Concepts
23 entries
| Entry | Summary | As of |
|---|---|---|
| AI agentsconcept | An AI agent combines a model with state, tools, and a control loop that selects actions toward a goal. Autonomy is a system property shaped by permissions, environment, retries, observation, and stop conditions—not only the base model. | 2026-07-13 |
| AI guardrailsconcept | Guardrails are layered controls intended to reduce unacceptable inputs, outputs, or actions in an AI system. Risk management also requires governance, measurement, monitoring, incident response, and human accountability. | 2026-07-13 |
| Artificial intelligenceconcept | Artificial intelligence is an umbrella term for computational systems performing tasks associated with intelligent behavior. The field predates modern neural networks and has repeatedly changed its dominant methods and evaluation questions. | 2026-07-13 |
| Context windowconcept | A context window bounds the tokens a model can directly condition on for a request. A larger advertised window does not guarantee equal recall, reasoning quality, latency, or price across its full length. | 2026-07-13 |
| Deep learningconcept | Deep learning uses layered parameterized functions to learn representations at multiple levels of abstraction. Depth is an architectural property; useful performance still depends on training data, objectives, optimization, and compute. | 2026-07-13 |
| Direct Preference Optimizationconcept | Direct Preference Optimization trains from preference pairs without fitting a separate explicit reward model. Its objective relates preference likelihood to a reference policy through a classification-style loss. | 2026-07-13 |
| Embeddingsconcept | An embedding maps discrete or complex objects into vectors whose geometry can support learned comparisons. Distance is meaningful only relative to a model, training objective, normalization choice, and downstream task. | 2026-07-13 |
| Fine-tuningconcept | Fine-tuning updates pretrained model parameters using task- or domain-specific examples. It trades additional training and maintenance for behavior that may be harder to achieve reliably through prompting alone. | 2026-07-13 |
| Foundation modelconcept | A foundation model is trained broadly and then adapted or prompted for many downstream uses. Broad reuse creates leverage but also propagates model-level limitations into multiple applications. | 2026-07-13 |
| Knowledge distillationconcept | Knowledge distillation trains a student model using targets or distributions produced by a larger teacher. The goal is behavior transfer, not a literal compression that preserves every capability and failure mode. | 2026-07-13 |
| Large language modelconcept | A large language model estimates distributions over token sequences using a high-capacity neural network. Fluent continuation is not the same operation as consulting a fact database, even when outputs appear knowledgeable. | 2026-07-13 |
| Machine learningconcept | Machine learning builds systems whose behavior is shaped by data and an optimization objective. Learning performance depends on data, representation, objective, optimization, and evaluation—not the algorithm name alone. | 2026-07-13 |
| Model evaluationconcept | Model evaluation measures defined behavior on specified tasks, data, prompts, and scoring rules. A benchmark score is evidence for that setup, not a universal ranking across cost, safety, latency, and real workloads. | 2026-07-13 |
| Model hallucinationfailure mode | Hallucination describes generated content that is unsupported or false despite being presented fluently. Retrieval, citations, constraints, and verification can reduce risk, but none makes every generated claim trustworthy by default. | 2026-07-13 |
| Model quantizationconcept | Quantization represents model values or computations with lower numerical precision. It can reduce memory and increase throughput, but accuracy and hardware behavior depend on the chosen quantization method. | 2026-07-13 |
| Model tool useconcept | Tool use lets a model request structured external operations such as search, computation, or API calls. The application remains responsible for authorization, schema validation, execution, and handling untrusted tool output. | 2026-07-13 |
| Positional encodingconcept | Positional information is added because basic attention does not inherently encode sequence order. The original Transformer used sinusoidal encodings, while later systems may use learned or rotary alternatives. | 2026-07-13 |
| Promptingconcept | Prompting steers a model through instructions, examples, context, and output constraints supplied at inference time. Prompt sensitivity means a useful evaluation records the exact prompt and settings rather than only a model name. | 2026-07-13 |
| Reinforcement learning from human feedbackconcept | RLHF uses human preference evidence to train a reward signal and optimize model behavior. The InstructGPT pipeline combines supervised demonstrations, preference comparisons, reward modeling, and policy optimization. | 2026-07-13 |
| Retrieval-augmented generationconcept | Retrieval-augmented generation conditions a generator on documents selected from an external corpus. Retrieval can improve provenance and freshness, but poor retrieval or unsupported synthesis can still produce incorrect answers. | 2026-07-13 |
| Self-attentionconcept | Self-attention computes content-dependent interactions among positions in the same sequence. Queries and keys determine weights while values carry the information combined at each position. | 2026-07-13 |
| Tokenizationconcept | Tokenization maps text into discrete units consumed and produced by a language model. Token boundaries affect sequence length, cost, multilingual behavior, and how apparently simple strings are represented. | 2026-07-13 |
| Transformer architectureconcept | The Transformer replaces recurrent sequence processing with attention and position-wise feed-forward computation. Parallel training and flexible token interactions helped the architecture become a foundation for language and multimodal systems. | 2026-07-13 |