Seven Career Paths for Programmers in the Age of AI
Honestly, over the past two years, few people in tech haven't felt anxious.
Before, learning one language or one framework could feed you for five years. Now? GPT-4 hasn't been out long, and Claude can already write complete projects. Cursor has turned coding into a conversation. People around you keep asking: "AI is so powerful, are programmers going to lose their jobs?"
My view is: no. But the meaning of the word "programmer" is undergoing a fundamental change.
In the past, it was "humans write code, machines execute it." The future is more like "humans state requirements, AI understands, calls tools, automatically executes, and continuously optimizes." AI isn't opening a new track next to the software industry; it is redefining the entire field of software engineering.
So the question isn't "will programmers be replaced," but rather — in this new world, what kind of work is still worth doing?
From my observations, there are roughly a few paths to take.
Path 1: Build the "Brain"
This is the most fundamental and hardcore direction. Not using AI, but creating AI itself.
People in this field design model architectures, train large-scale models, optimize performance, and research new algorithms daily. The ChatGPT, Midjourney, and Stable Diffusion you see are all backed by these people wrestling with Transformer architectures, figuring out distributed training, and battling CUDA.
This path has a high barrier to entry. A solid mathematical foundation is basic; many positions require a master's degree, and a PhD is not uncommon. If you already have an algorithms background and enjoy digging into underlying principles, this is your home turf. Job titles are usually Algorithm Engineer, Large Model Researcher, NLP/CV Engineer, and the like.
But honestly, this path is narrow. Nationwide, only a handful of companies can afford to develop their own large models from scratch. Most programmers don't need to, and don't have the conditions to, squeeze into this space.
Path 2: Use the "Brain" to Build Products (LLM Application Development)
This is probably the most pragmatic and most in-demand direction right now.
You don't train models; you take off-the-shelf models (OpenAI, Anthropic, APIs from major domestic companies) to develop things people actually use. Enterprise intelligent customer service, AI search, knowledge base Q&A, smart office assistants, AI programming tools — these are the areas where companies are currently spending real money to hire.
The tech stack isn't unfamiliar either: calling APIs, writing prompts, setting up RAG (Retrieval-Augmented Generation, allowing AI to query internal company documents), using vector databases (like Milvus, Pinecone, Weaviate), and leveraging orchestration frameworks like LangChain or LlamaIndex. Essentially, it's still software development, just using the "large model" as a new, powerful component.
If you come from a backend, frontend, Android, or web development background, this path has the lowest transition cost. Your existing engineering skills are fully applicable; you just need to supplement some AI-related knowledge and ways of thinking. Prompt Engineering and RAG tuning are the core skills in this direction.
Path 3: Let AI Really "Do the Work" (Agent Development)
Much of the "building products with large models" mentioned earlier still stays at the "you ask a question, AI gives an answer" stage. But what Agent development aims for is enabling AI to independently complete a task.
For example, you say, "Help me analyze today's server logs." A traditional LLM application might give you a paragraph of analysis. But what would an Agent do? It would read the log files itself, identify error patterns, search the internet for relevant information, generate a report, and even propose repair suggestions — all without needing you to guide it step-by-step.
Behind this are capabilities like task planning (Planning), tool invocation (Tool Calling), long-term memory (Memory), self-reflection (Self-reflection), and automated workflows. Reasoning paradigms like ReAct and CoT (Chain-of-Thought), along with Agent frameworks like AutoGPT and LangGraph, are the hot tech stacks in this direction currently.
Agent development is one of the fastest-growing areas right now, with many companies exploring it. Future software systems are likely to be heavily developed and maintained with the participation or even leadership of Agents. This path is a bit deeper than pure "LLM application development," but the rewards are also greater.
Path 4: Be AI's "Logistics Chief" (AI Infra / MLOps)
Algorithm engineers have built the model, but how do you make it run fast, stable, and cheap in a production environment? That's the problem AI Infra solves.
