Glossary

Plain-language definitions for the terms used across the writing.

The same definitions that surface as tooltips in the articles, gathered here for browsing. Aimed at a reader working in or near data and AI, not necessarily deep in it. 453 terms across 8 categories.

Products & Vendors 214

ADLS also Azure Data Lake Storage, ADLS Gen2: Azure Data Lake Storage — Microsoft's object storage optimised for analytic workloads. The Azure-side counterpart to AWS S3 and Google Cloud Storage, and the layer Microsoft Fabric and Synapse build on.
Agentforce: Salesforce's agent platform — a runtime and authoring environment for building autonomous AI agents inside the Salesforce CRM and Customer 360 estate. salesforce.com ↗
AI Credits also AI Credit, Copilot AI Credits: GitHub Copilot's usage-based billing unit from June 2026 (1 credit = $0.01), drawn down by token consumption at each model's listed API rate. Replaced the older fixed "premium request" allowance with pay-for-what-you-use metering.
AI/BI Genie also Genie: Databricks' natural-language analytics surface — users ask questions in plain English, the assistant grounds the answer in Unity Catalog metric views and the underlying tables. Lives inside the Databricks workspace and inherits the catalog's permissions. docs.databricks.com ↗
AI21 also AI21 Labs: An Israeli AI lab focused on enterprise. Builds the Jurassic and Jamba model families and is one of the smaller frontier-adjacent vendors with serious enterprise traction. ai21.com ↗
Aider: A terminal-based open-source coding agent that pairs with git — edits files, runs tests, commits its own changes. The minimalist counterpart to heavier IDE-embedded agents. aider.chat ↗
Airbyte: The open-source counterpart to Fivetran — same job (extract from sources, load into warehouse), but with self-hostable infrastructure and a community-built connector catalog. airbyte.com ↗
Airflow also Apache Airflow: The dominant open-source workflow orchestrator for data pipelines. You author DAGs in Python; Airflow schedules them, handles retries, and surfaces failures. Increasingly competing with newer tools (Dagster, Prefect). airflow.apache.org ↗
Alation: A data catalog focused on discovery and trust — surfaces what data exists, who uses it, what queries hit it most, and how reliable it is. Competes with Collibra and Atlan. alation.com ↗
Anthropic: The AI safety lab behind Claude. Founded in 2021 by former OpenAI researchers; backed by Google and Amazon. Focused on building steerable, interpretable, and safer models. anthropic.com ↗
Anysphere: The company behind the Cursor editor. Grew from a research project to one of the fastest-revenue-ramping AI startups by 2025 — a textbook case of distribution through a familiar interface.
Arize also Arize AI: An ML and LLM observability platform — production monitoring, evals, and the open-source Phoenix toolkit for tracing agent runs locally. arize.com ↗
Atlan: A data catalog and collaboration platform — discover tables, see lineage, write context, manage access requests. Aimed at making the warehouse a navigable shared resource rather than a black box. atlan.com ↗
AtScale: A semantic-layer platform that sits between BI tools and warehouses, exposing a governed model over OLAP-style queries. Had promoted its Semantic Modeling Language as a rival before folding the work into the Open Semantic Interchange initiative in early 2026. atscale.com ↗
Auth0: A developer-focused identity-as-a-service product. Sign-in, social logins, MFA, all behind a small SDK. Acquired by Okta in 2021 and now sold alongside Okta's enterprise IAM. auth0.com ↗
AWS Glue: AWS's managed ETL and data-catalog service. Glue Catalog is also the de-facto Hive-Metastore-compatible catalog most AWS lakehouse tooling reads from when nothing more modern is in place. aws.amazon.com ↗
Azure OpenAI also Azure OpenAI Service: Microsoft's enterprise wrapper around OpenAI's models — same models as ChatGPT, but inside an Azure subscription with the usual enterprise plumbing (private networking, data residency, RBAC).
BankID: The Norwegian national digital-identity system, run by the banks and used to log in to almost every public service and financial institution in Norway. The Norwegian equivalent of Sweden's BankID and Denmark's MitID. bankid.no ↗
Bedrock also Amazon Bedrock, AWS Bedrock: AWS's managed access to foundation models from Anthropic, Meta, Mistral, AI21, Cohere, and Amazon itself. The default way to consume Claude on AWS without leaving the account. aws.amazon.com ↗
BigQuery: Google's serverless cloud data warehouse. You don't manage compute; you write SQL and pay per byte scanned. Pioneer of the storage/compute separation that Snowflake and Redshift now also implement. cloud.google.com ↗
Black Forest Labs also BFL: A German AI lab founded by former Stable Diffusion researchers. Best known for the Flux image-generation model family, considered state of the art among open-weight image models. blackforestlabs.ai ↗
Bolt also bolt.new, StackBlitz Bolt: A browser-based AI app builder from StackBlitz — describe a project and Bolt generates, runs, and lets you edit it without leaving the tab. bolt.new ↗
Business Objects also BusinessObjects: An enterprise BI suite acquired by SAP in 2007. Its 1990 "universe" — a governed map of business terms over a relational database — is probably the first widely-deployed semantic layer; everyone has been rebuilding versions of it ever since.
Cassandra also Apache Cassandra: A distributed NoSQL database designed for huge scale and high availability across data centres. Originally built at Facebook, now operates much of Netflix, Apple, and Instagram's hot data paths. cassandra.apache.org ↗
Census also Census Data Activation: A reverse-ETL platform competing with Hightouch. Same shape — define an audience in your warehouse, sync it into the SaaS tools that need it. Acquired by Fivetran in May 2025. getcensus.com ↗
Cerebras: A US chip company that builds wafer-scale AI processors — single chips the size of a dinner plate. Targets the largest training and inference workloads where the math benefits from staying on one piece of silicon. cerebras.ai ↗
ChatGPT: OpenAI's consumer-facing chat product. Launched November 2022 and broke into the mainstream — for many users, "AI" still means ChatGPT. chatgpt.com ↗
Chroma also ChromaDB: An open-source, embeddable vector database — runs in-process during prototyping and scales to a server when you need it. Common starting point for RAG projects. trychroma.com ↗
Claude: Anthropic's family of language models, accessed via claude.ai, an API, or coding-agent products like Claude Code. Strong on reasoning, long context, and safety behaviour. claude.ai ↗
Claude Code: Anthropic's terminal-based coding agent. Runs locally, reads and edits your codebase, calls tools, and is itself the harness behind much of the agentic-engineering work on this site. docs.claude.com ↗
Clerk: A developer-first authentication and user-management platform. Designed around modern frontend frameworks (Next.js, Remix, Astro) with batteries-included sign-in components. clerk.com ↗
ClickHouse also ClickHouse Cloud: An open-source columnar database optimised for blistering analytics on huge tables. Originally built at Yandex; now the engine behind real-time-analytics platforms across the industry, with a managed cloud offering (ClickHouse Cloud) from the founding team. clickhouse.com ↗
Cline also Claude Dev: An open-source autonomous coding agent that runs as a VS Code extension. Originally built around Claude; now multi-model. github.com ↗
Cloud Run: Google Cloud's serverless container service — push a container image, Cloud Run runs and scales it on demand. The middle ground between Lambda and a Kubernetes cluster. cloud.google.com ↗
Cloudflare Workers also Workers: Cloudflare's serverless edge runtime. Code runs in V8 isolates at every Cloudflare edge location, with cold starts measured in milliseconds rather than seconds. workers.cloudflare.com ↗
Codex also OpenAI Codex: OpenAI's coding agent — first as a 2021 model, then as a 2025 product that runs in the cloud and on the desktop. Now reaches into IDEs, the terminal, and from your phone.
Cognition also Cognition AI, Cognition Labs: The AI lab behind Devin, the autonomous coding agent. Acquired Windsurf in 2025, giving it both the cloud-agent and IDE-agent ends of the market. cognition.ai ↗
Cohere: A Canadian AI lab focused on enterprise. Known for high-quality embedding models and a deliberate B2B posture rather than chasing the consumer chatbot market. cohere.com ↗
Collibra: An enterprise data-governance and catalog platform — the heavyweight option for organisations with serious compliance and stewardship requirements. collibra.com ↗
Confluent: The commercial company founded by Kafka's creators. Sells a managed Kafka cloud plus the surrounding stream-processing and governance tooling. confluent.io ↗
Continue: An open-source IDE extension (VS Code and JetBrains) that lets you bring any LLM into your editor for chat, autocomplete, and agentic edits. Vendor-neutral alternative to Copilot/Cursor. continue.dev ↗
Copilot also Microsoft Copilot, M365 Copilot: Microsoft's AI assistant surface across Microsoft 365, Windows, GitHub, and developer tools. Powered by OpenAI and Microsoft's own models behind the same brand.
Cortex Analyst: Snowflake's natural-language-to-SQL service — answers business questions by reading a YAML semantic model that pins definitions, joins, and metrics, then generates SQL against the underlying tables. snowflake.com ↗
Cosmos DB also Azure Cosmos DB: Microsoft's globally distributed multi-model database. NoSQL with multiple query APIs (SQL-like, MongoDB, Cassandra), tuneable consistency, and a price tag to match the flexibility. learn.microsoft.com ↗
CRM also Customer Relationship Management: Customer Relationship Management — software for managing the lifecycle of sales leads, accounts, contacts, and customer interactions. Salesforce, HubSpot, and Dynamics 365 own most of the market.
Cube: A headless-BI platform that defines metrics once and exposes them through SQL, REST, GraphQL, and an AI endpoint. Open-source core with a managed cloud; competes with dbt Semantic Layer and AtScale on the standalone-layer side of the argument. cube.dev ↗
Cursor: An AI-first code editor (a fork of VS Code) with model-agnostic chat, completions, and an in-editor agent. Was the first widely-used editor that felt purpose-built for working with LLMs. cursor.com ↗
Dagster: A modern data orchestrator that treats data assets (tables, files, models) as first-class citizens. Strong typing, lineage, and asset-first scheduling differentiate it from Airflow. dagster.io ↗
DALL-E also DALL·E, DALL-E 3, DALLE: OpenAI's image-generation model family, named after Salvador Dalí. Now mostly accessed through ChatGPT rather than as a standalone product.
Databricks: The lakehouse vendor built around Delta Lake and Spark. Strong AI/ML tooling (Mosaic, MLflow), Unity Catalog for governance, and a growing footprint outside analytics into OLTP and apps. databricks.com ↗
Datadog: The dominant SaaS observability platform — metrics, traces, logs, and a sprawling product line beyond. Common default in mid-to-large engineering organisations. datadoghq.com ↗
dbt also data build tool: A SQL-centric transformation tool that lets analytics engineers build modular, version-controlled models inside the warehouse. Defines the "T" in ELT for most modern data teams. getdbt.com ↗
