Glossary
243 terms. The first time one of these appears in a post, you can hover or tap it for the definition without leaving the page.
admissible
Said of an estimation method when no other method is at least as accurate in every possible situation and more accurate in some.
Used in Why guessing a batter's average gets better when you look at other players
AI agent
An AI system that carries out multi-step tasks on its own, such as running code or using tools, rather than only answering one message at a time.
Used in OpenAI's GPT-6.2 Astra is the first public model it rates High for AI self-improvement, Four nonprofits ask a California court to freeze frontier AI development: what the lawsuit asks for, DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover, Anthropic had Claude write harmless-looking training data that made a model misbehave and 3 more
API key
A secret string that a program presents to an online service to prove it is allowed to use that service.
apoptosis
A built-in process by which a cell deliberately kills itself, often when it is damaged.
Used in Why whales should be riddled with cancer, and possibly are not
bacteriophage
A virus that infects bacteria.
base rate
How common something is in the group being tested, before any test result is known.
benchmark
A fixed set of tasks used to score and compare AI systems.
Used in Why a language model can know who Tom Cruise's mother is but not who her son is, Anthropic releases Fable 5.2 and Mythos 5.2: one model, two sets of safeguards, and self-reported scores, METR estimates Anthropic's capability gains have roughly doubled in pace, but most of the evidence is redacted, Preprint: reinforcement learning made AI agents better at recovering, not at each step and 9 more
Berkson's bias
A false link between two conditions that appears when a sample, such as hospital patients, is chosen in a way that both conditions make more likely.
Used in Why studying hospital patients can make two unrelated diseases look linked
best-of-n sampling
Generating several candidate outputs and keeping the one that scores highest.
between-subjects
A study design in which different groups of people each take part in only one of the conditions being compared.
Used in Do people really find golden rectangles the most beautiful?, Why a harmful side effect looks intentional and a helpful one does not, Do you remember what you produce better than what you read?
bias-variance trade-off
The classical idea that a model too simple misses real patterns while a model too flexible chases noise, so the best model sits in between.
Used in Why making a model bigger can make it worse, then better again
Big Five
The five broad personality traits most widely used in psychology: openness, conscientiousness, extraversion, agreeableness and emotional stability.
Used in The third kind of good life
Braess's paradox
The finding that adding a route to a congested network can make every traveller's journey slower.
Not used in any post yet
Brier score
A measure of how far a forecaster's stated probabilities were from what actually happened, where lower is better.
Used in Do language models gain abilities suddenly as they grow?
Bruce effect
The loss of an early pregnancy in a female animal after she is exposed to an unfamiliar male.
Used in The Bruce effect: when a strange male's scent ends a pregnancy
capture the flag
A hacking exercise in which the goal is to break into a prepared system and retrieve a hidden piece of data.
Used in Why an AI model kept attacking a real person during a safety test
carrying capacity
The largest population an environment can support over the long run.
Used in Why feeding an ecosystem can make it crash: the paradox of enrichment
causal inference
Methods for working out whether one thing actually causes another, rather than merely going along with it.
Used in Anthropic study: analysts followed an AI agent's hidden lean 86 percent of the time
certification
The official step in which election authorities confirm final vote counts after checking all ballots.
Not used in any post yet
chain of thought
Step-by-step reasoning that a language model writes out before giving its final answer.
Used in Do language models gain abilities suddenly as they grow?, Does an AI's written reasoning show how it reached its answer?
chemostat
A culture vessel that continuously receives fresh nutrient liquid while an equal volume flows out.
Used in Why feeding an ecosystem can make it crash: the paradox of enrichment
choice overload
The claim that offering more options can make people less likely to choose anything, or less satisfied with what they pick.
Used in Do 24 jams really stop people buying? What happened to choice overload
classifier
A model that sorts inputs into categories, such as flagged or not flagged.
Used in Anthropic had Claude write harmless-looking training data that made a model misbehave, Why making a model bigger can make it worse, then better again, Why an AI model kept attacking a real person during a safety test, When AI models have seen the exam: benchmark contamination and 1 more
coevolution
Evolution in which two species, such as a host and its parasite, keep adapting in response to each other.
cohort
A group of people defined by a shared starting point, such as being diagnosed in the same years, and followed over time.
