Appearance
The JD List: Open Roles at Top Companies
This list answers a very specific question: if you opened a job board right now, what exactly are the open roles in trending AI fields asking for?
Let's set the boundaries first. This article reproduces no real job description verbatim, and it is not a real-time scrape—JDs are fluid; a role that's open today may be frozen next month. What it distills instead is the recurring skill combinations and role archetypes found in public JDs from major tech companies in China and abroad between 2023 and 2025, centered on trending AI concepts. For each role type, you get 2–3 sample JDs synthesized from public information to benchmark yourself against. What you should take away is a reusable analysis framework, not a job listing that will expire.
Data Cutoff and Disclaimer
This article is synthesized from job descriptions publicly visible around the time of writing (dataAsOf: 2025-06) on companies' official career sites and recruiting platforms (Maimai, BOSS Zhipin, Liepin, LinkedIn, etc.). All sample JDs are "illustrative descriptions distilled from common patterns"; they are no guarantee of any company's currently open roles, nor do they point to any specific company or position. For actual openings, locations, headcounts, and requirements, defer to what is live on the official career pages.
1. How to Read This List
1. Job Titles Drift; Skill Sets Are the Constant
Start with a real case of title drift: in 2021, teams everywhere were hiring "NLP algorithm engineers"; from 2024 onward, the very same departments began posting openings for "LLM application engineer," "RAG engineer," and "Agent engineer."
How one team's job titles evolved over four years (illustrative):
2021 NLP Algorithm Engineer / Search Algorithm Engineer
2022 Algorithm Engineer (LLM Track)
2024 LLM Application Engineer / LLM Algorithm Engineer
2025 RAG Engineer / Agent Engineer / Multimodal Algorithm Engineer
The skill core barely changed: Python + Transformer + LLM applications + engineering.
Only the title kept changing.The first rule of reading a JD: translate the title into a "skill vector" instead of memorizing titles. RAG application engineer, LLM application engineer, and AI engineer may come from the same team with about 70% overlap in skill requirements, while two identically named "LLM engineers"—one on a foundation model team, one on a business tech team—may have completely unrelated requirements.
2. Distinguish "Hard Requirements" from "Nice-to-Haves"
Almost every JD comes in two sections: one for "requirements" and one for "nice-to-haves / preferred qualifications."
| Hard requirements | Nice-to-haves (preferred) | |
|---|---|---|
| What they really mean | Cross this line or you won't get an interview | Bonus points if your resume has them; usually not a dealbreaker |
| Typical content | Degree, years of experience, must-know languages and frameworks | Domain experience, publications, competitions, open-source contributions |
| How they screen | A filter: miss one and you're out | A ranker: separates the candidates who already passed |
| Candidate strategy | Check each one; you need them all | Treat them as "differentiation material"—one or two with a good story is enough |
Three Signals for Telling Hard Requirements from Nice-to-Haves
- Wording: "proficient in," "expert-level," "X+ years of" are usually hard requirements; "prior experience with X is a plus" and "bonus points" mark nice-to-haves.
- Count: hard requirements usually number five or fewer and are independent of each other; a JD listing 12 "hard requirements" is really painting an ideal-candidate picture, and the real bar is lower than it looks on paper.
- Role type: campus recruiting JDs have few, broad hard requirements (they test fundamentals); experienced-hire JDs have many, narrow ones (they test fit).
3. How to Use This List Properly
① Pick 2–3 target roles → choose the closest match from Sections 2–7 of this article
② Check off against the knowledge-point breakdown → use [JD Knowledge-Point Breakdown](/career/knowledge-map) to mark every item
can do / can't do / sort of know it
③ Build a remediation list and set priorities → "can't do" items under hard requirements come first
"sort of know it" items under nice-to-haves come secondThe full methodology behind this "learn backward from the JD" approach is covered in Learning Paths; turning role requirements into an actionable study plan is its main thread.
2. LLM Algorithm Engineer (Pretraining / Fine-Tuning / Alignment)
Role overview: Building the capabilities of large language models—one of the "model-building" roles. Foundation model teams, research institutes, and open-source teams at China's big tech companies post these JDs year-round, spanning three sub-tracks: pretraining, instruction fine-tuning (SFT), and alignment (RLHF/DPO). The hard requirements generally run higher than for application roles.
