LLM Flashcards — 330+ Visual Cards on Transformers, Attention, RAG & Agents | LLMs Research
LLMs Research Lab
Visual LLM reference

LLM Flashcards: learn how language models actually work.

A set of 330+ cards covering the full LLM stack, from tokenization and attention through RAG, agents, and inference. A clean diagram, a short explanation, and nothing you have to wade through. Built by a working LLM research lab, not a content farm.

330+visual cards
22topics
3formats included

Get the cards → See what’s inside

One-time purchase, lifetime updates included. Rated 5.0 out of 5 across 45 ratings, 96% five-star.

LLM Flashcards cover showing visual cards on self-attention, query-key-value vectors, and the KV cache.

A few sample cards

Every card is self-contained: a diagram you could redraw on a whiteboard, plus a few sentences on what it means and why it matters. Scroll to browse →

Visual flashcard explaining the transformer architecture: input embedding through self-attention, add and norm, feed-forward, and add and norm, with residual skip connections, repeated N times.
What is a Transformer?
Visual flashcard explaining Retrieval-Augmented Generation: a query retrieving relevant document chunks from a vector store, added to the prompt before the model answers.
What is RAG?
Visual flashcard explaining RoPE, rotary position embedding, which encodes token position by rotating query and key vectors.
RoPE position embeddings
Visual flashcard explaining the KV cache: storing key and value tensors during autoregressive generation so they are not recomputed each step.
KV cache at inference
Visual flashcard explaining the ReAct agent framework: a loop of thought, action, and observation that lets a model use tools across multiple steps.
The ReAct agent loop
Visual flashcard explaining Mixture of Experts: a router sending each token to a small subset of expert feed-forward networks.
Mixture of Experts
Visual flashcard explaining Chain-of-Thought prompting: asking a model to reason step by step before answering, improving accuracy on multi-step problems.
Chain-of-Thought
Visual flashcard explaining Byte Pair Encoding tokenization: merging the most frequent character pairs into subword tokens.
Byte Pair Encoding
Visual flashcard explaining quantization: representing model weights with fewer bits to shrink memory and speed up inference.
What is quantization?
Visual flashcard explaining hallucination in language models: confident but factually wrong output, and why it happens.
Hallucination

What’s inside

330+ cards across 22 topics, ordered so each one builds on the last. Read straight through to build a mental model from first principles, or jump to the topic you need.

  1. I.Tokenization. BPE, byte-level, vocabularies, special tokens, fertility. 12 cards
  2. II.Embeddings & retrieval. Vectors, similarity, vector search, BM25, ColBERT, rerankers. 14 cards
  3. III.Transformer architecture. Attention, QKV, positional encoding, RoPE, the full block. 30 cards
  4. IV.Model architecture variants. MoE, MLA, Mamba, linear attention, SwiGLU, RMSNorm. 16 cards
  5. V.Training fundamentals. Objectives, loss, optimizers, activations, backprop, grokking. 18 cards
  6. VI.Distributed training. Data, tensor, and pipeline parallelism, ZeRO, FSDP. 10 cards
  7. VII.Scaling laws. Chinchilla, compute-optimal training, emergence, test-time scaling. 10 cards
  8. VIII.Fine-tuning. SFT, LoRA, QLoRA, DoRA, adapters, model merging. 15 cards
  9. IX.RLHF & alignment. Reward models, PPO, DPO, GRPO, RLVR, reward hacking. 19 cards
  10. X.Inference & decoding. KV cache, sampling, speculative decoding, batching, caching. 19 cards
  11. XI.Quantization & efficiency. INT8/INT4, GPTQ, AWQ, BitNet, GGUF, distillation. 12 cards
  12. XII.Prompting. Few-shot, chain-of-thought, tree of thoughts, prompt optimization. 19 cards
  13. XIII.Reasoning. Reasoning models, long CoT, verifier search, MCTS, tool use. 15 cards
  14. XIV.Context management. Lost in the middle, compression, YaRN, KV eviction, context rot. 10 cards
  15. XV.Retrieval-augmented generation. Chunking, HyDE, GraphRAG, Self-RAG, agentic RAG. 24 cards
  16. XVI.Agents & tools. Function calling, MCP, computer use, orchestration, agent eval. 22 cards
  17. XVII.Multimodal. Vision transformers, VLMs, fusion, speech, diffusion, documents. 8 cards
  18. XVIII.Advanced concepts. Multimodal LLMs, CLIP, synthetic data, continual pretraining. 6 cards
  19. XIX.Evaluation & benchmarks. Perplexity, MMLU, GPQA, SWE-bench, needle-in-a-haystack. 16 cards
  20. XX.Safety & ethics. Hallucination, bias, watermarking, unlearning, oversight. 17 cards
  21. XXI.Interpretability. Mechanistic interp, probing, logit lens, induction heads, SAEs. 7 cards
  22. XXII.APIs & practical. Chat completion, structured outputs, batch, embeddings, routing. 13 cards

Who it’s for

  • Engineers working with LLMs who want a clean visual reference to keep open while reading papers or model cards.
  • Anyone preparing for an AI or ML engineering interview. The cards map to the concepts that actually come up: transformer internals, attention variants, KV cache, RAG, fine-tuning, inference trade-offs. Revise them with the Anki set in the weeks before.
  • Students in NLP or deep-learning courses who think better through diagrams than dense paragraphs.
  • Self-taught learners who have used an LLM API and want to understand what is actually happening underneath.

How to use it

  • Read straight through, topic by topic, to build a foundation.
  • Import the Anki set and review on your commute with spaced repetition.
  • Print four cards per page for physical study, or a single card full size as a poster.
  • Keep the PDF open as a reference while reading papers or model cards.

Questions

What formats do I get?

Three. A multi-page PDF organized by topic, high-resolution enough to print four cards per page, an Anki set (.apkg) for spaced-repetition review on desktop or mobile, and a zip of every card as a separate image.

Are these for beginners or experts?

Both, but most useful in the middle. If you have used an LLM API and want to understand what is happening underneath, this is built for you. The diagrams are approachable, but the technical depth assumes some ML background.

Is it good for interview prep?

Yes. The cards cover the concepts that come up in LLM and ML engineering interviews, and the Anki set makes them easy to revise with spaced repetition in the weeks before.

How often does the set update?

New cards are added as new techniques and research land. Past buyers get every update free, with no expiry and no resubscription.

Who made it?

These cards started as study notes inside LLMs Research, an independent applied research lab publishing on KV cache compression, adaptive compute, and multi-agent systems. They grow alongside the research work. You can read more about the lab.

What buyers say

5.0 ★★★★★ 45 ratings on Gumroad · 96% five-star

Quotes from verified Gumroad ratings, unedited.

Get the cards →

Free PDF

Not getting the full set today?

Subscribe to the LLMs Research newsletter and we’ll send you 30 of the visual cards as a free PDF.

Free. Unsubscribe anytime.

One last step

We opened Substack in a new tab with your email filled in. Click Subscribe there (the free plan is all you need) and your PDF arrives by email. Not there in a minute? Check spam.