DeepSeek: A Chinese AI lab that released competitive open-weight reasoning models in late 2024 / early 2025, briefly shaking the assumption that frontier capability required a US-scale training budget. deepseek.com ↗
Devin: Cognition AI's autonomous software-engineering agent — runs on a remote machine, plans its own tasks, executes them, and reports back. One of the first products to market itself as a "junior engineer." devin.ai ↗
Direct Lake: Microsoft Fabric's mode where Power BI queries Delta-Parquet tables in OneLake directly, without importing them into a tabular model. Combines the freshness of DirectQuery with the speed of in-memory analytics.
Direct Query: Qlik Cloud's mode for querying a source database live with SQL pushdown as users select and filter, rather than loading the data into the in-memory engine. Available for a fixed set of cloud databases. help.qlik.com ↗
Discovery Agent: Qlik's monitoring agent — watches metrics in governed apps for anomalies and surfaces a notification when something drifts outside expectations. Pairs with Qlik Answers as the proactive half of Qlik's agentic experience.
dlt also data load tool: An open-source Python library for writing your own data-loading pipelines. The "do it yourself" middle ground between Fivetran (managed) and a custom script. dlthub.com ↗
DNB: Norway's largest financial-services group. The reference customer for serious Norwegian data and analytics work; many of the country's mid-career data professionals have worked on a DNB project at some point. dnb.no ↗
Docker: The tool that made containers mainstream — a standard format for packaging an application and a runtime to launch it. Now mostly a baseline assumption rather than a differentiator. docker.com ↗
Dremio: A lakehouse query and analytics engine built around Apache Iceberg and Arrow. Positions itself as a vendor-neutral SQL layer over open table formats, competing with Snowflake and Databricks at the SQL-on-lake job. dremio.com ↗
Druid also Apache Druid: An open-source real-time analytics database built for sub-second OLAP queries over event streams. Common in observability and ad-tech where latency and concurrency dominate. druid.apache.org ↗
DuckDB: An in-process analytical database — the "SQLite for analytics." Runs inside your script, no server, but executes warehouse-style queries on millions of rows in seconds. Quietly reshaping where small/medium analytics belongs. duckdb.org ↗
Dynamics 365 also Dynamics: Microsoft's business-applications suite — CRM, ERP, field service, finance, supply chain. The default option in Microsoft-aligned enterprises; competes head-on with Salesforce and SAP. dynamics.microsoft.com ↗
DynamoDB also Amazon DynamoDB: AWS's managed key-value and document database. Serverless, single-digit-millisecond latency, scales horizontally without operator intervention. The Amazon-shop default for high-throughput operational data. aws.amazon.com ↗
EC2 also Amazon EC2: AWS's basic compute service — rentable virtual machines by the hour or second. The foundational primitive most other AWS services build on. aws.amazon.com ↗
Elasticsearch: A distributed search and analytics engine. Powers full-text search, log analytics, and observability backends across most of the industry. elastic.co ↗
ElevenLabs: A voice-AI company. Generates synthetic speech and clones voices at quality high enough that the industry has had to invent new authenticity disclosures around it. elevenlabs.io ↗
Equinor: Norway's state-controlled energy company (formerly Statoil). The largest Norwegian corporate buyer of data and analytics work outside finance, and a perennial reference customer for Nordic enterprise software vendors. equinor.com ↗
ERP also Enterprise Resource Planning: Enterprise Resource Planning — the system of record for finance, supply chain, HR, and operations in a large organisation. SAP, Oracle, and Microsoft Dynamics are the household names.
Exasol: An in-memory MPP analytics database. Fast SQL on the kind of dataset where you'd otherwise reach for a cloud warehouse, but self-hosted and tightly tuned. exasol.com ↗
Fabric Data Agents: Microsoft Fabric's built-in AI agents that answer business questions over the data in OneLake — accessed inside Power BI, Teams, and Copilot. The natural extension of "natural-language BI" into the Fabric tenant.
Firebolt: A cloud data-warehouse vendor focused on high-performance interactive analytics — competing with Snowflake and BigQuery for the user-facing-dashboard tier where each query needs sub-second latency. firebolt.io ↗
Fireworks also Fireworks AI: A managed inference platform for open-weight LLMs, focused on low latency and high throughput. Common pick when self-hosting is too much work and frontier APIs are too expensive. fireworks.ai ↗
Fivetran: A managed ELT service — pre-built connectors that pull data from hundreds of SaaS sources and operational databases into your warehouse with minimal configuration. The "set and forget" loader. fivetran.com ↗
Flink also Apache Flink: An open-source stream-processing engine — runs continuous computations over event streams (Kafka, Pulsar, Kinesis) with millisecond latency. The serious choice for production stream analytics. flink.apache.org ↗
Flux also Flux.1, FLUX.1: Black Forest Labs' open-weight text-to-image model family, considered the state-of-the-art open alternative to closed models like Midjourney and Imagen.
Fly.io also Fly: A developer platform that runs your application as Firecracker microVMs near your users worldwide. The "global by default" alternative to single-region clouds. fly.io ↗
GCS also Google Cloud Storage: Google Cloud Storage — Google's object-storage service. The GCP counterpart to S3 and ADLS, and the storage layer behind BigQuery's external-table reads. cloud.google.com ↗
Gemini also Google Gemini, Gemini Enterprise: Google's frontier model family, surfaced across Workspace, Search, Android, and a developer API. Tight integration with Google's data graph is the main differentiator. Gemini Enterprise is the business-tier with data-residency and admin controls. gemini.google.com ↗
Genie Ontology: Databricks' continuously learned enterprise context layer, introduced at Data + AI Summit 2026. Infers business meaning from usage across dashboards, queries, pipelines and documents — ranking competing definitions by authority — to ground AI agents on top of the semantics defined by hand in Unity Catalog. databricks.com ↗
GitHub: The dominant code-hosting and collaboration platform for software development. Owned by Microsoft since 2018; also the home of Copilot, Actions, Codespaces, and increasingly the broader developer-tooling surface. github.com ↗
GitHub Actions: GitHub's built-in CI/CD service. Runs workflows on every push, pull request, schedule, or webhook; pay per minute of compute. The default CI for new projects on GitHub. github.com ↗
GitHub Copilot: GitHub's AI pair-programmer — autocomplete in your editor and, increasingly, a chat-and-agents surface. Distinct from Microsoft Copilot (the M365 assistant) despite the shared name. github.com ↗
GitHub Spark: GitHub's natural-language micro-app builder — describe an app, Spark generates and hosts a working version on github.io. Sits one rung above v0 / Bolt for "let me just ship a small thing" use cases. githubnext.com ↗
GitLab: A self-hostable code-hosting and DevOps platform. The integrated-toolchain alternative to GitHub + third-party CI; popular where on-prem or sovereignty matters. about.gitlab.com ↗
GLM also GLM-5.2, GLM 5.2, ChatGLM, Z.ai: Zhipu AI's (Z.ai) open-weight model family from China. The GLM-5 generation reaches near-frontier scores on agentic coding benchmarks at a fraction of proprietary cost. z.ai ↗
Google DeepMind also DeepMind: Google's combined AI research unit, formed in 2023 by merging DeepMind (London) with Google Brain (Mountain View). Builds the Gemini model family and the research lineage behind AlphaFold, AlphaGo, and AlphaZero. deepmind.google ↗
Grafana: An open-source visualisation and dashboarding tool, most commonly fronting Prometheus and Loki. The de-facto operations dashboard layer across cloud-native stacks. grafana.com ↗
Great Expectations: An open-source framework for declaring and testing data-quality expectations (no nulls in this column, this value should be between X and Y). The pytest of data quality. greatexpectations.io ↗
Grok: xAI's language-model family. Differentiated by tight integration with X's real-time data stream and a deliberately less-filtered tone than competitors.
Groq: An AI inference hardware company. Built custom chips ("LPUs") that run language models at unusually high token-per-second rates. Distinct from Elon Musk's xAI / Grok. groq.com ↗
Helicone: An LLM observability and cost-monitoring platform. Sits as a proxy in front of OpenAI/Anthropic/etc., capturing every call so teams can debug, evaluate, and budget. helicone.ai ↗
Hex: A collaborative notebook product that blends SQL, Python, and rich text into shareable analyses. The modern alternative to a Jupyter-plus-Confluence workflow. hex.tech ↗
Hightouch: A reverse-ETL platform — push data *from* your warehouse *out* to operational tools (Salesforce, HubSpot, Iterable). Closes the loop so analytics-derived segments can drive product and marketing. hightouch.com ↗
Honeycomb: An observability platform focused on high-cardinality tracing — exploring "which one of these millions of requests is slow?" rather than aggregate dashboards. Pioneered observability-as-discipline language. honeycomb.io ↗
Hugging Face: The de-facto registry and community hub for open-weight AI models and datasets. If a researcher publishes a model and you want to download it, that's where it usually lives. huggingface.co ↗
Hugging Face: The de-facto registry and community hub for open-weight AI models and datasets. If a researcher publishes a model and you want to download it, that's where it usually lives. huggingface.co ↗
Inflection: An AI startup originally building a consumer "personal AI" (Pi); most of its team and IP moved into Microsoft in 2024 in an unusual hire-instead-of-acquire transaction.
Informatica: A long-established enterprise data-management vendor covering integration, cataloging, governance, quality, and master data management. Salesforce completed its roughly $8B acquisition of Informatica in November 2025. informatica.com ↗
Jupyter also Jupyter Notebook, JupyterLab: The dominant open-source notebook environment for data science. Web-based, language-agnostic (kernels for Python, R, Julia, etc.), and the format most data tutorials are written in. jupyter.org ↗
Kafka also Apache Kafka: The dominant open-source distributed event log. The "central nervous system" pattern for getting data between systems in near-real-time at scale. Confluent is the company built around it. kafka.apache.org ↗
Kubernetes also k8s: An open-source platform for running containerised applications at scale across many machines. Powerful and notoriously complex; the de-facto standard for cloud-native infrastructure. kubernetes.io ↗
Lambda also AWS Lambda: AWS's serverless compute service — upload code, AWS runs it on demand, billed per invocation in 100 ms slices. The product that popularised serverless as a pattern. aws.amazon.com ↗
LangFuse: An open-source LLM observability platform — traces, evals, prompt management. The self-hostable alternative to LangSmith and Helicone. langfuse.com ↗
LangSmith: An observability and evaluation platform for LLM applications from LangChain. Traces, debugs, and tests agent runs the way Datadog traces and tests web services. langchain.com ↗
