Summary

The key dividing line repeatedly identified in this discussion is not demand, but whether unit economics can turn positive on every call. 朱哲清 estimates that 80%—90% of industry costs now sit in inference. A single Deep Research run exceeding 1M tokens costs roughly $3—4 on domestic models and $8—10 on top overseas models; Grok can run into the tens of dollars. He has “not seen a single company able to break even on costs” in a C-side use case. Survival routes include model routing, context compression, vertical fine-tuning and self-hosted serving. For C-side products, the question is also whether model providers can cut prices by roughly 80% while leaving room for profit; otherwise, continued cash burning eventually ends in collapse.
Cursor’s $10B valuation and the simultaneous backlash over limits on its $20 plan show that application-layer growth does not automatically become gross margin. 泓君 cited the company’s $900M financing round in June and the view that Cursor and GitHub together contributed roughly $1.2B of Anthropic’s approximately $4B revenue last year. 陈志博 put it bluntly: Cursor is “basically working for Anthropic.” With Claude, OpenAI, Gemini and xAI all treating coding as a primary battleground, Cursor must absorb upstream inference costs while competing head-on with its suppliers. 泓君 speculated that the pricing change may also be intended to improve gross margin ahead of the next financing round.
Poke AI is trying to turn the entire web’s toolset into the model’s action space, tackling tool discovery and context costs at the same time. A single model may support only tens of thousands to 1M tokens of context, while descriptions of all web tools could run to tens of millions of tokens. Traditional Agents replay the full history and tool information at every step, causing the same content to be processed 20, 30 or even 50 times. Poke wants developers to call an Agent as easily as they call the ChatGPT API, but model reliability still falls sharply when it faces thousands of highly similar tools.
The three companies are betting on three different moats: tool ecosystems, identity networks and design delivery. Second Me distinguishes between the “doing” of a profession and a person’s “being,” aiming to give everyone an identity Agent that represents them and connect those Agents into a network. Lovart is not competing with foundation models such as Nano Banana for point-generation capability; it uses a designer-style planner, aesthetic memory and an editable canvas to deliver personalized work. For Lovart, model capability is raw material: “Only water deep enough can carry a ship large enough.”
Training paradigms are shifting from SFT toward RFT and RL because of the data bottleneck, but “learning through use” remains far from solved. 朱哲清 believes the most expensive part of SFT is now licensed data and labeling for complex tasks. RFT/RLVR comes at the cost of longer, less stable training and weaker generalization. 陶芳波 summarized the shift as: “The SFT era was about finding data; the RFT era is about finding environments.” Their key disagreement is whether the next generation will enter an era of experience: 陶芳波 believes online use will make Agents smarter, while 朱哲清 stresses that few-shot feedback may be an outlier and that safe inference-time updates are “extremely difficult.”
B-side Agents are already close to breaking even, while C-side products are taking three different bets: subscriptions, network effects and outcome-based revenue sharing. Poke can cover costs on the B side; on the C side, subscriptions may cover inference but not labor. Lovart uses tiered credit subscriptions, testing delivery value through whether users keep buying. Second Me is choosing to charge as little as possible in order to build an identity network; the possibility of an AI-native advertising model is 泓君’s growth bet, and 陶芳波 agreed. Pine takes a cut of the amount it recovers in users’ billing disputes, offering the most direct answer: “Take a percentage of the final outcome.”
Trust in and the durability of high-priced KOL traffic are weakening, while community and the Agent kernel are becoming more important distribution and platform assets. Bill believes expensive KOL traffic decays quickly and carries weak trust; Runway and Luma had more durable early growth through co-creation on Discord. 陶芳波 sees the Claude Code SDK as a deeply underestimated “coding kernel” that could even become the CUDA of the Agent era. 泓君’s changed view of Cluely, Lovable and Replit also shows that value need not come from fully automated delivery: real demand can exist in live summarization, prototype communication and giving people more confidence.

