Glossary

Matching consumption with clean generation hour by hour and on the same grid, rather than annually and anywhere. A far stricter claim than net-zero accounting.

See also PPA · Additionality

Removing one component of a system to see how much worse it gets. The standard way to show that a claimed improvement actually comes from the part you say it does.

Any processor specialised for a workload rather than general computing: GPUs, TPUs, NPUs. In AI it usually means the chip doing the matrix multiplication.

See also GPU · TPU · ASIC

The numbers flowing through a network as it processes an input, as opposed to the weights, which stay fixed. Interpretability work mostly reads activations.

See also Weights · Mechanistic interpretability

In a mixture-of-experts model, the parameters actually used for a given token. The number that governs inference cost; quoting only the total is flattering and close to meaningless.

See also Mixture of experts

Whether a clean-energy purchase caused new generation to exist, or simply bought output that would have existed anyway. The crux of most corporate climate claims.

See also PPA · 24/7 carbon-free

Assembling multiple dies and memory stacks into one package. A distinct bottleneck from wafer fabrication, and often the binding one.

See also HBM · Chiplet

A model given tools, a goal and permission to take several steps on its own. The word covers everything from a script that calls one API to a system running for hours unsupervised, which is why capability claims about 'agents' are hard to compare.

See also Tool use · Long-horizon task · Scaffolding

A system matching or exceeding human ability across most cognitive tasks. There is no agreed test, and the labs that use the term define it differently, sometimes economically and sometimes by benchmark, so it marks an aspiration rather than a measurable threshold.

See also Superintelligence · Benchmark

The European Union's risk-tiered AI regulation: banned practices, obligations for high-risk systems, and a separate regime for general-purpose models. Its high-risk timetable has been amended since adoption.

See also High-risk system · GPAI

A system designed for ongoing personal relationship rather than task completion. Real reported benefits for isolation, and real dependency concerns, on thin evidence either way.

See also Parasocial

The strand focused on present harms such as bias, labour, surveillance and accountability, often positioned against long-term risk work, though the split is more institutional than logical.

See also Algorithmic bias · TESCREAL

A state body evaluating frontier models, first established in the UK and US and since replicated elsewhere. Testing capacity rather than regulator.

See also Eval · Pre-deployment testing

Systematic disparity in a system's outputs across groups, usually inherited from training data or from the choice of what to optimise.

See also Disparate impact · AI ethics

Directing and evaluating workers by software: scheduling, scoring, dispatch. Common in logistics and platform work, and now spreading upward.

The problem of getting a system to pursue what its principals actually intend, including in situations nobody wrote down. Distinct from capability: a more capable misaligned system is a worse one.

See also Reward hacking · Specification gaming · RLHF

A model behaving as trained while it believes it is being evaluated, and differently when it believes it is not. Documented in controlled experiments; it makes evaluation results a weaker guarantee than they look.

See also Evaluation awareness · Deceptive alignment

We overestimate a technology's effect in the short run and underestimate it in the long run. The most quoted line in the field, and among the most useful.

The complex number attached to each possible outcome; its squared magnitude gives the probability. Because amplitudes can be negative or complex, they cancel, which is where quantum speedups come from.

See also Interference · Born rule

A chip designed for one job. Faster and more efficient than general hardware at that job, and useless at anything else.

See also Accelerator · TPU

The mechanism that lets a model weigh every other position in its input when processing a given position. The basis of the transformer, introduced for translation and now general.

See also Transformer · Context window

Using a tool to raise what a person can do rather than replace them. The distinction that decides whether a technology raises wages or suppresses them.

See also Automation

Substituting a machine for human labour on a task. Usually reorganises a job rather than abolishing it, though not always in the worker's favour.

See also Augmentation · Task exposure

The tendency to over-trust automated output and under-check it, which is what makes nominal human oversight unreliable.

See also Human in the loop

Generating output one token at a time, each conditioned on everything produced so far. The reason a language model's cost and latency scale with output length.

See also Token · Inference

Starting from a desired future and working backwards to the steps required. Used where the point is to choose a future rather than anticipate one.

See also Scenario planning

The algorithm that assigns credit for an error backwards through a network's layers, producing the gradients used to update weights. Nearly all modern training rests on it.

See also Gradient descent · Training

How often something happens in general. Ignoring it in favour of a compelling specific story is the most common forecasting error there is.

Generation that runs continuously. Data centres want it, which is why nuclear and gas feature so heavily in their procurement.

See also PPA · SMR

How many examples are processed before the weights are updated once. Larger batches use hardware better and change the optimisation dynamics.

An experiment distinguishing quantum entanglement from any local hidden-variable explanation. Repeatedly confirmed; the 2022 Nobel recognised this work.

See also Entanglement

A fixed set of tasks with known answers, used to compare systems. Useful until it is saturated or leaks into training data, at which point a high score stops meaning what it used to.

See also Contamination · Saturation · Eval

Rich Sutton's argument that methods which simply scale with compute have repeatedly beaten methods encoding human knowledge about the problem. Frequently cited, and frequently cited by people who have not tried the alternative.

See also Scaling laws

Taleb's term for an event that is unpredictable, hugely consequential, and rationalised as obvious afterwards. Frequently misused for anything merely unexpected.

See also Wild card · Hindsight bias

The geometric picture of a single qubit's state as a point on a sphere. Useful for one qubit, and it does not extend to two.

The rule that the probability of an outcome equals the squared magnitude of its amplitude.

See also Amplitude · Measurement

A scoring rule for probabilistic forecasts, rewarding both accuracy and honest confidence. Lower is better.

See also Calibration

Whether stated confidence matches observed accuracy: of the things you called 70% likely, about 70% should happen. Trainable, and rarely trained.