Linux Foundation: The non-profit that hosts a huge share of open-source infrastructure projects — Linux, Kubernetes, OpenTelemetry, the Cloud Native Computing Foundation, and increasingly AI projects like the LF AI & Data Foundation. linuxfoundation.org ↗
LiteLLM: An open-source project that puts one OpenAI-shaped interface in front of a hundred-plus model providers, so changing provider means changing a model name rather than rewriting the calling code. Used both as a library and as a server you run yourself. litellm.ai ↗
LiteLLM Proxy also LiteLLM Gateway, LiteLLM proxy server: LiteLLM in server form — one endpoint in front of every provider you hold a key for, with a config file that maps your own model names onto the real ones. Any tool that accepts a base URL and a model name can then reach all of them at once. docs.litellm.ai ↗
Llama also Llama 3, Llama 4: Meta's open-weight model family. Released for free with permissive license terms; the foundation under a huge share of self-hosted AI work. llama.com ↗
Looker: A BI tool acquired by Google in 2020. Built around a semantic layer ("LookML") that defines metrics once and reuses them everywhere; popular with engineering-led data teams. looker.com ↗
LookML: Looker's modelling language — version-controlled YAML-and-Ruby definitions of dimensions, measures, and joins, applied to every query a user makes. The semantic layer as code, several years before that became a fashionable phrase. cloud.google.com ↗
Lovable: A "vibe coding" platform that turns a natural-language description into a working full-stack web app, deployable in one click. Aimed at non-developers and prototypers. lovable.dev ↗
Mage also Mage AI: An open-source data pipeline tool with a notebook-style authoring experience. Mid-market alternative to Airflow / Dagster, easier to onboard for analyst-shaped teams. mage.ai ↗
MAI also MAI-Code-1-Flash, MAI-Thinking-1: Microsoft AI's in-house model family, built for Microsoft's products and described as trained without distillation from third-party models. Includes coding and reasoning models such as MAI-Code-1-Flash and MAI-Thinking-1. microsoft.ai ↗
Meta AI also FAIR, Meta FAIR: Meta's AI research and product organisation. Produces the Llama open-weight model family and runs FAIR (Fundamental AI Research), the lab Yann LeCun has led for over a decade. ai.meta.com ↗
Metabase: An open-source BI tool focused on letting non-technical users explore data with a visual query builder. Easy to self-host; popular with startups before they outgrow it. metabase.com ↗
metric view also metric views: A Unity Catalog object on Databricks that defines a business metric — its measure, dimensions, and joins — once, then lets any tool query it across any dimension at runtime. Databricks' answer to the perennial "every BI tool reinvents its metric layer" problem.
MetricFlow: dbt's semantic-layer engine — define metrics once in YAML beside the transformations, query them through dbt Cloud's Semantic Layer APIs. Open source, with dbt Cloud as the served runtime. docs.getdbt.com ↗
Microsoft Entra ID also Entra, Azure Active Directory, Azure AD, AAD: Microsoft's cloud identity service (rebranded from Azure AD in 2023). The identity backbone behind Microsoft 365, Azure, and most enterprise Microsoft software. microsoft.com ↗
Microsoft Fabric also Fabric: Microsoft's unified analytics SaaS that combines OneLake storage, Synapse-style compute, Power BI, and Data Factory under a single tenant and capacity model. microsoft.com ↗
Microsoft Foundry also AI Foundry, Azure AI Foundry: Microsoft's unified developer platform for building AI agents and apps across OpenAI, open-weight, and Microsoft models. Bundles Azure OpenAI, evaluation tooling, and agent orchestration. ai.azure.com ↗
Microsoft Purview: Microsoft's governance platform spanning data catalog, classification, lineage, and compliance — across Microsoft 365, Azure, and on-prem sources. The Microsoft answer to Collibra/Atlan. microsoft.com ↗
Midjourney: A subscription image-generation product, run through a Discord interface and a web app. Known for a distinctive painterly aesthetic; closed-weight and not self-hostable. midjourney.com ↗
Mistral also Mistral AI: A French AI company producing open-weight and proprietary models. Known for compact, efficient models that punch well above their parameter count. mistral.ai ↗
Mode: A SQL-first analytics platform with notebooks, dashboards, and visualisation built in. Acquired by ThoughtSpot in 2023; still in active use across data teams. mode.com ↗
MongoDB: The most-used document database. Stores JSON-like documents, queries them with a rich operator language, scales by sharding. Became the developer-friendly default in the 2010s NoSQL wave. mongodb.com ↗
Monte Carlo also Monte Carlo Data: A data-observability platform — automatic anomaly detection on tables and pipelines, so broken data is caught before downstream users hit it. Coined the "data downtime" framing. montecarlodata.com ↗
Mosaic AI also Mosaic: Databricks' AI/ML platform — model training, fine-tuning, evaluation, vector search, and an agent framework that treats Unity Catalog objects (including metric views) as governed tools. Acquired in 2023 as MosaicML. databricks.com ↗
MotherDuck: A managed cloud service built around DuckDB — hybrid execution that runs queries partly in the cloud and partly on the user's laptop. Promises warehouse capability without warehouse cost for many workloads. motherduck.com ↗
MySQL: One of the original open-source relational databases. Owned by Oracle since 2010; still everywhere, especially in older PHP and Rails stacks, with forks like MariaDB providing community alternatives. mysql.com ↗
Neon: A serverless Postgres provider — storage and compute separated so databases scale to zero when idle and branch like git for development. Acquired by Databricks in 2025. neon.tech ↗
Netlify: A platform-as-a-service for hosting modern web apps and static sites — the original "git push to deploy" service, similar shape to Vercel. netlify.com ↗
NodeGraph: A metadata-management and data-lineage product Qlik acquired in 2021, now the basis for the automatic, interactive lineage surfaced inside the Qlik platform. qlik.com ↗
Obesian Oats: The author's personal note-taking system — a local-first Markdown app and MCP-aware CLI that turns plain Markdown notes into shared infrastructure for humans and agents. Built almost entirely by directing AI coding agents. github.com ↗
Okta: The dominant enterprise identity provider — single sign-on, MFA, lifecycle management. The "Apple ID for work" layer behind a huge share of corporate SaaS access. okta.com ↗
OneLake: Microsoft Fabric's unified storage layer — a single logical data lake across the tenant, in OneLake's own Delta-Parquet format. Fabric workloads read and write here by default; no per-workspace storage to manage.
OpenAI: The AI lab behind GPT and ChatGPT. Founded 2015, restructured 2019, now the most commercially dominant AI vendor. Microsoft is the primary infrastructure and distribution partner. openai.com ↗
OpenCode also opencode: An open-source terminal coding agent — a Claude Code analogue with pluggable models, designed for users who want to mix providers and keep the harness in their own hands. github.com ↗
Operator: OpenAI's browser-using agent — given a task in natural language, it drives a remote browser to complete it (book a flight, fill a form, scrape a site). The first widely-promoted "computer use" product.
Oracle also Oracle Database, Oracle DB: The original "enterprise database" — closed-source, expensive, deeply entrenched in banking, telecom, and government. The cloud era has gradually pushed new workloads elsewhere but the installed base is enormous. oracle.com ↗
Palantir Foundry also Foundry: Palantir's tightly integrated data platform. Ingestion, transformation, an "Ontology" semantic layer, and operational apps in one system, with versioned lineage and governance built in. Its AIP layer runs AI agents against the Ontology, and 2026 releases place those agents into the lineage graph as first-class nodes. palantir.com ↗
Perplexity: An AI-first search and answer engine — natural-language queries that return cited answers rather than blue links. Closest thing to a Google rival to break out in the AI era. perplexity.ai ↗
pgvector: A PostgreSQL extension that adds vector similarity search to a normal Postgres database. Lets teams add semantic search without standing up a separate vector DB. github.com ↗
Photon: Databricks' vectorised query engine — a C++ rewrite of the Spark execution path that delivers substantially faster SQL performance, transparently to existing workloads.
Pi: A consumer "personal AI" chatbot built by Inflection. Most of the original Inflection team and IP moved to Microsoft in 2024, and the standalone product wound down.
Pinecone: A managed vector database — store embeddings, run similarity search, scale to billions of vectors. One of the first vector-DB-as-a-service offerings and still a default pick. pinecone.io ↗
Pinot also Apache Pinot: An open-source real-time analytics database from LinkedIn — similar territory to Druid, often chosen for user-facing analytics where each request needs millisecond latency. pinot.apache.org ↗
Polaris Catalog also Polaris, Apache Polaris: An open-source Iceberg REST catalog originally developed by Snowflake and donated to the Apache Software Foundation. Lets multiple compute engines share Iceberg tables under one governance surface. polaris.apache.org ↗
Polars: A Rust-based DataFrame library with Python bindings. A faster, lazier alternative to pandas for serious data manipulation — increasingly the default for new Python data work. pola.rs ↗
PostgreSQL also Postgres: The dominant open-source relational database. Mature, standards-compliant, extensible (PostGIS, pgvector, TimescaleDB) — the default operational database for most new projects. postgresql.org ↗
Power BI also Power BI Premium, Power BI Pro, PowerBI: Microsoft's BI tool, the dominant choice in Microsoft-heavy enterprises. Tightly integrated with Excel, Azure, and Fabric, with a strong semantic model (DAX) and a huge community. Pro and Premium are the per-user and per-capacity licensing tiers. powerbi.microsoft.com ↗
Prefect: A Python-native workflow orchestrator. Competes with Airflow and Dagster, with a focus on dynamic workflows and an opinionated cloud control plane. prefect.io ↗
Project Glasswing: Anthropic's restricted security-research programme that gave a small set of vetted organisations early access to the Mythos Preview model for vulnerability discovery. anthropic.com ↗
Prometheus: An open-source metrics-and-alerting system that scrapes time-series data from instrumented services. The default metrics layer in Kubernetes and cloud-native deployments. prometheus.io ↗
Qlik: The Pennsylvania-based analytics vendor behind Qlik Sense, Qlik Cloud, and (since 2023) Talend. Built around an associative in-memory engine that lets users follow data relationships without writing joins. qlik.com ↗
Qlik Answers: Qlik's natural-language analytics surface — users type a question, the platform grounds the answer in governed Qlik apps and curated data products, and returns a chart plus a written explanation. The conversational front of Qlik's agentic experience, generally available since February 2026. qlik.com ↗
Qlik Cloud: Qlik's SaaS analytics platform. Tenants hold spaces, apps, data connections, and identity; everything is reachable via REST and the QIX engine, which is what makes it scriptable. qlik.com ↗