Sample JD 1: LLM Algorithm Engineer (Alignment)
| Dimension | Content |
|---|---|
| Role | LLM Algorithm Engineer (Alignment / RLHF) |
| Company type | Big tech foundation model team / research institute |
| Responsibilities | Build alignment data pipelines; implement and tune RLHF/DPO training workflows; design preference-data collection and quality-assurance standards; contribute to model safety and controllability evaluations |
| Requirement keywords | Expert-level PyTorch; familiarity with Fine-Tuning and PEFT and Alignment: RLHF and DPO; understanding of how PPO works; large-scale training experience a plus |
| Related pages on this site | Fine-Tuning and PEFT · Alignment: RLHF and DPO · LLM Evaluation and Benchmarks |
Jargon decoder: "Familiar with RLHF/DPO" = you can explain clearly "how the reward model is trained, why PPO is expensive, and what DPO simplifies"; "alignment experience preferred" is often a nice-to-have rather than a hard bar—70% of alignment work is data engineering, not algorithms, so highlight your data and evaluation skills when you apply.
Sample JD 2: LLM Algorithm Engineer (Fine-Tuning / Vertical Adaptation)
| Dimension | Content |
|---|---|
| Role | LLM Algorithm Engineer (Industry Fine-Tuning) |
| Company type | Big tech or AI companies in finance / healthcare / legal and other verticals |
| Responsibilities | Run vertical-specific instruction fine-tuning on open-source base models; build domain instruction datasets; create domain evaluation sets and keep iterating on them; work with inference optimization to bring down deployment costs |
| Requirement keywords | Python; PyTorch; Fine-Tuning and PEFT (LoRA/QLoRA); Evaluation and Benchmarks; understanding of domain data |
| Related pages on this site | Fine-Tuning and PEFT · Large Language Models (LLM) · Building an LLM Eval Suite |
Jargon decoder: "Expert in LoRA" = you can explain the low-rank assumption, how to choose rank, the trade-offs between LoRA and full-parameter fine-tuning, and "why LoRA works with so few parameters"; "familiar with evaluation" = you can design domain evaluation sets and weigh the cost trade-offs between human and model-based evaluation, not just run MMLU.
Sample JD 3: LLM Algorithm Engineer (Pretraining, Senior)
| Dimension | Content |
|---|---|
| Role | LLM Algorithm Engineer (Pretraining, Senior) |
| Company type | Leading cloud providers / high-profile startups |
| Responsibilities | Take part in pretraining models in the 10B–100B parameter range; optimize data mixtures and curriculum learning strategies; help optimize distributed training frameworks and failure recovery |
| Requirement keywords | Large-scale training experience (thousand-GPU class); deep understanding of Transformers and Attention; Megatron/DeepSpeed; data engineering |
| Related pages on this site | Transformers and Attention · Large Language Models (LLM) · Inference Optimization and Quantization |
Jargon decoder: "Participated in thousand-GPU-scale training" = a genuine hard bar, almost exclusively open to senior candidates/PhDs—new grads shouldn't benchmark themselves against it; "data mixture" = the combination of high-quality data ratios, deduplication, and curriculum learning—the most underrated responsibility in pretraining JDs.
Learn to Spot the "Fake Gates" in LLM Role JDs
LLM role JDs tend to read the most "intimidating" (RLHF, thousand-GPU training, Agent frameworks), but read closely and most of it is nice-to-haves. The real distribution: only pretraining roles genuinely require large-scale training experience; the core work of fine-tuning/alignment roles is usually data and evaluation; application roles only need "use models well + evaluate well + engineer well." How to tell: look at which org the role belongs to—foundation model team vs. business tech team. The former wants training capability; the latter wants application capability.
3. RAG Application Engineer
Role overview: Combining general-purpose models with retrieval into products like enterprise knowledge-base Q&A, AI search, and customer support. Demand has been heavy and openings plentiful since 2023—this is the workhorse of the application track. The core mandate is "get every stage of the RAG pipeline right, and be able to quantify it with evaluations."
