Thirteen years of deep learning, framed as a history of desperation

What it does

Daily-LLM is a bilingual (Chinese/English) curriculum repository tracing deep learning and large language models from 2012’s AlexNet through 2025’s DeepSeek R1. It organizes the field into six phases—visual intelligence, language transformers, scaling and multimodality, alignment, production systems, and capstone projects—each backed by historical context and code. The README presents a year-by-year timeline of what broke and what fixed it.

The interesting bit

The entire arc is framed as “被逼出来的历史”: history forced by constraints. Word2Vec happened because One-Hot vectors made “cat” as distant from “dog” as from “airplane”; ResNet solved the degradation wall beyond 20 layers; RLHF was necessary because GPT-3 would parrot toxicity without complaint. That narrative turns a standard tutorial collection into a story about problem-solving rather than capability theater.

Key highlights

• Explicit 2012–2025 timeline naming the bottleneck each breakthrough solved
• Six phased modules from neural-network prerequisites to production (RAG, Agents, vLLM, MLOps)
• Bilingual documentation with parallel Chinese and English READMEs
• End-to-end capstone projects framed as enterprise-grade systems
• Dependencies partitioned by phase, matching libraries to the era being studied

Verdict

Worth bookmarking if you are a Chinese- or English-speaking developer bootstrapping from CNN basics to LLM deployment. Skip it if you are looking for a single importable framework or novel research implementation.