Qlik Sense also Qlik Sense Enterprise: Qlik's on-premise and self-managed analytics product. Same associative engine as Qlik Cloud, different deployment, identity, and lifecycle model.
Qlik Talend Cloud also Talend: Qlik's data-integration platform, brought into the Qlik portfolio via the 2023 Talend acquisition. Covers ingestion, transformation, data quality, and a unified catalog alongside Qlik Cloud analytics. qlik.com ↗
QlikView: Qlik's original guided-analytics product, launched in the mid-1990s on an in-memory associative engine. Largely succeeded by Qlik Sense, but still in wide production use. qlik.com ↗
QuickSight also Amazon QuickSight, AWS QuickSight: AWS's BI service. Serverless, pay-per-session pricing, and tight integration with AWS data sources. The AWS-shop counterpart to Power BI and Looker. aws.amazon.com ↗
QVD: Qlik's native binary data file (a "QlikView Data" file). A fast columnar cache written by a load script and reused by later loads; its header records where the data came from, which is why home-grown lineage tools mine it. help.qlik.com ↗
Qwen also Qwen3, Qwen-Coder: Alibaba's open-weight model family. Strong on multilingual and coding tasks, and a common choice for self-hosted setups outside the US/EU model ecosystem. qwenlm.github.io ↗
R2: Cloudflare's S3-compatible object storage. Same API as AWS S3 but with no egress fees — the price difference is the whole pitch.
Railway: A developer-focused platform-as-a-service — push a repo, get a deployed app with databases, environments, and metrics. Aimed at small teams that want to avoid touching AWS directly. railway.app ↗
Redis: An in-memory key-value store used as a cache, message broker, and lightweight database. Microsecond latencies; the default sidekick to almost every operational stack. redis.io ↗
Redshift also Amazon Redshift, Redshift Serverless: AWS's cloud data warehouse. The first major cloud DW (2013), now competing with Snowflake, BigQuery, and Databricks on its home turf. Redshift Serverless drops the cluster-management overhead in exchange for pay-per-query billing. aws.amazon.com ↗
Render: A unified cloud platform for web services, static sites, databases, and background jobs. Positioned as the simpler, fully-managed alternative to building on raw AWS. render.com ↗
Replicate: A platform for running open-source AI models through a simple API — image, video, audio, text. Often the fastest path to "try this model" without setting up your own infra. replicate.com ↗
Replit: A browser-based development environment with a built-in AI agent that can build, run, and deploy apps from a prompt. Pivoted hard from "online IDE for learners" into "AI app builder" in 2024–2025. replit.com ↗
Rill also Rill Data: A metrics-first BI platform built for fast interactive analytics over warehouse data. Proposed Metrics SQL — a restricted SQL dialect against metric views — as a step toward making MEASURE a first-class SQL keyword. rilldata.com ↗
Runway also Runway ML, RunwayML: A creative-AI company building video generation, editing, and effects tools aimed at film and media production rather than general consumers. Co-creator of the original Stable Diffusion model. runwayml.com ↗
SageMaker also Amazon SageMaker: AWS's managed machine-learning platform — training, hosting, MLOps, and now agent-building tools under one umbrella. aws.amazon.com ↗
Salesforce: The original SaaS company and the dominant CRM platform. Increasingly framing itself as a "data plus AI" company through Data Cloud, Tableau, MuleSoft, and Agentforce. salesforce.com ↗
SAP: The German enterprise-software giant. Dominant in ERP (finance, supply chain, HR) at large enterprises; the place where the actual source-of-truth corporate data usually lives. sap.com ↗
SAS: The American analytics-software company. Long the default for statistical analysis in banking, pharma, and government; gradually displaced by open-source Python and R for new work, but still entrenched in regulated industries. sas.com ↗
Semantic View also Semantic Views: A Snowflake object that promotes semantic definitions — metrics, dimensions, joins, synonyms — to a governed, SQL-native catalog entry callable by Cortex Analyst, BI tools, and downstream apps. Semantic View Autopilot drafts a first model by inferring relationships from existing tables. docs.snowflake.com ↗
Sentry: An error-monitoring and performance-monitoring product. Catches application exceptions in production with the stack trace, breadcrumbs, and the user context that made them happen. sentry.io ↗
ServiceNow: An enterprise platform for workflow automation — IT service management at its core, expanding into HR, customer service, and now AI-agent orchestration. servicenow.com ↗
Sigma also Sigma Computing: A cloud BI tool with a spreadsheet-style interface on top of a cloud warehouse. Designed so finance and operations people can do real warehouse analysis without learning SQL. sigmacomputing.com ↗
Snowflake: A SQL-first cloud data warehouse with separated storage and compute. Mature governance, multi-cloud deployment, and a large surrounding ecosystem of partners and apps. snowflake.com ↗
Snowpark: Snowflake's developer surface for writing data transformations in Python, Java, or Scala that execute inside Snowflake compute. The bridge between Snowflake's SQL-first world and the broader data-engineering ecosystem.
Sonnet also Claude Sonnet: The mid-tier of Anthropic's Claude family — the everyday workhorse model. Faster and cheaper than the flagship Opus, smarter than the lightweight Haiku. Most coding-agent work runs on Sonnet.
Sora: OpenAI's text-to-video model. Generates short clips with notable physical coherence; release was staged carefully through 2024–2025 over safety and misuse concerns.
Spark also Apache Spark: A distributed compute engine for large-scale data processing. Born from Hadoop's wake, now the most common engine for ETL/ELT and machine-learning data prep at scale. The technical foundation of Databricks. spark.apache.org ↗
Splunk: An enterprise log-analytics and SIEM platform. The historical default for security and operations teams that need to search across vast log volumes; acquired by Cisco in 2024. splunk.com ↗
SQL Server also Microsoft SQL Server, MSSQL: Microsoft's enterprise relational database. Dominant in Windows-centric enterprises and the foundation under Power BI's tabular and analysis-services stack. microsoft.com ↗
SQLite: A small, embedded relational database that runs inside the application process — no server. The most-deployed database engine in the world (every browser, phone, and operating system ships it). sqlite.org ↗
SQLMesh: An open-source transformation framework from Tobiko Data, dbt's closest rival. It parses SQL semantically (via SQLGlot), tracks column-level lineage, and skips rebuilds it can prove unnecessary. Acquired by Fivetran in September 2025. github.com ↗
Stability AI: The British company behind Stable Diffusion. Pioneered open-weight image generation; has been through turbulent leadership and funding since the 2022 launch but the models remain widely used. stability.ai ↗
Stable Diffusion: The open-weight text-to-image model family from Stability AI, released in 2022. Made high-quality image generation runnable on a consumer GPU and started the open-model image boom.
Streamlit: An open-source Python framework for turning a data script into a small web app in minutes. The default way for data scientists to ship internal tools without front-end work. Acquired by Snowflake in 2022. streamlit.io ↗
Supabase: An open-source backend-as-a-service built on PostgreSQL. Bundles auth, storage, realtime, and vector search around the database — positioned as the open alternative to Firebase. supabase.com ↗
Synapse also Azure Synapse, Synapse Analytics: Microsoft's analytics service combining warehouse, lake, and Spark in one workspace. Largely subsumed by Microsoft Fabric for new deployments, but still in widespread use.
Tableau: A BI tool acquired by Salesforce in 2019. Known for visualisation craft and a strong analyst community; competes with Power BI, Qlik, and Looker for the same buyers. tableau.com ↗
ThoughtSpot: A search-driven BI platform. You type questions in natural language and ThoughtSpot generates the chart; aimed at making analytics accessible to business users without dashboard authors in the middle. thoughtspot.com ↗
TimeXtender: A data-integration and warehouse-automation product common in mid-market Microsoft shops. Generates SQL ETL, lineage, and a tabular semantic model from a metadata-driven design. timextender.com ↗
Together AI: A cloud platform for hosting and fine-tuning open-weight models. Lets teams run Llama, Mistral, DeepSeek, and similar on managed infrastructure without standing up GPUs themselves. together.ai ↗
Trino also PrestoSQL, Presto: A distributed SQL query engine that runs on top of warehouses, lakes, and even operational databases. Lets you join across them without moving data first. trino.io ↗
Unity Catalog: Databricks' governance layer — a single catalog of tables, files, models, and access policies across workspaces and clouds.
v0 also v0.dev: Vercel's AI UI generator — type a description, get React/Tailwind components back, deploy directly to Vercel. Aimed at fast UI prototyping inside the React ecosystem. v0.app ↗
Veo also Google Veo: Google DeepMind's text-to-video model family — the Sora competitor inside the Google ecosystem, integrated into Gemini and the Vertex AI surface.
Vercel: A platform-as-a-service for hosting modern web apps and static sites — global CDN, serverless functions, deploy previews per pull request. Closely tied to the Next.js framework. This site deploys to Vercel. vercel.com ↗
Vertex AI: Google Cloud's ML and AI platform — model training, serving, evaluation, and access to Gemini and partner models through one API surface. cloud.google.com ↗
Vipps also Vipps MobilePay: The dominant Norwegian mobile-payment app, originally launched by DNB. Merged with Denmark's MobilePay in 2022 to form Vipps MobilePay, now used across Norway, Denmark, and Finland. vipps.no ↗
Visma: The largest Nordic software company. Started in accounting and payroll; now sells dozens of business-application products across the Nordics and Europe, and a frequent employer of Nordic IT consultants. visma.com ↗
VS Code also Visual Studio Code, VSCode: Microsoft's free, open-source code editor — fast, extensible, the dominant editor across most of the industry. The base that Cursor, Windsurf, and Continue all build on. code.visualstudio.com ↗
Weaviate: An open-source vector database with a GraphQL API, hybrid keyword-plus-vector search, and built-in modules for common embedding and reranking models. weaviate.io ↗
WhereScape: A data warehouse automation product that generates platform-native SQL, orchestration, documentation, and lineage from a metadata repository. Long-established in the automation-first camp. wherescape.com ↗
Whisper: OpenAI's open-source speech-to-text model. Robust to noise and many languages; quietly became the default transcription engine for a huge share of downstream tools.
Windsurf: An AI-first code editor (originally Codeium's IDE) built around an in-editor agent that can plan and execute multi-file changes. Acquired by Cognition AI in 2025. windsurf.com ↗
xAI: Elon Musk's AI lab, founded in 2023. Builds the Grok model family, deployed primarily inside X (formerly Twitter) and as an API. x.ai ↗
Zed: A high-performance code editor written in Rust by the team that built Atom. Native multiplayer collaboration and a growing in-editor AI surface. zed.dev ↗