Sample JD 1: RAG Application Engineer (Enterprise Knowledge Base)
| Dimension | Content |
|---|---|
| Role | RAG Application Engineer (Enterprise Knowledge Base) |
| Company type | Big tech internal productivity teams / digital transformation departments at mid-to-large enterprises / AI startups |
| Responsibilities | Build enterprise document retrieval Q&A systems; design the full chain of document parsing, chunking, embedding, and reranking; build RAG evaluation sets; improve recall and answer accuracy |
| Requirement keywords | End-to-end Retrieval-Augmented Generation (RAG); Vector Databases and Semantic Search; Python; LangChain/LlamaIndex or in-house pipelines; an evaluation mindset |
| Related pages on this site | Retrieval-Augmented Generation (RAG) · Building a RAG App from Scratch · Vector Databases and Semantic Search |
Jargon decoder: "Expert in RAG" = you can explain chunking (how to split, how big, along what boundaries), embedding selection, the difference between retrieval and reranking, hybrid search, and—most critically—how to evaluate (retrieval hit rate + citation accuracy of generated answers); "familiar with vector databases" = you understand similarity computation, index types (HNSW, etc.), and metadata filtering, not merely having used a particular product.
Sample JD 2: AI Search / Q&A Product Engineer
| Dimension | Content |
|---|---|
| Role | AI Search Algorithm Engineer (RAG) |
| Company type | Search companies / content platforms / AI-native applications |
| Responsibilities | Improve retrieval quality and the citation accuracy of answers; implement reranking models; handle long documents and multi-turn conversations; build offline evaluation and online metric monitoring |
| Requirement keywords | Retrieval-Augmented Generation (RAG); retrieval ranking experience; LLM Evaluation and Benchmarks; online-metrics thinking |
| Related pages on this site | Retrieval-Augmented Generation (RAG) · LLM Evaluation and Benchmarks · Common Pitfalls and Antipatterns |
Jargon decoder: "Improve retrieval quality" = quantifying retrieval and reranking with metrics such as recall@k and MRR; "online metrics" = business metrics like citation click-through rate, off-topic answer rate, and retention—the bonus that wins RAG role interviews is "you not only have offline metrics, but also a closed loop of online monitoring."
Sample JD 3: Knowledge Graph + RAG
| Dimension | Content |
|---|---|
| Role | LLM Application Engineer (Knowledge Injection) |
| Company type | Enterprise software / industry knowledge companies |
| Responsibilities | Combine knowledge graphs with RAG to improve domain answer accuracy; build entity-relation extraction and graph query pipelines; optimize multi-hop Q&A scenarios |
| Requirement keywords | Retrieval-Augmented Generation (RAG); Knowledge Graphs and Knowledge Injection; Python; graph database experience a plus |
| Related pages on this site | Knowledge Graphs and Knowledge Injection · Retrieval-Augmented Generation (RAG) · Vector Databases and Semantic Search |
Jargon decoder: "Knowledge injection" = using RAG, knowledge graphs, fine-tuning, and other means to put external knowledge into a form the model can use—the keywords are "structured + traceable"; "multi-hop Q&A" = questions that take multi-step reasoning to answer, where interviews often probe how you decompose the retrieval path.
4. Agent Engineer
Role overview: A role that exploded after 2024, centered on getting the model to "plan on its own, call tools, and complete tasks." The JD keywords are Agent frameworks, tool calling, ReAct, task planning, and long-horizon stability. See AI Agents.
Sample JD 1: Agent Application Engineer
| Dimension | Content |
|---|---|
| Role | Agent Application Engineer |
| Company type | AI startups / big tech innovation business units |
| Responsibilities | Design and implement Agent workflows; integrate tools (APIs, databases, browsers); build the closed loop of task planning, tool calling, and result validation; handle multi-turn failure retries and fallbacks |
| Requirement keywords | AI Agents; Python; Function Calling / Tool Use; the ReAct pattern; hands-on skill at the level of Building an Agent from Scratch |
| Related pages on this site | AI Agents · Building an Agent from Scratch · Prompt Engineering |
Jargon decoder: "Familiar with ReAct" = you can explain the "reason → act → observe" loop, why it beats calling tools directly, and when it breaks down; "Function Calling" = the model outputs structured tool-call arguments, where the core is schema design, argument validation, and error handling; "fallback" = detecting when an Agent is stuck in a loop or drifting off course and exiting safely—the interview question Agent roles love most is "what do you do when your Agent hangs?"
