A couple of days ago, I helped a friend with her master's thesis. Her advisor asked her to redo the modeling validation for part of the research cited in her paper using code. This kind of requirement is common in research, but since she has absolutely no coding experience, she was completely stuck at this step.
She had already spent half a day trying various domestic AI tools, but the code either wouldn't run or produced completely irrelevant results. Eventually, she had no choice but to ask me for help.
When she reached out to me, she was already somewhat frustrated. I recommended she use ChatGPT and optimize her prompts, asking her to give it another try, but it still didn't work.
So I asked her to send me the materials. When I opened them, it was a PDF. At that point, I had a basic guess: the problem was likely not the model's capability, but the input format. So I threw the file into Antigravity and ran it with Claude inside. The model quickly gave a rough modeling plan, but also indicated that the current environment couldn't directly parse the PDF content. So it was actually searching online based on the title and inferring from common research methods. In other words, it hadn't actually read the paper.
So next, I did something very simple: I used PDFgear to export the PDF as images and re-input them to the model. After running the code a few times and fixing the bugs that surfaced, the mathematical model ran smoothly. If you take out my complaints and her emotional breakdown on the other end of the internet, the whole process took about half an hour.
This experience made me strongly feel a saying: The AI era is the best of times and the worst of times.
From an efficiency perspective, this is indeed a very good era.
If you rewind a few years, completing a mathematical modeling task often required reading the paper yourself, understanding the research methods, searching for materials, writing code, and repeatedly tuning parameters. Even for an experienced person, it could take an entire day. Now, the process is much simpler: let the model handle understanding the paper's structure and generating the code framework, while humans only need to do some checking and debugging. Many tasks that were once only doable by professional researchers can now be accomplished by ordinary people with the help of tools. As long as you know how to ask questions, verify results, and fix a bit of code, you can break down complex tasks and solve them step by step.
At the same time, this is also a somewhat bad era.
My friend was stuck simply because of an "input format issue." The model didn't read the paper's content and made things up, so the entire process naturally couldn't proceed. Before she came to me, she had already spent a lot of time on various "magical" AI tools, but none of them clearly pointed out the real problem. Only Claude reminded her that the current environment couldn't parse the PDF. Actually, it only required a simple step: having the user convert the PDF to images and re-input it, or directly generate a PDF parser, and the problem would be solved. But most tools didn't proactively explain this step.
This experience increasingly makes me feel that the real gap between first-tier models may not necessarily lie in intelligence itself. People love discussing model parameters, rankings, and benchmarks, as if being "smarter" would bring a completely different experience. But in real-world use, the experience gap often appears in engineering details.
Whether a system can recognize input format issues, explain the limitations of its current reasoning, and provide actionable repair suggestions—these details are crucial for users. Often, users don't need more complex reasoning abilities; they just need a clear prompt: "Your input might have a problem."
Recently, many people like to talk about AI Agents, imagining a system that can automatically understand needs, call tools, and complete complex tasks. It sounds great, but in my view, reality is still far from that stage.
In this small case, the entire process still required the human to first judge where the problem might be, then choose the tool, adjust the input format, verify the model's results, and finally fix the code. AI indeed makes many steps faster, but control over the process remains in human hands. In the short term, this is unlikely to change fundamentally.
What's more interesting is that the truly important abilities in the AI era haven't actually changed. They are still understanding problems, breaking them down, judging the boundaries of tool capabilities, and combining different tools to solve problems.
Many early internet users are actually familiar with this way of thinking. In the era when internet speeds were only tens of KB per second, if you wanted to find an image, you might have to scour forums, FTPs, and various resource sites. Back then, everyone naturally realized that tools have limitations, so you had to keep trying different methods and combine various tools together.
But today's internet environment is completely different. The vast majority of Generation Z grew up in recommendation algorithms and streaming platforms, where content is pushed to them automatically, and they just need to click to play. All tools are perfectly packaged and ready to use out of the box. Over time, people's understanding of tools has actually decreased.
Tools have become more powerful, but many people are increasingly unfamiliar with their boundaries. When they encounter a problem not covered by mainstream tools, they often either get angry or simply give up.
So I increasingly feel that the real gap between people in the future may not be who has AI, but who understands tools better. More and more people will be able to use a single tool, but those who can combine multiple tools to solve problems will still be few.
In the AI era, this ability will become even more important. After all, AI itself is just one of many tools, and tools never solve problems on their own. It is always humans who truly solve problems.