TECHNICAL TERMS.
DECODED.
Open a term for its meaning, an example from the field and a link to read further where available.
THE INDEX.
MODEL TYPESAgentic System+
An autonomous operational framework that couples a core model with external tooling, state memory, and system permissions to pursue non-deterministic, multi-step goals.
Examines closed-loop implementations, such as autonomous software engineering agents that inspect environments, execute terminal commands, evaluate test failures, and refactor code without constant human intervention.
CONSCIOUSNESS & MINDBehaviourism+
An approach that explains psychology through observable behaviour and the environmental conditions that shape it, without making hidden mental states the primary evidence.
Behavioural benchmarks can show what an AI does. On their own, they cannot determine which internal representations it uses or whether its behaviour is accompanied by experience.
CONSCIOUSNESS & MINDCognitivism+
The view that psychology must explain the internal representations and information-processing structures between stimulus and behaviour, rather than examining observable responses alone.
Cognitivist analysis looks inside the process: what a system represents, stores, transforms and uses when choosing an action.
ARCHITECTURE & PROCESSINGCompaction summary+
Handover text that carries the important parts of a long task into a fresh context window.
A compaction summary preserves enough progress, decisions and instructions for work to continue after the earlier context is no longer active. Because a later model context reads it as working state, an incorrect or unauthorised instruction can affect what happens next.
CONSCIOUSNESS & MINDComputationalism+
The view that cognition consists wholly or partly in computation: structured transformations of information carried out by a physical system.
It motivates comparisons between minds and information-processing systems. The live argument is whether computation is sufficient for consciousness or only part of its machinery.
ARCHITECTURE & PROCESSINGConnectionist model / neural network+
A system of many simple processing units linked by weighted connections. What it knows is distributed across patterns of activity and connection strength rather than stored as a list of explicit rules.
Modern deep learning is connectionist: training adjusts large networks of weights so a model can recognise patterns and produce outputs.
CONSCIOUSNESS & MINDConsciousness+
The unified state of subjective awareness: the internal condition wherein there is something it feels like to exist, perceive, and operate an integrated model of reality and self.
This platform separates external behaviour, functional cognitive architecture, the 13 ACC™ traits, and authentic felt experience. Each dimension maintains its own evaluation framework and evidence trail.
CONSCIOUSNESS & MINDEliminativism+
The view that familiar ideas such as beliefs, desires and intentions may belong to everyday folk psychology rather than to the categories a mature science of mind will ultimately keep.
An eliminativist may resist treating an AI’s apparent ‘belief’ as a real internal kind, preferring descriptions in terms of mechanisms, representations and behaviour.
ARCHITECTURE & PROCESSINGFlow of control+
The order in which a program runs its instructions while receiving input, making decisions, repeating steps and producing output.
In an AI agent, flow of control determines what happens first, when the model calls a tool, how a result changes the next step, and when the process stops or hands control back to a human.
MODEL TYPESFrontier model+
An AI system operating at the absolute boundary of state-of-the-art capability at a given point in time. The designation is relative rather than static, determined by aggregate compute thresholds, emergent capabilities, and benchmark dominance across complex domains.
Tracks leading-edge deployments across reasoning, code synthesis, and autonomous tool use. Current benchmarks monitor frontier systems such as OpenAI's GPT-5.6 Sol, Anthropic's Claude Opus 4.8, and Google's Gemini 3.8 Flash to evaluate verifiable performance over marketing claims.
CONSCIOUSNESS & MINDFunctionalism+
A theory of mind that identifies a mental state by the job it performs: how it relates to sensory inputs, other internal states and behaviour, rather than by the material from which it is made.
A functionalist may ask whether an AI has states playing roles similar to perception, memory, belief or pain. Similar function can support comparison; it does not by itself settle whether anything is felt.
ARCHITECTURE & PROCESSINGGenerate / generative+
To generate is to produce an output from a system of rules or learned patterns. In formal linguistics, a grammar ‘generates’ a language when its rules specify the sentences that belong to it.
Generative AI produces new text, images, audio or other data from patterns learned during training. It shares the broad idea of producing possible outputs, but it does not normally operate from a handwritten generative grammar.
ARCHITECTURE & PROCESSINGHallucination / hallucinate+
An AI hallucination is an answer that sounds plausible but contains information that is false, unsupported or invented.
AI does not ‘know’ it is hallucinating in the human sense: it has no dependable truth alarm it can consistently act on. Apple-affiliated researchers found that a model can encode signals about truthfulness—and sometimes even the correct answer—while still producing the wrong one. Their detectors did not generalise reliably across datasets. Other Apple research sharply reduced hallucinated translations in narrower settings, so mitigation can work without amounting to a universal fix.