Sample JD 2: Agent Platform / Orchestration Engineer
| Dimension | Content |
|---|---|
| Role | Agent Platform Engineer |
| Company type | Big tech AI platform departments / cloud providers |
| Responsibilities | Build Agent orchestration and runtime platforms; provide visual workflows, tool registries, state management, and monitoring; support multi-Agent collaboration and parallel scheduling |
| Requirement keywords | Backend engineering skills; AI Agents patterns; distributed task scheduling; observability |
| Related pages on this site | AI Agents · Common Pitfalls and Antipatterns · Deployment and Inference Optimization in Practice |
Jargon decoder: "Orchestration platform" = turning the Agent from a "one-off conversation" into an "operable system": tool registries, state persistence, task queues, failure retries, audit logs; "multi-Agent collaboration" = the scheduling problem of splitting up tasks, assigning them, and merging results—your answer has to land on engineering, not on "role-playing" concepts.
Sample JD 3: Agent Products / Embodied AI
| Dimension | Content |
|---|---|
| Role | Agent Algorithm Engineer (Multimodal Interaction) |
| Company type | Embodied AI / robotics / smart hardware companies |
| Responsibilities | Combine visual perception with language models to build a "see → think → act" closed loop; generate simulation-environment data; deploy models to real hardware |
| Requirement keywords | AI Agents; Multimodal Models; Python; reinforcement learning or imitation learning a plus |
| Related pages on this site | Multimodal Models · AI Agents · Diffusion Models and Generative AI |
Jargon decoder: "See → think → act" = the closed loop of perception, planning, and execution—the standard paradigm for embodied AI Agents; "simulation data" = generating training data in simulated environments because real-robot data is expensive and risky—a reminder that data engineering skills matter just as much in Agent roles.
5. AI Infra / Inference Optimization Engineer
Role overview: Making models run fast and run cheap. One of the fastest-growing role types after 2024. JD keywords: quantization, KV Cache, inference frameworks (vLLM), deployment, cost optimization, performance profiling.
Sample JD 1: Inference Optimization Engineer
| Dimension | Content |
|---|---|
| Role | LLM Inference Optimization Engineer |
| Company type | Cloud providers / big tech AI platforms / high-profile startups |
| Responsibilities | Optimize LLM inference latency and throughput; implement quantization, KV Cache, and continuous batching; build performance benchmarks; support business scaling |
| Requirement keywords | Inference Optimization and Quantization; Python + C++/CUDA; vLLM/TensorRT; deep understanding of Transformers and Attention |
| Related pages on this site | Inference Optimization and Quantization · Deployment and Inference Optimization in Practice · Transformers and Attention |
Jargon decoder: "KV Cache" = reusing past Keys/Values during generation—be ready to explain why it saves compute, how much VRAM it costs, and how to manage it; "continuous batching" = dynamic slot scheduling that solves the problem of "one long sequence stalling the entire batch"; "quantization" = INT8/INT4 principles, where precision loss comes from, and calibration methods—see Inference Optimization and Quantization.
Sample JD 2: Model Deployment / Serving Engineer
| Dimension | Content |
|---|---|
| Role | Model Serving Engineer |
| Company type | Big tech platform departments / SaaS companies |
| Responsibilities | Take models to production as services; elastic scaling and canary releases; monitoring, alerting, and fault diagnosis; multi-model resource scheduling |
| Requirement keywords | Containers and cloud native; Python/Go; Deployment and Inference Optimization in Practice; monitoring and observability |
| Related pages on this site | Deployment and Inference Optimization in Practice · Inference Optimization and Quantization · Common Pitfalls and Antipatterns |
Jargon decoder: "Serving" = packaging a model into a highly available API—the test is QPS/latency budgets, timeouts and retries, VRAM sharing; "canary release" = running old and new models side by side and shifting traffic in proportion—the test is how you guarantee "quality doesn't fall off a cliff after the switch," which requires taking Evaluation and Benchmarks online.