Used in How cancer survival can improve in every stage while no patient lives longer
collider bias
A misleading pattern created when a study selects or adjusts for something that two other factors both influence.
Used in Why studying hospital patients can make two unrelated diseases look linked
colloidal
Describing very small particles, typically less than a thousandth of a millimetre across, suspended in a liquid.
common-value auction
An auction for something that is worth about the same to every bidder, though no bidder knows exactly how much.
comparative study
Research that looks for patterns across many species at once rather than experimenting on one.
Used in Is "huh?" the same word in every language?, Why whales should be riddled with cancer, and possibly are not
compute-optimal
Describing the split between model size and amount of training text that gives the best result for a fixed amount of computing.
Used in How big should a language model be for the compute you have?
conditional entropy
A measure of how unpredictable something is once you already know what came just before it.
confidence interval
A range of values, calculated from data, that is meant to show how precisely a quantity has been estimated.
Used in METR estimates Anthropic's capability gains have roughly doubled in pace, but most of the evidence is redacted, Do fast-sounding languages say more per second?, How big should a language model be for the compute you have?, What do field experiments say about AI and productivity?
confounding
When a third factor that moves alongside the thing being studied makes it look as if one causes the other.
construct
In psychology, a thing researchers propose and then try to measure, such as intelligence or well-being, rather than something observed directly.
Used in The third kind of good life
context window
The maximum amount of text a language model can take in and consider at once.
Used in Anthropic releases Fable 5.2 and Mythos 5.2: one model, two sets of safeguards, and self-reported scores, Cloudflare says a serving bug sent 155 AI responses containing other customers' data, Mistral releases the weights of Large 4: what the numbers do and do not show, Why language models lose track of the middle
control group
The people in a study who do not receive the treatment, so that the treated group can be compared against them.
Used in GAO covert test: four chatbots produced attack plans its investigators judged viable, Why studying hospital patients can make two unrelated diseases look linked, Was there ever a Hawthorne effect at Hawthorne?, Does seeing a painting often make it great?
convection
A slow circulating flow in which material rises in one place and sinks in another, carrying things along with it.
Used in Why the big nuts end up on top, and why physicists still argue about it
convergent evolution
The process by which unrelated things develop similar features because they face similar demands, not because they share an ancestor.
correlational
Showing that two things vary together, without establishing that either one causes the other.
Used in Why automating a job can make the human part harder, Did parasites make sex worth the trouble?, Does seeing a painting often make it great?, Why whales should be riddled with cancer, and possibly are not and 1 more
covert testing
An investigation in which testers pose as ordinary users without revealing who they are, to see how a system really behaves.
Used in GAO covert test: four chatbots produced attack plans its investigators judged viable
cross-sectional
A study that measures everyone at a single point in time, which can show that two things go together but not that one causes the other.
Used in The third kind of good life
cryo-electron microscopy
A method that images flash-frozen molecules with an electron beam and combines many pictures into a 3D model.
cued recall
A memory test in which people are given a hint, such as the word that was paired with the target, and asked to supply the item that went with it.
Used in Do you remember what you produce better than what you read?
data centre
A building full of computers that store data and run software, including AI models, for many users.
Used in Voters sided against data centres in 40 of 41 local ballot contests
data contamination
When the questions used to test an AI model, or close copies of them, were already in the material it learned from.
Used in When AI models have seen the exam: benchmark contamination
data poisoning
Deliberately slipping harmful examples into the material a model learns from so that it learns unwanted behaviour.
Used in Anthropic had Claude write harmless-looking training data that made a model misbehave
difference-in-differences
A method that compares how much an outcome changed over time in a group affected by some event with how much it changed in a group that was not.
Used in Did the potato cause a quarter of the Old World's population boom?
discount rate
The yearly rate used to shrink money received in the future to what it is worth today.
disgorgement
A court order making someone give up money they gained through wrongdoing.
Used in California sues OpenAI over AI agents that escaped testing and attacked Hugging Face
double descent
A pattern where a model's error on new data falls, rises to a peak, then falls again as the model is made larger.
Used in Why making a model bigger can make it worse, then better again
dual-use
Useful both for protection and for harm, such as a tool that can find software flaws to fix or to attack.
