Summary

In the first half of 2025, AI’s center of gravity shifted from pre-train scaling to post-train and reinforcement learning, while model competition spilled over into an Agent arms race. DeepSeek R1 brought RL-based reasoning into the public eye; OpenAI o3-pro, Claude 4, and Gemini 2.5 Pro took turns topping the leaderboards, while China saw the R1 0528 release, Qwen 3, and Doubao 1.6. Meta plans to invest approximately $15B in Scale AI and restructure its AI division. Model iteration did not slow as expected; it has instead become an all-out arms race.
The dividing line between an Agent and an ordinary chat application is not another dialogue box, but the ability to perceive its environment, plan autonomously, and change that environment through tool use. Browsers, search, sandboxes, Computer Use, MCP, and coding tools mean models are no longer merely “talking on paper”; their capability frontier is thought to have expanded by more than an order of magnitude. But the “everything can be an AI Agent” boom has also blurred the boundaries between model vendors, application companies, and so-called wrappers.
AI coding was the first to validate PMF—not because code is inherently suited to generation, but because it has a complete and verifiable feedback loop. Code can be generated, verified, compiled, executed, and presented in a virtual environment, while mature compilers and DevOps provide clear reward signals. That is why products such as Cursor established themselves first; embodied intelligence faces a much harder challenge in mapping code to the physical world and obtaining real-world feedback.
The long-term technical bet favors RL-driven Agents, but near-term commercial products will still rely heavily on workflows. OpenAI Deep Research and Kimi Researcher represent the end-to-end RL route, jointly optimizing reasoning, context, memory, tool use, and multi-step chains of thought within a single model. Products such as Manus look more like engineered systems that combine different model capabilities, RAG, and search tools. “The war is far from over”; which architecture can win over the long term remains unproven.
The capabilities worth compounding in Agent startups include vertical know-how, user relationships, and high-density feedback environments—not an advantage built solely on today’s engineering flourishes. The Bitter Lesson reminds teams that “the technology frontier will expand”: browsers, IDEs, and professional tools that retain their GUI can continuously collect context and user feedback. Products should first achieve PMF while laying the groundwork for future learning-based iteration.
Agent monetization has yet to close the loop; subscription, usage-based, and outcome-based pricing each carry their own paradoxes. Users cannot realistically pay $20 per month for every low-frequency Agent. Token pricing would make an inefficient Agent using 10K tokens cost ten times as much as an efficient one using 1K tokens. Outcome-based pricing is shaped by the subjective value of tasks such as travel planning and industry research, while multi-Agent collaboration introduces internal settlement problems. The ideal end state is “more work, more pay,” but measuring outcomes may initially be a problem for sociology.
Over the next 1-2 years, investment preferences will center on vertical To B Agents, niche and personalized long-chain To C tasks, and AI hardware. To B products can evolve from prompts and workflows to agentic workflows and fully managed delivery, extending the chain from tools to services and outcomes. To C products need not pursue a general-purpose platform from day one: niche needs multiplied by a huge user base can create entirely new categories. Travel, long-form content generation, companionship, and consumer electronics are potential directions, but the 2 investors also caution that many views could be rapidly revised or even overturned within 6 months or by next year.

Deep dive

1. Models and Applications Entered an Arms Race in H1 2025

刘鹏琦 sees DeepSeek R1 as the acceleration point: it brought RL-based reasoning models into the public eye and pushed every major vendor into a new competitive arena. 颜前航, meanwhile, summarizes the main shift as a move from pre-train scaling to post-train scaling.
Overseas model development did not slow as previously expected. OpenAI has iterated to o3-pro, Anthropic launched the Claude 4 series, and Google joined the leaderboard rotation with Gemini 2.5 Pro—even floating the concept of an AI OS.
China’s catch-up effort has been equally intense. The 0528 release of DeepSeek R1 significantly improved coding capabilities, while the market continues to await V4 and the R2 built on it. Qwen 3, Doubao 1.6, and other vendors continue to iterate, although some of the “Six Little Dragons” have already abandoned model training.
Meta is facing a different kind of pressure. After DeepSeek disrupted the open-source path, the company is preparing to restructure its AI division and plans to invest approximately $15B in Scale AI. The program argues that its impact on the industry remains worth watching.

