Under the Crisis of AI Replacing Humans: A Programmer's Path to Overcoming Anxiety

[baoshi.rao]

2023 was a magical year for me. I unexpectedly transitioned from a programmer to an 'internet celebrity' in the AI news circle. By the end of the year, my readership on platform X exceeded 100 million, and on Weibo, it surpassed 1 billion. Many people learned about the latest AI news, tutorials, and Prompt usage tips through my Weibo or X. And all of this actually began with my struggle with AI anxiety.

In this article, I will share my story with you—how I developed AI anxiety, how I overcame it, and how I successfully turned AI into my powerful assistant, becoming someone who is adept at using AI.

When OpenAI's ChatGPT was first released, I started using it immediately and found it to be completely different from any AI product I had used before. Not only did it excel in language understanding, but it could also generate high-quality content and even demonstrate a certain level of reasoning ability. This piqued my great interest, and I began using it more and more to assist with my daily work. At the same time, in the field of image generation, open-source tools like Stable Diffusion and commercial products like Midjourney also showed astonishing progress. AI became a very hot topic, so I started paying more and more attention to the AI field.

But as I paid more attention, I found myself becoming increasingly anxious because the field of AI is developing at an incredibly fast pace. New AI products are launched every day, and major upgrades occur frequently, such as the release of GPT-4, Midjourney V5, text-to-video generation, multimodal capabilities, and more. As a result, I spend a significant amount of time every day keeping up with various AI updates, afraid of missing any important information. This has led to difficulty concentrating and poor sleep.

Moreover, I’ve noticed that many others share this anxiety, especially due to the widely circulated saying: 'It’s not AI that will replace you, but people who are good at using AI.' This has caused many, like me, to develop AI anxiety, fearing they’ll be left behind if they don’t keep up with this wave of AI. I also learned a new term: FOMO (The Fear of Missing Out), which refers to the fear of missing important opportunities.

I began to explore the root of my anxiety. My AI anxiety may stem from several sources:

One source stems from fear of AI—worrying about being replaced by AI or by those who use AI, leading to job loss. Especially in the current tough job market, media often exaggerates AI's capabilities, which can mislead those unfamiliar with AI, plunging them into anxiety.

Another source is social pressure. With AI being such a hot topic, everyone is talking about it. If I don't understand AI, will I fall behind? Will I be isolated?

A third source is the fear of missing out. My generation witnessed the internet wave around 2000 and the mobile internet wave around 2010. Many lucky individuals seized these opportunities and achieved great success. But I didn't catch any of those waves. Now, with AI booming, does this mean the AI wave is starting? Having missed previous opportunities, I can't afford to miss the AI wave!

These reasons have led me to develop AI anxiety. To overcome it, I need to address the root causes. Based on my situation, I've prescribed myself a few remedies:

Learning AI, understanding AI

Our fear of AI largely stems from a lack of understanding about it.

I used to think AI was an esoteric field requiring exceptional mathematical skills or years of specialized systematic learning, so I never attempted to understand it. But do we ordinary people really need to learn so much foundational knowledge to effectively use AI?

The turning point came earlier this year when I saw someone create an application like this: upload a document, and you could conduct Q&A based on its content—it was like magic, absolutely amazing! But this time, I didn't shy away; instead, I tried to learn how it worked.

Fortunately, much information is publicly available nowadays, and there are abundant open-source projects. I quickly found similar open-source implementations and followed the clues to locate relevant technical documentation, trying to understand their principles (see Figure 1). I learned about professional concepts like Retrieval-Augmented Generation (RAG) and Embedding, discovering the existence of vector storage and similarity search. This learning process was somewhat challenging but much easier than I initially imagined, and it turned out to be highly rewarding.

However, I also set a boundary for myself: to focus on application-level knowledge without delving too deeply into underlying principles. It's similar to learning front-end development—understanding how to use React without needing to know its underlying implementation. On one hand, I'm not arrogant enough to think that just a few months of study would allow me to grasp profound AI principles. On the other hand, for me, understanding and being able to use the technology is sufficient.

Later, when many similar document Q&A products emerged, I not only stopped feeling anxious but also helped educate others about the principles behind them. I realized that for most people, it's unnecessary to learn all the underlying AI knowledge. Understanding the basics and focusing on how to apply it is more than enough.

Why not give it a try?

If anxiety about AI document Q&A products stems from a lack of understanding, then seeing someone demonstrate how to create a flashy web game using natural language reignited my fear of being replaced by AI.

However, I believe that instead of worrying, it's better to try it out and see if it's truly as powerful as claimed. So, I attempted to create a simple game, only to find that the actual results were far from ideal. It turns out that those who showcased impressive demos were using pre-trained classic game codes. For large language models, these games have been repeatedly trained on, making it easy to generate relevant code. But for a game that has never been trained before, achieving satisfactory results is much more challenging.

I've also tested natural language generation for frontend pages (see Figure 2) and understood that its underlying principle relies on large models to generate HTML, CSS, and JS code based on requirements. For simple demo pages, this works fine, but for pages with specific requirements or complex sites, there's still a significant gap. Current large models have limitations, such as restricted context length, which prevents generating large amounts of code or modifying existing code extensively. Additionally, the quality of generated code heavily depends on how the prompt is written. If you're proficient in frontend development and can provide detailed requirements, the results are better. However, without frontend knowledge, it's challenging to effectively guide the AI to complete tasks.

