Model Is Good Enough: In 2026, Applications Are Scarcer Than Bigger Models

Recently, I re-turned to a draft lecture by Limu in SJTU, which I was most impressed with not by the size of a parameter, nor by the size of a model, but by a simple judgment:Model is good enough.

This statement, of course, is not to say that the study of models is over, let alone that the bottom capacity is no longer relevant. It's a different and more uncomfortable fact:For the vast majority of researchers, entrepreneurs and application teams, the main focus continues to be on Better Model, with marginal returns that are not so high.

Over the past few years, our most important concern has been: can models be stronger? Today's more scarce issues begin to become:Where exactly are these capabilities to be put, how they are organized into real tasks and how they are deposited into a continuous data and feedback loop?

By 2026, it was more serious to think:After the model was enough, did we use it in a place that was worth it?

It's the Second Half talking about the same turn, but it's different.

The one before me. Model shift and uncertainty modelling after RL Agent to LLM Agent: The Second Half And I've written about the line of Shanyu Yao: when models, algorithms and RLs have shown some generalization, bottlenecks will shift from Better Model to Better Evaluation, Better Dataset, and the definition of problems closer to the real world.

In my view, the lecture was about the turn of the same era, but it was a step forward.

The Second Half More emphasis: The model is strong enough to force us to rewrite benchmark and evaluation.
Limu 's Reminder More emphasis: The model has crossed the threshold of adequacy, and then it is time to find what applications are worth doing, what data is worth building up, what post-training can bring capacity to reality.

They are not a conflict, but a back-to-back.The assessment is a bridge and the application is a destination. We're rewriting the assessment not to create a more beautiful leaderboard, but to migrate model capabilities to the reality scene. One step further, and in many cases the definition of the assessment is precisely the application itself.

And that's what I'm feeling now:The Second Half reminds us that problem definitions are more important than continuing to paint models, and that LiMu ' s reminders are more straightforward: there are no established applications, and a better definition of problem might end up as a separate set of research circles.

Why, 2026, it's time to focus on the app.

Today's AI is hard to describe as "doing nothing." It writes codes, does math, analyses documents, searchs information, calls tools, stabilizes systems over and above a set of benchmarks. Capacity is certainly moving up, but “capability” is not the most informative issue in itself.

The awkward reality is on the other side:The model is becoming stronger and does not automatically become a synchronized re-engineering of life and working methods.

We already have a powerful set of models, but the daily lives of most people are not completely rewritten as a result. For many, AI still remains more at the level of chat windows, search for enhanced, and occasionally writing something; the deep embedded in long-term work streams and life-organisation is still very small.

This is not because capacity does not exist, but because it is not organized into a sufficiently suitable application interface.

The word “appropriate” is more harsh than many think. Not every demand is suited to today's model, nor does every direction that seems to be in demand generate a stable product. Without a sufficiently clear mission boundary, sufficiently stable feedback and a return flow of sedimentary data, the model is stronger and often remains in demo, trial and short-term shock.

I'd rather be a little more sharp now:AI is not lacking capacity, it is missing interfaces; it is not missing answers, it is missing the way that the organization is being made to work.

What is scarce in the next phase is not a three-point model stronger than the previous generation, or a first one that gets back to a particular game with gold medals and leaderboard, but rather a pattern of tasks that are real, stable and can build up feedback on a continuous basis.

Why is application not a commercial issue, but a technical one?

Application is easily described as a matter of product manager preference as if it were a matter of commercial landing, market selection or user operation. But if we put it back in today's big model technology vault, it's a very hard technical issue.

Because applications are both an export of modelling capabilities and an entry point for subsequent data and training signals.

More precisely, there is a clear chain in it:

Apply first-defined task boundaries. Are you doing code changes, web operations, medical aids, experimental planning, or personal affairs management?
The mission boundary determines what data can be taken. Can you get track, results, type of error, manual correction, long-term preferences, environmental consequences?
Data quality determines whether post-training is meaningful. If there is no steady feedback, the so-called scenario optimization is often left with Prompt repeated error (of course, there are many applications that are limited to it, such as the fire 's Skylls, which are almost optimized in the Prompt Black Box).
The feedback structure determines whether the model is more useful for sustainability. Even with a stronger base model, it is difficult to form a closed ring without clear rewards, no verifiable results, no reusable experience.