Sample JD 3: Training Infrastructure Engineer (Infra-Focused)
| Dimension | Content |
|---|---|
| Role | LLM Training Infrastructure Engineer |
| Company type | Leading cloud providers / foundation model companies |
| Responsibilities | Build and schedule distributed training clusters; fault detection and resuming from checkpoints; network topology and storage optimization; workload colocation and utilization improvements |
| Requirement keywords | Distributed systems; GPU clusters; Python + a systems language; understanding the Large Language Models (LLM) training process |
| Related pages on this site | Large Language Models (LLM) · Inference Optimization and Quantization · Deployment and Inference Optimization in Practice |
Jargon decoder: "Checkpoint recovery" = a thousand-GPU training run goes for weeks, and when it crashes mid-run you resume from a checkpoint—the test is systems design (state saving, distributed consistency); "colocation" = training and inference jobs sharing GPUs—the test is resource isolation and scheduling.
6. AI Product Manager
Role overview: Defining "what problems AI can solve and how to measure success." JD keywords: model capability assessment, prompt design, evaluation frameworks, cost accounting, the data flywheel. You don't have to build models, but you have to understand them.
Sample JD 1: AI Product Manager (Consumer Apps)
| Dimension | Content |
|---|---|
| Role | AI Product Manager (Conversational / Content Generation) |
| Company type | Big tech innovation units / AI-native consumer apps |
| Responsibilities | Define the shape of AI conversational products; design prompts and interaction flows; build effectiveness evaluation frameworks; track user feedback and model iterations; assess competitors |
| Requirement keywords | Prompt Engineering; LLM Evaluation and Benchmarks; data analysis; product methodology; understanding the limits of RAG/Agent capabilities |
| Related pages on this site | Prompt Engineering · LLM Evaluation and Benchmarks · Retrieval-Augmented Generation (RAG) |
Jargon decoder: "Understands LLM capability boundaries" = knowing which tasks suit models and which they handle poorly, and judging "can prompting solve this need, does it need fine-tuning, or does it need a different approach altogether"; "builds evaluation frameworks" = turning "is it good?" from a gut feeling into a reproducible process—the single most valuable skill for an AI PM, see Building an LLM Eval Suite.
Sample JD 2: AI Product Manager (B2B / Enterprise Services)
| Dimension | Content |
|---|---|
| Role | Enterprise AI Product Manager |
| Company type | Enterprise software companies / big tech B2B businesses |
| Responsibilities | Research customer scenarios and define AI features; design knowledge-base Q&A / workflow automation products; account for token costs and ROI; drive data security and compliance solutions |
| Requirement keywords | Requirements analysis; understanding of Retrieval-Augmented Generation (RAG) capabilities; cost modeling; AI Safety and Governance awareness; B2B delivery experience |
| Related pages on this site | Retrieval-Augmented Generation (RAG) · AI Safety and Governance · Prompt Engineering |
Jargon decoder: "Token cost accounting" = converting model calls into per-user/per-request costs to judge whether the business model holds up; "data security and compliance" = enterprise data never leaves the domain, content compliance, auditability—the hidden gate for B2B AI products and the core of the AI Safety and Governance page.
Sample JD 3: AI Product Manager (Platform / Tools)
| Dimension | Content |
|---|---|
| Role | AI Platform Product Manager |
| Company type | Cloud providers / big tech platform teams |
| Responsibilities | Design the product shape of model APIs and developer platforms; define pricing and quota policies; build developer experience and documentation; analyze model usage and cost data |
| Requirement keywords | Platform product thinking; understanding of the Large Language Models (LLM) ecosystem; data analysis; cost awareness from Inference Optimization and Quantization |
| Related pages on this site | Large Language Models (LLM) · Inference Optimization and Quantization · LLM Evaluation and Benchmarks |
Jargon decoder: "Developer experience" = documentation, SDKs, error messages, debugging tools—letting developers "hook up to the model in one line of code"; "quotas and pricing" = rate limiting, billing, prioritization—the test is whether you can translate technical metrics into business rules.
7. Multimodal Algorithm Engineer
Role overview: Modeling tasks that span text + images + speech + video, in two directions: multimodal understanding and multimodal generation. JD keywords: vision encoders, image-text alignment, diffusion models, cross-modal retrieval, video understanding.
