Popular

Reinforcement Learning from Human Feedback — a training technique used to align language models with human preferences. Human raters compare model outputs and choose the better response; these preferences train a reward model which then guides further fine-tuning via reinforcement learning.

“RLHF is the key step that turns a raw language model into a helpful, harmless assistant.”

Describing AI systems capable of autonomous action, planning, and decision-making. An agentic AI can break down tasks, use tools, and work toward goals with minimal human intervention.

“The new release moves toward more agentic workflows where the AI can complete multi-step tasks independently.”

The ability to understand the internal state of a system from its external outputs — logs, metrics, and traces. A highly observable system lets engineers diagnose production issues, understand performance bottlenecks, and predict failures without needing to redeploy or add new instrumentation.

“Poor observability meant it took hours to find the root cause of the outage.”

A publicly available technical specification that anyone can implement, use, and extend without royalty obligations or proprietary restrictions. Open standards like HTTP, JSON, and OpenAPI enable interoperability between different vendors and communities, reducing lock-in and fostering innovation.

“MCP is an open standard, meaning any AI vendor can implement it to connect their models to tools.”

A set of practices, tools, and cultural philosophies that unite software development (Dev) and IT operations (Ops) teams. DevOps breaks down silos, automates repetitive tasks, and instills shared responsibility for the full software lifecycle from code to production monitoring.

“After adopting DevOps, their release cycle went from monthly to daily.”

The process of connecting an AI model's outputs to verified, real-world information sources. Grounding reduces hallucination by anchoring responses to retrieved documents, databases, or live data rather than relying purely on the model's learned parameters.

“Grounding the chatbot in our product database eliminated the fabricated feature claims.”

Safety constraints and filters applied to AI systems to prevent harmful, offensive, or out-of-scope outputs. Guardrails can be implemented at the model level (via training), prompt level (system instructions), or application level (output classifiers) to keep AI behavior within acceptable boundaries.

“The guardrails blocked the model from providing detailed instructions on dangerous activities.”

To excessively compliment or praise someone, often to the point of being insincere or sycophantic. Can also mean to admire someone with glazed-over eyes.

“Stop glazing that streamer so hard, they're not going to notice you in chat.”

The process of converting raw text into discrete units called tokens that a language model can process. Tokens are typically subword units — common words become single tokens while rare words split into multiple tokens. All LLM pricing and context limits are measured in tokens, not characters or words.

“The word "unbelievable" tokenized into three pieces: "un", "believ", "able".”

A technique for controlling the frequency of requests a client can make to an API or service within a given time window. Rate limiting protects systems from abuse, prevents overload, and ensures fair resource allocation among consumers. Responses typically include headers indicating current usage and remaining quota.

“The API returned a 429 Too Many Requests error once rate limiting kicked in at 100 calls per minute.”

A specialized database optimized for storing and querying high-dimensional vector embeddings. Vector databases power semantic search, recommendation systems, and RAG architectures by efficiently finding the most similar vectors to a given query.

“We stored all our documentation as embeddings in a vector database so the AI could find relevant passages instantly.”

A training methodology developed by Anthropic where a set of guiding principles (a "constitution") is used to self-supervise and refine AI outputs. The model critiques and rewrites its own responses according to the constitution, reducing the need for human labelers for harmful content.

“Constitutional AI lets the model identify and self-correct its own harmful outputs using defined principles.”

A computing paradigm that processes data at or near its source — at the "edge" of the network — rather than sending it all to a central cloud datacenter. Edge computing reduces latency, lowers bandwidth costs, and enables real-time processing for users around the globe.

“Serving the API from edge nodes cut response times from 200ms to 20ms for international users.”

An API that exposes data as a graph of interconnected nodes and edges, allowing clients to traverse relationships and fetch exactly the data they need in a single request. GraphQL is the most common implementation, replacing multiple REST endpoints with a flexible query language.

“The graph API let the client fetch a user, their posts, and each post's comments in one request instead of four.”

The coordination and sequencing of multiple AI agents, services, or steps in an automated workflow. An orchestrator determines which tools to invoke, in what order, and how to pass outputs between steps to complete a complex task end-to-end.

“The orchestration layer decided to call the search tool before invoking the summarization agent.”

A prompting approach where a small number of input-output examples are included in the context to guide model behavior on a new task. Few-shot prompting helps models understand the desired format, tone, or logic without any weight updates.

“We gave the model three few-shot examples of our data format and it immediately understood the pattern.”

Retrieval-Augmented Generation - An AI architecture pattern that combines a language model with external knowledge retrieval to provide more accurate and up-to-date responses.

“We implemented RAG to give our chatbot access to the latest product documentation.”

The ability of a model to perform a task it has never been explicitly trained or shown examples for. Zero-shot learning relies on the model's generalized understanding from pretraining to handle novel tasks based on instruction alone.

“The model classified customer sentiment zero-shot without any labeled training examples.”

The property of an operation where performing it multiple times produces the same result as performing it once. Idempotent API endpoints are critical for safe retries — if a network error occurs, the client can re-send the request without fear of duplicating side effects like charges or database records.

“Pass an idempotency key with payment requests so retries don't charge the customer twice.”

Artificially generated data that mimics the statistical properties of real-world data, used for training or testing AI models. Synthetic data can be created by generative models, rule-based systems, or simulations, and is especially valuable when real data is scarce, sensitive, or expensive to collect.

“We generated synthetic medical records to train the model without risking patient privacy.”