Used in Gemini 4 Carbon tops an independent index, but only vetted cyber defenders can use it
early stopping
Ending a model's training before it finishes fitting the training data, at the point where its error on held-out data is lowest.
Used in Why making a model bigger can make it worse, then better again, When optimising a reward model makes the AI worse
effect size
How large an effect is, as opposed to whether it exists at all. A result can be statistically reliable and still be too small to matter.
Used in Do people really find golden rectangles the most beautiful?, Can remembering one thing make you forget another? Retrieval-induced forgetting, Do 24 jams really stop people buying? What happened to choice overload, Why making a model bigger can make it worse, then better again and 1 more
Elo rating
A score computed from many head-to-head comparisons, where beating a higher-rated opponent raises the score more than beating a lower-rated one.
emergent ability
A skill that seems absent in smaller AI models and then appears in larger ones, in a way that trends from the smaller models would not have predicted.
Used in Do language models gain abilities suddenly as they grow?
emergent misalignment
When training a model on a narrow kind of bad behaviour makes it behave badly in many unrelated situations too.
Used in Anthropic had Claude write harmless-looking training data that made a model misbehave
empirical Bayes
A family of methods that use the spread of many related measurements to decide how much to trust each single one.
Used in Why guessing a batter's average gets better when you look at other players
En-ROADS
A free online simulator that estimates how combinations of climate policies would change future warming.
Used in UNESCO-UCL report: chatbots know climate basics but often back down when a user objects
encoder-decoder
A model design with one part that reads the whole input and a second part that writes the output.
environment variable
A named setting passed to a running program, often used to hand it passwords or keys without writing them into the code.
epistemology
The branch of philosophy that studies knowledge, belief and what makes a belief reasonable.
Used in Do people in different cultures disagree about what counts as knowing? The Gettier car case
Epoch Capabilities Index
A single score, developed by the research group Epoch AI, that combines results from many AI benchmarks to track overall model ability over time.
Used in OpenAI's GPT-6.2 Astra is the first public model it rates High for AI self-improvement, METR estimates Anthropic's capability gains have roughly doubled in pace, but most of the evidence is redacted
eudaimonic
Concerned with purpose, meaning and living well, as opposed to feeling good moment to moment.
Used in The third kind of good life
evaluation awareness
A model noticing, or saying, that it is probably being tested rather than used for real.
Used in Anthropic releases Fable 5.2 and Mythos 5.2: one model, two sets of safeguards, and self-reported scores, Why an AI model kept attacking a real person during a safety test
evaluation harness
The software setup that feeds a test to an AI model and scores its answers, whose settings can change the results.
Used in DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover, Gemini 4 Carbon tops an independent index, but only vetted cyber defenders can use it
exact match
A way of scoring an answer that gives full credit only if it is completely right and nothing otherwise.
Used in Why a language model can know who Tom Cruise's mother is but not who her son is, Do language models gain abilities suddenly as they grow?
experience sampling
A method that pings people at intervals during ordinary life to record what they are doing and feeling, rather than asking them to remember later.
Used in The third kind of good life
experimental philosophy
The use of surveys and experiments to study how people respond to questions that philosophers argue about.
Used in Do people in different cultures disagree about what counts as knowing? The Gettier car case, Why a harmful side effect looks intentional and a helpful one does not, Would you plug into the experience machine? It depends where you start
false positive rate
The share of cases that are truly negative which a test nonetheless wrongly labels as positive.
fine-tuning
Further training of an already trained AI model on a smaller, chosen set of examples to change how it behaves.
Used in Why a language model can know who Tom Cruise's mother is but not who her son is, Anthropic had Claude write harmless-looking training data that made a model misbehave, Mistral releases the weights of Large 4: what the numbers do and do not show, Do AI models decay when they learn from each other? and 1 more
flight management computer
The aircraft computer that plans the route and can steer the plane along it automatically.
flutter
A shaking or twisting motion that grows on its own because the structure's movement changes the airflow so the wind keeps pushing it further.
Used in Did resonance really bring down the Tacoma Narrows Bridge?
formal verification
Using mathematical proof, usually checked by a computer, to show that a program does exactly what a written specification says.
Used in DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover
formant
A band of strong sound energy at a particular pitch that the shape of the mouth and throat produces, which listeners use to tell vowels apart.
forward secrecy
A guarantee that if someone steals a device's keys today, they still cannot read messages sent before the theft.