And that's why I'm increasingly agreeing with the simple but important words in the draft:Pre-training and post-training are as important as most ordinary researchers and application teams, and post-training tends to be closer to real leverage.

Pre-train determines whether a system has broad a priori or sufficient universal base capacity; but post-train determines whether these capabilities can enter into specific scenarios and become reusable, deliverable, sustainable optimization. Post-train is not worth doing, is useful, and depends fundamentally on the application itself providing stable and high-quality data.

Application is not an issue outside the model. It is increasingly like whether modelling capacity can continue to sink as a precondition for real capacity.

A little more:Without good applications, data closed rings cannot rise; without data closed, post-training can easily become partial fixes; without post-training, modelling can be difficult to enter into real production.

Two applications that are taking place: Claude Code and OpenClaw

If the shift is to be applied only in abstracto, conclusions can easily appear empty. So I'd like to put in two of the samples that are happening here, not to compare the products, but to put the judgement ahead into the real task structure.

Claude Code: Why is coding anent the first running applications?

Anthropic, official. Claude Code Defined as a chat model for real development processes: it is not a chat model that simply answers code questions, but that can read the code library, edit files, run commands, access development tools and work on systems that actually develop surfaces such as terminals, IDE, desktops and web pages. And then to the outside, Claude Code, official files and... Agent Skills Documents Mechanisms such as instruation, hoops, MCP, skills have been clearly placed within its capacity boundaries, indicating that its goal is never more like chat, but more like a workstream interface.

This is important. Claude Code is not a model that suddenly gets smarter, but a type of model. agentic coding workflow It's starting to run.

It is no coincidence that it became the first application of value realization:

Environmental high digitization. Codes, terminals, tests, logs, version controls are all in machine-readable environments.
Success or failure is relatively verifiable. The code is not running, tested, constructed successfully, diff is not reasonable and can give some clear feedback.
The tool chain exists naturally. The development process was inherently dependent on editors, shell, Git, CI, issue tracker; angent was not flat-rised, but plugged into existing infrastructure.
Human audit is clear. A lot of things can be done by antgent, then by an engineer, and not by giving control to the whole thing from the beginning.
Data flow fast. A failed fix, a rejected patch, a passed test, a code review could quickly form the next round of optimization signals.

And that explains why the coding anent will be one of the most persuasive front lines of application in the past year. It is not successful because programmers prefer to taste it scarcely, but because the mission space naturally corresponds to the advantages of today ' s model: strong language understanding, strong tool use, strong local planning, rapid exposure to errors and high external feedback density.

So Claude Code says:Once mission space is sufficient to digitize, feedback is clear enough and workflow interfaces mature enough, today ' s big model can stabilize the generation of outputs.

And that's why I prefer to see it as an application level signal, not as a mere product signal. Claude Code is not just about Anthropic making a good tool, but about the proof that it:In some scenarios, AI can already be organized into continuous productivity.

OpenClaw: Why is it harder to do a PAI intelligence closer to the future?

On the other side, OpenClaw represents another more attractive imagination.

OpenClaw Official GitHub README describes it as a personal use of the scene. personal AI assistantI'm sorry. In its public direction, it is not a second mobile phone that talks, App, but rather an assistant who enters a personal digital life: to take on tasks such as alerting, searching, controlling, recording, coordinating and long-term companionship through chats, voice, equipment access, skills expansion, automation and personal knowledge organizations.

If Claude Code is the one that enters the professional stream, OpenClaw is the direction that is closer to the final form of AI in many minds:It's not that you go to a tool to find it, but it starts following you, understanding you, serving you, and getting into your life interface.

And that's why it's more direct than coding anent to the question of how AI affects everyday life; after all, not everyone needs to write codes, and it's a small demand for Coding to reach the limit.

But the problem is precisely here: the closer to life, the harder it is.

Compared to the code anent, PIS faces another task structure:

Targets are more subjective. “Is this reminder timely” “is it a considerate arrangement” “is it really helpful to have this proposal”, and often there is no uniform answer.
Feedback is more delayed. It may take hours, days or even longer to know whether a single individual arrangement has been successful.
Environment is more isomeric. Cell phones, messages, calendars, mail, positioning, home equipment, browsers, voice input, third-party services, all of which are different systems.
The boundaries of authority are more sensitive. The closer it lives, the more it has to deal with high-risk areas of privacy, payment, contact, equipment control, and long-term memory.
The criteria for success are not clear. The test results are available to the coding agent, and more often the personal assistant can only see if the user feels more comfortable, which is difficult to stabilize.