Sample JD 1: Multimodal Understanding Algorithm Engineer
| Dimension | Content |
|---|---|
| Role | Multimodal LLM Algorithm Engineer (Understanding) |
| Company type | Big tech foundation model teams / content platforms |
| Responsibilities | Train and fine-tune image/video understanding models; align vision encoders with language models; build multimodal evaluation sets; image-text retrieval and visual question answering |
| Requirement keywords | Multimodal Models; Transformers and Attention; Python/PyTorch; data and evaluation |
| Related pages on this site | Multimodal Models · Transformers and Attention · LLM Evaluation and Benchmarks |
Jargon decoder: "Aligning the vision encoder with the LLM" = "translating" image features into the text feature space (e.g., CLIP-style contrastive learning)—the test is whether you understand the trade-offs among alignment approaches such as Q-Former, Adapters, and direct projection; "multimodal evaluation" = beyond accuracy, also checking image-text consistency and hallucination (the model "seeing" things that aren't there).
Sample JD 2: Image / Video Generation Algorithm Engineer
| Dimension | Content |
|---|---|
| Role | Generative AI Algorithm Engineer (Image/Video) |
| Company type | AI creative tool companies / big tech content platform teams |
| Responsibilities | Improve text-to-image/text-to-video quality with diffusion models; accelerate sampling and inference; controllable generation (ControlNet-style); build aesthetics and consistency evaluations |
| Requirement keywords | Diffusion Models and Generative AI; image/video generation; Inference Optimization and Quantization; Python/PyTorch |
| Related pages on this site | Diffusion Models and Generative AI · Image Generation Case Study · Video Generation Case Study |
Jargon decoder: "Sampling acceleration" = DDIM, DPM-Solver, and similar methods that cut the number of diffusion sampling steps—the core "quality vs. speed" trade-off; "ControlNet" = steering generation output with extra conditions (pose, edges, depth)—the test is whether you understand conditional injection mechanisms; "consistency evaluation" = measuring whether generated images match the text description (e.g., CLIP Score)—essential quantitative literacy for generation roles.
Sample JD 3: Multimodal Algorithm Engineer (Speech)
| Dimension | Content |
|---|---|
| Role | Multimodal Algorithm Engineer (Speech) |
| Company type | Speech technology companies / big tech smart hardware divisions |
| Responsibilities | LLM-ify speech recognition and synthesis; joint speech-text modeling; end-to-end dialogue systems; low-resource optimization |
| Requirement keywords | Multimodal Models; speech technology; Transformer architectures; data augmentation |
| Related pages on this site | Multimodal Models · Speech AI Case Study · Transformers and Attention |
Jargon decoder: "LLM-ification" = migrating speech tasks from traditional HMM/Attention architectures to unified Transformer/LLM frameworks; "end-to-end" = speech straight to text/semantics, no longer chaining multiple independent modules—a classic multimodal fusion interview question.
8. Master Table of High-Frequency JD Jargon
Here is a quick-reference table consolidating the jargon from the six role types above—when an interviewer writes one line, you should be thinking of three things:
| What the JD says | Literal meaning | What the interview actually expects you to explain |
|---|---|---|
| Expert in RAG | Has built knowledge-base Q&A | How to chunk, embedding selection, retrieval vs. reranking, how to evaluate RAG |
| Familiar with Agents | Has used an Agent framework | The ReAct loop, tool-call schemas, failure fallbacks, long-task stability |
| Expert in LoRA | Has fine-tuned a model | The low-rank assumption, rank selection, trade-offs vs. full-parameter fine-tuning |
| Familiar with RLHF/DPO | Understands alignment | Reward model training, why PPO is expensive, what DPO simplifies |
| Familiar with inference optimization | Has heard of quantization | INT8/INT4 principles, KV Cache, continuous batching, latency/throughput trade-offs |
| Familiar with vector databases | Has used some vector DB | Similarity metrics, HNSW, metadata filtering, how chunking affects retrieval quality |
| Has built an evaluation framework | Has run a few metrics | Evaluation set design, human vs. model evaluation, mapping metrics to business goals |
| Understands multimodal AI | Has read multimodal papers | Vision-encoder-to-LLM alignment approaches, multimodal hallucination, cross-modal retrieval |
| Familiar with diffusion models | Has used text-to-image tools | Forward/reverse processes, sampling acceleration, conditional control, consistency evaluation |
| Has data engineering experience | Has processed data | Cleaning, mixture ratios, deduplication, quality checks, data pipelines—the hidden hard requirement in almost every role |
The One-Line Takeaway
For every technical term in a JD, the interview assumes you can "explain the principle + do the work + articulate the trade-offs." At the literal "used it / heard of it" level, you might as well have never learned it. If you can't see through any row of this table, go back to the corresponding page on this site and fill the gap.