Through hands-on experimentation, I've found that programmers don't need to worry about being replaced in the short term. AI currently can't fully generate a complete project from a given requirement. Programmers are still needed to analyze requirements, design architectures, break them into modules, generate code, and handle testing and deployment. While AI might assist in generating or optimizing certain code modules, programmers are essential for compiling and debugging errors.

Avoiding Extremes in AI Perception

I've noticed that some people, like myself, initially approach AI with anxiety-driven curiosity, only to swing to the opposite extreme of dismissing it entirely when its performance falls short of expectations.

While I agree that AI's current capabilities have limitations, maintaining consistent learning and practice with AI represents a more rational approach.

In my view, cutting-edge models like GPT-4 already demonstrate remarkable language proficiency and preliminary reasoning abilities - significant achievements indeed. For most people, language competence combined with basic reasoning can already accomplish previously impossible tasks.

Another point is that the scaling laws of large models (Scaling Laws) have not yet failed, meaning the more training data and parameters in the neural network, the stronger the model's capabilities. The current scale of models has not reached its limit, indicating there is still room for improvement in large models. If there are further technological breakthroughs, it suggests that true Artificial General Intelligence (AGI) may not be far away.

For these reasons, I remain optimistic about the future of AI. It will likely follow a trajectory similar to Moore's Law in the PC era, with significant leaps in AI technology occurring periodically. During this process, if we continue to learn and leverage AI effectively, we can not only avoid anxiety but also use AI to accomplish more valuable tasks.

Understanding AI's Strengths and Limitations

Facing even the most formidable opponent, knowing both yourself and your adversary ensures victory in every battle. By understanding AI's strengths and limitations, we need not fear being overtaken by AI; instead, we can harness it to serve our purposes.

Taking large language models as an example, I've summarized their strengths as follows:

However, there are also some obvious limitations:

By understanding both their strengths and limitations, I can not only avoid anxiety but also leverage their advantages while mitigating their weaknesses.

Teacher Zou Xin (Chief Content Advisor of CSDN & "The New Programmer") gave a great suggestion: Having a social circle for communication is also a good way to avoid anxiety.

Upon closer observation, I found that many people suffer from AI anxiety like me. Since everyone is anxious, it actually makes the anxiety feel less intense. We can also exchange AI news and learning experiences regularly, which is very helpful for improving our skills in using AI.

Over the years, I've developed a habit of writing down and sharing the knowledge I acquire daily. This practice not only helps me better understand vague details but also allows me to receive valuable feedback. The same applies to my journey in learning AI. I document and share valuable insights and knowledge I come across, especially since AI attracts so much attention nowadays. Many find my shared content highly beneficial, leading to positive feedback—ranging from gratitude to corrections—which has greatly enriched my learning and significantly eased my AI-related anxiety.

Since the saying "What replaces you isn't AI, but people who leverage AI" causes me anxiety, the best way to counter it is to become someone who effectively uses AI.

Starting early this year, I actively began integrating AI into my daily work and life. For instance, I've experimented with AI to enhance efficiency in standardized writing, programming, and translation tasks.

For daily emails, work task descriptions, and documentation, I rely on ChatGPT for drafting. Particularly, as my English isn't perfect, I used to spend considerable time proofreading grammar and spelling for formal emails and documents. Now, with ChatGPT, I can briefly describe what I need in a mix of Chinese and English, and it generates high-quality English content quickly and accurately.

When it comes to coding, AI is limited by context length and can't handle complex projects entirely, but it excels at assisting with code generation, small modules, or functions. I use GitHub Copilot the most—it acts like a "copilot" while coding. With good comments, it can generate code for me, especially for test cases I used to dislike writing. Now, with AI's help, these tasks are completed effortlessly. It even assists with complex algorithms, greatly boosting my development efficiency.

The biggest help from AI, however, has been in translation.

I often come across first-hand articles or videos in English, but many of my Weibo followers prefer high-quality Chinese translations. Before ChatGPT, I couldn't do this efficiently because tools like Google Translate required extensive proofreading. After GPT-4 was released, I found its translation quality superior to professional services like Google Translate, with strong customization options. This led me to experiment more with AI-powered translations.

First, for text translation, I noticed that ChatGPT's initial output isn't perfect and still carries machine translation traces. However, if I ask ChatGPT to polish the translation once more, the result becomes much smoother and nearly indistinguishable from human translation. Using this method, I can now translate large volumes of high-quality English content into Chinese with minimal proofreading (see Figure 3).

Next comes subtitle translation. Subtitles not only contain text but also timestamps. Translating English subtitles requires not only converting the English text to Chinese but also readjusting the timing and text segmentation of the Chinese subtitles based on the translated word order and length. In the past, tasks like subtitle translation were handled by entire teams—some specialized in translation, while others adjusted the timing. Today, large language models combine linguistic and reasoning capabilities, enabling them to both translate and adjust timestamps, significantly improving the efficiency of subtitle translation. Over the past year, with the help of AI, I’ve personally translated nearly 100 instructional videos—something that would have been unimaginable before.

Unknowingly, I have transitioned from someone who knew nothing about AI and was filled with anxiety towards it, to someone who no longer fears AI and extensively applies it in daily life.

In November 2023, OpenAI's board suddenly fired CEO Sam Altman, leading many to speculate that it was because OpenAI had developed a highly intelligent AI system. This incident once again sparked anxiety and even fear about AI among many people. However, I no longer feel anxious about it because, through my learning and practice, I have come to clearly understand the principles and limitations behind current AI technology.

I believe that each of us can learn about and understand AI. There's no need to worry about it replacing us; instead, we can use AI as a tool to enhance our productivity. Both you and I can become proficient users of AI.