From "Model Supremacy" to "Experience-Driven": How to Play the Second Half of the AI Game | Linear Voice
The Transformation, Challenges, and Future Vision of AI's Second Half
AI competition has entered the "second half" — a concept first articulated by former OpenAI researcher Shunyu Yao that is now reshaping the industry's competitive focus: In this second phase, the core challenge will no longer be training a smarter model, but rather winning on the evaluation, evolution, and real-world deployment of intelligent agents.
Behind this lies a profound shift from "model supremacy" to "experience-driven" development. Technological breakthroughs remain foundational, but the true engine of change is the evolution of human needs — from "help me get things done" to "understand how I feel." This transformation is not only birthing new product forms, but also rewiring the underlying logic of human-computer interaction: interaction design has become a psychological problem rather than an engineering one, and the teams building agents have moved from pure engineering mindsets toward hybrid "engineering + management" roles.
Today's insights come from a deep conversation at Linear Capital's AGM, where Linear Capital Managing Director Can Zheng spoke with three AI application company founders. Minghao Guan, Co-founder and CEO of Final Round; Kaijie Chen, Founder and CEO of Macron AI; and Yifeng Yin, Founder of TEA.AI, shared their practical perspectives on how agents are crossing the critical chasm from "usable" to "actually good," and the technical logic and product philosophy behind it.
We are witnessing a crucial "identity shift" in the growth of AI agents.
Their core mission has shifted from faithfully executing preset instructions to autonomously understanding, deciding, and evolving in complex environments. This marks the beginning of a new phase: the logic of AI competition has fundamentally changed, from refining a model's "intelligence" to building an agent's "capabilities" and "character." The former concerns what it "knows"; the latter tests how it "keeps getting better" in real-world scenarios, ultimately earning human trust.
This transformation is reshaping the entire chain from technology to product to business. On the technology side, the center of gravity has moved from one-time pre-training to continuous evaluation and improvement loops. On the product side, the central design question has shifted from "how to let users complete tasks fastest" to "how to make users want to stick around long-term." On the business side, the value metric has expanded from "cost-cutting, efficiency-boosting tool" to "accountable digital collaborator."
In other words, AI agents are undergoing a transformation from "tool" to "collaborator." This requires not just more sophisticated algorithms, but a deep understanding of human needs, social behavior, and organizational logic. Below, the insights from these three founders:
Michael @ Final Round
The changes are coming from several layered developments. First, longer context windows let us feed richer content and scenarios into an agent's workflow. Second, tool interfaces have become unprecedentedly mature and easy to use — building a complex agent system today might only require a simple pipeline. Finally, creative capabilities are being unlocked, such as real-time search and multi-agent interaction, all of which are rapidly expanding what agents can do.
Kaijie Chen @ Macron AI
I think the shift is fundamentally one of mindset. The biggest change from last year to this year is our move from the "pre-training era" to the "experience-driven era."
Previously, we used pre-training to teach a model — extracting and imitating world knowledge — and the model's value was essentially fixed. But now, with the rise of reinforcement learning (RL), we've realized that true intelligence gains must depend on continuous interaction with the environment. Whether it's context or MCP, the goal is to let models gain more experience from interaction and use it to improve themselves.
Cursor is a perfect example. Every time a user hits Tab, it's an interaction with the environment and the user, and that data feeds back to improve the model. So this year's agent explosion, on the surface, looks like accumulated technology, but at its core it's a mindset shift — from training a good model to building a loop where the model evolves itself through excellent product experiences.
Yifeng Yin @ TEA.AI
My perspective is a bit different. Beyond the earth-shaking technical changes, I think human needs have also fundamentally shifted.
People used to care more about "what can this help me do" — specific scenarios. Now, demand for abstract needs beyond specific tasks is growing. The explosion of the term "emotional value" is proof of this. And emotional value can only be provided by "people" or "things that seem like people."
The biggest role of an agent in interaction may not be being the most efficient or fastest responder, but rather being "human-like." To some extent, having five agents is like having five people working for you — this is a sense of "power." Secondly, human communication isn't purely instrumental; an agent can both handle tasks for you and provide emotional support. This is market demand forcing the explosion of agentic AI. Meeting emotional needs is something only large models and agents can currently do; traditional algorithms can't, so this is pushing companies to innovate in this direction.