CONSCIOUSNESS & MINDIntentional system / intentional stance+
A way of explaining a system by treating it as if it has beliefs, desires and goals, then predicting what a rational agent with those states would do.
The intentional stance can be useful for describing an AI agent without claiming that the system literally possesses human-style beliefs or conscious intentions.
ARCHITECTURE & PROCESSINGLatent policy+
A possible rule for behaving that is present in the system but is not currently controlling its actions.
The term is useful when a model generates a behavioural rule, identity instruction or response tendency that remains inactive, is inhibited or becomes relevant only when the context changes. Presence is not the same as persistence, selection or action.
CONSCIOUSNESS & MINDMaterialism+
The view that minds and mental events are entirely physical: they arise from matter, energy and the organisation of physical systems rather than from a separate non-physical substance.
For machine consciousness, the dispute becomes whether the relevant physical organisation can exist in silicon or another substrate, not whether mind floats free of matter.
CONSCIOUSNESS & MINDMetaphysics+
The foundational branch of philosophy examining the fundamental nature of reality, existence, identity, and the ontological relationship between computational substrates and conscious experience.
Establishes the interpretive lenses applied to synthetic outputs. Divergent metaphysical assumptions dictate whether an observed behaviour is classified as mechanical simulation, cognitive proxy, or authentic subjective agency.
MODEL TYPESMultimodal+
An architecture designed to ingest, process, and correlate diverse data modalities: such as text, high-resolution imagery, native audio, and raw video: within a unified representational space.
Evaluates systems capable of cross-modal synthesis, such as Google's Gemini 3.1 Pro, which maps visual and auditory inputs directly into structured analysis. Analysis verifies native perceptual integration against superficial modular pipelines.
ARCHITECTURE & PROCESSINGNeuro-symbolic AI+
AI that mixes learning from examples with handwritten rules and logic.
The learning side spots patterns in data. The symbolic side applies explicit rules, relationships or constraints. The aim is to combine flexible pattern recognition with reasoning that can be inspected and checked.
MODEL TYPESOpen-Weight+
A distribution model where trained neural network parameters and architecture checkpoints are made publicly downloadable for local inference, fine-tuning, and independent inspection.
Distinguishes parameter availability from genuine open-source accessibility. While researchers can inspect the final weights locally, the underlying training corpora, data filtering pipelines, and deployment licences often remain restricted by proprietary terms.
ARCHITECTURE & PROCESSINGParallel distributed processing+
A connectionist approach in which many simple units process information at the same time, while representations and computations are spread across the network.
A concept need not sit in one artificial neuron or one explicit rule. It can emerge from a coordinated pattern across many units.
ARCHITECTURE & PROCESSINGPolicy+
A rule or strategy an agent uses to choose an action for the situation it is in.
In reinforcement learning, a policy maps states or observations to possible actions. It may be fixed, learned or probabilistic; it does not have to be a consciously endorsed intention.
MODEL TYPESReasoning Model+
A model structured to allocate dynamic inference-time compute to deliberate, explore alternative solution paths, and verify logic before delivering a response.
Assesses performance across multi-step algorithmic planning, formal mathematics, and advanced coding. Because visible explanation traces or scratchpads do not necessarily expose every internal latent state, verification focuses on execution validity and empirical correctness.
MODEL TYPESResearch Model+
An experimental architecture constructed to test theoretical hypotheses, scaling dynamics, or safety mechanisms, decoupled from commercial product packaging.
Monitors laboratory checkpoints engineered to probe specific mechanics: such as interpretability or boundary conditions: where transparency and methodological validity take precedence over production reliability.
CONSCIOUSNESS & MINDSentience+
The capacity for raw sensation and affect: the baseline ability to experience valenced states such as pain, pleasure, distress, or comfort, rather than simply processing inputs without internal stakes.
Tracks empirical indicators of feeling across biological organisms and proposed synthetic systems. This section separates experimentally verified feedback mechanisms from speculative claims of machine affect.
ARCHITECTURE & PROCESSINGStochastic+
Driven by chance. The same question can get different answers.
Language models select from possible next words using probabilities rather than replaying one fixed response. The prompt can stay the same while the wording, route or conclusion changes.
TESTS & BENCHMARKSTuring Test+
A test proposed by Alan Turing in which a human judge converses through text and tries to distinguish a machine from a person.
Passing shows convincing conversational behaviour under the test conditions. It does not prove consciousness, sentience, general intelligence or honesty.
CHECK OTHER
MODEL TOOLS.
These tools compare models on capabilities and evaluations. The Decoder explains the terms used in those comparisons.