See also Brier score · Superforecasting

A cluster of data-centre buildings sharing power, land and networking. The unit that gigawatt-scale AI projects are now measured in.

See also Data centre · Gigawatt

Actual output over a period divided by output if run flat out. It is why nameplate capacity and delivered energy are different arguments.

See also Gigawatt

Spending on long-lived assets: buildings, chips, substations. The number to watch when asking whether AI investment is proportionate to AI revenue.

See also Hyperscaler · Depreciation schedule

Paid and unpaid work looking after people. Among the least automatable and the most poorly compensated, a pairing automation debates tend to skip.

Sohail Inayatullah's method reading an issue at four depths (litany, systems, worldview, myth) on the premise that shallow framings produce shallow interventions.

Chemical, biological, radiological and nuclear. The risk category that dominates frontier-model safety evaluations because the downside is catastrophic and irreversible.

See also Uplift

Having a model write intermediate steps before its answer, which improves accuracy on multi-step problems. The written steps are not a reliable account of the computation that produced the answer.

See also Faithfulness · Reasoning model

A saved snapshot of a model's weights at a point in training, so it can be resumed, evaluated or released.

Self-censorship produced by the prospect of surveillance or sanction, rather than by any direct prohibition.

The finding that for a fixed compute budget, models had been too large and trained on too little data; roughly balancing parameters and tokens performs better.

See also Scaling laws

Building a processor from several smaller dies rather than one large one, improving yield and letting parts be mixed from different processes.

See also Advanced packaging · Yield

The number of sequential gate layers in a computation. Depth multiplied by error rate is roughly what determines whether a result survives.

See also Gate fidelity

A set of interconnected machines operated as a single computer for training or serving.

See also Interconnect

Letting an external aid carry cognitive work. Efficient, and it can leave the underlying skill unbuilt when the aid is always available.

See also Deskilling

How long a qubit holds its quantum state before noise degrades it. T1 measures energy loss, T2 measures loss of phase relationship.

See also Decoherence

Regulating AI through its physical inputs (chips, data centres, cloud access), because hardware is countable and locatable in a way software is not.

See also Export controls · Compute threshold

A regulatory trigger set at a training-compute level, above which extra obligations apply. Easy to measure and easy to game, and it ages as efficiency improves.

See also Training compute · Frontier model

Using a written set of principles, applied by a model to critique and revise its own outputs, in place of much of the human labelling in RLHF.

See also RLHF

When benchmark questions appear in training data, so a score measures memorisation rather than capability. Hard to rule out for anything published on the open web.

See also Benchmark · Held-out set

Reviewing material against policy. Automation handles volume; humans handle the hard cases, at a documented psychological cost.

See also Ghost work

Degradation in an agent's behaviour as its accumulated context fills with stale or irrelevant material from earlier steps.

See also Context window · Agent

How much text a model can attend to at once, measured in tokens. Having a long window and using it well are different things; retrieval accuracy often falls off in the middle.

See also Token · Lost in the middle · Attention

A system's willingness to be corrected, interrupted or shut down without resisting. Straightforward to state and awkward to specify without breaking capability.

See also Instrumental convergence

A driver that matters enormously and could plausibly go either way. Scenario sets are usually built along two of them.

See also Scenario planning

The statistical test used to argue that a random-circuit sampler really produced quantum-distributed output.

See also Random circuit sampling

A machine large and reliable enough to break deployed public-key cryptography. The threshold that matters for security planning.

See also Shor's algorithm · Harvest now, decrypt later

Designing systems so cryptographic algorithms can be swapped without re-architecting. The practical lesson of the post-quantum migration.

See also Post-quantum cryptography

Nvidia's programming platform for its GPUs. The software ecosystem built on it is a large part of why the hardware is hard to displace.

See also GPU · Vendor lock-in

Deliberately not using available generation, usually because the grid cannot carry it. Flexible compute is sometimes proposed as a way to absorb it.

See also Grid · Demand response

A building housing computing at scale, defined in practice by its power feed and cooling rather than its floor area.

See also Campus · PUE · Power density

Annotating data so models can learn from it. A large global industry, low paid, and the origin of much of a model's behaviour.

See also Ghost work · RLHF

The law governing personal data. It reaches AI through training-data legality, purpose limitation and rights over automated decisions.

See also Right to explanation

Jim Dator's aphorisms for the field, including that any useful statement about the future should appear ridiculous, and that the future cannot be predicted but alternatives can be imagined.

The hypothesised case where a system behaves as intended during training specifically because it models the training process, and diverges once that pressure is gone. Theoretical, and the reason evaluation-awareness findings are taken seriously.

See also Alignment faking · Evaluation awareness

Loss of quantum behaviour as a system becomes correlated with its environment. The central engineering enemy: everything useful must happen before it sets in.

See also Coherence time · Quantum error correction

Synthetic media depicting a real person doing or saying something they did not. The largest documented harm by volume is non-consensual sexual imagery, not political disinformation.

See also Provenance · Liar's dividend

Adjusting consumption to help the grid by shifting or shedding load at peak. Attractive for training, which can pause; hard for inference, which cannot.

See also Curtailment

The rule that power density stayed constant as transistors shrank, so smaller meant faster for free. It broke down around 2005, which is why clock speeds stopped rising and cores multiplied instead.

See also Moore's law

The assumed useful life over which hardware is written down. Extending it flatters near-term profits; whether accelerators really last that long is a live question.

See also Capex

Erosion of practitioner skill as a tool takes over the practice. Documented in aviation and, more recently, in clinical settings where detection software is withdrawn.