Deep dive

1. Agent valuations take off first; unit economics come to collect later

泓君 used three tracks to sketch the heat of 2025: coding Agents such as Cursor, the $5B legal Agent Harvey AI, and general-purpose Agents such as Manus, Genspark and ChatGPT. She said the current lifespan of these foundation-model and coding-Agent companies is roughly 2 years. Cursor raised $900M at a $10B valuation as early as June—roughly 3x the valuation she cited for 月之暗面—but the shift “from dazzling success to deep crisis” can be just as fast.
The three founders represent three product paths: 朱哲清 (Bill) is building Poke AI as a general-purpose, tool-calling Agent model alongside a tool ecosystem; 陶芳波’s Second Me is trying to build an AI-native identity network; and 陈志博’s Lovart integrates multimodal foundation models so ordinary users and professional designers can create higher-quality visual work together.
朱哲清’s background shaped his technical orientation. At Meta, reinforcement learning was used for ad-creative rewriting and budget pacing; the former lifted ad CTR by more than 6.7%, while the latter corresponded to several hundred million dollars in revenue. He started his company in late 2024 because he believed “reinforcement learning is probably about to take off, so it was time to take the plunge.”

2. Poke compresses the web’s tools into the model’s action space

朱哲清 described the basic contradiction: mainstream models support anywhere from tens of thousands to 1M tokens of context, but putting descriptions of every internet tool into a prompt could require tens of millions of tokens. A truly autonomous Agent should not need to know in advance whether a customer uses Salesforce or HubSpot, nor should it require users to first discover a specific beauty tool before removing acne from a face.
Poke is therefore building more than workflows. It wants to turn tools from prompt descriptions into “part of the model’s action space,” allowing the Agent to discover and call tools across the web on its own. The end goal is to make building an Agent “as easy as building a language model like ChatGPT,” while avoiding the repeated submission of the entire history and tool list on every round.
In a traditional architecture, the same information passes repeatedly through an expensive model across 20, 30 or 50 consecutive calls. 朱哲清 cited Perplexity, saying its losses are roughly comparable to its revenue and that it loses several tens of millions of dollars a year, with tool calling and context management both dependent on expensive models at the core. Poke’s second problem is therefore not capability, but pushing the cost down to a level where developers can build Agents “without losing money.”
For the podcast workflow 泓君 proposed, Poke can already edit a video and upload it to YouTube with a single prompt. It can also handle TikTok, Instagram, Facebook, LinkedIn and Twitter, check and respond to comments hourly, or pull down videos, generate transcripts and redistribute them. The selling point is not any single tool, but continuous cross-platform execution.

3. Ecosystem openness and human-in-the-loop determine the order of commercialization

The openness of domestic and overseas platforms creates a clear boundary. 朱哲清 said YouTube, Facebook and Instagram require partnerships but are broadly accessible; 哔哩哔哩, 小红书 and 知乎 are essentially impossible to integrate. The team is speaking with platform C-suites, but he believes front-end websites may ultimately converge toward chatbots, and that platforms opening themselves to Agents is “just a matter of time”—“an irreversible trend.”
Tool volume remains a hard ceiling for models. Once each platform contains many operations and the total reaches thousands of similar tools, call reliability falls sharply; the more similar the tools, the harder they are to distinguish. 朱哲清 did not present “all web tools” as a solved problem, explicitly acknowledging that “models still need to improve, and they’re all very expensive.”
Product priorities are determined first by real enterprise demand, then by whether the ecosystem already has an API and whether revenue sharing is involved, and finally by how much human participation the task requires. As a PLG company, Poke will not immediately customize for a single customer; it looks for common problems shared by “20 companies.”
When 泓君 asked whether a general-purpose Agent should first focus on publishing, transcription or content generation, 朱哲清 identified human-in-the-loop as the next major pain point. Most Agents generate and publish a video, then leave users with only “so what, how do I change it?” Lovart allows users to continue interacting with and adjusting the output, which is precisely the kind of design he believes is worth preserving.

