This glossary explains AI terms in an easy-to-understand way the DeepSeek R1.
A
Adapters – Small updates to an AI model to help it specialize in a specific task without needing a complete overhaul.
Example: Think of it as adding a plugin to your browser to get extra features.
Artificial Intelligence (AI) – Technology that enables machines to think and make decisions like humans.
Example: AI helps recommend your favorite Netflix shows based on your watching history.
B
BF16 (BFloat16) – A way of storing numbers in AI to keep things accurate while using less memory.
Example: It’s like using shorthand in notes to save space but still understanding the meaning.
C
Chain-of-Thought (CoT) – A method where AI explains its reasoning step-by-step before giving an answer.
Example: Instead of just answering “5,” the AI shows how it solved “2+3.”
Cold-Start Data – Basic information used to train an AI model when it has no prior knowledge.
Example: When a new employee joins a company, they first go through training before working.
D
DeepSeek-R1 – A powerful AI model designed to improve reasoning and problem-solving through learning from feedback.
Example: Like a chess player improving after every game by analyzing mistakes.
DeepSeek-R1-Zero – A version of DeepSeek-R1 that learns only through trial and error, without prior training data.
Example: A baby learning to walk without being taught, just by trial and error.
Distillation – Teaching a smaller AI model to be as smart as a larger one while using fewer resources.
Example: Summarizing a long book into a few pages without losing the key ideas.
F
FP8 (Float8) – A number format that sacrifices some accuracy to make AI faster and more efficient.
Example: Using an approximate recipe instead of measuring every ingredient perfectly.
Fine-Tuning – Customizing an AI model to improve its performance on specific tasks.
Example: Training an AI to recognize medical terms by feeding it medical textbooks.
I
Inference – The AI’s ability to generate responses based on what it has learned.
Example: Like a student answering questions on a test based on prior study.
Inference Providers – Online platforms that allow AI models to process information and provide answers.
Example: Think of them as a cloud service that lets you use AI without needing a powerful computer.
L
Large Language Model (LLM) – A type of AI trained on massive amounts of text to understand and generate human-like language.
Example: ChatGPT, DeepSeek-R1, and GPT-4 are all LLMs.
Language Mixing – When an AI unexpectedly blends multiple languages in its responses.
Example: Answering a question in English but suddenly switching to Spanish mid-sentence.
M
MMLU (Massive Multitask Language Understanding) – A test used to check how well AI understands various topics.
Example: Like an SAT exam for AI.
Monte Carlo Tree Search (MCTS) – A strategy AI uses to make the best decisions by exploring different options.
Example: Trying multiple routes in a video game to find the quickest way to win.
Q
Quantization – Shrinking an AI model to make it run faster and use less memory.
Example: Compressing a large video file so it takes up less space on your phone.
Qwen & Llama – Open-source AI models used to create versions of DeepSeek-R1 that run efficiently.
Example: Think of them as different car brands built on similar technology.
R
Reinforcement Learning (RL) – A way for AI to learn by trial and error using rewards and penalties.
Example: Like training a dog by giving treats when it follows commands.
Reward Modeling – Teaching AI which responses are good or bad by giving it feedback.
Example: A teacher marking answers right or wrong to guide a student’s learning.
S
Safetensors – A secure format for storing AI models to prevent hacking or corruption.
Example: Think of it as a tamper-proof file that keeps AI safe.
Self-Evolution – The ability of AI to improve itself without extra training data.
Example: Learning from past mistakes and improving automatically.
Supervised Fine-Tuning (SFT) – Training AI using labeled examples to improve accuracy.
Example: Showing an AI correct and incorrect answers so it learns faster.
T
Tensor Type (BF16-FP8-F32) – Different ways AI stores numbers to balance speed and accuracy.
Example: Using different battery modes on a phone – power-saving vs. high-performance.
Test-Time Scaling – AI improving its reasoning by taking extra steps during processing.
Example: Double-checking an answer before submitting it.
W
Win-Rate Benchmarks – Comparing AI models to see which performs best in competitions or real-world tasks.
Example: AI models competing in a knowledge quiz to see which one is the smartest.