Landmark AI Papers
All 14 cards in this deck
One-line recall for the papers everyone cites — from Attention to R1. Answer each question before you open it — then run the review above, which schedules every card with spaced repetition so it comes back right before you would forget it.
“Attention Is All You Need” (2017) — the contribution?
The Transformer: replace recurrence with self-attention so sequences process in parallel — the architecture under every modern LLM.
GPT-3 (2020) — the headline idea?
In-context learning at scale: a 175B model does new tasks from a few examples in the prompt, no fine-tuning.
InstructGPT (2022) — the recipe?
RLHF: supervised fine-tune, train a reward model on human preferences, optimize the policy with PPO — how raw LMs became assistants.
Constitutional AI (2022) — the replacement for human harm labels?
A written constitution: the model critiques and revises itself against principles, then RL runs on AI feedback (RLAIF).
Chain-of-Thought (2022) — the trick?
Prompting the model to show intermediate reasoning steps dramatically improves multi-step problem solving in large models.
Self-Consistency (2022) — the upgrade to CoT?
Sample many diverse reasoning paths, majority-vote the final answer — correct answers converge, scattered errors disagree (+17.9 GSM8K).
Tree of Thoughts (2023) — beyond chains?
Search over intermediate “thoughts”: propose, self-evaluate, expand with BFS/DFS, backtrack — Game of 24 went 4%→74%.
LoRA (2021) / QLoRA (2023) — the fine-tuning unlock?
Train tiny low-rank adapters on a frozen base (LoRA); keep that base in 4-bit NF4 (QLoRA) — 65B fine-tuned on one GPU.
DPO (2023) — what did it remove from RLHF?
The reward model and RL loop: optimize directly on preference pairs with a closed-form loss targeting the same objective.
GRPO / DeepSeekMath (2024) — the critic-free move?
Sample a group of answers, use the group’s mean reward as the baseline — PPO’s signal at half the memory, the engine of R1.
DeepSeek-R1 (2025) — what emerged from reward alone?
Chain-of-thought reasoning and self-correction (“wait — let me rethink”) from pure RL on answer correctness, no supervised reasoning data.
MemGPT (2023) — the analogy?
Context window as RAM, external storage as disk, the LLM as its own memory manager — paging facts via function calls.
Titans (2025) — memory where?
In weights, at test time: a neural memory module keeps training during inference, writing what surprises it — 2M-token recall.
DeepSeek-V3 (2024) — the cost shock?
Frontier-class 671B MoE (37B active) trained for ≈$5.6M via MLA, FP8, MTP and aux-loss-free balancing.