CS Career Guide in the AI Era: Large Models, Agents, and Embedded Systems

Under the AI wave, computer science students are more anxious than ever—training bootcamps are fueling panic, job postings are raising the bar, and social media is full of "CS is dead" rhetoric. A Bilibili creator offered a counterintuitive take in just two minutes: the point isn't to work harder, but to pick the right direction.

According to this creator's analytical framework, CS students' career paths can be divided along two axes—education level and interests—which is far more useful than generic "keep pushing" advice.

Direction 1: Large Models—The Channel for Cashing In on Educational Advantages

If you have a double first-class key university undergraduate degree or a master's/PhD background, with solid math and algorithm fundamentals, the creator's first recommendation is crystal clear: prioritize large models.

Corresponding roles include large model algorithm engineer and algorithm researcher. These positions have high technical and educational thresholds, but salaries are in the top tier among technical roles. In other words, if you've already earned a strong educational background, don't waste it—large models are currently one of the few technical tracks where your educational advantage can be directly monetized.

The logic behind this isn't complicated: training, fine-tuning, and aligning large models all require deep mathematical foundations and engineering capabilities, and companies naturally apply educational filters when hiring. For qualified students, this is the battlefield with the most favorable supply-demand dynamics.

What is a "large model"? A large model (Large Language Model, LLM) typically refers to a neural network with parameters ranging from billions to trillions, pre-trained on massive text corpora. Representative products include GPT, Claude, Qwen, DeepSeek^Ad, and others. Its core capabilities come from the Transformer architecture and self-attention mechanism—learning statistical patterns of language during pre-training, then aligning with human preferences through supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), thereby gaining general capabilities in conversation, reasoning, and code generation.

What does a large model algorithm role actually do? Positions typically split into several tracks: pre-training handles data cleaning, distributed training, and compute scheduling; fine-tuning and alignment adapt base models to specific tasks or domains; inference optimization handles quantization, distillation, KV Cache, and other acceleration methods to reduce deployment costs. These roles demand far more in linear algebra, probability and statistics, and optimization theory than typical development positions—which is precisely why companies often use education and publication records as the first screening threshold during resume review. This is the technical basis for the article's description of it as a "channel for cashing in on educational advantages."

Direction 2: Agents—The Pragmatic Choice for Regular Undergraduates

For students from regular undergraduate programs, private universities, or those who upgraded from associate to bachelor's degrees, the creator's direction is AI applications and agent development.

There's a key clarification here: "prioritize" doesn't mean "only option." Agent development isn't about training a large model from scratch—it's about using large models' capabilities to solve real business problems. Enterprise deployment scenarios are already quite specific—intelligent customer service, office assistants, data analysis assistants, enterprise knowledge bases—and demand in these areas is growing.

More noteworthy is the policy signal. The creator mentions that agent development-related occupations have begun entering the national vision of new professions and new job categories. This means the direction has both market demand and institutional recognition, making it worth attention from CS students at all educational levels.

From a tech stack perspective, agent development emphasizes engineering implementation capabilities rather than pure algorithm research—you need to understand API calls, RAG architecture, workflow orchestration, prompt engineering, and how to embed large model capabilities into business processes. The barrier to acquiring these skills is far lower than training large models, but their direct value to enterprises may be higher.

What's the relationship between agents and RAG? An agent refers to a system with a large model as its decision-making core, capable of autonomously calling tools and executing multi-step tasks to achieve goals. Its typical structure is "large model + tool calling + memory + planning." RAG (Retrieval-Augmented Generation) is one of the most commonly used technical components: first retrieving relevant document fragments from an enterprise knowledge base, then feeding those fragments as context to the large model to generate answers—solving the problems of outdated model knowledge and "hallucinations."

Workflow orchestration and prompt engineering. Workflow orchestration refers to using code or visual tools to string together steps like "retrieval—reasoning—API calls—validation—output" into a stable process. Common frameworks include LangChain, LlamaIndex, and Dify. Prompt engineering stabilizes model output quality through structured instructions, few-shot examples, and chain-of-thought methods. The essence of these skills is treating the large model as a highly capable but uncertain component, using engineering methods to constrain and combine it—thus requiring significantly less algorithmic theory than training large models and being closer to the skill stack of traditional backend development and system integration.

Direction 3: Embedded Systems—The Bridge for AI into the Physical World

If you have a natural interest in hardware, microcontrollers, or integrated circuits, the third path is embedded systems, extending toward robotics, low-altitude economy (drones), and intelligent driving.

The creator used a precise metaphor: large models are AI's brain, but for AI to enter the real world, it needs embedded systems to give it hands and feet. Robots need to move, drones need to fly, smart cars need to perceive and control—none of this is possible without underlying hardware and embedded systems.

The unique value of this direction: it sits at the intersection of AI and the physical world and is unlikely to be replaced by pure software solutions in the short term. As embodied intelligence and autonomous driving advance, demand for embedded engineers will only increase.

Embedded systems and embodied intelligence. Embedded systems refer to computing systems customized for specific functions with constrained resources, typically running on microcontrollers (MCUs) or embedded Linux, directly controlling sensors, motors, communication modules, and other hardware. The difference from general software development: it requires handling real-time constraints, power consumption, memory, and hardware timing—debugging often requires oscilloscopes and logic analyzers.

Low-altitude economy and intelligent driving. The low-altitude economy refers to the industrial form built around drones, eVTOL, and other low-altitude aircraft, and has been written into government work reports in multiple regions. Intelligent driving relies on onboard embedded platforms to complete the closed loop of perception, decision-making, and control. What these two directions share: AI models must ultimately land on specific chips and mechanical structures to create value—the intermediate adaptation, real-time scheduling, and safety redundancy are all handled by embedded engineers. This is why the article judges that this direction is "unlikely to be replaced by pure software solutions in the short term"—algorithms can iterate in the cloud, but control execution must happen locally on physical devices.

The Real Risk: It's Not AI Stealing Jobs, But Choosing the Wrong Direction

The creator's conclusion is worth remembering: AI replaces those who only write repetitive code; what's truly hard to replace is people who understand large model architecture, agent design, and embedded systems at a low level.

What these three directions share: they all require systematic knowledge structures, not just coding proficiency. Each corresponds to different educational backgrounds and interest preferences, but they share one common premise—pick a direction, then build solid technical skills and projects.

Rather than competing with AI, think about how to make AI work for you. For CS students, the most urgent thing right now isn't持续 anxiety, but completing direction selection and investing in practice as soon as possible.

What's your educational background? Which direction are you planning to pursue? This might be a more worthwhile question to answer first than "should I change careers."

One-line Summary

CS career choices should be directed by education level: large models, agents, and embedded systems each have their fitting audience.

Paragraph Summary

Drawing on a Bilibili creator's analytical framework, the article offers anxious CS students a career path stratified by education and interests. First, those with double first-class undergraduate degrees or master's/PhD backgrounds and solid math and algorithm foundations should prioritize pursuing large model algorithm engineer and researcher positions—the track where educational advantages are most directly monetized. Second, students from regular undergraduate programs, private universities, and associate-to-bachelor's upgrades are better suited for AI applications and agent development, focusing on using large models to solve real enterprise problems rather than training models from scratch—and related occupations have entered the national vision of new professions. Third, those interested in hardware and microcontrollers can take the embedded route, extending toward robotics, low-altitude economy, and intelligent driving, becoming the hands and feet for AI to enter the physical world. The article ultimately emphasizes that AI replaces those who only write repetitive code; what's truly scarce are systems-oriented talents who understand architecture, design, and low-level fundamentals. The priority now is to choose a direction and invest solidly in practice.