Large Language Models Trivia Questions
- ❓ 102+ questions
- 🗂️ AI
- 🎚️ Easy to expert
- ✨ Free to play
massive neural networks for text. Play Large Language Models trivia solo to sharpen your knowledge, or challenge a friend head-to-head in Trivia Tango — every question comes with an explanation so you learn as you play. Questions span every level, from easy warm-ups to expert-level stumpers, so there's a real challenge here however much you already know.
Already playing?
Sample Large Language Models Quiz Questions
A mix of easy, medium and hard — questions run from warm-up to expert, so there's a real challenge at every level. Think you know the answers? Play to find out.
The "T" in GPT stands for this neural network architecture introduced in 2017 that revolutionized natural language processing through self-attention mechanisms.
Difficulty: Easy- Tensor
- Transformer
- Transfer
- Token
This attention variant adds linear bias terms based on position distance instead of learned embeddings, enabling extrapolation to longer sequences without additional training.
Difficulty: Medium- ALiBi
- RoPE
- Sinusoidal encoding
- NoPE
This initialization scheme scales weights by 1/√d to prevent output variance from exploding with layer depth, standard for attention projections.
Difficulty: Hard- Xavier/Glorot initialization
- He initialization
- Orthogonal initialization
- Lecun initialization
When you type a message to an AI assistant, each word or word-piece is converted into one of these basic units that the model actually processes.
Difficulty: Easy- Embeddings
- Neurons
- Parameters
- Tokens
This OpenAI paper introduced in-context learning by demonstrating that GPT-3's 175 billion parameters enabled task performance through examples alone, without gradient updates.
Difficulty: Medium- Scaling Laws paper
- Language Models are Few-Shot Learners
- Training Language Models paper
- GPT-4 Technical Report
This research paper demonstrated that attention patterns in middle layers often attend strongly to the first token regardless of content, acting as a "no-op" attention pattern.
Difficulty: Hard- Induction heads paper
- Attention sink paper
- Grokking paper
- Scaling monosemanticity paper
This technique uses human ratings of AI responses to train a reward model, which then guides the AI to produce more helpful and safer outputs.
Difficulty: Easy- RLHF
- RAG
- LoRA
- PEFT
This technique used in RLHF training constrains how far the updated policy can deviate from the reference policy at each step, preventing catastrophic forgetting of pre-trained capabilities.
Difficulty: Medium- Reward modeling
- Value clipping
- KL penalty
- Entropy bonus
Think You Know Large Language Models?
Play a free round in your browser, learn something new with every answer, and challenge your friends to beat your score.
Or get the app