So the point of OpenClaw is not whether it's done with the PA problem today, but whether it's done with the problem sufficiently directly:The closer life is, the more we can do more than just model it. The challenge is how to define the mandate, how to manage the authority, how to organize the long-term memory, how to integrate the decentralized environment into an actionable interface, and how to extract credible data and feedback from these interactions.

That is why I am both good at, and vigilant at, this direction. Look, because it's really closer to "AI into life" and because it's too easy to be miswritten as "just to change the model for stronger." The hardest part of PIS intelligence is probably not at all in the model itself, but in the application structure itself.

What do these two samples tell you?

Put Claude Code and OpenClaw together, and it's not who cares who's more advanced, but they almost put two of today's applications in the immediate future.

Claude Code, who is a member of the National Council of Women, is a member of the National Council of Women.High digitized, highly validated, highly functional - Yeah. The scene has begun to stabilize the realization of value.
OpenClaw is a representative of:High proximity, complexity of the environment, sensitivity of the authority The scenes are extremely tempting in directions, but far from running.

These two points together are one thing:The difficulty of application arises first from the task structure, not from the model parameter volume. A single application is often difficult not because models are not rewritten and not sufficiently reasoned, but because the environment is too fragmented, feedback too slow, objectives too vague, mandates too much, and evaluation too subjective.

And that's why “better application” becomes a question of “better data”. The application will continue to be stronger only if the mission is structured in a way that allows for stable feedback; if there is stable feedback, there is a greater likelihood that it will be meaningful post-training; and if there is real value in the post-training.

A real application worth betting on, at least on what terms.

If the previous discussion were to be organized into a more reusable set of judgements, I would say that the next stage would be a worthy application, at least for the following conditions:

HF. The task does not happen once in a while, but it can generate real needs on a continuous basis.
Relatively stable. The environment and objectives cannot be completely rewritten every day, otherwise it is difficult to sink experience.
There is a section to validate feedback. It is not necessary to be as perfect as a code test, but at least there are some clear results that can be used to correct the system.
It can generate data flow back. User corrections, environmental consequences, implementation trajectory, long-term preferences are recorded.
Could embed an existing process. The good application is not to force users to change all habits, but to insert competencies into existing work or life streams.
Continuous improvement through training or workover. It is difficult to establish a system that is not able to fit the scene with the accumulation of its use.

Looking back at this standard, it is clear why the coding anent was successful first: it was high frequency, digitized, verifiable, reviewable, revolving. Why is personal observer more attractive and more difficult to understand: It is certainly high-frequency, but feedback is subjective, decentralized, sensitive and the most difficult layer is not in the model itself.

And that is why I am increasingly not convinced of the optimistic judgment that “as long as the model continues to grow stronger, all applications will grow”.Many demo have been demo not because the model is only a little short of last capacity, but because they do not have a sufficiently good task structure from the outset.

The technology world is the easiest to overestimate linear extrapolation of model progress, while underestimating the hard constraints of the application structure itself. While the growth of model capabilities is of course important, the upper values are increasingly determined not by the amount of parameters but by the design of the mission, the feedback design, the rights design and the workflow design.

Conclusion: the most important issue in 2026 is not stronger, but more useful

If the Second Half makes you realize that the focus of AI's research will be on changing the definition of evaluation and real problems from doing better; if the line about envirronment and reward is becoming more and more convinced that the feedback structure determines the direction of experience learning, then the other thing that Limu reminded me of is that it is more modest, but difficult:Without an established application, neither of the preceding would automatically be of social value.

I'm now making a clear judgment that what matters is not just better model, but more than better evaluation. Organize capacity for real application and allow applications to generate better data in turn.

The model will certainly continue to grow stronger. But for most people, the more worthwhile place to be bet may be to go not to the next abstract best model, but to find an application interface that can enter the work stream, enter life, enter the long-term feedback loop.

That's after 2026, AI's most real main battlefield.

References

Model shift and uncertainty modelling after RL Agent to LLM Agent: The Second Half
AEnviron:Agent needs a unified environmental layer?
Anthropic Docs, Claude Code overview
Anthropic Docs, Agent Skills overview
OpenClaw, GitHub repository