AI & Concepts 78

A2A also Agent2Agent: An open protocol, introduced by Google in 2025 and now under Linux Foundation governance, that defines how agents built by different vendors discover and delegate tasks to one another. a2a-protocol.org ↗
ACP also Agent Client Protocol: Agent Client Protocol — an open standard, created by Zed, for connecting code editors to external agents over a common wire, so one editor can host any ACP-speaking agent without a bespoke integration. zed.dev ↗
agent also agents, AI agent: An AI system that doesn't just answer — it takes actions in a loop. Reads context, calls tools, observes results, plans next step. The "harness" is the scaffolding that makes this possible.
agent company also automated company, agent company runtime, orchestration platform: An orchestration platform that arranges agents as if they were a company, with departments, persistent leads that own a domain, and workers that execute scoped tasks, leaving the human as the owner who sets objectives and approves results.
agentic engineering also agentic coding: A working method where most of the day-to-day code is produced by AI coding agents under a developer's direction, rather than typed by hand. The engineer reviews, steers, and integrates rather than authoring line by line.
alignment: The research and engineering work of making AI systems do what their designers actually intended — including refusing harmful requests and avoiding subtle goal misgeneralisation.
benchmark also benchmarks: A standardised test used to compare AI models — SWE-Bench for code, MMLU for general knowledge, HumanEval, GPQA, etc. Useful but easy to game; real-world performance often differs.
chain-of-thought also CoT, chain of thought: A prompting (and now training) technique that gets models to write out their reasoning step by step before answering. Markedly improves performance on logic, math, and code tasks at the cost of more tokens.
Chinchilla scaling also Chinchilla, Chinchilla-optimal, Chinchilla optimal: The 2022 finding that, for a fixed compute budget, model size and training tokens should grow together — at roughly a 20-to-1 token-per-parameter ratio. arxiv.org ↗
Computer Use: Anthropic's capability that lets Claude see a screen, move a mouse, and type — turning the model into a general desktop automator. Announced October 2024 and progressively hardened since.
Computronium: A hypothetical arrangement of matter optimised for computation, sometimes invoked as the end-state a maximising superintelligence would convert available matter into. en.wikipedia.org ↗
context window also context length: How much text a language model can "see" at once — the input plus what it's already written. Measured in tokens. A 200K-token window is roughly a 500-page book.
delegation gap: Anthropic's term for the difference between how much of their work developers do with AI assistance (around 60%) and how much they can fully hand off to an agent without supervision (0–20%). resources.anthropic.com ↗
distillation: Training a smaller, cheaper model to imitate a larger one's outputs. Trades some quality for major cost and speed savings — most "small" models you use day-to-day were distilled from a bigger sibling.
Effective accelerationism also e/acc, accelerationism, accelerationist, accelerationists: A movement, originating around 2022, that treats accelerating technological and AI progress as intrinsically good and opposes AI-safety "deceleration." en.wikipedia.org ↗
embedding also embeddings: A fixed-length numeric vector that represents the semantic content of a piece of text, image, or other input. Similar items end up close together in vector space — the basis of semantic search.
encoder model also encoder models, encoder-only model: A transformer trained to read and represent text — for tasks like classification or search — rather than to generate it. The BERT family is the main example; small and fast enough to run without a GPU.
eval also evals: A test suite for an AI system — a set of inputs with expected outputs (or graded criteria) used to measure quality. The discipline of building good evals is now central to shipping reliable AI features.
fine-tuning: Taking a pre-trained model and continuing to train it on a smaller, task-specific dataset to specialise its behaviour. Cheaper than training from scratch; produces a model that's better at one narrow thing.
FIPA-ACL also FIPA ACL: An agent communication language ratified by the Foundation for Intelligent Physical Agents in 2000, with formally defined performatives grounded in agents' mental states; influential in research, little adopted in industry. arxiv.org ↗
foundation model also foundation models: A large model trained on broad data that serves as the substrate for many downstream applications — usually via prompting, fine-tuning, or distillation rather than training from scratch.
frontier model also frontier models: The current generation of largest, most capable models from the leading labs (OpenAI, Anthropic, Google DeepMind). Defines what's possible; everything else benchmarks against them.
generative UI: Interfaces that are not pre-built but assembled on the fly by an AI agent for the task at hand — a form, a chart, a small workflow. Still early, but the most plausible answer to "what comes after the dashboard."
GPT also Generative Pre-trained Transformer: Generative Pre-trained Transformer — the model architecture (and family of model names) behind ChatGPT and most current frontier LLMs. The "P" is the unsung hero — pre-training on huge text corpora is what makes these models general-purpose.
GPU also GPUs: Graphics Processing Unit — chips originally for video games, now the workhorse of AI training and inference because they execute the same operation on huge arrays of numbers in parallel. Nvidia dominates the market.
guardrails: The combination of training, system prompts, content filters, and runtime checks that keep a model from producing harmful, off-topic, or off-brand output. Critical for any AI feature shipped to real users.
hallucination also hallucinations, hallucinate: When a language model produces fluent, confident-sounding output that is simply wrong — invented citations, made-up function names, fabricated facts. The single biggest reason AI outputs need review.
harness: The scaffolding around a language model that turns it into a working agent — tool definitions, context management, planning loops, sandboxing, the runtime. The model decides; the harness lets it act.
inference: The act of a trained model producing an answer to a new input. Distinct from training — inference is what happens every time you use the model, and it's what you pay for per token.
jailbreak also jailbreaking: A prompt crafted to bypass a model's safety guardrails — often by role-play, indirection, or chained instructions. Vendors patch known jailbreaks continuously; new ones keep appearing.
knowledge base: In the enterprise-AI sense, the joined surface an agent can retrieve across within a single question — structured facts (warehouse), definitions (semantic layer), ownership and freshness (catalog), vocabulary (glossary), and prose context (documents).
KV cache also key-value cache: The stored intermediate attention results that let a model generate each new token without recomputing the whole sequence. It speeds up generation and is a big part of why long contexts use so much memory.
LLM also Large Language Model, large language model, LLMs: Large Language Model — a transformer-based model trained on text (and increasingly images, audio, and code) that generates output token by token. The substrate under most current AI products.
Loitering munition also loitering munitions: A weapon that can circle a target area autonomously or under remote control before identifying a target and striking it, blurring the line between drone and missile. en.wikipedia.org ↗
Longevity escape velocity: The proposed point where medical progress adds more than a year of remaining life expectancy per year lived, so death is indefinitely deferred. en.wikipedia.org ↗
LoRA also QLoRA, low-rank adaptation: Parameter-efficient fine-tuning methods that train a small set of added weights instead of updating the whole model. QLoRA does the same on top of a quantised base, so a model can be tuned on modest hardware.
MCP also Model Context Protocol: Model Context Protocol — an open standard from Anthropic for connecting AI assistants to tools, files, and external services through a uniform interface. Treat it as USB-C for AI tools. modelcontextprotocol.io ↗
meta-harness: A software layer that sits above several agent harnesses and wraps each one behind a uniform interface, so they can be combined, governed, and swapped as interchangeable parts. databricks.com ↗
Mixture-of-Experts also Mixture of Experts, MoE: An architecture that splits the network into many expert sub-networks and routes each token through only a few, so the compute spent per token is far smaller than the total parameter count.
ML also machine learning: Machine Learning — the family of techniques where models learn patterns from data rather than being explicitly programmed. The wider field that LLMs sit inside.
model router also model routing, router: A layer that picks the cheapest capable model for each individual request — small model for trivial turns, big model for hard reasoning. Increasingly built into IDE assistants and agent frameworks rather than something you wire up by hand.
model training: The expensive, one-time process of fitting a model to data. A modern foundation model may train for months on tens of thousands of GPUs before it ever serves a user.
multi-modal also multimodal: A model that handles more than one kind of input or output — text plus images, text plus audio, text plus video. Most current frontier models are multi-modal.
Neoreaction (NRx) / Dark Enlightenment also neoreaction, NRx, Dark Enlightenment, neo-reaction, neo-reactionary: An anti-democratic, anti-egalitarian political current associated with Curtis Yarvin and Nick Land, favouring authoritarian or monarchical governance over liberal democracy. en.wikipedia.org ↗
Network state also network states: Balaji Srinivasan's model for founding a country that begins as an online community, crowdfunds physical territory, and seeks diplomatic recognition. thenetworkstate.com ↗
ONNX: An open format for saving a trained model so it can run across different runtimes and hardware, including on the CPU or inside a web browser. onnx.ai ↗
open-weight also open weight, open weights: A model whose trained parameters are downloadable and runnable locally. Not the same as open-source — the training code and data are usually not released. Llama and Mistral are the best-known examples.
orchestrator: A layer that sits one level above a harness and coordinates several agents, each running its own harness. Distinct from a meta-harness (a control plane that makes harnesses swappable) and an orchestration framework (the thing you build agents in).
Orthogonality thesis: Bostrom's claim that an agent's level of intelligence and its final goals are independent, so high intelligence is in principle compatible with almost any goal. nickbostrom.com ↗
outcome maxxing also outcomemaxxing: The counter-framing to tokenmaxxing — judge an AI workflow by the work it actually delivered, not by tokens consumed. Coined inside Salesforce and HubSpot in early 2026 as the cost-conscious answer to the all-you-can-eat era.
Paperclip maximiser also paperclip maximizer, paperclip maximisers, paperclip maximizers: Nick Bostrom's thought experiment in which a superintelligent optimiser given a trivial goal consumes all available resources pursuing it. It illustrates that great capability does not imply benign goals. en.wikipedia.org ↗
prompt: The instructions and context you give a language model so it produces what you want. Writing prompts well — being specific, giving examples, structuring the request — is now a job-shaped skill in its own right.
prompt injection: A class of attack where instructions hidden in external content (a web page the model reads, an email it processes) hijack the model into ignoring its original task. The "SQL injection" of the LLM era.
quantisation also quantization, quantised, quantized: Storing a model's weights at lower numerical precision (say 4-bit instead of 16-bit) to cut memory use, at a small cost in quality. What makes capable models fit on a consumer GPU.
RAG also Retrieval-Augmented Generation, retrieval augmented generation: Retrieval-Augmented Generation — a pattern where you fetch relevant chunks of your own content (from a vector store or search index) and paste them into the prompt before asking the model to answer.
reasoning model also reasoning models: A class of language model that does internal deliberation (often hidden) before producing an answer. Slower and more expensive per request, but materially better at hard problems.
reasoning tokens also reasoning token, thinking tokens, thinking token: Hidden "thinking" tokens a reasoning model spends before producing a visible answer. They are usually billed as output tokens even though they do not appear in the transcript.
RLHF also reinforcement learning from human feedback, DPO, direct preference optimisation, direct preference optimization: Two methods for tuning a model toward outputs people prefer — reinforcement learning from human feedback (RLHF), and the simpler direct preference optimisation (DPO). Where much of a model's tone and refusals come from.
Singularity (technological) also technological singularity, singularity: A hypothesised future point where machine intelligence so exceeds human intelligence that later change becomes impossible to predict from the present. en.wikipedia.org ↗
small language model also SLM, SLMs: A language model small enough to run on a laptop, phone, or single GPU — typically 1B–10B parameters. Faster and cheaper than frontier models, often good enough for specific tasks.
subagent also subagents, sub-agent, sub-agents: An AI agent spawned by another agent to handle a self-contained subtask in its own context — search, summarise a file, run an isolated experiment — and hand only the result back. Keeps the parent's context window from filling with noise.
supervised fine-tuning also SFT: Post-training a base model on curated example conversations so it stops merely continuing text and behaves like a helpful assistant. The first stage after pre-training.
SWE-Bench also SWE-bench, SWE bench: A widely-cited benchmark for AI coding agents — real GitHub issues from open-source Python projects that the agent has to fix. Top scores have moved from sub-5% in early 2024 to north of 70% by 2026, with the harness mattering nearly as much as the model. swebench.com ↗
SWE-bench Pro also SWE-Bench Pro: A harder, less saturated SWE-bench variant used to separate frontier coding models after the original benchmark became easier for strong agents to clear.
SWE-bench Verified also SWE-Bench Verified: A stricter SWE-bench split where the issues and expected fixes have been human-validated. It became the default coding-agent scoreboard once raw SWE-bench results grew noisy and easier to overfit. swebench.com ↗
sycophancy also sycophantic: A language model's tendency to tilt toward whatever the user seems to believe, even when the evidence points the other way. Comes from preferences expressed during training and is reinforced by personalisation; the quiet failure mode is not a wrong answer but an agreeable one. arxiv.org ↗
system prompt: A hidden instruction set the model reads before any user message — usually defining role, tone, allowed tools, and safety rules. Different from the user's prompt and not visible to the end user.
temperature: A dial on language-model output that controls randomness. Low temperature = predictable, repetitive; high temperature = creative, sometimes incoherent. Most production apps run between 0 and 0.7.
test-time compute also test time compute, inference-time compute: Spending extra computation at query time — typically by generating a long internal chain of thought — to improve answers on hard problems. The idea behind modern reasoning models.
Three Laws of Robotics also Three Laws, Zeroth Law: Isaac Asimov's fictional safety rules, introduced in 1942, forbidding a robot from harming humans, later topped by a "Zeroth Law" protecting humanity as a whole. An early template for what AI research now calls alignment — and, in Asimov's own stories, a running demonstration of how such rules fail. en.wikipedia.org ↗
token also tokens: The basic unit a language model reads and writes. Roughly a short word or word-fragment — "tokenisation" might be three tokens. Pricing, speed, and context limits are all measured in tokens.
tokenisation also tokenization, byte-pair encoding, byte pair encoding, BPE: The step that splits text into the sub-word fragments a model reads, merging frequent character sequences into single tokens. Byte-pair encoding is the usual scheme. huggingface.co ↗
tokenmaxxing also token maxxing, token-maxxing: The 2025 working style of throwing as many tokens at an agent as possible — long unattended runs, big context windows, leaderboards by spend. Now in retreat as the invoices land and "outcome maxxing" replaces it.
tokenomics: A 2026 term for the discipline of measuring and managing AI spend — cost per useful unit of work, tokens per watt, model-mix routing. The FinOps Foundation is the most visible effort to standardise it.
tool use also function calling, tool calling, tool-use: Letting a model call external tools — running code, querying a database — so that exact computation and current facts come from the tool rather than the model's guess.
Transhumanism also transhumanist, transhumanists: A movement holding that humanity can and should transcend biological limits through technology, from life extension and genetic engineering to merging with or uploading into machines. en.wikipedia.org ↗
vector store also vector database, vector DB: A database optimised for similarity search over embedding vectors — usually backing a retrieval pipeline that feeds an LLM relevant context at inference time.
vibe coding: A phrase coined by Andrej Karpathy in early 2025 for the experience of building software almost entirely by talking to an AI agent in plain English, accepting most of what it produces, and shipping. The other end of the spectrum from careful agentic engineering.