4. Second Me bets on “being,” not another task tool

陶芳波 defines a professional Agent as “doing something”: roles such as accountants and programmers exist because of their functions. But people scroll through social feeds, eat with friends and chat casually not to obtain a service, but because they care about another person’s “being.” That is the state of existence Second Me wants to model and display.
His analogy is that Facebook profiles moved offline identity online, while AI can absorb 20 or 30 years of a person’s internet history into its parameters and form an identity Agent more interactive than a profile page. “It seems like everyone is asking whether an Agent can do what Google does.” Second Me asks whether an Agent can do Facebook again.
Once network nodes become AI rather than static profiles, 2 identity Agents can discover common topics and connect and interact with each other. Users could also create avatars like Sun Wukong—to accompany their children and parents, or discuss problems with colleagues. 陶芳波 would not claim this is “the next generation,” but insists it is a separate line from task Agents: not bringing the world to people, but “bringing me to the world.”
Identity modeling initially depends on existing social histories, but Chinese platform data is difficult to obtain. Offline information could come from Always On hardware. 陶芳波 mentioned the Lucky and “零宇宙” hardware built by his friend 孙杨, and observed that users are already recording conversations and daily life directly with their phones. As glasses and other devices become more common, this input “may become somewhat more natural,” rather than being a solved problem.

5. The ideal of a personal model is expensive; Lovart treats models as raw material

Second Me’s current online approach uses multi-level prompt abstraction. For an individual’s experiences spanning tens or even hundreds of millions of tokens, a conversation with a new friend may not require all meeting records; it may need only a concise personality representation, such as a Big Five classification. 陶芳波 believes multi-level task-specific context abstraction can reduce costs, but admits that “it’s still very expensive.”
The more aggressive open-source version of Second Me maintains unique parameters for each person and continuously trains them, because people themselves are constantly changing. The team estimates that updating each user’s model alone would cost several hundred dollars per month. This version has not launched and exists only in the open-source system; it should not be conflated with the current online product.
陈志博 positions Lovart as a delivery layer above tools. APIs such as Nano Banana and ChatGPT continue to improve image and video generation, but highly subjective work still needs “a designer’s brain.” Lovart uses a planner to select tools, puts the user’s profession, experience, aesthetic and stylistic preferences into memory and context, and organizes creation according to a professional designer’s workflow.
Professional users can rapidly iterate in batches on the canvas and inject their own judgment, because the output of a single chatbot “will most likely be some distance from the final deliverable.” Ordinary users can receive work approaching professional quality directly. For Lovart, foundation-model upgrades are a tailwind, as 陈志博 put it: “Only water deep enough can carry a ship large enough.”

6. Inference has swallowed most of AI companies’ costs

朱哲清 estimates that 80%—90% of the cost base at AI companies is now inference, while training costs continue to fall. Complex delivery still typically requires top-tier models such as Gemini Pro and Claude 4. Gemini Flash and smaller OpenAI models are cheaper, but struggle to complete high-quality coding tasks end to end.
A single complex deep-research run typically exceeds 1M tokens: roughly $3—4 on domestic models, $8—10 on top overseas models, and potentially the tens of dollars on Grok. 泓君 noted that ChatGPT’s basic subscription is $20 per month and that the C-side conversion rate is below 5%. Even if OpenAI’s own token costs are lower, external Agents accessing similar capabilities can still face a model in which “the more users you have, the higher the cost.”
陶芳波 said Second Me’s inference costs have already exceeded its traditional cloud costs. It optimizes in 2 ways: assigning summaries, context compression and similar tasks to cheaper models such as Gemini Flash; and applying multi-level abstraction to personal histories, RAG and tool results by task, sending only genuinely relevant information into the model.
陶芳波 offered a conjecture about why GPT-5 uses a router: users tend to “let the better model handle even the simplest task,” leaving the platform no choice but to proactively route requests to smaller models. This was explicitly presented as speculation, not a confirmed product mechanism.

