The AI terms that
actually matter.
Clear, technical definitions of 113 key concepts — written by AI engineers, not marketers.
Canary-Qwen 2.5B
NVIDIA's Canary-Qwen 2.5B is a Speech-Augmented Language Model that holds the #1 spot on the Hugging Face Open ASR Leaderboard with a 5.63% average word error rate.
Read definitionCatastrophic Forgetting
Catastrophic forgetting is the tendency of a neural network to abruptly lose previously learned knowledge or capabilities when it is trained on new data, because the same shared weights that encoded the old task get overwritten by gradient updates for the new one. It is the central obstacle to continual learning and a real risk in every LLM fine-tuning or RLHF pass.
Read definitionChain-of-Thought Prompting
Chain-of-Thought (CoT) prompting is a technique that instructs a language model to externalise its reasoning as a sequence of intermediate steps before producing a final answer, dramatically improving accuracy on arithmetic, logic, and multi-step tasks.
Read definitionChunk Overlap
Chunk overlap is the practice of letting adjacent retrieval chunks share a window of tokens or characters at their boundary, so a sentence or fact split across a fixed-size cut still appears intact in at least one retrieved chunk. It's the standard, low-effort fix for boundary information loss in fixed-size and recursive text splitters.
Read definitionColBERT (Late Interaction Retrieval)
ColBERT is a retrieval architecture that encodes queries and documents independently into one vector per token, then scores a pair at search time with a cheap operator called MaxSim, keeping token-level precision without paying the cost of a cross-encoder, and forming the basis of the late-interaction retrieval family (ColBERTv2, PLAID, ColPali, ColQwen).
Read definitionComposable AI
Composable AI is an architectural paradigm that replaces monolithic, rigid AI systems with a modular ecosystem of interoperable, reusable components—such as specific models, agents, and data pipelines—that can be assembled and scaled independently.
Read definitionContext Rot
Context rot is the tendency for an LLM's reliability on a task to degrade as the input context grows longer, even well before the model's advertised context limit, and even when every piece of information the model needs is technically present in that context.
Read definitionContinual Learning
Continual learning (also called lifelong learning) trains a single model on a sequence of tasks or data distributions over time, aiming to acquire new knowledge without catastrophically forgetting what it already learned, unlike standard training which assumes a fixed, i.i.d. dataset seen all at once.
Read definitionCorrective RAG (CRAG)
Corrective RAG (CRAG) inserts a lightweight evaluator between retrieval and generation that scores how well the retrieved documents actually answer the query, then routes the pipeline down one of three corrective paths: refine and keep, discard and fall back to a live web search, or both at once. It was introduced in a January 2024 paper and is now a standard node in agentic RAG graphs built with LangGraph and similar frameworks.
Read definitionCross-Encoder vs. Bi-Encoder
A bi-encoder encodes a query and a document separately into vectors compared by a fast distance metric, while a cross-encoder feeds the query and document into the model together so it can attend across both; the tradeoff between the two, speed and scale versus accuracy, shapes almost every retrieval and reranking pipeline.
Read definitionDeepSeek V4
DeepSeek V4 is an open-source Mixture-of-Experts model family released in April 2026, with the flagship V4-Pro reaching 1.6 trillion total parameters and matching closed-source frontier models on coding and reasoning benchmarks at a fraction of the inference cost.
Read definitionDense Retrieval vs. Sparse Retrieval
Dense retrieval finds documents by comparing learned vector embeddings for semantic similarity, while sparse retrieval finds documents by matching weighted terms (like BM25) in an inverted index; modern search and RAG systems increasingly fuse both rather than picking one.
Read definitionStay ahead of the curve
Weekly newsletter on agentic AI, LLMs, and what we're building at Superteams — straight to your inbox.
Ready to ship AI in production?
We deploy fractional AI teams that deliver production-grade systems in 30–90 days. No fluff, no obligation.