9. Trend Observations and Job-Seeking Advice
The observations below are directional judgments as of dataAsOf (2025-06), not precise statistics.
1. Supply and Demand Across the Six Role Types
| Role | Demand heat | Supply competition | Advice |
|---|---|---|---|
| LLM Algorithm Engineer (Pretraining) | Medium | Extremely high | The bar is training scale and publications; not recommended as a primary target for most people |
| LLM Algorithm Engineer (Fine-Tuning/Alignment) | High | High | Differentiate with data + evaluation skills |
| RAG Application Engineer | Very high | Medium-high | Huge number of openings with clear boundaries—one of the most beginner-friendly entry points |
| Agent Engineer | High and rising | Medium | What's scarce is people who "have built real pipelines," not people who "have read the papers" |
| AI Infra / Inference Optimization | High and rising | Medium | A systems engineering background is hard currency; the supply-demand gap is obvious |
| AI Product Manager | High | Medium-high | Evaluation frameworks + cost awareness are scarce skills |
| Multimodal Algorithm Engineer | High | High | Understanding and generation are splitting into two tracks—pick one and go deep |
2. Three Structural Observations Worth Noting
- Evaluation skill is scarce across every role. Nearly every JD mentions "evaluation," yet few people can genuinely "design evaluation sets + weigh human vs. model evaluation + map metrics to business outcomes." It's the easiest place to leapfrog the competition; the entry point is LLM Evaluation and Benchmarks.
- Data engineering is the hidden hard requirement. JDs often don't state it directly, but "knowing data cleaning, mixture ratios, and quality checks" largely decides whether you can do real work in your first month on the job. See Dataset and Tool Directory.
- Don't fixate on big tech. Big tech JDs use standardized wording and fine-grained division of labor, which makes them good for calibrating your standards—but the offer you actually land may well come from mid-size companies, unicorns, or vertical-domain players. Take your standards from big tech; find your opportunities in the whole market.
Further Reading
Continue reading on this site
- Module Overview and Role Landscape — the full picture of the seven role types and their capability profiles; locate yourself before reading JDs
- JD Knowledge-Point Breakdown — turn this page's JD skill terms into a personal remediation checklist
- Capability Benchmarking: What Your Resume Should Highlight — rewrite your project experience against JDs and tell the story interviewers want to hear
- Interview Question Bank — high-frequency questions and answer frameworks; self-test last
- Learning Paths — the complete main line of learning backward from JDs and interview questions
- Retrieval-Augmented Generation (RAG) · AI Agents · Multimodal Models — the three knowledge pillars of the application track
- Inference Optimization and Quantization · Deployment and Inference Optimization in Practice — the two mainstays of the infrastructure track
- Glossary — one consistent terminology baseline, to avoid "sort-of-know-it" understanding
References
- Official career sites, Chinese companies: Alibaba (talent.alibaba.com), Tencent (careers.tencent.com), ByteDance (jobs.bytedance.com), Baidu (talent.baidu.com), Meituan (zhaopin.meituan.com), and other official career sites
- Official career sites, overseas companies: Google Careers (careers.google.com), Meta Careers (careers.meta.com), OpenAI Careers (openai.com/careers), Anthropic Careers (anthropic.com/careers), Microsoft Careers (careers.microsoft.com)
- Recruiting platforms: Maimai (maimai.cn), BOSS Zhipin (zhipin.com), Liepin (liepin.com), LinkedIn (linkedin.com)—useful for observing role distribution and JD wording frequency
- levels.fyi — reference for overseas role types and compensation levels; unofficial data
- Hugging Face Jobs — a bellwether for roles in the open-source AI ecosystem, often leading the way on new technology directions
- State of AI Report — an industry-wide report that includes talent market analysis
One Bottom Line
All JDs in this article are illustrative examples synthesized from public information—not scraped in real time and not pointing to any specific company's openings. Always defer to the JDs published live on official channels—treat every JD you read as training data; this article only teaches you how to read them.