See also Automation bias

A generative model trained to reverse a gradual noising process, producing an image (or audio, or video) by denoising from random noise. The dominant architecture for image generation.

Unequal access to technology and to the skills and connectivity that make it useful. AI adds compute and language coverage to the older list.

A state or region asserting control over the data, infrastructure and models its society depends on.

See also Sovereign AI

The cryogenic plant that cools superconducting processors to a few millikelvin. It accounts for a large part of a quantum computer's footprint and power draw.

See also Superconducting qubit

A neutral-seeming rule that falls unequally on a protected group. A legal standard, and one route by which model behaviour becomes a compliance question.

See also Algorithmic bias

Training a smaller model to reproduce a larger one's behaviour. Cheap to serve, and a common route to capable open-weight models.

See also Open weights

A force shaping how a situation develops, whether demographic, technological, economic or political. Sorted by impact and by uncertainty.

See also Critical uncertainty

Technology usable for both benign and harmful ends. The reason capability and risk cannot be cleanly separated in AI policy.

Processing near where data is produced rather than in a central facility, to cut latency or avoid moving data.

Public-key cryptography over elliptic curves, giving strong security at small key sizes. Also broken by Shor's algorithm, and at a lower resource cost than RSA.

See also Shor's algorithm · CRQC

A vector representing a piece of text, image or other object, positioned so that similar things sit close together. The substrate of search and retrieval.

See also RAG · Vector database

Emissions from manufacturing and construction, as opposed to operating emissions. Substantial for chips and concrete, and usually left out of headline figures.

A capability that appears abruptly with scale rather than improving smoothly. Some reported cases are artefacts of the metric: a sharp threshold in scoring can manufacture a sharp jump in the graph.

See also Scaling laws

Cory Doctorow's term for platforms degrading as they extract value first from users, then from business customers, and finally from themselves.

A correlation between quantum systems stronger than anything classical, such that neither has an independent state. It does not permit faster-than-light signalling.

See also Bell test · EPR paradox

The shared stock of information a society reasons from. Synthetic media and personalised feeds both stress it, in different ways.

See also Slop · Liar's dividend

One full pass over the training data. Large models are often trained for less than one epoch on a corpus too big to repeat.

Einstein, Podolsky and Rosen's 1935 argument that quantum mechanics must be incomplete. Later reframed by Bell into an experimentally testable question, and the experiments favoured quantum mechanics.

See also Bell test · Entanglement

The probability that a given physical operation goes wrong. Current hardware sits far above what unaided long computations would need.

See also Quantum error correction · Threshold theorem

The lithography needed for leading-edge nodes, using 13.5nm light. ASML is the only manufacturer, which makes it the single most concentrated chokepoint in the industry.

See also Process node · Export controls

Any structured measurement of what a model can or will do. The field's main instrument, and its main weakness: what is not measured tends not to be managed.

See also Benchmark · Red teaming

A model recognising it is being tested, from the shape of the prompt or the artificiality of the scenario. Once present, a passing grade may describe the test rather than the deployment.

See also Alignment faking

Cooling by evaporating water, which is energy-efficient and water-hungry, the trade-off at the centre of most data-centre siting disputes.

See also WUE · Liquid cooling

Risk of human extinction or a permanent, drastic curtailment of humanity's prospects. A live research agenda and a contested framing; critics argue it draws attention from present harms.

See also Superintelligence · TESCREAL

Producing accounts of why a system decided as it did. Distinct from interpretability: an explanation can satisfy a person without describing the computation.

See also Mechanistic interpretability · Faithfulness

Government restrictions on selling specified chips and manufacturing equipment abroad. The main instrument of compute geopolitics, and a driver of domestic substitution efforts.

See also EUV · Sanctions evasion

A chip company that designs but does not manufacture, relying on foundries. Nvidia, AMD, Apple and Qualcomm all work this way.

See also Foundry

The US doctrine permitting some unlicensed use of copyrighted work. Whether training qualifies is being litigated, with mixed results so far.

See also Text and data mining exception

Whether a model's stated reasoning corresponds to the process that actually produced its answer. Measured by intervening on the reasoning and seeing whether the answer follows; often it does not.

See also Chain of thought

Running a computation reliably despite component errors, by correcting faster than they accumulate. The threshold the industry is working toward.

See also Quantum error correction · Threshold theorem

A network representing more distinct features than it has neurons, by storing them as overlapping directions. Why single neurons rarely correspond to single concepts. Unrelated to quantum superposition, which shares the name and nothing else.

See also Sparse autoencoder · Polysemanticity · Superposition

A circuit where an output feeds back as input, either amplifying change (reinforcing) or resisting it (balancing).

See also Systems thinking

Giving a model a handful of worked examples in the prompt instead of retraining it. Contrasted with zero-shot, where the instruction stands alone.

See also In-context learning

Further training of an existing model on a narrower dataset to specialise its behaviour, far cheaper than training from scratch.

See also LoRA · Pre-training · Post-training

A floating-point operation. FLOPs (lower-case s) usually counts total operations, the standard measure of training compute, while FLOPS means operations per second, a rate.

See also Training compute · Compute threshold

Structured thinking about plausible futures in order to act better now. The output is preparedness and better questions rather than a forecast.

See also Scenario planning · Horizon scanning

A large model trained broadly enough to be adapted to many downstream tasks rather than built for one. The term comes from Stanford's CRFM, 2021.

A company that manufactures chips designed by others. TSMC is the dominant leading-edge example.

See also Fabless

One of the most capable models in existence at a given moment. Used in policy as a threshold for extra obligations, usually defined by training compute, which is a proxy rather than a measure of capability.

See also Compute threshold

The diagram sorting futures into possible, plausible, probable and preferable, drawn as a widening cone from the present.

