AI Systems Bootcamp in China

AI Systems Bootcamp in China | Generated by AI

Home 2026.07

Question: Is starting an AI bootcamp in China for high school / university students (model training, agents, etc.) a good idea?

Answer:

Yes — but only if you don’t position it as “AI education”. That market is already saturated with tutorial-style courses. The viable version is:

“AI systems bootcamp: build + deploy real agents and small-scale models in 2–4 weeks.”

You’re not teaching theory. You’re compressing real production workflows (like what you already do with GPU training + CLI agents) into hands-on shipping.

1. Where the real opportunity is

Most China bootcamps:

teach Python basics
wrap HuggingFace pipelines
stop at notebook-level demos

Gap you can exploit:

end-to-end systems thinking
- data → training → eval → deployment → agent loop
real infra exposure
- multi-GPU training, not Colab toys
agent engineering
- tool use, memory, planning, CLI-first workflows

This is closer to:

“mini OpenAI applied lab”
not “AI course”

2. What to teach (tight, production-first)

Week 1 — LLM + inference core

tokenization, attention (only intuition-level math)
run local LLM (vLLM / llama.cpp)
build CLI chat interface

python chat.py --model Qwen2.5-7B

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B")

prompt = "Design a trading agent architecture"
inputs = tok(prompt, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=200)
print(tok.decode(out[0]))

Week 2 — Fine-tuning small models

LoRA / QLoRA
dataset building (critical skill most ignore)
eval loops

accelerate launch train_lora.py \
  --model meta-llama/Llama-3-8B \
  --dataset data.jsonl

Key idea: students learn data quality > model choice

Week 3 — Agents (core differentiator)

tool calling
memory (vector DB)
planning loop

Minimal agent loop:

while True:
    obs = env.observe()
    action = llm(f"Obs: {obs}\nWhat next action?")
    env.step(action)

Add:

web tools
code execution
file system tools

Week 4 — Capstone system

Examples:

research agent (paper → summary → repo)
trading simulator agent
WeChat content generator pipeline
code review bot for GitHub repos

Must deploy:

FastAPI backend
simple web UI or CLI (CLI preferred if you want engineering purity)

3. Stack (keep it realistic)

Training: PyTorch + HF + DeepSpeed (optional)
Inference: vLLM / llama.cpp
Agents: pure Python loops (avoid frameworks first)
Vector DB: FAISS (no overengineering)
Deployment: Docker + simple VPS

Avoid:

“no-code AI platforms”
heavy SaaS abstractions

4. Business reality in China

Strong signals

parents pay for “future-proof skills”
universities want “AI practical training”
strong interest in agents + LLM apps

Weak points

regulatory sensitivity (content + minors)
high churn unless output is visible (projects)

Pricing structure

2–4 week intensive: 3k–10k RMB/student
cohort-based (20–50 ppl)
strong leverage if you show GitHub output

5. Your real moat

Given your background (training models, GPU infra, agent CLI tools), your edge is:

you can show real training runs, not slides
you can teach “how systems break in production”
you can design curriculum from first principles

That’s rare. Most instructors are 2 layers removed from real infra.

6. Failure modes (important)

Becoming tutorial factory
- students copy notebooks → no retention
Too much theory
- kills perceived value fast
No artifact output
- must force GitHub repos every week
Over-frameworking agents
- LangChain-style abstraction early → confusion

7. MVP suggestion

Don’t start with a school.

Start with:

1 cohort (10–15 people)
2-week “agent sprint”
goal: everyone ships 1 working agent repo

If retention + output quality is high → expand.

References:

Back Donate