The Story of
Modern AI
From AlexNet to agents — 51 papers across 14 years. Tap any year to go deeper. See what was stuck, what broke through, and why it matters now.
Chapter
The Reasoning Era
Thinking at inference time unlocks a new frontier. Reasoning is the new scaling.
Chapter
The Proliferation
AI leaves the lab and enters everything. Open source, agents, deployment.
Tool use, RAG, and multi-step reasoning turn AI from oracle into collaborator. SWE-bench exposes where models still fail at real engineering. Llama 3 and Phi-3 prove you don't need 70B parameters to be useful. The era of deployment begins.
LLaMA leaks, then Llama 2 ships openly. GPT-4 raises the frontier ceiling. Mixtral shows sparsity beats scale. The gap between closed frontier models and what a developer can run locally narrows from years to months. Agents become real with AutoGen, Voyager, Reflexion.
Chapter
The Alignment Era
Power without control is dangerous. The field turns inward.
InstructGPT shows RLHF can make models follow human intent instead of just predicting text. Constitutional AI offers a path to self-correction. ChatGPT launches in November and reaches 100M users in 60 days. Every AI lab in the world pivots. Alignment stops being academic.
Chapter
The Scale Hypothesis
Bigger models don't just improve — they unlock abilities nobody planned for.
CLIP links images and text in a shared embedding space. DALL-E generates images from text descriptions. The multimodal era begins quietly. LoRA shows that you can adapt massive models with a fraction of the compute. The future of efficient AI starts here.
GPT-3 has 175B parameters and can write code, answer questions, and draft essays with a few examples and no fine-tuning. The "few-shot learning" paper changes what PMs think AI products can do. Scaling laws show this is systematic, not accidental.
Chapter
The Transformer
One architecture replaces everything. Attention is all you need.
GPT-2 generates text so coherent that OpenAI initially withholds it — worried it'll be misused. The decision is later reversed, but the moment lingers. Scale is clearly the variable that matters. Nobody yet knows how far it goes.
No papers catalogued for this year yet.
BERT proves that unsupervised pretraining on unlabeled text, then fine-tuning on small labeled datasets, beats everything built task-by-task. The "train once, adapt everywhere" model replaces 10 years of task-specific engineering.
"Attention Is All You Need" ships in June. No recurrence. No convolution. Just attention. Every major model you use today — GPT-4, Claude, Gemini — runs on this architecture. This is the moment the modern AI era begins.
Chapter
The Foundation
Deep learning proves itself. The field learns to see, hear, and generate.
RNNs and LSTMs hit a wall with long contexts. The field knows something needs to change architecturally. AlphaGo beats Lee Sedol. Everyone outside the lab starts paying attention. Inside the lab: quiet urgency.
No papers catalogued for this year yet.
ResNet solves the vanishing gradient problem with skip connections. 152-layer networks become trainable. Depth = power becomes the orthodoxy. LSTM-based models rule sequences. Everything is getting better — just slowly.
No papers catalogued for this year yet.
GANs make AI creative for the first time. Seq2Seq teaches models to translate whole sentences, not just words. For the first time, AI goes from classifying the world to producing new parts of it.
No papers catalogued for this year yet.
Word2Vec gives words coordinates in vector space. "King − Man + Woman = Queen" isn't a trick — it's the beginning of the idea that meaning can be computed. The field starts to believe that language is just another pattern.
No papers catalogued for this year yet.
AlexNet halves the ImageNet error rate overnight. The AI winter thaws. Researchers who spent decades on hand-crafted features suddenly have to reckon with the fact that raw compute + data + depth beats everything.
No papers catalogued for this year yet.
What's next
You've seen the history.
Now read the papers that made it. Deep dives, simulators, and action plans for every breakthrough.
Explore the library