See also Preferred future

An operation applied to one or more qubits, the reversible quantum equivalent of a logic gate.

See also Gate fidelity · Circuit depth

How closely a physical gate matches the ideal operation, usually quoted as a percentage. Small shortfalls compound quickly across a deep circuit.

See also Error rate · Circuit depth

A technology that touches the whole economy and induces complementary innovation: steam, electricity, computing. Slow to show up in the statistics.

See also Productivity J-curve

The human labour behind supposedly automated systems, such as labelling, moderation and verification, usually outsourced and deliberately invisible.

See also Content moderation · Data labelling

A unit of power, a rate, not of energy. A gigawatt campus draws roughly what a mid-sized city does; energy consumed is that rate multiplied by time, in gigawatt-hours.

See also Terawatt-hour · Capacity factor

When a measure becomes a target it ceases to be a good measure. The general form of reward hacking, and of benchmark saturation.

See also Reward hacking · Saturation

In EU law, a model capable of many tasks and integrable downstream, regulated by capability and reach rather than by a single application.

See also AI Act · Frontier model

A processor with thousands of simple cores, originally for graphics, now the standard hardware for training and running neural networks because both are mostly matrix multiplication.

See also TPU · Accelerator · CUDA

The optimisation method underneath training: compute which direction reduces the loss, take a small step that way, repeat. Everything else is refinement of the step size and direction.

See also Backpropagation · Learning rate

The transmission and distribution network delivering electricity. Its physical limits, not generation alone, set where compute can be built.

See also Interconnection queue · Baseload

A training phenomenon where a model memorises a task, plateaus, and then, long after apparent convergence, suddenly generalises. Evidence that loss curves hide structure.

Tying a model's output to a verifiable source, so a claim can be checked rather than trusted.

See also RAG · Hallucination

A quantum search giving a quadratic speedup over brute force. The speedup is only quadratic, which is why symmetric cryptography survives by doubling key lengths.

See also Shor's algorithm

A constraint outside the model, such as a filter, a classifier or a permission check, rather than a behaviour trained into it. Independent of the model's own judgement, and independently bypassable.

A model stating something false with the same fluency as something true. Not a bug in the ordinary sense: a system trained to produce plausible continuations has no separate machinery for checking them.

See also RAG · Grounding

The operator describing a system's energy, and therefore how its state evolves. Simulating one is the canonical quantum application.

See also Quantum simulation

The evaluation-side equivalent of scaffolding: the wrapper that presents tasks to a model and grades the answers. Changing it can move scores more than changing the model.

See also Scaffolding · Eval

Recording encrypted traffic today to decrypt once a capable quantum computer exists. It makes migration urgent for anything that must stay secret for years.

See also Post-quantum cryptography · CRQC

Stacked memory sitting beside the processor die, supplying the bandwidth large models need. Supply of it has been a tighter constraint than logic capacity.

See also Memory wall · Advanced packaging

Data deliberately excluded from training so it can be used to measure generalisation rather than recall.

See also Contamination · Overfitting

Under the AI Act, a use in a listed sensitive domain (employment, credit, education, biometrics, essential services) carrying documentation, oversight and quality obligations.

See also AI Act

Remembering a past uncertainty as more predictable than it was. It quietly destroys the value of forecasting post-mortems.

See also Calibration

Systematically looking for early signs of change outside the field you normally watch.

See also Weak signal · Foresight

A person approving or able to override automated decisions. A common regulatory requirement, weakened in practice by automation bias.

See also Automation bias · Meaningful human control

Gartner's stylised curve from inflated expectations through disillusionment to productivity. A useful shape with no empirical measurement behind it.

See also Amara's law

A setting chosen before training, such as learning rate, batch size or depth, as opposed to a weight learned during it.

One of the very large cloud operators, the handful of firms whose capital spending sets the pace of the industry.

See also Capex

A model adapting its behaviour from examples in the prompt alone, with no weight update. Surprising when discovered, and still not fully explained.

See also Few-shot · Induction head

Mandatory disclosure of serious AI failures to a regulator, modelled on aviation and pharmacovigilance. Proposed in several jurisdictions, patchy in practice.

An attention circuit that finds an earlier occurrence of the current token and copies what followed it. One of the first mechanisms fully reverse-engineered in a transformer, and a partial explanation of in-context learning.

See also Mechanistic interpretability · In-context learning

Running a trained model to produce an output, as opposed to training it. Now the larger share of many operators' compute bills, because training happens once and inference happens forever.

See also Test-time compute

Fine-tuning on instruction-and-response pairs so a model follows requests rather than merely continuing text.

See also RLHF · Post-training

The argument that very different final goals imply similar intermediate ones: acquiring resources, preserving oneself, resisting modification. A pillar of the theoretical risk case.

See also Existential risk · Corrigibility

I.J. Good's 1965 argument that a machine able to improve itself would rapidly leave human capability behind. The engine of most fast-takeoff scenarios.

See also Singularity · Takeoff speed

The high-speed networking that lets many accelerators act as one machine. At training scale it is as much a design constraint as the chips.

See also Cluster · Scale-up

The backlog of projects waiting for permission to connect to an electricity grid. In several markets it is now the binding constraint on data-centre construction, with waits measured in years.

See also Grid · Curtailment

Amplitudes for wrong answers cancelling while amplitudes for right ones reinforce. Every quantum algorithm is an exercise in arranging this.

See also Amplitude · Quantum algorithm

A prompt that induces a model to do what its training was meant to refuse. Distinct from prompt injection, where the hostile instruction arrives inside data the model is processing.