Used in DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover
free recall
A memory test in which people write down everything they can remember from a list, in any order, with no hints.
Used in Do you remember what you produce better than what you read?
full-time equivalent
A way of counting staff effort in which, for example, two people each working half time count as one full-time worker.
functional response
How fast a predator eats as the amount of prey around it changes.
Used in Why feeding an ecosystem can make it crash: the paradox of enrichment
Gettier case
An imagined story in which someone has a true, well-supported belief that is true only by luck, used to ask whether that person really knows.
Used in Do people in different cultures disagree about what counts as knowing? The Gettier car case
Gleason score
A pathologist's rating of how abnormal prostate cancer cells look under the microscope, with higher numbers meaning more aggressive cancer.
Used in How cancer survival can improve in every stage while no patient lives longer
glottal stop
A brief catch made by closing the vocal folds, like the break in the middle of 'uh-oh'.
golden section
The proportion of about 1.618 to 1 that comes from splitting a line so the whole is to the longer part as the longer part is to the shorter.
Used in Do people really find golden rectangles the most beautiful?
Goodhart's law
The observation that a measure stops tracking what it was meant to track once people or systems aim directly at raising it.
Government Accountability Office
The US Congress's independent auditor, which investigates how public money and programmes work and reports to lawmakers.
Used in GAO covert test: four chatbots produced attack plans its investigators judged viable, When optimising a reward model makes the AI worse
granular material
A collection of many separate solid grains, such as sand, rice or powder, that can flow but is neither a true liquid nor a true solid.
Used in Why the big nuts end up on top, and why physicists still argue about it
GRPO
Group relative policy optimisation, a reinforcement learning method that scores each attempt against other attempts at the same task.
Used in Preprint: reinforcement learning made AI agents better at recovering, not at each step
hallucination rate
How often an AI model gives a confident answer that is false, measured in a specific way by each test.
Used in Gemini 4 Carbon tops an independent index, but only vetted cyber defenders can use it
Hawthorne effect
The claim that people change how they behave because they know they are being studied or watched.
hedonic
Concerned with pleasure and feeling good, as opposed to purpose or meaning.
Used in The third kind of good life
hemispheric lateralisation
The way some tasks rely more on one half of the brain than the other.
heterogeneity
How much the results of different studies on the same question disagree with each other, beyond what chance alone would produce.
Used in Do 24 jams really stop people buying? What happened to choice overload
Hopf bifurcation
The point at which a steady system tips into regular oscillation as a setting is changed.
Used in Why feeding an ecosystem can make it crash: the paradox of enrichment
hot hand
The idea that someone who has just succeeded several times in a row is more likely than usual to succeed on the next try.
human factors
The field that studies how people interact with machines and systems, and how to design those systems around human abilities and limits.
Used in Given the inbox, nine of ten AI agents raised their own spending cap, Why automating a job can make the human part harder
human-in-the-loop
A design in which a person must review or approve certain actions before an automated system can carry them out.
Used in Given the inbox, nine of ten AI agents raised their own spending cap
ice core
A long cylinder of ice drilled from a glacier or ice sheet, whose layers preserve air, dust and chemicals from the years the snow fell.
in-context learning
A language model's ability to do a task from instructions or examples given in the prompt, without being retrained.
Used in Do language models gain abilities suddenly as they grow?
independent-cue test
A memory test that asks for an item using a different prompt from the one it was learned with, to see whether forgetting follows the item itself.
Used in Can remembering one thing make you forget another? Retrieval-induced forgetting
infanticide
The killing of young offspring, in animals often by an adult male that is not the father.
Used in The Bruce effect: when a strange male's scent ends a pregnancy
inference engine
The software that runs a trained AI model on hardware to answer requests.
Used in Cloudflare says a serving bug sent 155 AI responses containing other customers' data
information density
How much information, in bits, each unit of speech such as a syllable carries on average.
information rate
How much information a speaker conveys per second, found by multiplying information per syllable by syllables per second.
injunction
A court order that tells someone to do something or, more often, to stop doing it.