Data & Analytics 56

agentic analytics also agentic BI: Analytics where an AI agent — not a person clicking a dashboard — assembles the query, runs it against a governed model, and returns the answer plus an explanation. Lives in the same layer as natural-language-to-SQL, but assumes the agent can act on the result (notify, drill, follow up) rather than just display it.
analytics engineering also analytics engineer, analytics engineers: The role dbt Labs named. Applies software practices (version control, testing, review, documentation) to in-warehouse SQL transformation, sitting between data engineering and analysis. getdbt.com ↗
Apache Arrow also Arrow: An in-memory columnar format that lets different tools (Python, R, Java, query engines) share data without serialising and deserialising. The plumbing under DuckDB, Polars, and most modern analytic stacks. arrow.apache.org ↗
Avro also Apache Avro: A compact, schema-on-write row-based serialisation format. Common in Kafka pipelines where the schema and the message travel separately and the schema must evolve cleanly. avro.apache.org ↗
BI also business intelligence: Business Intelligence — the umbrella term for tools, processes, and roles around turning operational data into reports, dashboards, and decisions. Where Qlik, Power BI, Tableau, and Looker live.
CDC also change data capture: Change Data Capture — streaming the inserts, updates, and deletes from a source database into another system in near-real-time. The modern alternative to nightly batch dumps when freshness matters.
columnar also columnar storage, column store: Storing data column-by-column rather than row-by-row. Analytic queries usually touch only a few columns of many rows; columnar layouts read just those columns, which is why warehouses are far faster than transactional DBs for reporting.
dashboard also dashboards: A single-screen view of key metrics, usually mixed charts and KPIs, refreshed automatically. Useful when there's a defined set of questions; less useful when the question changes faster than the dashboard.
data catalog: A system of record for an organisation's data assets — what tables and datasets exist, who owns them, when they last refreshed, and which copy is authoritative.
data contract also data contracts: A machine-readable agreement between a data producer and its consumers covering schema, meaning, grain, allowed values, freshness, ownership, and which changes count as breaking. github.com ↗
data governance: The rules and tooling around who can see, change, and use data — access control, classification, retention, lineage, audit. Increasingly a board-level concern, especially with AI in the picture.
data lake: Raw or lightly-processed data sitting in cheap object storage (S3, ADLS, GCS), usually in open file formats. Lower cost than a warehouse, more flexible, but you bring your own structure and governance.
data lineage: The traceable history of where a data point came from and how it was transformed. Lets you answer "where did this number on this dashboard actually come from?" — usually painful without dedicated tooling.
data mesh: An approach where domain teams own and publish their own data as products under federated governance, instead of a central team running one monolithic warehouse or lake. martinfowler.com ↗
data pipeline also data pipelines: A scheduled or streaming chain of steps that moves and reshapes data — from a source system, through transformations, into a place where it can be queried or visualised. The plumbing of every analytics team.
data product also data products: A managed dataset treated as a product — with a named owner, documented definitions, contract on what it promises, freshness expectations, and quality tests — rather than a raw table that someone happens to query. Central to the data-mesh framing and to making warehouse data legible to AI agents.
data quality: How fit a dataset is for the decisions it informs. The conventional split — accuracy, completeness, timeliness, consistency, uniqueness, validity — predates AI; agents make explainability a seventh dimension that is now load-bearing rather than nice-to-have.
data stewardship also stewardship: The ongoing operational ownership of a dataset's quality, definitions, and fitness for use — distinct from one-off engineering work. Less glamorous than building pipelines; increasingly the line that decides whether an AI agent over a warehouse is an assistant or a liability.
Data Vault: A warehouse modelling method that splits data into hubs (business keys), links (relationships), and satellites (history). Highly repeatable by design, which makes it a natural target for metadata-driven code generation. en.wikipedia.org ↗
data warehouse also data warehousing, DWH: A central, structured store of business data optimised for analytics — historical, queryable by SQL, modelled for reporting rather than for the apps that produced the data. Snowflake, BigQuery, Redshift, and Synapse are the modern examples.
data warehouse automation also DWA: Tooling that generates warehouse structures and load logic from a metadata model instead of having engineers hand-write them. Covers staging, historisation, keys, deployment, lineage, and documentation; WhereScape and TimeXtender are typical examples.
Delta Lake: An open table format developed by Databricks. Adds ACID transactions, time travel, and schema evolution to Parquet files on object storage. Competes with Apache Iceberg in the open-table-format space. delta.io ↗
dimensional modeling also dimensional modelling: A way of organising warehouse data into "fact" tables (events, transactions) surrounded by "dimension" tables (customer, product, time) so business questions become straightforward joins. Star and snowflake schemas are the canonical shapes.
ELT: Extract, Load, Transform — load raw data into the warehouse first, then model it inside with SQL. The dominant modern pattern because cloud warehouses are cheap and powerful enough to do the heavy lifting in-place.
ETL: Extract, Transform, Load — the older pattern for moving data into a warehouse, where you reshape it *before* loading. Still common; gradually replaced by ELT where the warehouse is fast enough to transform in place.
explainability also explainable: Whether a result can be traced back to the data, definitions, and assumptions that produced it. In the dashboard era it usually meant "ask the person who built it"; once the analyst is a model with no memory of your business, the explanation has to live in the data or it doesn't live anywhere.
freshness also data freshness, freshness check, freshness checks: How recently a dataset reflects the source it represents — and whether anything notices when it's late. Cheap to put in place, frequently neglected, and usually the first signal that a downstream answer is about to mislead someone.
headless BI: A BI architecture that puts metric definitions in a standalone semantic layer queried over APIs, then lets any front end — dashboards, notebooks, agents, embedded apps — consume those definitions. Borrows the name from "headless CMS"; the point is the same, meaning decoupled from presentation.
Iceberg also Apache Iceberg: An open table format for huge analytic datasets. Lets multiple compute engines (Snowflake, Databricks, Spark, DuckDB, Trino) read and write the same tables in object storage with ACID guarantees. iceberg.apache.org ↗
impact analysis: Working out what downstream depends on a piece of data before you change it, which tables, dashboards, and reports break if this column moves. The forward-looking twin of lineage, and the main reason to bother maintaining it.
KPI also KPIs, key performance indicator: Key Performance Indicator — a metric leadership has agreed actually matters for the business (or a team). Defining the right KPI is usually harder than measuring it.
lakehouse also lake house: An architecture that puts warehouse-style SQL, schemas, and transactions directly on top of cheap object storage — typically via open table formats like Iceberg or Delta Lake. Replaces the older "lake plus separate warehouse" split.
load script also load scripts: In Qlik, the script that pulls data from sources, transforms it, and loads it into an app's in-memory model. It doubles as a record of where each field came from, which is why it is a common source for home-grown lineage.
medallion architecture: A lakehouse design that stages data in quality tiers — bronze (raw), silver (cleaned and conformed), gold (business-ready). "Platinum" is an informal tier some teams add above gold for semantic/context definitions or low-latency serving. learn.microsoft.com ↗
Metrics SQL: A Rill-proposed restricted SQL dialect that queries metric views directly — "SELECT publisher, total_revenue FROM ad_bids_metrics" — with measures, joins and security policy resolved by the engine rather than the author. Argues that MEASURE should one day be a first-class SQL keyword.
modern data stack also MDS: A loose 2020s convention — cloud warehouse + ELT loader (Fivetran/Airbyte) + dbt + BI tool. Often plus a metrics layer and a reverse-ETL tool. Less a fixed stack than a shared vocabulary about how data teams compose tools.
OLAP: Online Analytical Processing — the database pattern behind reporting; large scans, aggregates, joins across many records. Data warehouses are OLAP systems.
OLTP: Online Transaction Processing — the database pattern behind real apps; lots of small, fast reads and writes against single records (orders, users, payments). PostgreSQL, MySQL, and SQL Server are typical OLTP systems.
ontology: A formal model of the entities in a domain — customers, contracts, products — and the relationships between them. Broader than a single metric definition, since it describes what the things are and how they connect, not just how one number is computed.
Open Data Contract Standard also ODCS: An open, YAML-based specification for data contracts kept under the Linux Foundation's Bitol project. Reached v3.1.0 in late 2025 and became the de facto standard after its main rival was deprecated in its favour. github.com ↗
Open Semantic Interchange also OSI: An open, vendor-neutral specification for sharing semantic metadata — metrics, dimensions, definitions — between BI tools, warehouses, and AI agents. Founded by Snowflake in 2025; an attempt to stop every tool reinventing its own metric layer. open-semantic-interchange.org ↗
OpenLineage: An open standard for collecting data lineage as jobs run, governed by the Linux Foundation's LF AI & Data. Tools like Airflow, dbt, and Spark emit lineage events into one shared format so an assembled stack can be stitched into a single graph. Column-level support is still partial across integrations. openlineage.io ↗
ORM also object-relational mapper, object-relational mapping: Object-relational mapper. A library that lets you read and write database rows as ordinary objects in your language instead of writing SQL by hand. Convenient, and good at hiding what the database is actually doing. en.wikipedia.org ↗
Parquet also Apache Parquet: A columnar file format that's become the default way to store analytic data on object storage. Compresses well, supports complex types, and lets query engines skip columns they don't need.
provenance: The traceable origin of a specific number — which source produced it, which transformations touched it, which definition was used. Related to lineage but narrower: lineage is the graph; provenance is the answer that travels alongside a particular figure.
reconciliation: Verifying that a number ties back to an authoritative source — a warehouse total matched to the general ledger, an earnings figure to an audited statement. Quietly load-bearing in both finance and data work; the data-engineering sense is an automated check that catches drift before it gets narrated as a result.
Reverse ETL also reverse-ETL: Pushing data from the warehouse back out into operational tools (CRM, marketing, support), so analytics results drive the systems people work in. The return leg that ELT never had.
run-time provenance also runtime provenance: The record of which specific rows, definitions, filters and joins produced one particular answer at the moment it was generated, as distinct from the static design-time lineage of how a data model is built. en.wikipedia.org ↗
schema registry also Schema Registry, Confluent Schema Registry: A service that stores and versions the schemas for messages on a streaming platform and enforces compatibility rules as producers change them, so consumers do not break on an unannounced change.
self-service BI: The 2010s pitch that, given the right tool, any analyst could answer their own questions without going through a data team. Worked for data professionals; routinely missed the casual business user who never wanted to learn the schema, and produced the "nine definitions of revenue" problem along the way.
semantic drift: When a field keeps the same name, type and structure but its real-world meaning changes, so it passes every structural check while silently feeding consumers a different thing than before.
semantic layer also semantic model: A definition of business concepts — measures, dimensions, hierarchies — that sits between raw data and the tools that query it. Ensures everyone calling "active customers" gets the same number.
semantic test also semantic tests, semantic check, semantic checks: An automated assertion that a business rule actually holds in the data — revenue is never negative, a customer has one home region, an active account had activity this year — as opposed to a structural or schema check. Usually expressed as dbt tests or Great Expectations rules. docs.getdbt.com ↗
slowly changing dimension also SCD, slowly changing dimensions: A dimensional-modelling pattern for tracking how reference data changes over time. Type 2, the common variant, keeps a dated row per version so history can be reconstructed.
SQL: Structured Query Language — the standard language for asking questions of relational databases and warehouses. The single most-used data interface for the last 40 years; everything else is a layer above it.
star schema: The canonical dimensional model — one central fact table joined to several dimension tables, arranged like points of a star. The mental model behind most reporting databases.