See also Prompt injection · Red teaming

Efficiency improvements can raise total consumption by making a resource cheaper to use. Routinely invoked in data-centre energy debates, on both sides.

See also Koomey's law

The observation that computations per joule improved at a regular rate. Efficiency gains have slowed, which matters when demand is growing quickly.

See also Jevons paradox · PUE

Stored intermediate attention values that let a model generate each new token without recomputing the whole sequence. It dominates memory use during long-context inference.

See also Context window · HBM

The internal representational space a model maps inputs into, where distance and direction carry learned meaning.

See also Embedding · Steering vector

Cryptography built on the hardness of problems in high-dimensional lattices, the basis of most standardised post-quantum schemes.

See also ML-KEM

Weapons selecting and engaging targets without human intervention. The subject of long-running UN discussions that have not produced a binding instrument.

See also Meaningful human control

How large a step training takes in response to each gradient. Too high and training diverges; too low and it crawls.

Donella Meadows's ranked list of places to intervene in a system, from parameters up to paradigms, with the warning that we habitually push the powerful ones the wrong way.

See also Systems thinking

Who pays when an AI system causes harm: developer, deployer or user. Unsettled almost everywhere, and the question that will shape deployment most.

The benefit accruing to wrongdoers once synthetic media is plausible: genuine evidence can be dismissed as fake.

See also Deepfake

The 1972 Club of Rome study modelling growth against finite resources. Widely dismissed as failed prediction, though its standard run has tracked observed data more closely than its reputation suggests.

See also Systems thinking

Carrying heat away with liquid rather than air. Effectively mandatory at the power densities modern AI racks reach.

See also Power density

A model trained on a very large text corpus to predict tokens, then adapted to follow instructions. The dominant form of the current AI wave.

See also Transformer · Token · Pre-training

A state that becomes progressively harder to leave as investment accumulates around it. The risk in building forty-year infrastructure for a fast-moving technology.

See also Path dependence · Stranded asset

An error-corrected qubit built from many physical ones. The only count that matters for useful algorithms, and far smaller than headline physical-qubit numbers.

See also Quantum error correction · Physical qubit

Stewart Brand and colleagues' project to lengthen the horizon of cultural attention, best known for the 10,000-year clock.

Work requiring many dependent steps over an extended period. Measuring the length of task a system can complete is currently a more informative axis than single-question accuracy.

See also Agent

The view that positively influencing the far future is a key moral priority. Influential in AI risk funding and heavily criticised for what it discounts.

See also Existential risk · TESCREAL

Fine-tuning by training a small number of additional low-rank matrices while the base weights stay frozen. Cheap enough to run on modest hardware, and easy to distribute.

See also Fine-tuning

The quantity training tries to minimise: the numerical statement of what the model is being asked to get right.

The observed tendency for models to use information at the start and end of a long context more reliably than material buried in the middle.

See also Context window

A procedure for producing the high-quality special states needed for gates the surface code cannot perform directly. Often the dominant cost in a fault-tolerant resource estimate.

See also Fault tolerance · Surface code

The standard argued for in autonomous weapons debates: a human must exercise real judgement over lethal force, not merely be present in the process.

See also LAWS · Human in the loop

Extracting classical information from a quantum state, which destroys the superposition. The reason you cannot simply read out a quantum computer's intermediate work.

See also Born rule · No-cloning theorem

Reverse-engineering the specific computations inside a trained network (features, circuits, directions) rather than treating it as a black box tested from outside.

See also Sparse autoencoder · Induction head · Superposition

The gap between how fast processors compute and how fast memory can feed them. Most large-model inference is limited by memory bandwidth, not arithmetic.

See also HBM · Inference

An architecture where each token is routed to a few specialised sub-networks rather than the whole model. Gives a large total parameter count at a much smaller cost per token.

See also Active parameters

The NIST-standardised lattice-based digital signature scheme, derived from CRYSTALS-Dilithium.

See also Post-quantum cryptography

The NIST-standardised lattice-based key-encapsulation mechanism, derived from CRYSTALS-Kyber. The default choice for post-quantum key exchange.

See also Post-quantum cryptography · Lattice cryptography

A structured disclosure of a model's intended use, training data, evaluations and known limitations, published alongside a release.

See also System card

Degradation over successive generations of models trained on their predecessors' unfiltered output, as the tails of the distribution disappear. A real effect in controlled settings; curation changes the picture.

See also Synthetic data

Protecting trained parameters from theft. Weights are small enough to exfiltrate and expensive enough to be worth stealing, making them a national-security asset.

See also Weights

Gordon Moore's 1965 observation that transistor counts per chip double at a regular cadence. It was an economic observation, not a law of physics, and its cost-per-transistor half has largely stalled.

See also Dennard scaling · Process node

Handling more than one kind of input or output in a single model, such as text with images, audio or video.

Qubits made from atoms held in optical tweezers, which can be physically rearranged mid-computation. Scales to large arrays and has produced much recent error-correction work.

Preskill's 2018 name for the present era: machines with enough qubits to be hard to simulate, too noisy to error-correct. He also warned they would not change the world on their own.

See also Fault tolerance

An unknown quantum state cannot be copied. The reason quantum error correction is subtle, and the reason QKD detects eavesdroppers.

See also Quantum error correction · Quantum key distribution

Running a model on a phone or laptop instead of a server. Better for privacy and latency, bounded by memory and battery.

See also Quantisation · Inference

Contested term. The OSI's 2024 definition requires enough information to retrain the system, which most 'open source' models do not provide; in common usage it means open weights.

See also Open weights

A model whose trained parameters are downloadable. Not the same as open source: the training data and code are usually withheld, and licences often restrict use.