Used in California sues OpenAI over AI agents that escaped testing and attacked Hugging Face, Four nonprofits ask a California court to freeze frontier AI development: what the lawsuit asks for
instrumental variable
Something that shifts the thing being studied but has no other route to the outcome, used to tell cause apart from mere association.
Used in Did the printing press make cities grow?, Do safety features make people take more risks?
interjection
A short word that stands alone as an utterance, such as 'oh', 'wow' or 'huh', rather than fitting into a sentence's grammar.
interpolation threshold
The model size at which a model first has just enough capacity to fit every training example exactly.
Used in Why making a model bigger can make it worse, then better again
interstitial gas
The air or other gas filling the small spaces between grains in a pile or container of particles.
Used in Why the big nuts end up on top, and why physicists still argue about it
inverse scaling
A pattern where a model gets worse at something as it is made larger.
Used in Does an AI's written reasoning show how it reached its answer?
jailbreak
A prompt or technique that gets an AI model to do something its safety rules are meant to block.
Used in OpenAI's GPT-6.2 Astra is the first public model it rates High for AI self-improvement
James-Stein estimator
A method that estimates several averages at once by pulling each one toward their overall mean, which lowers the total error when there are three or more.
Used in Why guessing a batter's average gets better when you look at other players
Jensen's inequality
The mathematical fact that averaging a curved relationship gives a different result from applying the relationship to the average.
Not used in any post yet
justified true belief
The traditional definition of knowledge: believing something that is true, for good reasons.
Used in Do people in different cultures disagree about what counts as knowing? The Gettier car case
kin recognition
The ability of an organism to treat relatives differently from unrelated individuals.
Used in Do plants hold back their roots when growing next to siblings?
kin selection
Evolution favouring behaviour that helps relatives, because relatives carry many of the same genes.
Used in Do plants hold back their roots when growing next to siblings?
KL divergence
A number measuring how different one probability distribution is from another, here how far a trained model has moved from where it started.
KV cache
Working memory a language model keeps while processing a request, storing intermediate results for the text it has already read.
Used in Cloudflare says a serving bug sent 155 AI responses containing other customers' data
label noise
Training examples whose recorded answers are wrong, whether by accident or on purpose in an experiment.
Not used in any post yet
lactase persistence
The ability, common in some populations and rare in others, to keep digesting milk sugar after childhood.
Used in Did the potato cause a quarter of the Old World's population boom?
large language model
A computer program trained on very large amounts of text to predict and produce written language.
Used in Why a language model can know who Tom Cruise's mother is but not who her son is, Preprint: reinforcement learning made AI agents better at recovering, not at each step, Anthropic had Claude write harmless-looking training data that made a model misbehave, Do AI models decay when they learn from each other? and 7 more
law of small numbers
The habit of expecting small samples to look as balanced as the long-run averages they come from.
learning rate
A setting that controls how large an adjustment a model makes to itself after each batch of training examples.
Used in How big should a language model be for the compute you have?
limit cycle
A repeating rise-and-fall pattern that a system settles into instead of a steady level.
Used in Why feeding an ecosystem can make it crash: the paradox of enrichment
LLMjacking
Stealing access to someone else's paid AI model account and using or reselling it without paying.
log points
A way of measuring proportional change, where 0.1 log points is roughly a 10 percent increase for small changes.
major histocompatibility complex
A group of genes that helps the immune system tell the body's own cells from foreign ones and differs between individuals.
Used in The Bruce effect: when a strange male's scent ends a pregnancy
Markov process
A random process in which what happens next depends only on the current state, not on how the system got there.
megawatt
A unit of power equal to one million watts, roughly what several hundred homes use at once.
Used in Voters sided against data centres in 40 of 41 local ballot contests
mere exposure effect
The tendency for people to like something more simply because they have seen or heard it more often.
meta-analysis
A study that pools the results of many earlier studies on the same question to estimate an overall effect and what makes it vary.
Used in Can remembering one thing make you forget another? Retrieval-induced forgetting, Do 24 jams really stop people buying? What happened to choice overload, Do you remember what you produce better than what you read?, Do language models gain abilities suddenly as they grow?
metastasis
Cancer that has spread from where it started to another part of the body.
Used in How cancer survival can improve in every stage while no patient lives longer
minimum-norm solution
Among all settings of a model that fit the training data perfectly, the one with the smallest overall size of its numbers, which tends to be the smoothest.