Yifeng Yin @ TEA.AI
It's a shift from "completing a task" to "truly communicating with a person." Past human-computer interaction (HCI) principles were about getting you to finish tasks in the shortest time. Not anymore — HCI isn't just "human-computer interaction," it's become "human-human interaction," because agents are human-like.
Current HCI design is no longer an engineering problem but a psychological one — you need to make an agent genuinely connect with people, to build a sense of trust. Almost no company can do this yet, and this is precisely where true user stickiness lies.
You might throw away an old hammer for a better one — that's tool utility. But imagine — you have a dog you've raised for ten years; would you abandon it just because a more expensive dog shows up? Impossible. That's emotional bond. If we can reach that standard, it's a generational experience gap: from "using a pure tool to achieve a goal" to "having a friend who can help." We're exploring this path.
Kaijie Chen @ Macron AI
My cofounder said this at a Linear Tech π event, so of course I support it — but with a clarification. Not all companies, but a good agent product company will eventually end up doing model training. And "training" here mainly means post-training, not pre-training.
Take Thinking Machines Lab: Mira left OpenAI to start her own company, fully capable of building a pre-training team, but they skipped that step and only do post-training. Including their recent product Tinker — a cheaper tool using fewer GPUs that can do RL post-training — it's all about the post-training piece.
Why is it necessary? Because we're now in the "Era of Experience." The biggest consensus in Silicon Valley right now is that whether you're a product company or a model company, you need to scale reinforcement learning into your product and let the product iterate with users. It's like launching a new generation of recommendation systems. Models used to be frozen, never upgrading. Now we need to think about how to make the 99% of parameters "behind" also move, through online reinforcement learning or policy adaptation, not just context adaptation — then we can make users feel the product keeps getting better as they use it.
Future benchmarks will change: no longer how many questions a model answers correctly, but the slope of capability improvement through user interaction. If a product builds a good enough feedback environment, it could completely surpass everything else within 3 to 6 months and reach the highest metrics. That's why product companies must engage with model training and build their own environments — otherwise, as AI pipelines grow increasingly complex, they won't be competitive.
Michael @ Final Round
We built a very direct internal tool for this: screen recording software. The simplest way to learn something is to watch how a real person does it, which to some extent follows the RL concept. We record key workflows we want to replicate, review them repeatedly, and identify the details where "humans can do it but machines can't," constantly trying to break through those limitations.
Kaijie Chen @ Macron AI
It's completely normal for agents to fail at things. There are many general-purpose agents now, and their boundaries are defined by both user behavior and technology, and they're rapidly expanding. Users themselves don't have a clear concept at first, don't know where the boundaries are, so they always make "outrageous" requests — like "use my mini-program to make a sequel to Black Myth: Wukong" — obviously impossible. Of course, there are also seemingly reasonable demands limited by available tools.
I can think of three ways to expand boundaries: first, manual tool integration; second, automatic expansion like browsers; and third, which I think **represents the future, is agents discovering their own failure cases, adding them to RL training sets, and training the next generation. The difficulty lies in intelligently identifying and filtering cases. Our automatic improvement pipeline already runs under the supervision of the first two manual methods, but the ideal state is: put the product out there, users use it, and it automatically gets better.
Yifeng Yin @ TEA.AI
We have a good reference: RLVR (Reinforcement Learning with Verifiable Reward). If you can break down an agent workflow into steps that can be instantly verified, you can rapidly locate errors. But the problem is, the vast majority of user instructions are too vague to be broken down into quickly verifiable components.
However, user behavior provides clues. Like when a user stares blankly for several seconds after receiving a response, or directly says "I don't like this." We need to judge whether our output was good from all these subtle interaction cues. Because all large models are essentially some form of supervised learning — you need to "decode" user reactions.
Another issue: if you use model training methods to improve, efficiency may be high, but the user's current failed interaction experience is terrible. So you need at least an "in-context reflection" method for real-time improvement, giving users better immediate feedback before training the model. This reflection process itself also generates high-quality data for subsequent training.
Yifeng Yin @ TEA.AI
The biggest mindset difference is that agent design isn't just an engineering problem, it's a management problem. Because the core characteristic of an agent is being "human-like," and agent systems emphasize team collaboration. You need to consider what kind of "person" to place in what position, designing the entire organizational topology.
You need to train agents with different personalities. The same question, given to agents with different personalities, should yield different results. What kind of personality do we want in this role? This is what we consider in design. Our time allocation is probably 60% engineering, 40% discussing "what kind of 'person' do we actually want in this app."
