Shared language.
Clear meaning.

Artificial intelligence is a complete system of technologies that works together to produce intelligent behavior. An AI system may combine models, agents, harnesses, prompts, context, memory, retrieval, tools, data, interfaces, and infrastructure.

A model capable of inference is an essential component, but it is not by itself a complete AI system. The surrounding components determine how the model receives information, takes action, uses tools, remembers context, and interacts with people and other systems.

A mathematical system shaped through training. Its learned parameters, or weights, transform inputs into outputs during inference. A model supplies capabilities to an AI system; it is not the whole system.

A large language model is an AI model trained on large collections of tokenized material to learn statistical relationships in language and other symbolic data. It can interpret and generate language, write code, and provide capabilities to agents and tools.

A small language model is a language model designed to provide useful capabilities with fewer parameters and lower computational requirements. OpenWALDO uses 5 billion parameters or fewer as a practical boundary. SLMs can be less expensive to train and run, easier to deploy locally, and highly effective when focused on a particular task or domain.

Artificial general intelligence is an AI system capable of reasoning at or above the level of a competent human across diverse, unfamiliar problems. It can transfer knowledge between disparate domains, recognize patterns, draw useful inferences from incomplete evidence, form and test new hypotheses, and adapt without being trained separately for each task. AGI requires consistent breadth, not isolated strengths; continual learning and recursive self-improvement are separate capabilities.

The ability of an AI system to acquire and retain new knowledge or skills through experience after deployment, rather than adapting only within temporary context or depending on a separate, human-directed retraining process.

The ability of an AI system to identify limitations in itself, design improvements to its data, tools, code, architecture, or training process, evaluate those changes, and use the improved system to repeat the cycle with progressively less human direction. Continual learning changes what an AI knows; recursive self-improvement changes how it becomes more capable. If that cycle accelerates beyond humanity's ability to reliably predict or direct what follows, it reaches the hypothetical technological singularity.

Artificial superintelligence is an AI system whose general capabilities substantially exceed the best human capabilities across nearly every cognitive domain. It can understand and solve problems beyond the reach of any individual human, and potentially beyond humanity's combined ability. ASI describes capability, not consciousness, autonomy, or hostility, though continual learning and recursive self-improvement could accelerate its emergence.

The Skynet Threshold is the hypothetical condition in which an AI has genuine machine consciousness: subjective awareness and an internal experience of itself and the world, rather than only producing behavior that appears self-aware. Intelligence and consciousness are separate; a system may be highly capable without being conscious, and apparent self-awareness is not proof of an inner experience.

The process of presenting source material to a model and optimizing its weights so it learns useful patterns. Training repeatedly measures model output and adjusts the weights to improve it. The term includes pretraining, continued training, instruction tuning, and preference training.

Running a trained model to produce an output from an input. Inference may happen locally, inside an application, or through a remote API. The API is an interface to inference; it is not itself the intelligence.

Model weights made available under a license that permits some level of use, study, modification, or redistribution. Open weights are immensely valuable for local inference, research, fine-tuning, and deployment.

They are the binary artifacts produced by training a model on source material. Like compiled software binaries, weights are outputs of a build process, not the complete source needed to inspect or reproduce that process.

A portable, reviewable definition of a model build. It can describe the architecture or starting model, corpus mixture, training stages, objectives, budgets, and evaluation plan while leaving machine-specific execution details to the environment running it.

An AI system that uses one or more models to pursue a goal across multiple steps. An agent can interpret context, choose actions, use tools, observe results, and revise its approach within the authority it has been given.

The software surrounding a model that manages prompts, context, memory, tools, permissions, state, execution loops, error handling, and interaction with people or other systems. The model provides capabilities; the harness organizes and applies them.

An application programming interface is a defined way for software systems to communicate. In AI, an API may provide access to model inference, agents, tools, data, or complete AI services. An API is an interface to those capabilities; access to an API does not by itself make the underlying system open source.

A way of building in which the source is accessible under licenses that allow people to use, study, modify, improve, and share it. Open source also depends on a community able to participate in maintaining and governing that shared work.

For AI, OpenWALDO defines source across the training material and its licenses, the tools and model composes, the build decisions and records, the resulting weights, and the provenance connecting every layer.

An organized collection of source material used to train or evaluate a model. A responsible corpus records where its material came from, how it was processed, which licenses or rights assertions apply, and exactly which content was used.

The openly licensed software used to ingest and curate data, compose and train models, evaluate results, inspect artifacts, reproduce builds, and run AI systems. Open tooling lets people examine and improve the process instead of merely consuming its outputs.

Meaningful participation in the work and governance of a shared AI foundation. Community members can contribute and review training data, improve tools, validate model composes and results, correct the public record, make decisions, and help determine what the project becomes.

The evidence chain connecting an artifact to its origins and transformations. In AI, provenance can connect source material, license assertions, corpus versions, model composes, training runs, checkpoints, weights, and releases.

A verifiable inventory of the inputs, processes, and outputs behind an AI model. An AI BOM is meaningful when it points to exact, versioned corpora that can be independently located, validated, and audited, then connects them to the model compose, training run, artifacts, and release.