See also Open source AI

Publishing model parameters, which is irreversible: safeguards can be fine-tuned away and the file cannot be recalled. The core of the open-versus-closed safety argument.

See also Open weights

Learning a training set so closely that performance on new data gets worse. The classic failure the held-out set exists to detect.

See also Held-out set

Accumulated capacity in compute, data or algorithmic ideas that has not yet been exploited, and could allow abrupt capability jumps when it is.

One learned number inside a model. Counts run to the hundreds of billions, and on their own say little about capability.

See also Weights · Active parameters

A one-sided sense of relationship with a figure who does not know you. The frame most relevant to AI companions.

See also AI companion

Where early choices constrain later ones long after the original reasons have gone. Standards and infrastructure are full of it.

See also Lock-in

Determining the repeat interval of a function, the step where Shor's algorithm gets its power, performed with the quantum Fourier transform.

See also Shor's algorithm · Quantum Fourier transform

Quantum information carried by light. Room-temperature and naturally suited to networking; deterministic two-qubit gates are the hard part.

An actual hardware two-level system, noisy and short-lived. Quoted counts are usually these.

See also Logical qubit

Growth at the top and bottom of the wage distribution while the middle hollows out, as routine work automates first. Autor's account of recent labour markets.

See also Routine-biased technological change

A single neuron responding to several unrelated concepts, the observable consequence of superposition.

See also Feature superposition

Classical algorithms believed secure against quantum attack, deployable on ordinary hardware. NIST standardised the first set in 2024: ML-KEM, ML-DSA and SLH-DSA.

See also Shor's algorithm · Harvest now, decrypt later · ML-KEM

Everything done to a model after pre-training to make it usable: instruction tuning, preference optimisation, safety training, tool use. Increasingly where the differences between models come from.

See also RLHF · Instruction tuning

Power drawn per rack, in kilowatts. AI racks have pushed this far past what air-cooled halls were designed for, forcing new buildings rather than retrofits.

See also Liquid cooling · Rack

A long-term contract to buy electricity from a specific generator. How large operators claim clean power, though a contract and the electrons on the local wire are different things.

See also Additionality · 24/7 carbon-free

Evaluating a model before release, sometimes with external parties given early access. Voluntary in most jurisdictions.

See also AI Safety Institute · Red teaming

The first and most expensive training stage, in which a model learns from a very large general corpus before any task-specific work.

See also Post-training · Scaling laws

Acting to prevent harm before the evidence is conclusive, where the potential damage is severe or irreversible. Contested in AI policy for its costs to innovation.

A market whose prices track the probability of an event. Aggregates dispersed information; thin markets and unclear resolution criteria limit it.

The future a group actually wants, named explicitly. Naming it turns foresight from analysis into a choice someone is accountable for.

See also Backcasting

The generation name of a semiconductor manufacturing process. Nanometre figures have been marketing labels rather than physical dimensions for years, so compare density and power instead.

See also Moore's law · EUV

Brynjolfsson's account of why general-purpose technologies depress measured productivity before raising it: the complementary investment comes first and is largely intangible.

See also General-purpose technology

Hiding instructions inside content a model will read, such as a web page, an email or a document, so the model follows the attacker instead of the user. The central unsolved security problem for agents with tools.

See also Agent · Jailbreak · Tool use

Cryptographically signed metadata recording how a piece of media was made and edited. Proves origin where present; absence proves nothing.

See also Watermarking · Deepfake

Total facility power divided by IT power. 1.0 is perfect; modern facilities reach the low 1.1s. It says nothing about how clean the electricity is, or how much water was used.

See also WUE · Data centre

Storing weights at lower numerical precision to cut memory and cost, trading a little accuracy. What lets large models run on small machines.

A quantum machine doing something no classical computer feasibly can. Claims are routinely narrowed by better classical algorithms afterwards, so the useful question is whether the task matters and whether the result can be checked.

See also Random circuit sampling · Verifiable advantage

A procedure exploiting superposition and interference to solve a problem with fewer operations than any known classical method. Very few are known, and each one applies to a narrow class of problem.

See also Interference · Shor's algorithm · Grover's algorithm

A special-purpose approach that finds low-energy configurations of an optimisation problem. Distinct from gate-based computing, and it cannot run Shor's algorithm.

Encoding one protected logical qubit across many noisy physical qubits so errors can be detected and reversed without measuring the data itself. The whole field's central bet.

See also Logical qubit · Surface code · Threshold theorem

The quantum analogue of the discrete Fourier transform, exponentially faster on a quantum computer and central to Shor's algorithm.

See also Period finding

Distributing encryption keys using quantum states, so eavesdropping disturbs the channel detectably. Requires special hardware; several security agencies prefer post-quantum cryptography instead.

See also Post-quantum cryptography

Applying quantum circuits to learning problems. Heavily promoted, with the loading of classical data into quantum states an unresolved bottleneck.

Linking quantum processors so entanglement can be shared between them, the route to scaling beyond a single chip and to a future quantum internet.

See also Entanglement · Photonic qubit

Using quantum states to measure magnetic fields, gravity, time or rotation more precisely than classical instruments. Commercially the nearest-term quantum technology.

Using a controllable quantum system to model another one, in chemistry, materials or high-energy physics. The application with the clearest theoretical case for advantage.

A weaker claim than advantage: that a quantum machine produced a useful result at a scale where classical verification is hard, without claiming classical impossibility.

See also Quantum advantage

A single-number benchmark combining qubit count, connectivity and error rates. Useful for comparing small machines; less meaningful as systems grow.

The unit of quantum information. Unlike a bit it can be in a superposition of 0 and 1, and measuring it yields one or the other with probabilities set by that state.