2. The Agent Boom Broke Down the Old Boundary Between Model Vendors and Application Companies

Beginning in January, OpenAI rolled out Operator and Deep Research in succession, which 刘鹏琦 sees as the formal launch of the Agent race. Model vendors including MiniMax and Kimi soon released Agent products directly as well; the expected division of “models as models, applications as applications” never materialized.
Cursor has already validated part of the PMF. Windsurf went through OpenAI’s acquisition process and Anthropic’s decision to stop supplying frontier-model APIs, while products such as Lovable and Replit also grew rapidly. Anthropic’s own move into coding is further evidence that “the war is far from over.”
Teams in China and overseas, including Manus, Lovart, and Genspark, competed to launch the “world’s first” Agent for one category or another, pushing the market toward “everything can be an AI Agent.” Some have even described ChatGPT as “an Agent company wrapped around a language model.”
颜前航 acknowledges that 2 years ago, his team believed applications had to keep their distance from model vendors or they could “be swallowed up at any time.” Today, teams operating extremely close to the model layer—once dismissed as wrappers—are gaining attention through user experience and market positioning. It is a direct challenge to that earlier view.

3. The Qualitative Shift in Agents Is From Answering Questions to Changing the Environment

刘鹏琦 divides AI applications into 3 stages. The first is the ChatGPT-style prompt-answer interaction: the user asks, and the model responds. By his estimate, more than 80% of AI applications across the market may still be stuck in this simple interaction.
The second is workflow. People predefine nodes and paths, then engineer together search, private-data retrieval, processing, and model capabilities. Dify, Coze, and LangChain have provided versions of this architecture, and most commercially mature products today still belong to this stage.
The third-stage Agent must “autonomously perceive the environment, autonomously make decisions, and then execute the corresponding tasks.” It reads instructions, context, and long-term memory, but also plans steps, calls tools, and changes its environment—unlike a fixed workflow, which is controllable but less flexible and general.

4. Tool Use Means Models Are No Longer Merely “Talking on Paper”

刘鹏琦 believes the most significant incremental capability of Agents this year, relative to ChatGPT or o4, is not reasoning alone but tool use. Browsers, search, sandboxes, Computer Use, coding, and the increasingly widespread MCP have collectively opened the door to execution.
Pre-trained models are constrained by training cutoff dates and public data. Once AI can autonomously search, retrieve external information, and execute operations, its effective capabilities “will certainly expand by more than an order of magnitude.” The program borrows a human analogy: tool use is one of the most important abilities distinguishing humans from other organisms.
Coding was the first domain to close the loop for a reason. Code combines language properties with highly structured training data, while the compilers, testing systems, and DevOps tools built over the past 20-30 years allow an Agent to generate, verify, compile, run, and present results inside a virtual feedback loop.
颜前航 emphasizes that early ChatGPT-generated code could “look right” but often did not work once placed in a compiler. Today’s toolchain exposes hallucinations immediately. Embodied intelligence must bridge the gap between code and physical execution, making real feedback and a closed loop much harder to establish.

5. Verifiable Feedback Loops Determine the Speed of Agent Evolution

Programming tasks have clear right-and-wrong answers and reward signals, making them suitable both for product delivery and further reinforcement learning. Both speakers see this as the common reason coding can form PMF quickly while many subjective tasks continue to progress slowly.
颜前航 uses AlphaZero to describe the ideal loop: the Agent decides how to play, evaluates the result, plays repeatedly, and reflects. If an Agent can think, execute, verify, and receive feedback within an environment, it re-enters the classical reinforcement-learning framework.
RL cannot replace foundational capabilities. 颜前航 compares it to tennis: a teacher must first teach the correct swing, after which practice can optimize it. If the starting motion is wrong, extensive reinforcement may simply harden the error. The capability ceiling is therefore determined by both the base model and reinforcement learning.

