Landmark AI research
Follow selected primary-source milestones from machine intelligence to modern generative and scientific systems.
Steps
8 steps-
Computing Machinery and Intelligence
Start with an early operational question about machine intelligence.
Alan Turing's 1950 paper asks whether machines can think and reframes the question through an imitation game. The paper also anticipates objections and discusses learning machines rather than defining intelligence through one internal mechanism.
-
The Perceptron
Meet an early learning model and its historical context.
Rosenblatt's perceptron paper presents a probabilistic model for information storage and organization. The perceptron became foundational to the history of learned linear decision rules and neural-network research.
-
Learning representations by back-propagating errors
See the training method that made multilayer learning practical.
The 1986 Nature paper demonstrates backpropagation for learning internal representations in multilayer networks. Its importance lies in efficiently assigning error signals through composed differentiable layers.
-
ImageNet Classification with Deep Convolutional Neural Networks
Connect large-scale data and accelerated computation to the deep-learning revival.
AlexNet reports a deep convolutional network trained for large-scale ImageNet classification. Its results helped demonstrate the combined value of learned representations, GPUs, large datasets, and regularization.
-
Attention Is All You Need
Read the architecture shift that established the Transformer.
The Transformer paper proposes an encoder-decoder architecture based on attention without recurrent or convolutional sequence layers. Its design enables greater training parallelism while modeling token-to-token interactions directly.
-
Language Models are Few-Shot Learners
Examine scale and in-context task behavior in a landmark language-model paper.
The GPT-3 paper studies autoregressive language-model scaling and task performance from instructions or examples in context. The work reports few-shot behavior without gradient updates for each evaluated task.
-
Denoising Diffusion Probabilistic Models
Add the diffusion lineage behind a major class of generative models.
The DDPM paper presents high-quality image synthesis through learned reversal of a diffusion process. Its formulation connects denoising score matching with a probabilistic latent-variable objective.
-
Highly accurate protein structure prediction with AlphaFold
Close with a scientific application and its distinct evidence standard.
The AlphaFold paper reports a neural system achieving high accuracy in the CASP14 protein-structure prediction assessment. The architecture incorporates evolutionary, geometric, and physical reasoning signals while outputting confidence estimates.