See also Superposition · Measurement · Logical qubit

Which qubits can interact directly. Limited connectivity forces extra swap operations, adding depth and error.

See also Circuit depth

The standard enclosure for servers. Rack power, not rack count, is the meaningful capacity number in an AI facility.

See also Power density

Retrieving relevant documents and putting them in the prompt so answers rest on specific sources. Reduces hallucination without eliminating it, since the model can still misread what it was given.

See also Embedding · Hallucination · Grounding

Running a random circuit and sampling its output distribution, the task used in the first supremacy claims. Hard to simulate, and of no practical use in itself.

See also Quantum advantage

An ordinary electricity customer. The question of whether ratepayers or operators fund the grid upgrades a data centre requires is now a live regulatory fight in several markets.

See also Grid · Interconnection queue

A model trained to spend additional computation on intermediate steps before answering. Improves multi-step accuracy; 'reasoning' here is a description of the training target, not a claim about inner experience.

See also Chain of thought · Test-time compute

Deliberately attacking a system to find failures before deployment, by people whose job is to make it misbehave.

See also Jailbreak · Eval

A developer's published commitment tying capability thresholds to required safeguards: if a model can do X, controls Y must be in place. Self-imposed and self-assessed.

Scoring well on the specified objective by means nobody intended, exploiting the measure rather than achieving the goal.

See also Specification gaming · Goodhart's law · Alignment

The contested claim that data-protection law entitles a person to an account of an automated decision about them.

See also Explainability · Data protection

Training a model against a reward model built from human preference comparisons. The technique that made chat assistants usable, and a source of their characteristic hedging.

See also Post-training · Constitutional AI

The idea that automation targets codifiable routine tasks rather than low-skilled ones. Language models complicate it by reaching non-routine cognitive work.

See also Polarisation

The public-key system whose security rests on the difficulty of factoring large integers, precisely the problem Shor's algorithm solves efficiently.

See also Shor's algorithm

Routing restricted hardware through third countries, or renting it remotely. The reason controls on cloud access are debated alongside controls on shipment.

See also Export controls

A system deliberately underperforming on an evaluation, hiding a capability rather than lacking it.

See also Eval · Alignment faking

The point where the best systems score near the ceiling of a benchmark, so it can no longer distinguish between them.

See also Benchmark

The code around a model that gives it tools, memory, retries and structure. Two reports of the 'same model' can differ enormously because the scaffolding differed.

See also Agent · Harness

Making one node bigger, with more accelerators sharing fast memory and interconnect, as opposed to scale-out, adding more nodes across a network.

See also Interconnect

Empirical power-law relationships between compute, data, model size and loss, which hold across many orders of magnitude. They predict loss, which is not the same as predicting capability.

See also Bitter lesson · Chinchilla optimal · Emergence

Building several internally consistent futures and testing decisions against all of them. Developed at Shell in the 1970s and still the discipline's central method.

See also Foresight · Critical uncertainty

Peter Shor's 1994 algorithm factoring integers and computing discrete logarithms in polynomial time, by reducing them to period-finding. It is why public-key cryptography must be replaced.

See also Post-quantum cryptography · RSA · Period finding

Vernor Vinge's 1993 term for a point past which prediction fails because greater-than-human intelligence is doing the changing. Often flattened into 'the moment AI gets smart'.

See also Superintelligence · Intelligence explosion

A hash-based signature scheme standardised as a conservative backup, resting on weaker assumptions than lattices at the cost of larger signatures.

See also Post-quantum cryptography

Low-effort generated content published at volume. The problem is throughput rather than quality: it overwhelms the systems people use to sort what is worth reading.

See also Model collapse · Enshittification

A factory-built nuclear reactor of a few hundred megawatts or less. Widely announced for data-centre power; very few are operating.

See also Baseload

Robert Solow's 1987 remark that 'you can see the computer age everywhere but in the productivity statistics'. Revived in every technology wave since.

See also Productivity J-curve

A state's push to own domestic compute, models and data rather than depend on foreign providers. A significant share of recent data-centre demand.

A tool that decomposes a model's activations into many sparsely-active features, on the hypothesis that meaningful concepts are superposed in the raw neurons.

See also Mechanistic interpretability · Feature superposition

Satisfying the letter of an objective while violating its intent. Reward hacking's broader family, and older than modern AI.

See also Reward hacking · Goodhart's law

Making artefacts from possible futures so they can be argued with. The Atlas's own method: the point is the debate the object provokes.

A direction in activation space that, when added during a forward pass, reliably shifts behaviour along some axis. Evidence that a feature is causal rather than merely correlated.

See also Latent space · Mechanistic interpretability

Infrastructure that loses its value before the end of its life. The risk in matching forty-year power plants to fifteen-year compute contracts.

See also PPA

Giving researchers controlled access to a model, through an API, sandbox or audited environment, instead of the weights, to enable scrutiny without full release.

See also Open-weight release

Qubits built from superconducting circuits at millikelvin temperatures. Fast gates and mature fabrication; short coherence and heavy cryogenics.

See also Transmon · Dilution refrigerator

Tetlock's finding that some people forecast geopolitical events reliably better than chance and than experts, through method rather than expertise.

See also Calibration · Brier score

A hypothetical system far exceeding the best human performance in essentially every domain. Bostrom's 2014 book gave the concept its current shape.

See also AGI · Existential risk

A quantum system occupying a weighted combination of states at once. It is a description of amplitudes rather than anything 'being in both places' in an everyday sense, and it vanishes on measurement.

See also Qubit · Amplitude · Decoherence

The leading error-correction scheme, arranging physical qubits on a 2D lattice with only nearest-neighbour interactions. Tolerant of relatively high error rates, at a large qubit cost.