7. Cursor shows how high growth can amplify application-layer risk

陈志博 called Lovart a more standard C-side company and used Cursor to summarize the industry’s predicament: “Cursor is basically working for Anthropic.” His solution is to divide labor across smaller models, shorten inputs, and fine-tune and self-serve models in vertical scenarios. Because the tool space and user-expression space are limited, costs can fall “exponentially” versus external APIs.
泓君’s rebuttal was that long-term cost declines do not solve short-term cash flow. After raising $900M at a $10B valuation, Cursor changed its original $20 monthly plan to a token cap with additional charges above the limit, triggering a developer backlash. She cited the view that Anthropic generated roughly $4B in revenue, with Cursor and GitHub contributing about $1.2B combined, suggesting that the growth may sit on top of enormous upstream spending.
The problem is more acute because coding is precisely a primary battleground for every foundation-model company. Anthropic launched Claude Code, while OpenAI, Gemini and xAI continue to strengthen their coding capabilities. 陈志博 believes Cursor’s market is vulnerable to a scenario in which Anthropic’s new model is materially stronger than the model it sells externally, with an additional inference-cost advantage. The upstream company would then hold a clear edge in competition with the application. Lovart’s design-delivery layer is not the base model’s primary battlefield, leaving more room for the application.
泓君 doubts that Cursor training its own coding model would work, because code is an important weight and benchmark across every major model’s training process. She is more inclined to interpret the limit adjustment as a financing-driven financial move: if gross margin does not improve, “it will not be able to raise the next round.”

8. The data bottleneck pushes training from SFT toward RFT and RL

朱哲清 believes the most expensive part of SFT is no longer token generation but data. Social and content platforms have begun licensing content for legal and commercial reasons. Citing Reddit, he said data licensing to Claude and OpenAI was one of its largest revenue sources last year, showing that data is becoming SFT’s core bottleneck.
Increasing task complexity is also making human labeling ineffective. To label deep research, should a person browse 3,000 websites and then decide whether the Agent browsed correctly? His view is that reinforcement fine-tuning and RLVR are rising not because SFT has suddenly lost its value, but because the pure-supervision path “has reached the edge of getting stuck.”
The costs of RFT and RL are equally clear: training takes longer, is less stable and is difficult to conduct safely directly at inference time. Reinforcement-learning training can also trigger Modality Collapse, weakening generalization; after deployment to a specific use case, additional fine-tuning may still be required. Compute costs may therefore be materially higher than under the original path.
朱哲清 said the RL fine-tuning investment for Grok 4 has already approached the investment in its SFT training, and that 2.1 may require even more. His investment mapping was direct: “Cloud computing will continue to thrive,” citing Oracle’s 30% rise at the time as an example from the room.

9. The “era of experience” remains stuck at safe online learning

陶芳波 looked back at the team’s earlier MindOS as a workflow Agent, while Manus and Lovart represent the next generation driven by Reasoning Models. He summarized the paradigm shift this way: “In the SFT era, I had to find data; in the RFT era, we may have to find environments.” Agent founders need to define environments that are interactive and verifiable, not merely accumulate corpora.
Inspired by Richard Sutton’s speech that day, 陶芳波 believes the next step may be an “era of experience”: RL would not happen only during training, but would continuously accumulate experience through real-world use. The more Lovart users there are, for example, the smarter the model would become. He also admitted that “today I believe this does not exist,” at least in the sense that the model itself does not yet train automatically through use.
朱哲清’s reservation is that exploration has been discussed for years but rarely deployed because margins are too thin. If Meta sacrifices revenue for exploration and spends large amounts of compute to improve performance by only 2%—3%, it is difficult to win organizational support. Exploration is becoming important again because SFT is hitting data and capability barriers; without new data, greedily repeating the known optimal strategy causes the model to stop learning.
He still places exploration at training time: collect new data, update the policy centrally, then redeploy it to the real environment. Feedback from only 3 people on site cannot establish whether it is an outlier. Requiring a model to update immediately and safely from a few-shot signal is the hardest part of inference-time learning: “This is extremely difficult.”