6. Reinforcement Learning Is Seen as the Long-Term Path to Higher Agent Intelligence

刘鹏琦’s core view is that “Agents iterated through reinforcement learning may be a path for future AI applications to reach so-called ultimate intelligence.” An Agent is fundamentally an actor in the classical RL framework, adjusting its behavior according to environmental rewards and penalties in pursuit of long-term goals.
He cites the textbook Reinforcement Learning: An Introduction and notes that its 2 authors received the 2024 Turing Award. RL also develops alongside cognitive science, psychology, and neuroscience, so it is not merely a computational method but something like a “universal law of evolution.”
The clearest analogy comes from learning itself: reading and listening to lectures resemble pre-training, while a teacher’s worked examples resemble SFT. Genuine mastery depends on the feedback provided by homework and exams. If an Agent is to surpass competitors or even humans, it must ultimately explore for itself rather than wait for humans to tell it step by step what to do.

7. Two Architectural Paths Define the Capability Boundary Between Model Vendors and Application Companies

OpenAI Deep Research and Kimi Researcher represent the end-to-end route: memory, context, tool use, and multi-step chains of thought are trained through RL by a single model across the full process. This is close to model post-training and, for now, is mainly being pursued by model vendors.
The advantage of end-to-end systems is that user needs and context remain on one continuous chain. The challenges are potential context explosion and the lack of proof that they can reliably complete ultra-long chains and complex tasks.
The alternative is to modularize Agent capabilities. R1 can handle reasoning and planning, Claude can handle coding, memory can be provided by RAG or other external mechanisms, and search and other tools can be added on top. Manus may broadly fit this engineered framework. RL here mainly improves individual capabilities rather than globally optimizing the Agent as a whole.
颜前航 therefore believes end-to-end Agents require teams that understand both model post-training and products, raising the bar significantly. Modular systems instead connect different capabilities through external software and engineering frameworks.

8. The Bitter Lesson Points Toward Learning-Based Systems Without Negating Short-Term Workflows

颜前航 cites Richard Sutton’s 2019 essay The Bitter Lesson: long-term AI progress comes from compute, data efficiency, and general methods, rather than continuously embedding human priors and feature engineering. The program also says OpenAI researchers read it aloud every day and “pray to the God of Optimization.”
This experience also explains the focus on unified architectures such as decoder-only models: they are thought to make better use of data and compute for scaling up. RL-driven Agents, relying on their own exploration, fit this long-term pattern more closely.
刘鹏琦 adds 2 conditions. An Agent cannot be a capability-free “novice”; vertical know-how remains essential. The environment must also provide sufficient context, a reasonably well-defined reward function, and enough feedback. Without those conditions, more compute cannot discover the optimal path.
Neither speaker therefore dismisses workflows. In the short term, startups must first find the intersection between the technical frontier and product demand—“get to PMF first.” Teams should simply keep one eye on the future and reserve the data and product structure needed for RL and learning-based evolution.

9. The Real Product Moat Is a High-Density Feedback Environment

刘鹏琦 rejects reducing a general-purpose Agent to a language UI. Language is a compressed form of information and may not accurately express needs or results. Design and drawing tasks still require selections, local edits, and visual feedback, making the GUI “not so easy to replace.”
The value of browsers and IDEs is not just operational convenience; they provide rich context. Cursor did not discard the IDE, allowing it to collect large volumes of user feedback signals.
If a product receives one request and returns one result, feedback becomes excessively sparse, and the Agent may not even know whether its delivery was effective. Applying The Bitter Lesson at the product layer first requires an environment in which models and users can interact continuously and iterate on their own.
The “AI application version of The Bitter Lesson” from both speakers is that engineering techniques can approach today’s technical ceiling, but teams “cannot forget that the technology frontier will expand.” Without accumulating user data and feedback, today’s engineering polish may be rapidly淘汰 when the next model or product arrives.

