Large language models
Neural networks trained on large text corpora to predict tokens, then adapted to follow instructions and produce text or actions.
Concept record →Atlas library · 18 records
Concepts with boundaries.
Each term is defined for this Atlas, linked to its source set, and—where necessary—paired with an explicit statement of what it does not establish.
Transformers
A sequence architecture built on attention rather than recurrence or convolution.
Concept record →Scaling laws
Empirical relations between training compute, parameters, data, and loss.
Concept record →Pre-training
The large-scale self-supervised phase in which a model learns from an unlabelled corpus.
Concept record →Post-training
Adaptation after pre-training — supervised fine-tuning, preference optimization, reinforcement learning from human or model feedback.
Concept record →Inference-time compute
Additional computation spent at generation time — longer reasoning traces, sampling with verification, or search over candidate outputs.
Concept record →Retrieval-augmented generation
Conditioning generation on documents fetched at inference time from an external store.
Concept record →Tool use
Model-initiated invocation of external functions — search, code execution, APIs — with results returned into context.
Concept record →Agent loop
An iterated cycle of reasoning, acting through tools, and observing results, continuing until a stopping condition.
Concept record →Coding agents
Agent systems operating on real codebases through file, shell, and test interfaces.
Concept record →Long-horizon task completion
The ability to carry a task requiring many dependent steps to completion, measured here by the time a human would take on tasks a system completes at a given success rate.
Concept record →Reliability and compounding error
The tendency of per-step failure probabilities to accumulate across a multi-step trajectory.
Concept record →Evaluation
The practice of measuring model or system behaviour against defined tasks and metrics.
Concept record →Construct validity
Whether a measurement actually measures the property it is claimed to measure.
Concept record →Benchmark contamination
Presence of evaluation data, or close variants, in training corpora.
Concept record →Local and on-device inference
Running models on user-controlled hardware rather than a hosted API.
Concept record →AI forecasting
Projection of future capability from historical measurements, expert elicitation, or trend extrapolation.
Concept record →Governance and accountability
Mechanisms for disclosure, reporting, and responsibility around deployed models.
Concept record →