AI Tech Beacon
What each tool does, what it costs, and what it does with your data — grouped by task rather than by vendor.
Glossary
See all →Agentic coding
The line an assistant crosses when it stops proposing text and starts acting.
Read → glossaryAI Risk Management Framework
A voluntary NIST framework for managing the risks of AI systems.
Read → glossaryCode provenance
Verifiable information about a software artifact stating where, when and how it was produced.
Read → glossaryCode completion
In editor tooling this is a protocol operation before it is a product feature.
Read → glossaryCode review automation
A systematic review of 119 papers on the automation of code review tasks.
Read → glossaryContent provenance
Disclosure, by whoever created or edited an asset, of how it was made and what was changed — carried with the asset rather than asserted beside it.
Read → glossaryDiffusion model
A class of latent variable models inspired by nonequilibrium thermodynamics.
Read → glossaryEU AI Act
Regulation (EU) 2024/1689, the European Union's law on artificial intelligence.
Read → glossaryGDPR
Regulation (EU) 2016/679, the European Union's general regime for processing personal data.
Read → glossaryIdempotency
An operation that can be repeated without changing the outcome.
Read → glossaryImage-to-image
Starting from an existing image instead of a description. Posed as a general problem rather than a set of separate tasks: the network learns the
Read → glossaryNegative prompt
Stating what should not appear, as a distinct input rather than a phrase inside the prompt.
Read → glossaryModel Context Protocol
An open protocol for connecting applications built on language models to external data sources and tools, so that each integration is not rebuilt per vendor.
Read → glossaryPrompt injection
A vulnerability that occurs when user input alters a model's behaviour or output in unintended ways.
Read → glossaryInpainting
Generating the contents of an arbitrary region of an image, conditioned on everything around it.
Read → glossaryRetrieval-augmented generation
Combining two memories rather than enlarging one. A pre-trained sequence-to-sequence model supplies the parametric memory; a dense vector index,
Read → glossaryModel card
A short document released alongside a trained model to report what it does and where it was evaluated.
Read → glossarySandboxing
Running an untrusted application inside a tightly controlled environment where its permissions are cut back to an essential set — typically no access
Read → glossaryReal-time transcription
Speech transcribed as it is spoken rather than from a finished recording.
Read → glossaryRate limit
The cap a server places on how many requests a client may make in a time window.
Read → glossarySeed
The bit string given as input to a pseudorandom process. Its usefulness follows from one property: different seeds produce different pseudorandom
Read → glossarySystem prompt
The instructions that steer a model's output according to what the surrounding application needs.
Read → glossaryTest generation
Deriving test cases from named design techniques rather than writing them ad hoc.
Read → glossaryText-to-image
Producing an image from a description. The approach that generalised the task treats text and image as one autoregressive stream of tokens, rather
Read → glossaryTraining opt-out
A rightholder's refusal that their works be mined to train models.
Read → glossaryUpscaling
Raising the resolution of an existing image by learning the transformation rather than interpolating pixels.
Read → glossaryWorkflow orchestration
The layer that runs a multi-step process rather than performing its steps.
Read →