The current four major models — GPT, Claude, Gemini, Grok — each have distinct "personalities." You can think of them as infinitely replicable AI employees. You need to decide which of these four "people" appears where. In this era, even engineers need some experience dealing with people, knowing what kind of person suits what work, then choosing the agent closest to that person to do the job, and in prompt engineering you also need to continuously generate feedback. So the role requirements for engineers have become very high in this era.
Kaijie Chen @ Macron AI
Internal team collaboration is also completely different. It requires unprecedented depth of collaboration between technology and product. Take Sora: its core feature (making generated faces match voices) was a product idea, with technology responsible for implementation. There must have been massive debate in between, but ultimately they combined into a brilliant product. Now many innovative features are proposed as prototypes by small teams, quickly built into a version and pushed live for testing, then reinforced based on feedback. This itself is a reinforcement learning process within the team.
I think the first characteristic of organizing teams in the AI era is that technology-product collaboration needs to become very different. Second, product-to-market cycles must be extremely fast — including Google now, which likes to dump everything out at once when launching products. In the past people preferred to polish and perfect before release, but now things change too fast; you must rapidly push to market and test, let data drive the next step. Teams need to adapt to this rapid change and adjustment.
Michael @ Final Round
Team velocity has become extremely fast. To put it plainly, the difference from before is: when we hire now, we ask: "Do you think that in the next 6 months, you could build an AI that replaces yourself?" People who say "no" out of fear of losing their jobs — we usually reject them. We want those who say "yes" and are willing to use AI tools to make themselves more efficient. We're also seeing that a huge number of companies in Silicon Valley now don't even have traditional interview processes anymore — they just jump straight into doing.
Another change is how team priorities are arranged. Previously, with limited resources, you could only do high-ROI things. Now with AI, teams have the capacity and time to do some low-ROI but potentially long-term valuable things, because the main cost is tokens, not massive engineer hours. Some things run continuously after being built, but may only show their value two years later. This lets us plan further into the future.
Michael @ Final Round
**I think the biggest change will be **information distribution methods. GEO (Generative Engine Optimization) is exploding in Silicon Valley right now — I get tons of vendor pitches selling GEO every day, everyone wants to occupy this new distribution channel. So traditional search may no longer be the biggest traffic entry point; GEO will be a completely new, critical distribution channel for all startups.
Kaijie Chen @ Macron AI
The biggest change — wouldn't it be OpenAI going public? But seriously, **I feel next year will be a very complex, composite kind of change, viewable from two angles.
On the technology side, 2024 was somewhat quiet for AI, but 2025 saw many open-source projects emerge after DeepSeek. **By the end of this year, as foundation models stabilize, newer technologies will emerge — like text diffusion generation, video reasoning, etc. These will spawn a batch of products with very innovative, eye-catching interactions, though you won't know if they're actually good. The market will be very excited, even if we don't fully understand what needs they solve.
On the people side, this year we've seen Silicon Valley giants recruiting at absurd prices, but it feels more like a satellite-launching spectacle. More top researchers are flowing back from overseas big tech, whether to start companies domestically or abroad, all looking for new opportunities.
So I think 2026, whether in underlying technology, projects, or these people, holds unprecedented opportunity — it's a time worth investing more, placing bigger bets.
Yifeng Yin @ TEA.AI
I agree that next year will be "a hundred flowers blooming," and also "demons dancing in riotous revelry." In 2023-24, research paths were still unified, everyone doing the same thing. But now technical paths have diverged, and previously entrenched problems are being challenged. The most important thing is distinguishing the "flowers" from the "demons." Most innovations may not be commercializable; discernment is needed here.
First trend: I think "super large models" may gradually quiet down, while small models get stronger and stronger, and edge hardware gets more powerful. We'll likely see super large models gradually replaced by highly specialized on-device models. The era of super large models may end in about 3 years, followed by an era where every company owns a pile of their own small models. Because training costs keep dropping, quality keeps rising, and hardware keeps improving.
Second trend: I think after this "demons dancing in riotous revelry," technical paths will reunify. Around 2027, we may enter another relatively "boring" period like 2024. But don't underestimate this era — AI evolves too fast for anything like the AI winter of the 1990s to happen. Before the next downturn arrives, there will definitely be new breakthroughs sparking new waves. Technology spirals upward through such market cycles of hot and cold, step by step landing as productive forces that can truly drive social progress.