10. The Embodied-Intelligence Debate: General Foundation Models or Multi-Species Collaboration

颜前航 believes embodied intelligence remains in the early research stage, with even the scaling law for its foundation models still unknown. Compute is not the main issue; data is the bottleneck. He wants the industry to target a general foundation model that can scale up rather than retreat into small vertical models because the problem is difficult.
刘鹏琦 agrees with a data-driven path but is more cautious about “general.” RL’s ceiling is determined by the environment. Even if all organisms live on Earth, they differentiate according to their environments and relationships with other species rather than evolving into one form capable of handling every task.
The long-term outcome could therefore be multi-Agent collaboration. A single person entering a natural environment might “not survive 2 days”; humanity’s advantage comes from large-scale organizations such as families, companies, and countries. Different Agents or embodied forms may likewise develop specialized strengths and combine them through collaboration.

11. The Technical Ceiling and Moats of Agents Have Yet to Take Shape

Products have improved visibly, but real-world performance remains mixed. How to maintain long contexts, manage memory, and determine whether reasoning capabilities trained on objective data such as mathematics and coding can handle subjective judgments and random outcomes in society all remain unresolved.
Tool use is also mature only in environments such as browsers, search, and coding. Once Agents enter complex internal enterprise software or the physical world, the gap in interacting with those environments remains substantial. The Agent ceiling is therefore not particularly high at this stage.
刘鹏琦 links the long-term moat to network and scale effects: whether more users and usage can feed back into better product experience and model capability. That loop “is still far from being there” and may only begin to emerge in H2 2025 or the following year.

12. Subscription, Token, and Outcome-Based Pricing Each Carry Paradoxes

The most common model today is a subscription of approximately $20 per product per month. But if large numbers of low-frequency vertical Agents emerge, users cannot pay for each one individually. If subscriptions are not the end state, extrapolating ARR from current subscriptions also becomes questionable.
Token-based pricing is equally unnatural. A stronger Agent might complete a task using 1K tokens, while a weaker one consumes 10K. Users may not be willing to pay ten times as much simply because the latter is inefficient; compute consumption is not equivalent to user value.
Outcome-based pricing appears more logical, but value is highly subjective. One user may care about travel planning while another values industry research. Even if the tasks differ by an order of magnitude in token usage, that cannot determine the price. When a general-purpose Agent calls multiple vertical Agents, complex settlement problems emerge as well.
颜前航 expands this into a question of production relations. Humanity has not fully solved how to value human labor, making it even harder to price AI’s independent delivery immediately. If organizations gradually become human-AI collaborations, pricing, incentives, management, and even enterprise management itself may all need to be reworked.

13. Investment Opportunities Lie in Vertical To B and Niche To C, Not Only in General-Purpose Platforms

刘鹏琦 favors vertical To B Agents. Industry know-how is not contained in foundation models, and these teams are closer to users. Products can start with prompts and workflows, evolve into agentic workflows as models improve, and ultimately deliver fully managed outcomes.
Beyond the idea that “one person can build a unicorn,” he still advises founders to leverage their existing experience and accumulated advantages while extending the service chain as far as possible—from tools to services to delivered outcomes. The more complete the chain, the greater the opportunity to accumulate user relationships and feedback.
颜前航 adds that To C does not have to mean general-purpose or platform-based. A large user base multiplied by niche, personalized needs can still form a major market. Short-form text tasks are already approaching the boundaries of OpenAI, Kimi, and MiniMax; more promising areas include travel, long-form content, and companionship.
The combination of AI and consumer electronics is another priority. Just as DJI and Insta360 built new categories from vertical products, AI hardware may expand outward from specific needs rather than trying to solve everything from day one. The final judgment remains cautious: “Many views may be rapidly revised or overturned within 6 months or by next year.”

Vol.67 The BP for a Hot Project May Be Fake Quark on Foundation Models + Gaokao: Bringing AI to Tens of Millions