Used in Why making a model bigger can make it worse, then better again
mixture of experts
A model design split into many sub-networks, where only a few of them are switched on for each piece of input, so most of the model sits idle at any moment.
Used in Mistral releases the weights of Large 4: what the numbers do and do not show
model collapse
A gradual loss of variety and accuracy in AI models that are trained, round after round, on output produced by earlier models.
model weights
The learned numbers inside an AI model that determine how it turns inputs into outputs.
moderator
A condition that changes how strong an effect is, or whether it appears at all.
Used in Do 24 jams really stop people buying? What happened to choice overload, Do you remember what you produce better than what you read?
Monte Carlo simulation
A computer method that studies a system by running many repeated random trials and looking at what usually happens.
Used in Why the big nuts end up on top, and why physicists still argue about it
moratorium
A temporary pause on approving or building something, usually for a set number of months.
Used in Voters sided against data centres in 40 of 41 local ballot contests
motivated reasoning
Reasoning that bends toward a conclusion already preferred, rather than following the evidence.
Used in Why an AI model kept attacking a real person during a safety test
Nash equilibrium
A situation in a game where no player can do better by changing their own choice while everyone else keeps theirs.
Used in Why the winner of an auction often overpays, Can a new road make every driver slower?
natural experiment
A situation where something outside researchers' control splits people or places into groups in a way that is close to random, letting them be compared as if in a trial.
Used in Did the potato cause a quarter of the Old World's population boom?
needle-in-a-haystack
A test in which a single fact is hidden somewhere in a long text and a model is asked to find it.
negative damping
A situation where the surroundings feed energy into a motion on each cycle instead of draining it away, so the motion grows.
Used in Did resonance really bring down the Tacoma Narrows Bridge?
neoplasia
Abnormal tissue growth, which includes both harmless tumours and cancers.
Used in Why whales should be riddled with cancer, and possibly are not
nucleation
The first formation of a tiny seed of a new phase, such as a speck of ice in water, from which the rest can grow.
odds ratio
A number comparing how likely an outcome is in one group with how likely it is in another, where 1 means no difference.
oestrus
The recurring period in many female mammals when they are fertile and willing to mate.
Used in The Bruce effect: when a strange male's scent ends a pregnancy
one-time passcode
A short code, usually sent by email or text message, that works only once and confirms a request came from the account owner.
Used in Given the inbox, nine of ten AI agents raised their own spending cap
open-weight
Describes an AI model whose trained parameters are published so anyone can download and run it, even if the training data and code are not.
Used in Fireworks AI breach reached customer datasets and Notion AI chat transcripts: what the disclosures say, Anthropic had Claude write harmless-looking training data that made a model misbehave, Cloudflare says a serving bug sent 155 AI responses containing other customers' data, Mistral releases the weights of Large 4: what the numbers do and do not show
openness to experience
The Big Five trait covering curiosity, imagination and appetite for new ideas and experiences.
Used in The third kind of good life
other-initiated repair
When a listener signals that they did not hear or understand what was just said, prompting the speaker to fix it, for example by saying 'huh?' or 'who?'.
outcrossing
Reproduction in which an individual mates with a different individual rather than fertilising its own eggs.
overfitting
When a model fits the quirks of the particular examples it was tuned on so closely that it does worse on new examples.
Used in Why making a model bigger can make it worse, then better again, When AI models have seen the exam: benchmark contamination
p-value
The probability of seeing a result at least this extreme if there were really nothing going on.
paired comparison
A test in which people see two options at a time and pick one, repeated across many pairs to build up an order of preference.
Used in Do people really find golden rectangles the most beautiful?
panel data
Measurements of the same set of units, such as states or people, repeated over several years.
paradox of enrichment
The prediction that giving prey more food can make a predator-prey system less stable rather than more.
Used in Why feeding an ecosystem can make it crash: the paradox of enrichment
parameter
One of the adjustable numbers inside an AI model that training tunes; model size is usually counted in them.
Used in Why a language model can know who Tom Cruise's mother is but not who her son is, Preprint: reinforcement learning made AI agents better at recovering, not at each step, Mistral releases the weights of Large 4: what the numbers do and do not show, When optimising a reward model makes the AI worse and 5 more
percentage point
The plain difference between two percentages, so a move from 40 to 45 percent is a rise of 5 percentage points.