10. B-side breaks even first; C-side business models split into 3 paths

Poke was not yet formally charging at the time and was even giving users unlimited credits, yet some people still paid voluntarily. 朱哲清 jokingly called it “powering the project out of love.” B-side revenue can already cover costs relatively easily. C-side subscriptions can roughly cover inference, but once labor is included there is “absolutely no profitability at all.” The outlook still depends on whether model providers are “generous enough.”
The B-side boundary is avoiding customization that turns the company into a consultancy. Poke sticks to PLG and uses community feedback to find common needs; SLG, by contrast, may do “whatever it takes” to win a large customer. Sending a Forward-Deployed Engineer directly into a customer’s company can ease concerns about data leaving the premises, but it implies extremely high labor costs.
Second Me is choosing the opposite path and trying not to charge. If an identity social network sets a high barrier to entry, it will lack enough nodes to create network value. 泓君 noted that earlier internet companies could delay monetization and build network moats, and imagined a future AI-native advertising model similar to feed ads. She called it her “bet,” and 陶芳波 agreed.
Lovart uses the most direct form of tiered subscription: different plans correspond to different credit levels—“If you want to use it, buy it; if you don’t, leave.” 陈志博 declined to disclose revenue and noted that the mix of high- and low-tier plans mainly reflects user profiles: professionals and AI enthusiasts are heavier users, while general creators are lighter users. The mix primarily shows differences in user types.

11. Product strength and community are replacing expensive traffic

Lovart has moved from invitation codes and KOL distribution to a formal launch. Its marketing team is based in San Francisco, but the core growth engine remains product experience and word-of-mouth “organic” traffic. 陈志博 reduces the logic to a simple formula: keep improving the quality of delivered work until users want to share it themselves.
陶芳波 observed that AI founders have to tell science-fiction stories themselves because early attention is a form of momentum. One shareholder joked: “When every AI founder is telling science-fiction stories, science-fiction writers have no novels left to write.” Founders are consequently forced to appear on camera and manage their image, even taking on highly specific burdens such as losing weight.
陶芳波 disclosed that one Second Me launch video took only half a day to shoot and cost roughly $1,500 end to end. The real problem is not production cost, but increasingly homogeneous industry marketing and capital pushing KOL prices to “astronomical levels.” If distribution relies entirely on paid KOLs, traffic may fall sharply and outsiders may conclude that the product is merely “pure traffic acquisition.”
He has more faith in community. In the early days of Runway and Luma, founders invited users to join Discord at events in San Francisco, collected problems directly, solved needs and iterated continuously. Once core users “can no longer do without you,” distribution begins to happen naturally. The logic is consistent with Poke’s PLG approach of using representative users to set product priorities.

12. Product surprises in 2025 extend from outcome-based pricing to the Agent kernel

One guest mentioned Pine’s outcome-based pricing: the Agent disputes bills with a bank or telecom company on behalf of the user, then takes a cut of the disputed amount. This changes not only the pricing unit, but also forces the product to deliver a “commercially meaningful outcome,” rather than charging users for process, tokens or incomplete tasks.
陶芳波 believes the Claude Code SDK is far more underestimated than Claude Code itself. He speculated that Anthropic may be trying to abstract the “kernel” of its strongest coding Agent into a foundational layer for all Agents, with applications then calling Claude’s base model. If true, it could be to Nvidia what CUDA is; 泓君 compared the idea to Stripe using SDKs and APIs to lower the barrier to payment development.
泓君 initially doubted that products such as Lovable and Replit could deliver useful outcomes. She later found that they let operations and product leads directly create UIs or developer-product prototypes. “I say it looks like this, you say it also looks like this”—verbal agreement can conceal a completely different mental image. A prototype turns communication error into something visible and editable.
泓君 also reversed her view of Cluely, once known as a “cheating tool.” At an English-language reinforcement-learning conference, she did not use interview-answer prompts; she used only live captions and key-point summaries, which gave her more confidence in understanding the discussion and organizing a follow-up outline. Its value was not answering for her, but making a meeting she had “not been very confident about” feel more manageable.

Opendoor's CEO on the Greatest Turnaround in Tech Uber CEO Dara Khosrowshahi on self-driving's future, changing business model, job displacement