Many people learn AI like this: today a “GPT magic trick”, tomorrow a “newest model round-up” — half a year later the folder is full but they still cannot build anything themselves. The problem is the reversed order: chasing hype before the map is built; when the hype shifts, all is lost.
This guide gives a no-detour order, each step mapped to “who to watch, what to do”.
Step 1 · Build the cognitive map: know what AI looks like
The goal is not one function but “what AI contains and in what order I should learn it”.
Watch: waytoagi (Towards AGI)
The most systematic Chinese AI learning knowledge base and roadmap, covering entry-to-advanced paths, tools, and concepts in one place.
Action: spend a day or two on its roadmap; sketch the layers “prompt -> application -> development -> algorithm” in your head. After this map, all later learning has something to hang on.
Step 2 · Patch basics: the base matters more than hype
Basics split into three blocks; choose by goal, but touch all a bit.
Coding and ML base: Mofan Python
Short, punchy Python and ML entry, covering data processing, sklearn, neural nets.
Action: do Python basics and sklearn first; running a small model yourself beats memorising ten tool names.
Theory intuition: Tongji Zihao and Gengzhi Ge
The former reads AI papers and clarifies CV and LLM frameworks; the latter explains LLMs and frontier research to ordinary people in plain style.
Action: treat these two as a “frontier radar” to build intuition for “what an LLM is and what is being researched”, not swayed by marketing words.
Math derivations: shuhuai008
The “ML whiteboard derivation” series derives algorithm math from scratch, rigorous and gradual.
Action: when you truly want an algorithm, follow the whiteboard once; skip if you only need to use it, must-watch to go deep.
Step 3 · Use tools: apply in real scenarios
With basics, touching tools will not degrade into “only clicking buttons”.
AI-art entry: Qiuye aaaki
The one-click SD packs that brought Stable Diffusion to Chinese creators were built to lower the barrier.
Action: start with his intro and packs, then follow the model leaderboard to skip environment-setup detours.
New-tool tests: Lao Mai’s Tool Library
First-hand tests of the newest AI tools at top speed.
Action: when a new model drops, watch his test before deciding to spend time — skip your own pitfalls.
Tool practice: Qiuzhi 2046
AI tool reviews and hands-on teaching, covering both “how to use” and “where the industry goes”.
Action: pick a tool she reviewed, do it hands-on once, connecting step-two basics with step-three tools.
Step 4 · Build projects: run what you learned
Watching more beats nothing; building one usable thing beats watching more.
Project samples: Lin Yi LYi
Hardcore-yet-playful AI projects, from training an AI to beat games to running an AI company — showing how an idea becomes engineering.
Action: decompose one project’s implementation path; understand the full flow from need to deployment.
Practice thinking: Laojiang Jukepao
Practice-oriented AI application teaching, stressing using tools in real workflows.
Action: pick a real task on your desk (auto-organise docs, batch-generate copy), force an end-to-end AI pass. One real project beats ten videos.
Step 5 · Follow trends: keep learning from going stale
AI iterates fast; the last step is building “news” channels without being led by them.
Industry and new models: Qiuzhi 2046 and Lao Mai’s Tool Library
The former for industry direction, the latter for tool timeliness.
Action: scan both weekly as “intel update” only; do not chase every hype. Principle as base, tools as chase — steady division.
Three common mistakes
- Chasing new models, skipping principle. Hype changes quarterly; principle lasts years.
- Watching without practising. AI is hands-on; one project beats ten videos.
- Believing “quick mastery”. “Master LLMs in three days” is mostly clickbait; systems engineering needs time.
To find more reference accounts in this niche, see 10 AI-learning creators worth following or browse all indexed AI creators.