Architecture foundations
Read AlexNet → ResNet → Transformer → Vision Transformer to trace convolution, residual connections, and attention as model backbones.
Curated reading path
20 papers that shaped modern AI
This is not a mechanical citation ranking. These papers changed model architectures, training methods, or the products people could build. Every paper is being explained in Chinese, English, and Japanese.
To understand modern AI systematically, begin with AlexNet, ResNet, and the Transformer for the architecture story. Continue through BERT, GPT-3, RAG, and InstructGPT for language models, then add diffusion, multimodal systems, and agents. The 20 papers below are chronological and each entry states the change it introduced.
Choose a reading path for your goal
Large language models
Read BERT → Scaling Laws → GPT-3 → RAG → InstructGPT → Chinchilla for pre-training, scale, retrieval, alignment, and compute-efficient training.
Generation and agents
Read VAE → GAN → DDPM → Latent Diffusion, then Chain-of-Thought and ReAct to connect generation, reasoning, and tool use.
- 01
ImageNet Classification with Deep Convolutional Neural Networks
ImageNet Classification with Deep Convolutional Neural Networks: It joined GPUs, large datasets, and deep convolutional networks into a scalable vision recipe.
Original - 02
Auto-Encoding Variational Bayes
Auto-Encoding Variational Bayes: It made probabilistic latent-variable models trainable with ordinary backpropagation.
Original - 03
Generative Adversarial Networks
Generative Adversarial Networks: It reframed generation as a game between a generator and a discriminator.
Original - 04
Deep Residual Learning for Image Recognition
Deep Residual Learning for Image Recognition: Residual connections made substantially deeper networks practical to optimize.
Original - 05
Attention Is All You Need
Attention Is All You Need: The Transformer replaced recurrence with attention and became the shared backbone of modern foundation models.
Original - 06
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding: It established bidirectional pre-training followed by task fine-tuning as a general NLP recipe.
Original - 07
Scaling Laws for Neural Language Models
Scaling Laws for Neural Language Models: It showed that model loss follows predictable power laws across scale, data, and compute.
Original - 08
Language Models are Few-Shot Learners
Language Models are Few-Shot Learners: It demonstrated broad few-shot task learning through prompting at unprecedented scale.
Original - 09
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks: It combined parametric memory with retrievable external knowledge, shaping modern knowledge assistants.
Original - 10
Denoising Diffusion Probabilistic Models
Denoising Diffusion Probabilistic Models: It turned iterative denoising into a stable route to high-quality generation.
Original - 11
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale: It showed that a pure Transformer over image patches could outperform convolutional networks at scale.
Original - 12
Learning Transferable Visual Models From Natural Language Supervision
Learning Transferable Visual Models From Natural Language Supervision: It aligned images and text at web scale, enabling strong zero-shot visual transfer.
Original - 13
LoRA: Low-Rank Adaptation of Large Language Models
LoRA: Low-Rank Adaptation of Large Language Models: It adapted large models through small low-rank updates instead of retraining every weight.
Original - 14
High-Resolution Image Synthesis with Latent Diffusion Models
High-Resolution Image Synthesis with Latent Diffusion Models: It moved diffusion into a compressed latent space, making high-resolution text-to-image generation practical.
Original - 15
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models: It showed that examples with intermediate steps can unlock much stronger reasoning in large models.
Original - 16
Training language models to follow instructions with human feedback
Training language models to follow instructions with human feedback: It established a practical recipe for aligning language models with human instructions and preferences.
Original - 17
Training Compute-Optimal Large Language Models
Training Compute-Optimal Large Language Models: It showed that many large models were under-trained and reset the compute-optimal balance of parameters and data.
Original - 18
ReAct: Synergizing Reasoning and Acting in Language Models
ReAct: Synergizing Reasoning and Acting in Language Models: It interleaved reasoning with tool actions, providing a direct blueprint for modern agent loops.
Original - 19
Segment Anything
Segment Anything: It turned segmentation into a promptable, transferable foundation capability backed by a large dataset.
Original - 20
Visual Instruction Tuning
Visual Instruction Tuning: It connected a vision encoder and a language model through visual instruction tuning to build a conversational assistant.
Original