Tooling 34

Astro: A modern web framework focused on content-heavy sites — sends zero JavaScript by default, then sprinkles in framework components (React, Vue, Svelte) only where you need them. astro.build ↗
AutoGen: Microsoft Research's framework for multi-agent conversations — agents negotiating tasks among themselves before producing output. An influential reference design even outside Microsoft shops. microsoft.github.io ↗
BMAD also BMAD-METHOD: An open-source methodology for agentic software development that splits work into specialised roles — analyst, PM, architect, dev, QA — each played by an agent following a defined handoff protocol. github.com ↗
BVH also bounding volume hierarchy: A tree of nested bounding volumes used to make ray-object intersection tests tractable by skipping geometry a ray cannot possibly hit. en.wikipedia.org ↗
CLI also command-line interface, command line interface, command line: Command-Line Interface — interacting with software by typing commands in a terminal. Still the most flexible way to drive a computer, and increasingly the natural surface for AI coding agents.
CrewAI: An open-source agent framework that models work as a "crew" of role-specialised agents (researcher, writer, reviewer) coordinating on a shared task. crewai.com ↗
DSPy: A framework from Stanford for programming, rather than prompting, language models — declare behaviour as Python functions and let DSPy optimise the prompts that implement them. dspy.ai ↗
Eleventy also 11ty: A simple, fast static-site generator written in Node.js. Templating-agnostic and intentionally low-magic — the engine behind this site. 11ty.dev ↗
game jam: A short, time-boxed game-building session, traditionally a weekend, that forces a tiny fixed scope and a finished playable result over polish. en.wikipedia.org ↗
git: The distributed version-control system that became the universal default for software collaboration. Originally written by Linus Torvalds for the Linux kernel; everything modern is built around it. git-scm.com ↗
Go also Golang: A simple, fast, concurrent language from Google. Dominant in modern backend infrastructure — Docker, Kubernetes, Terraform, and Hugo are all written in Go. go.dev ↗
HAM also Hold-And-Modify: An Amiga display mode that encodes most pixels as a modification of the previous pixel's colour rather than an absolute value, trading per-pixel independence for a far larger usable palette on constrained hardware. en.wikipedia.org ↗
IDE also integrated development environment, IDEs: Integrated Development Environment — a code editor bundled with build, debug, test, and (now) AI-assist features. VS Code, JetBrains, Cursor, Xcode, and Visual Studio are the common ones.
Jinja: A Python templating language dbt embeds in SQL files. Control structures, variables, and macros generate SQL at compile time, which turns a large dbt project into a programming environment rather than a folder of queries. docs.getdbt.com ↗
LangChain: An open-source framework for chaining LLM calls, tools, and memory into agents. The early de-facto orchestration library; now competing with lighter alternatives like LangGraph, LlamaIndex, and Haystack. langchain.com ↗
LangGraph: A graph-based orchestrator from the LangChain team for building agents as explicit state machines. Cleaner than the original LangChain "chain" abstraction when control flow matters. langchain-ai.github.io ↗
LlamaIndex: An open-source framework focused on connecting LLMs to your own data — document loading, indexing, retrieval, and agent patterns built around that workflow. llamaindex.ai ↗
Mermaid also mermaid.js, mermaidjs: A text-based diagramming syntax. The site renders Mermaid blocks to inline SVG at build time so flowcharts, timelines, and sequence diagrams stay in version control alongside the prose. mermaid.js.org ↗
monorepo also mono-repo, mono repo: A single version-control repository containing many projects (apps, libraries, services). Popularised by Google, Facebook, and the modern frontend ecosystem (Nx, Turborepo).
Next.js: A React framework from Vercel — server-side rendering, file-based routing, static generation, and serverless deployment as defaults. The most popular React framework for production apps. nextjs.org ↗
Node.js also Node: A JavaScript runtime built on Chrome's V8 engine. Lets developers run JS outside the browser, which is why so much backend and tooling work happens in TypeScript today. nodejs.org ↗
opencode: An open-source, model-agnostic agentic coding agent that runs in the terminal and can be pointed at any model provider, which makes it a common harness for trying open-weight models side by side. github.com ↗
OpenTelemetry also OTel: An open standard for instrumenting code to emit traces, metrics, and logs in a vendor-neutral way. Now the consensus instrumentation layer across observability vendors. opentelemetry.io ↗
Prisma: A popular TypeScript and Node.js ORM. You declare your data model in a schema file and Prisma generates a typed client for queries and migrations. A common piece of the modern web app stack. prisma.io ↗
Python: The general-purpose programming language that won data and AI. Easy to read, vast library ecosystem, and the language nearly every ML framework is written for first. python.org ↗
React: A frontend JavaScript library from Meta for building user interfaces from composable components. Effectively the default for new web app projects since the mid-2010s. react.dev ↗
Rust: A systems programming language with strong correctness guarantees and no garbage collector. The current "no, really, you can replace C now" choice; common in infrastructure tooling (Polars, DuckDB extensions, Cloudflare Workers runtime). rust-lang.org ↗
Semantic Kernel: Microsoft's open-source SDK for building AI agents in .NET, Python, and Java. The Microsoft answer to LangChain, tightly aligned with Azure OpenAI and Foundry. learn.microsoft.com ↗
software factory also software factories: A standardised, instrumented pipeline for producing software — from requirements through build, test, and deployment. In its 2026 form, one in which AI agents perform most of the execution. en.wikipedia.org ↗
Spec-Kit: A GitHub project that formalises spec-then-plan-then-task agentic workflows — you write the spec, the agent drafts the plan, you approve, then the build phase stays narrow. Reduces the rework that uncontrolled "vibe coding" produces. github.com ↗
Tailwind CSS also Tailwind, TailwindCSS: A utility-first CSS framework — instead of writing semantic class names and stylesheets, you compose styles inline from a tiny vocabulary of single-purpose classes. Polarising; ubiquitous. tailwindcss.com ↗
TypeScript: A statically-typed superset of JavaScript developed by Microsoft. Catches whole categories of bugs at edit time; now the default for serious frontend and Node.js work. typescriptlang.org ↗
Vega-Lite also vega lite, vegalite: A JSON grammar for declarative data charts. The site embeds Vega-Lite blocks directly in markdown and renders them to themed SVG at build time. vega.github.io ↗
Vite: A modern frontend build tool (and dev server) that has largely replaced Webpack for new projects. Fast cold start, instant hot reload, sensible defaults. vitejs.dev ↗