Model deployment, GPU scheduling, inference acceleration, service architecture, monitoring and alerting... these tasks are somewhat similar to backend operations in the traditional internet era, but they face AI-specific challenges: How to allocate video memory? How to reduce inference latency to milliseconds? How to use tools like TensorRT, vLLM, and Triton Inference Server? How does Kubernetes manage GPU clusters? How to implement model compression techniques like Quantization and Distillation?
If you are already familiar with Docker, K8s, and Linux and are interested in the system level, this path is a great fit. Algorithm engineers are responsible for "building the rocket"; you are responsible for "making the rocket fly."
Path 5: Be the "Designer" of AI Products
Technology ultimately has to land in products, in the hands of users.
AI writing tools, AI drawing software, AI learning assistants, AI office apps — the core competitiveness of these products isn't just "using AI," but "how AI is appropriately integrated into the product." Successful future products won't be existing software with "an AI feature added on"; they will be designed around AI's capabilities from day one.
This requires you to understand technology, understand users, and make product decisions. Full-stack developers, entrepreneurs, and tech leads often have an advantage on this path.
Path 6: Dive Deep into an Industry (AI + Vertical Domain)
This might be a path many overlook, but it holds the greatest long-term value.
No matter how powerful AI is, it must solve real-world problems. Healthcare, finance, education, manufacturing, law... every industry has its own knowledge barriers, business processes, and compliance requirements. Many people understand AI, but people who understand both AI and a specific industry are very rare.
For example, in the medical field, AI-assisted diagnosis and medical knowledge assistants cannot be developed deeply without a medical background. In finance, risk control models and robo-advisors are just theoretical exercises if you don't understand the business logic. The same goes for education and manufacturing.
If you have already been deeply cultivating a certain industry for years, don't rush to switch tracks. Add AI as a tool, and you might instead become the most sought-after person.
Path 7: "Feed" AI Good Data (AI Data Engineering)
This last path might not sound as glamorous, but it is extremely important.
The upper limit of AI's capability depends heavily on data quality. Data collection, cleaning, labeling, management, quality control — these dirty, tiring jobs determine whether a model is "smart" or "dumb." Large AI companies invest heavily in this area; data engineers, ETL engineers, and data platform developers are all in rigid demand.
If you are good at dealing with data, and SQL, Python, data pipelines, and feature engineering are your comfort zone, this path is very stable.
So How Does an Ordinary Person Choose?
After listing seven paths, you might be even more confused. It's actually not that complicated.
Have an algorithms background and like researching fundamentals — go build brains, do model R&D. You need to chew on hard bones like Transformers, attention mechanisms, and distributed training.
Have an engineering background, done backend/frontend/App work — the smoothest route is: first learn large model application development (Prompt Engineering + RAG + Vector Databases), then go deeper into Agents (Tool Calling + Memory + Planning), and if you have spare energy, expand towards AI product direction. This is the route with the lowest transition cost and fastest results.
Familiar with system operations and have a feel for performance optimization — go do AI Infra. Focus on learning model deployment, inference optimization, GPU scheduling, and the MLOps toolchain.
Rooted in a certain industry for many years — don't switch industries, bring AI into your industry. Combining Fine-tuning with domain knowledge is your moat.
Good at data processing — data engineering will always have food on the table. ETL, data labeling pipelines, and feature stores are all core skills.
A Few Final Words
AI won't simply "replace" programmers. But it is indeed changing what the word "programmer" represents.
The core competitiveness of the future is no longer "how many lines of code can I write," but "can I use AI to build systems that solve complex problems." The essence of software engineering is shifting from "the relationship between humans and code" to "engineers designing systems, coordinating AI Agents, and completing complex tasks."
In other words, code itself is becoming less and less valuable, but the ability to solve problems using code (and AI) is becoming more and more valuable.
Don't panic. Go learn, go try, go use. The real danger isn't that AI is stronger than you, but that you refuse to admit AI is stronger than you, and then stand still.
Top 1 from juejin.cn, machine-translated. The original thread is authoritative.
How to build a side hustle