Less jargon. More “oh, I get it.”
The words you keep bumping into, explained with a little breathing room.
Training builds a model; inference uses the finished model
Training changes a model by learning patterns from many examples and takes substantial computing work. Inference means running that trained model on a new request, which still uses computing power but does not repeat the full training process.
Read the sourceOpen weights, open source, and closed models differ
Open weights means people can download the learned model files, while open source may also include code, data, and broad reuse rights. Closed models keep key parts private and are usually accessed through an app or online service.
Read the sourceA benchmark measures selected tasks, not every real use
A benchmark tests models on a defined set of questions, tasks, prompts, and scoring rules. Its results help compare those conditions, but they do not guarantee the same order for different jobs, users, tools, or unseen data.
Read the sourceMore reasoning effort spends more work, without guaranteeing correctness
Reasoning effort tells a supported model how much work to spend before answering, which can change speed, cost, and results. A higher setting may help difficult tasks, but it does not prove intelligence or guarantee a correct answer.
Read the sourceAI is broad; machine learning and LLMs are narrower
Artificial intelligence is the broad field of making computers perform tasks associated with human intelligence. Machine learning builds systems from data, and a large language model is one kind of machine-learning system built mainly to work with language.
Read the sourceAI
Artificial intelligence: computer systems designed to perform tasks such as recognizing patterns, predicting outcomes, or generating content.
Machine learning
A way to build AI by learning patterns from examples instead of writing a rule for every situation.
LLM
A large language model: a model trained on language to predict and generate sequences of tokens.
Token
A small piece of text a language model processes. It can be a word, part of a word, or punctuation.
Transformer
A neural-network architecture that uses attention to relate different parts of an input.
Scaling laws
Observed relationships between training resources—such as data, model size and compute—and model performance. They are empirical patterns, not guarantees.
Inference
Running an already trained model to produce an answer or prediction.
Training
Adjusting a model’s internal numbers using data to improve its performance on a task.
Embedding
A list of numbers representing meaning or features. Similar ideas can end up close together in this numerical space.
Preprint
A research manuscript shared publicly before formal journal or conference peer review. Its claims need evaluation.
Hallucination
A generated answer that is false or unsupported, even when it sounds confident.
Agent
A system that uses a model to choose actions and tools in pursuit of a task. Its autonomy and reliability depend on its design.
Diffusion
A family of generative methods that learns to recover structure from noise, used for images and scientific modeling.
Alignment
Methods intended to make an AI system’s behavior better match human goals, instructions or values.
Benchmark
A defined set of tasks used to evaluate a system. A high score does not imply strong performance on every real-world task.
Open weights
Model parameters that are available to download. Licenses and access to training data or code can still vary.