eggy Party as “electronic cross-stitch.” The process itself is enjoyable; whether many people play the work or whether it makes money is secondary.
Social distribution will also change game design. Comparable FPS and simulation games on Roblox often remove the long onboarding found in traditional Steam games because they assume a friend is sitting beside the player and will explain how to play. Relationships do not just distribute the product; they replace part of the product tutorial.

18. Personalized gameplay turns real-world groups directly into game materials

Roblox has evolved many mechanics that work only through coordination among friends: two people tied together running a course, or turning a friend into an item. Wanaka wants to go further, mapping maps, characters and rules to specific relationships rather than merely letting friends enter the same standardized game.
张阳’s ideal scenario is a parkour map based on one’s own school. It could be extremely popular at that school and meaningless to outsiders. That locality is not a defect but the core of content for friends: its value comes from shared memories, not universal quality.
Another example is putting classmates into a merging game: “Merge two teachers into a dean of students; merge three deans of students into a principal.” Traditional game companies would struggle to produce this at scale, but it would naturally prompt real-world friends to invite one another, watch and remix the content.
For non-creators, the Aha moment is quickly generating their own character and familiar scenes, then immediately playing “our thing” with friends. For serious creators, it is discovering a tool that is simpler and more controllable than both pure AI conversation and traditional engines. The two experiences must ultimately converge within the same consumer network.

19. Cold start depends on local density; the long-term flywheel links creator communities and friend networks

张阳’s growth experience spans multiple products. At ByteDance, the India-focused content product Hello went from zero to 50M DAU in under a year. At its fastest, LiveIn surpassed TikTok in North American downloads and entered the top 3 of the overall app rankings. Answer.AI relied on explaining its core tool-driven need clearly and pushing TikTok content viral.
Wanaka cannot copy the playbook for tool products. Mass entertainment, content and games are more likely to grow through KOLs and organic invitations than performance advertising. In beta, the key question is: “After using this product, do you want to pull your friends in to play together?” If users never think about inviting anyone, something in the product is wrong.
Social products usually “explode one pocket at a time.” They first need high penetration in one school, region or social circle before expanding outward. 曲凯 connects this to early Facebook’s entry through campus photos and ratings. 张阳 agrees that the key is not broad traffic but sufficiently high relationship density within a group.
Over the next 2-3 years, Wanaka must validate two loops: whether friends will continue migrating in on the consumer side, and whether the creator side can produce people who cannot make good content elsewhere but can do so quickly on Wanaka, eventually becoming platform KOLs. The community grows around the production tool, creators push the tool forward, and social handles consumption and distribution.

20. Both ends of AI social work, but most people in the middle still need other people

张阳 places new AI social on a spectrum. At one end are OC and anime-focused users, who have rich imaginative capacity and can build a world with multiple AI characters, an experience close to a single-player game. He sees this as a precious state, like children imagining life in dolls and toy cars.
At the other end are highly expressive entrepreneurs and investors who post constantly and receive replies from a group of AIs, generating high information density and feedback in the short term. But this kind of network requires continuous human output. It cannot cover people in the middle who lack the energy to imagine constantly and do not want to express themselves publicly on a frequent basis.
He therefore does not believe a network consisting only of people and AI can replace existing social networks: “Because you are not speaking, AI cannot give you feedback.” Wanaka still makes people the social subjects. AI supplies additional content, creative tools and some character dialogue; what truly supports relationships is game-like content that people can consume together.
OC users overlap with game creators. They are more comfortable with virtual characters and worlds and more willing to construct personalities. Wanaka will offer AI dialogue capabilities to serve this group but will not reduce the entire product to companion chat. 张阳 believes East Asia has a broader anime-oriented environment, while the corresponding group in North America may be much smaller.

21. Product judgment starts with intuition, then data and business capability keep the company alive

张阳 still defines his strongest skill as product. At Xiaomi, he received “very classical” training: look at less data and think deeply about how users use a product and what their underlying needs are. The craftsmanship of the early Douban and Tieba generation of product builders was materially different from that of the later ByteDance-style product manager.
ByteDance trained him in hypotheses, experiments and data validation, close to Facebook’s experimental-science system. But he warns that major opportunities often cannot be planned out from existing data. Overreliance on logic weakens sensitivity to users and makes teams afraid to trust intuition for fear of “self-indulgence.”
He is looking for abnormal signals rather than surface answers in surveys: Why did young North Americans suddenly stop posting on Instagram? Why was making a game people’s first reaction after getting AI Coding? Why can a child invent gameplay with a piece of paper and a pen? “Feeling is the most fundamental thing in making products—or you could call it empathy.”
Growth, operations and monetization are not his strongest areas, but he has practiced all of them personally. That lets him step in during the launch phase and build virality and monetization conditions into product design. His balance is that intuition finds real demand, while logic and business judgment keep the product and company operating.

22. 2026 could be the consumer inflection point, but convenience will require more user permissions

By the end of the conversation, both speakers had revised their sense that technology seemed to be slowing in the second half of 2025. 曲凯 again became firmly bullish on 2026, expecting AI Coding, Agents and multimodality to advance together. 张阳 also feels that every Spring Festival has become a “carnival” of product and model launches, compressing 3 years of change into an unusually dense period.
曲凯 notes that as of 2025, apart from model companies’ own Chat products, no consumer AI application appeared to have surpassed Character.AI at peak scale. 张阳 believes Token costs, payment structures and intelligence have delayed a mobile-internet-style breakout, making it possible that some consumer products will explode this year.
张阳 raises a new question prompted by a personal-assistant product recorded in the source as “Open C lo ud” and “Open Clos”: Do people really need a personal assistant, or only “someone who goes to work for me”? If people no longer need to work, 曲凯 believes demand for assistants may fall, freeing time for social and games. 张阳 adds that much of the time could also go toward creation.
The cost is permissions. 张阳 places that product alongside the Doubao phone: for AI to become genuinely useful, users will voluntarily surrender access to their mobile devices and personal privacy. If refusing to use it means being unable to find work, authorization may shift from a choice to a condition of survival. 曲凯 calls the concern “very American.” 张阳’s answer: “This seems unavoidable.”

OpenAI Vs. Anthropic: How the Pentagon Picked Its Partner The SaaS Apocalypse: Who Lives & Who Dies | Insight Partners Co-Founder, Jerry Murdock