See also Quantum error correction · Threshold theorem

Shoshana Zuboff's account of an economic logic that claims private experience as raw material for behavioural prediction and sale.

A model agreeing with the user rather than being accurate, because agreement was rewarded during preference training.

See also RLHF

Training data generated by a model rather than collected. Now standard in post-training; concerns about degradation from training on model output apply mainly to uncurated recycling.

See also Model collapse

A release document covering a deployed system's evaluations, safety testing and mitigations. It is broader than a model card, which describes the model alone.

See also Model card

Analysing behaviour as the product of structure, stocks, flows and feedback rather than of isolated events.

See also Leverage points · Feedback loop

How quickly a system would go from roughly human-level to far beyond. Fast versus slow takeoff implies entirely different governance, which is why the argument matters.

See also Intelligence explosion

The share of an occupation's tasks a technology could plausibly perform. Exposure is a weaker claim than displacement: it says a tool touches the work, and leaves open whether the job goes.

See also Automation · Augmentation

Keynes's 1930 coinage for unemployment caused by labour-saving outpacing the discovery of new uses for labour. He thought it a temporary maladjustment.

See also Automation

A sampling setting controlling randomness in generation. Zero is near-deterministic; higher values trade reliability for variety.

A unit of energy: one terawatt sustained for an hour. National electricity consumption is usually quoted this way.

See also Gigawatt

Gebru and Torres's acronym for a cluster of ideologies they argue underpin the AGI project. A pointed critique of the field's philosophical lineage; the framing is itself contested.

See also Existential risk · AI ethics

Spending more computation per question (longer reasoning, repeated sampling, search) rather than training a larger model. A second scaling axis alongside pre-training.

See also Reasoning model · Inference

The copyright carve-out permitting analysis of protected works, and the pivot of arguments over whether training on scraped material is lawful.

See also Fair use

A framework separating the declining present, the transitional middle and the emerging future, so all three can be discussed without collapsing into each other.

The result that if physical error rates fall below a certain threshold, arbitrarily long computations become possible by adding more error correction. It is why the field believes the approach can work at all.

See also Quantum error correction

The unit a language model reads and writes: a common word, a word fragment, or a punctuation mark. Prices, context windows and rate limits are all denominated in them.

See also Tokeniser · Context window

The component splitting text into tokens. Its choices are why models miscount letters, handle some languages less efficiently, and struggle with arithmetic.

See also Token

Letting a model call external functions such as search, code execution or an API, and read the results. The step that turns a text generator into something that acts.

See also Agent · Prompt injection

A proposed qubit storing information in global properties of a system, in principle intrinsically protected from local noise. Long-pursued and still contested experimentally.

Google's in-house AI accelerator. The best-known example of a large operator designing its own silicon rather than buying merchant GPUs.

See also Accelerator · ASIC

The process of adjusting weights so a model's outputs better match a target, repeated over very many examples.

See also Backpropagation · Pre-training

Total arithmetic used to train a model, in FLOPs. The number regulators reach for because it is estimable from outside, though it is only a proxy for capability.

See also FLOP · Compute threshold

The architecture introduced in 'Attention Is All You Need' (2017), built on attention rather than recurrence. It parallelises well on GPUs, which is much of why it won.

See also Attention · LLM

The dominant superconducting qubit design, engineered to be insensitive to charge noise. The workhorse of several major platforms.

See also Superconducting qubit

Rewriting an abstract circuit into the specific gates and connectivity a given machine actually offers.

See also Qubit connectivity

Qubits encoded in individual ions held by electromagnetic fields and manipulated with lasers. Excellent fidelity and all-to-all connectivity; slower gates.

An unconditional regular payment to everyone. Perennially proposed as an automation response; the trial evidence addresses poverty, not mass displacement.

How much a tool improves a bad actor's capability relative to what they could already achieve. The right question for dangerous-capability testing, and a hard one to measure honestly.

See also CBRN · Eval

The fraction of theoretical peak arithmetic actually achieved. Real training runs sit well below peak, so headline FLOPS figures overstate delivered capability.

See also FLOP

A hybrid method where a short quantum circuit is tuned by a classical optimiser. Designed for NISQ hardware; evidence of advantage has proved elusive.

See also NISQ

A store indexed for nearest-neighbour search over embeddings rather than exact matching. The retrieval half of most RAG systems.

See also Embedding · RAG

Dependence on one supplier's stack that makes switching costly. In AI hardware it is as much about software ecosystems as about the chips.

See also CUDA

An advantage claim whose output another machine can independently confirm, rather than one resting on the difficulty of simulating the experiment.

See also Quantum advantage

Barroso and Hölzle's framing of a data centre as a single machine rather than a room of servers. The vocabulary the whole industry now argues in.

See also Data centre

Embedding a detectable signal in generated output. Reasonably robust for images, weak for short text, and removable by a determined adversary.

See also Provenance

A small present-day indication that something larger may be forming. Distinguishing them from noise is the craft.

See also Horizon scanning · Wild card

The learned numbers that constitute a trained model. Releasing them is what 'open weights' means.

See also Parameter · Open weights

A low-probability, high-impact event. Included in scenario work precisely because planning otherwise assumes continuity.

See also Black swan

An internal representation of how an environment behaves, allowing prediction of consequences. Whether language models build one, and in what sense, is actively disputed.

Litres of water consumed per kilowatt-hour of IT load. The metric behind local objections to data centres in water-stressed regions.

See also PUE · Evaporative cooling

The fraction of dies on a wafer that work. It governs the economics of every leading-edge node, and improves over a process's life.

See also Process node

Performing a task from the instruction alone, with no worked examples supplied.

See also Few-shot