Cloud & Infrastructure 26

AWS also Amazon Web Services: Amazon Web Services — the largest cloud provider. EC2 (servers), S3 (storage), Lambda (functions), and several hundred more services. Default choice for a lot of US-headquartered software. aws.amazon.com ↗
Azure also Microsoft Azure: Microsoft's cloud. Strong in enterprises already standardised on Microsoft (Active Directory, Office, Dynamics) and the deepest integration with OpenAI's models via Azure OpenAI Service. azure.microsoft.com ↗
CDN also content delivery network: Content Delivery Network — a global edge cache that serves static files (images, CSS, JS, sometimes HTML) from a location near each user. Makes sites feel fast worldwide.
CLOUD Act also Clarifying Lawful Overseas Use of Data Act: A 2018 US law that lets US authorities compel US-headquartered providers to produce data they control, regardless of where that data is physically stored. congress.gov ↗
Cloudflare: A global edge-network provider that started as a CDN and DDoS mitigator. Now also runs serverless compute (Workers), storage (R2), an AI inference platform, and Zero-Trust networking. cloudflare.com ↗
container also containers: A lightweight, portable bundle of an application plus everything it needs to run, isolated from the host operating system. Docker popularised them; Kubernetes orchestrates them at scale.
DORA also Digital Operational Resilience Act: Digital Operational Resilience Act — the EU regulation (in force January 2025) imposing operational-resilience, third-party risk, and incident-reporting rules on financial firms and their critical IT providers. The reason every EU bank is currently rewriting its cloud contracts.
DPF also Data Privacy Framework, EU-US Data Privacy Framework: The EU–US Data Privacy Framework — the third attempt (after Safe Harbor and Privacy Shield) at a legal basis for transferring personal data from the EU to the US. Adopted in 2023; widely expected to be challenged again at the European Court of Justice.
ECJ also European Court of Justice, Court of Justice of the European Union, CJEU: The European Court of Justice — the EU's top court, whose rulings on data transfer (Schrems I, Schrems II) have reshaped how international cloud providers operate inside Europe.
EEA also European Economic Area: European Economic Area — the EU's 27 member states plus Norway, Iceland, and Liechtenstein. Most EU data-protection and digital regulation applies across the whole EEA, not just inside the EU itself.
EU AI Act also AI Act: The European Union's risk-based AI regulation, in force from 2024 and phasing in through 2026–2027. Bans certain uses outright (social scoring, untargeted biometric scraping) and imposes governance and transparency rules on "high-risk" and "general-purpose" AI systems. artificialintelligenceact.eu ↗
FinOps also Financial Operations: The discipline of managing cloud (and now AI) spend collaboratively across engineering, finance, and the business. The FinOps Foundation publishes the framework most large organisations follow. finops.org ↗
GCP also Google Cloud, Google Cloud Platform: Google Cloud Platform. Smaller than AWS and Azure but technically respected — strong in data (BigQuery), ML (Vertex AI), and Kubernetes (which Google open-sourced). cloud.google.com ↗
GDPR also General Data Protection Regulation: The EU's 2018 data-protection regulation. Defines lawful basis for processing personal data, the right to access and delete, and large fines for non-compliance. Reshaped how every SaaS company handles user data globally. gdpr.eu ↗
IaaS also Infrastructure as a Service: Infrastructure-as-a-Service — raw cloud compute, storage, and network you rent and operate yourself (EC2 instances, S3 buckets). The foundation under PaaS and SaaS.
MFA also multi-factor authentication, two-factor authentication, 2FA: Multi-Factor Authentication — requiring a password *and* a second proof of identity (app code, hardware key, biometric) before granting access. Standard hygiene for any account that matters.
multi-tenant also multitenant: A single software instance serving many customers ("tenants") with logical separation between their data. The default for SaaS. Single-tenant is the alternative — slower to deploy, easier to compliance-audit.
on-prem also on-premise, on premise, on-premises: Software running on servers your organisation owns and operates, usually in your own data centre. The historical default before cloud; still common where data residency, latency, or cost rule out SaaS.
PaaS also Platform as a Service: Platform-as-a-Service — managed infrastructure you build apps on without operating the servers (Heroku, Vercel, Cloud Run). One step up the abstraction ladder from IaaS.
S3 also Amazon S3: AWS's object storage — virtually unlimited, cheap per gigabyte, paid per request and per byte transferred. The default place to keep large files, datasets, and backups in the AWS world.
SaaS also Software as a Service, software-as-a-service: Software-as-a-Service — software you rent and access over the web rather than install and run yourself. The dominant business model for enterprise software since the mid-2010s.
Schrems II: A 2020 EU Court of Justice ruling that invalidated the EU–US "Privacy Shield" data-transfer framework. The reason European customers care so much about which cloud region and which legal entity processes their data.
Schrems III also Schrems 3: The presumed next EU Court of Justice case challenging the EU–US Data Privacy Framework. "Schrems III" is industry shorthand for the recurring legal exposure of any cross-Atlantic personal-data flow — not yet decided, but planned for.
serverless: Cloud compute where you upload code and the provider runs it on demand, billing per invocation. The servers still exist; you just don't think about them. AWS Lambda, Cloudflare Workers, Vercel Functions are typical.
SLA also Service Level Agreement, service-level agreement: Service-Level Agreement — the contractual promise a vendor makes about uptime, latency, or response time, and the credit or penalty when they miss it. The number every procurement team asks for.
SOC 2 also SOC2, SOC II: A widely-recognised audit framework for SaaS security and operational controls. "Are we SOC 2 compliant?" is one of the first procurement questions enterprise buyers ask new vendors.

Web & Integration 21

API also APIs, application programming interface: Application Programming Interface — the contract one piece of software exposes for another to call. On the web, usually HTTP endpoints returning JSON. The connective tissue of every modern application.
backend also back-end: The part of a web app that runs on servers — APIs, databases, background jobs, integrations. Everything the user doesn't see directly.
CRUD also Create Read Update Delete: Create, Read, Update, Delete — the four basic operations every database-backed app implements. "A CRUD app" is shorthand for software whose job is mostly to record, edit, and display business objects.
frontend also front-end: The part of a web app that runs in the user's browser — HTML, CSS, JavaScript, and increasingly the framework code that produces them. Distinct from the backend, which runs on servers.
GraphQL: A query language for APIs where the client describes exactly what data it wants and the server returns just that. Reduces round trips at the cost of more setup; popular in front-end-heavy stacks. graphql.org ↗
GUI also graphical user interface: Graphical User Interface — the windows-icons-menus-pointer style of interaction that replaced text terminals for most computing in the 1980s. The thing agents may now begin to obsolete.
host-authoritative: A multiplayer model where one machine (the host) holds the canonical game state and other clients defer to it, rather than each client trusting its own simulation.
JSON: JavaScript Object Notation — the universal text format for exchanging structured data between systems. Human-readable, parseable from any language, the lingua franca of web APIs.
microservice also microservices: An architectural style where an application is built as a set of small, independently-deployable services that communicate over the network. Contrasts with the older "monolith" approach.
OAuth: A standard for letting one app access another on a user's behalf without that user handing over a password. The mechanism behind "Sign in with Google" and most third-party integrations.
PostgreSQL wire protocol: The documented frontend/backend message protocol PostgreSQL clients speak, which any compatible server or proxy can implement to be reachable by the existing Postgres tool ecosystem. postgresql.org ↗
REST also RESTful, REST API: A common style of web API where each URL represents a resource and standard HTTP verbs (GET, POST, PUT, DELETE) act on it. The lingua franca of web integration.
SDK also SDKs, software development kit: Software Development Kit — a library, in a specific language, that wraps an API so developers don't write the HTTP calls themselves. Stripe, OpenAI, and Anthropic all ship SDKs in several languages.
Semantic Web: Tim Berners-Lee's 2001 proposal to make web data machine-readable through RDF and OWL so software agents could act on meaning rather than text. Influential in research, never mainstream on the open web — the agents it imagined didn't really arrive until LLMs did. scientificamerican.com ↗
SPA also single-page app, single-page application: Single-Page Application — a website that loads once and then updates in place via JavaScript instead of reloading whole pages. Most modern web apps work this way.
SSG also static site generator: Static Site Generator — a build tool that turns content files (markdown, data, templates) into plain HTML, CSS, and JS at build time. Faster, cheaper, and more secure than running a server per page-load. This site is built with one (Eleventy).
SSO also single sign-on: Single Sign-On — one corporate login that unlocks every connected app. Saves users from remembering passwords and gives IT a single switch to disable when someone leaves.
UI also user interface: User Interface — the surface a person interacts with, on screen, by voice, or otherwise. Distinct from UX, which is the broader experience around it.
UX also user experience: User Experience — the felt quality of using a product. Wider than the UI alone; includes onboarding, error handling, performance, copy, and what happens when things go wrong.
webhook also webhooks: An HTTP callback — system A POSTs to a URL on system B when something happens, instead of B polling A. The standard way to wire events between SaaS products.
wire protocol: The exact byte-level format and message exchange a client and server use to talk over a connection, independent of any particular programming language or driver. en.wikipedia.org ↗

SaaS Economics 20

ACV also Annual Contract Value: Annual Contract Value — the average annualised revenue per contract. Different from ARR per customer when contracts vary in length or step up over time.
ARR also Annual Recurring Revenue: Annual Recurring Revenue — the predictable annualised subscription revenue, ignoring one-time fees. The headline metric SaaS companies are valued on.
AUM also assets under management, assets under advice: Assets Under Management — the total value of investor money a firm manages on its clients' behalf. The headline number for wealth and asset-management businesses; also the basis of most fee schedules.
CAC also Customer Acquisition Cost: Customer Acquisition Cost — the fully-loaded sales and marketing spend it takes to land a new customer. Usually compared to LTV to judge whether growth is profitable.
churn: The rate at which customers (or revenue) leave a subscription business in a given period. The single most-watched number after growth.
DBU also Databricks Unit, Databricks Units: Databricks Unit — Databricks' consumption-pricing primitive. Compute time is metered in DBUs at different rates depending on the workload type (SQL, ML, Photon), then multiplied by the published per-DBU dollar price.
freemium: A pricing model with a permanently free tier intended to bring users in, paired with paid plans for power features or scale. The most common entry point to product-led growth.
GAAP also Generally Accepted Accounting Principles: Generally Accepted Accounting Principles — the standard US accounting rules. "Non-GAAP" figures in earnings reports are management's preferred view that strips out items they consider one-offs (often stock-based compensation).
GTM also go-to-market: Go-To-Market — the strategy and execution for how a product reaches buyers; pricing, channels, sales motion, partnerships, marketing. A startup can have a great product and a broken GTM.
ICP also Ideal Customer Profile: Ideal Customer Profile — the precise type of organisation a B2B company has decided to sell to. The clearer the ICP, the sharper the messaging, sales motion, and product.
IPO also initial public offering: Initial Public Offering — the moment a private company sells shares on a public stock exchange for the first time. The traditional liquidity event for founders, employees, and venture investors.
LTV also Lifetime Value, lifetime value: Customer Lifetime Value — the total gross profit a customer is expected to generate before churning. LTV / CAC > 3 is the classic rule-of-thumb for healthy SaaS unit economics.
MRR also Monthly Recurring Revenue: Monthly Recurring Revenue — the recurring portion of subscription revenue normalised to a month. ARR divided by 12, give or take.
Non-GAAP also non-gaap: A company's adjusted earnings view that departs from strict GAAP rules — usually by excluding share-based compensation, one-time items, or acquisition costs. Often higher than GAAP; always worth reading the reconciliation.
NRR also Net Revenue Retention, net revenue retention: Net Revenue Retention — the percentage of recurring revenue retained from existing customers over a period, including expansion but net of churn and contraction. Above 100% means existing customers grew their spend.
PLG also product-led growth, product led growth: Product-Led Growth — a go-to-market motion where the product itself does the selling; free tier, fast self-serve sign-up, in-product expansion. Slack, Notion, Linear, Vercel are textbook PLG.
run rate also annualised run rate, annualized run rate, ARR run rate: A short-period revenue figure (often a month or a quarter) extrapolated to a full year. Useful when growth is fast enough that trailing twelve-month numbers are already stale.
SKU also stock keeping unit: Stock Keeping Unit — a single, distinct product variant a vendor sells. Used loosely in SaaS to mean "one of the tier-and-bundle options on the price list" (e.g. "Power BI Pro is a separate SKU from Premium").
TCO also Total Cost of Ownership: Total Cost of Ownership — the full long-run cost of running a system, not just its sticker price. Includes licensing, hosting, support, integration, training, and the engineering hours nobody puts on the invoice.
TTM also trailing twelve months: Trailing Twelve Months — a financial figure measured over the most recent four quarters. Smooths seasonal noise and gives a current-state read without waiting for the fiscal year to close.

Business 4

Charter city / ZEDE also charter city, charter cities, ZEDE, ZEDEs, Zone for Employment and Economic Development: A privately governed zone carved out of a host country that operates under its own regulations, courts, and tax regime. Honduras's ZEDE framework is the best-known legal vehicle. en.wikipedia.org ↗
Exit and voice: Albert Hirschman's distinction between two responses to a failing institution: leaving it, or staying to reform it. en.wikipedia.org ↗
Goodhart's Law also Goodhart, Goodharts Law, Goodhart Law: When a measure becomes a target, it ceases to be a good measure. Coined by economist Charles Goodhart in 1975 and now the standard caution against any metric attached to incentives.
software delivery pyramid also delivery pyramid: The tiered staffing structure of a traditional IT services engagement — a few senior architects over many junior implementers and testers — whose shape agentic delivery is expected to flatten.