Large Language Models Are Already Smart. Why Do TheyStill Struggle with So Many Small Tasks?

— Capability Gaps AI Has Yet to Close, Seen Through Long-Term Use

By Zhu Weisha – Contributing Writer

I spend a great deal of time every day using AI to discuss ideas, revise articles, and prepare

manuscripts. The more I use it, the more impressed I become by the remarkable advances in large language models—and the more clearly I see some of their shortcomings.

One striking phenomenon is that AI may repeat the same mistake again and again. Take the preparation of a manuscript according to clearly stated formatting requirements: Chinese body paragraphs should begin with a two-character indent, while English paragraphs should be flush left. After being corrected, the AI understands the requirement, yet it may make the same mistake on the next similar assignment.

Users who know how to write Skills may try to fix the problem themselves, but most people are

unfamiliar with creating, configuring, or invoking them. Asking users to handle all this can be more complicated than using conventional internet software. Users may not directly perceive the difference made by the computing-power arms race, but they immediately notice whether a product is easy to use.

Ease of use is an important condition for expanding the market.

As an integrated system, AI still has shortcomings in practical applications.

Ten Capability Gaps Revealed by Practical Use

First, difficulty identifying the right starting point for a conversation. When a user asks a question, AI often answers immediately without first determining what the user is actually trying to accomplish. Asking a clarifying question may matter more than supplying an instant answer. With an unfamiliar subject, the proper starting point may be to survey the facts and competing views rather than immediately apply an existing framework. However fluent the answer, if the problem is framed incorrectly at the outset, everything that follows may go in the wrong direction.

Second, insufficient depth of understanding. AI is good at summarizing what an article explicitly says, but it may fail to grasp ideas the author has not fully articulated. Sometimes the most important part of an article lies in its implicit assumptions, judgments not yet developed, or deeper relationships behind the facts. Understanding the words and understanding the ideas are not the same thing.

Third, limited ability to identify important questions independently. A large language model can readily generate dozens of research questions. But generating questions is not the same as discovering which question really matters. When we studied AI agents, new facts repeatedly changed the direction of our inquiry. The original question was not necessarily wrong; it simply was not always the one most worth pursuing. AI often needs the user to recognize this shift, rather than independently noticing that the research question itself should change.

Fourth, innovation remains inconsistent. AI is very good at connecting knowledge, structuring arguments, and proposing solutions, but a novel combination does not necessarily amount to a valuable new idea. Genuine innovation requires identifying facts that existing explanations cannot account for, developing a more powerful explanation, and subjecting it to counterexamples and practical tests. Large 2 language models have demonstrated some of this ability, but they remain far from consistently and independently discovering important problems.

Fifth, weaknesses in abstraction and rule formation. Human experience involves more than remembering what happened. More importantly, it involves discovering patterns and forming rules for judgment. AI may correct an error in one manuscript without identifying the common cause behind similar errors and abstracting it into a rule that applies to future tasks.

Sixth, the lack of a reliable self-checking loop. This is an especially visible weakness. AI can write an article and it can review an article, but it does not necessarily review its own work before delivery. When challenged by the user, it may even identify precisely what was wrong with its earlier concepts, criteria, or evidence. The capacity to check is present, yet no dependable procedure ensures that the check takes place. Being able to do something and doing it reliably every time are two different things.

Seventh, failure to improve consistently from user feedback. One important way humans learn is by correcting mistakes. AI can understand a correction without necessarily turning it into an ability it can reliably use next time. In fact, even without changing the base model’s parameters, a system can improve through memory, rules, programs, Skills, and regression tests. For reversible work that can be tested repeatedly, this approach is already technically feasible, but its adoption is not yet sufficiently widespread or reliable.

Eighth, valuable feedback does not reliably reach the people and systems responsible for improvement. Errors discovered by individual users may be relevant to many others. Yet those observations may never be classified, abstracted, verified, and conveyed to the systems responsible for improving products, Skills, and base models. The connection between an individual correction and improvement of the wider system remains incomplete. Of course, upward feedback must respect privacy: a user’s particular preference should not become a general rule without proper validation.

Ninth, inadequate personalization. AI may remember what a user has said without truly understanding that person’s long-term goals, experience, rules of judgment, and changing needs. Conventional internet accounts mainly record identity, transactions, and preferences. In the AI era, a personal account should be able to develop a revisable model of the individual’s needs through ongoing conversation. Every interaction should offer an opportunity to understand the user better, rather than starting over each time. In terms of providing stable, long-term personal service, some AI products offer account capabilities that fall short of mature internet services and fail to realize AI’s potential advantage.

Tenth, unclear concepts and functional boundaries for Agents and Skills. Different companies define an AI agent differently, and the functions of Skills and Agents may overlap. This is not entirely a cognitive defect of the model. It also reflects the speed of technological change and the evolving ways software is organized. The contributions made by models, runtimes, tools, and Skills should be assessed through real work. When an Agent or Skill was not written by the user, it is often harder to assess its execution process, permission boundaries, and the consequences of invoking it. Of these ten issues, some concern cognitive abilities that base models have yet to master. Others arise because application systems lack complete execution procedures; still others involve organizational mechanisms or the design of user services. Treating them all as the same kind of problem makes it harder to find the right remedies

What Is the Computing-Power Arms Race Actually Solving?

Today, enormous sums are being invested in AI computing infrastructure around the world. The central aim is to push back the cognitive limits of large language models through training, algorithmic improvements, and inference-time computation.

More powerful models may become better at understanding, reasoning, abstraction, planning, and innovation. They may also be able to perform work that previously required several programs working together. But more computing power will not automatically maintain personal accounts, guarantee a review before every deliverable, or preserve a user’s corrections as verified Skills.

There are two paths that can reinforce each other. One is to keep advancing the cognitive abilities of base models. The other is to organize abilities we already have into dependable work systems capable of continuous improvement. Some researchers have warned that the computing-power arms race could accelerate AI research further and might even trigger a so-called “intelligence explosion.” That forecast remains uncertain, but competition to improve base-model capabilities is unlikely to stop in the near term. By contrast, there are already many practical steps we can take to organize existing capabilities into work systems that improve over time. AI capabilities are still advancing rapidly.

Work that requires assistance from external programs today may be handled directly by a model tomorrow. But real-world data, transactions, authorizations, responsibilities, and personal needs will not resolve themselves simply because the model has become smarter. Many of these improvements do not require waiting for the next generation of large language models. We already have conversational capabilities, programming, memory, tool use, and automated checking. What is often missing is a mechanism for organizing those capabilities reliably.

AI progress should be judged not only by how many new capabilities it may acquire in the future, but also by how much of what it can already do is being used reliably today. Rather than rushing to pursue an all-powerful super-agent, we should first make AI good at the smaller tasks it can already handle.

It should learn to clarify proactively, check its work carefully, absorb experience, improve continuously, and genuinely understand the people it serves. These seemingly ordinary improvements, taken together, may be more valuable than continually adding new names for new features.

In short: first, do the small things we can already do—and do them well.