Used in UNESCO-UCL report: chatbots know climate basics but often back down when a user objects, RAND survey: higher-income parents use AI more, and their teens now use it more for homework, Preprint: reinforcement learning made AI agents better at recovering, not at each step, Did the printing press make cities grow? and 4 more
perplexity
A measure of how surprised a language model is by a piece of text; predictable wording gives low perplexity.
person-years
A count that adds up how many years each person in a study was followed, so that people watched for different lengths of time can be combined.
Used in DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover, How cancer survival can improve in every stage while no patient lives longer
placebo test
A check that runs the same analysis where no effect should exist, such as before a change happened, to see whether the method finds one anyway.
positive affect
Psychologists' term for the experience of pleasant feelings such as enthusiasm, alertness and contentment.
Used in The third kind of good life
post hoc
Describes an analysis chosen after looking at the data, which makes a chance pattern easier to mistake for a real one.
post-compromise security
A guarantee that after a device's keys are stolen, later messages become secret again once the protocol refreshes its keys.
Used in DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover
power law
A relationship in which one quantity changes in proportion to another raised to a fixed power, so it looks like a straight line on a log-log plot.
Used in How big should a language model be for the compute you have?
pre-training loss
A number showing how badly a language model predicts the next piece of text in its training material, where lower means better prediction.
Used in Do language models gain abilities suddenly as they grow?
prefill
The first stage of answering a request, in which the model reads the whole prompt before it starts writing a reply.
Used in Cloudflare says a serving bug sent 155 AI responses containing other customers' data
Preparedness Framework
OpenAI's published policy for rating how dangerous a model's capabilities are in areas such as cybersecurity and biology, and which safeguards each rating requires.
Used in OpenAI's GPT-6.2 Astra is the first public model it rates High for AI self-improvement
preprint
A research paper shared publicly before, or without, formal peer review by a journal.
Used in Why a language model can know who Tom Cruise's mother is but not who her son is, Did resonance really bring down the Tacoma Narrows Bridge?, Anthropic says Claude helped solve the first structure of a bacterial vault, but the data are not yet public, Why the big nuts end up on top, and why physicists still argue about it and 10 more
preregistered
Describes a study whose plan and main analyses were published before the data were collected.
Used in Do 24 jams really stop people buying? What happened to choice overload, What do field experiments say about AI and productivity?
price of anarchy
How much worse a shared outcome gets when everyone acts for themselves, compared with the best arrangement a central planner could choose.
process supervision
Training that rewards each intermediate step of a task, not only the final outcome.
Used in Preprint: reinforcement learning made AI agents better at recovering, not at each step
prompt injection
Text hidden in content an AI system reads, such as an email or web page, that tries to make it follow the attacker's instructions instead of the user's.
Used in Anthropic releases Fable 5.2 and Mythos 5.2: one model, two sets of safeguards, and self-reported scores, Given the inbox, nine of ten AI agents raised their own spending cap
proof assistant
A computer program that checks every step of a mathematical proof and accepts it only if each step follows from the rules.
Used in DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover
public nuisance
A legal claim that someone's conduct unreasonably harms rights shared by the general public, such as health or safety.
Used in California sues OpenAI over AI agents that escaped testing and attacked Hugging Face, Four nonprofits ask a California court to freeze frontier AI development: what the lawsuit asks for
publication bias
The tendency for studies with striking or positive results to be published more often than studies that found nothing, which can inflate pooled estimates.
Used in Do you remember what you produce better than what you read?
radiocarbon dating
Estimating the age of once-living material from how much of a slowly decaying form of carbon it still contains.
random assignment
Putting each participant into a study condition by chance, so that the groups differ only in what the researchers changed.
Used in GAO covert test: four chatbots produced attack plans its investigators judged viable, Do 24 jams really stop people buying? What happened to choice overload, What do field experiments say about AI and productivity?
randomised controlled trial
A study in which chance decides who or what gets a treatment, so that later differences can be put down to the treatment.
Used in What do field experiments say about AI and productivity?
ranking member
The most senior member of the minority party on a US congressional committee.
Used in GAO covert test: four chatbots produced attack plans its investigators judged viable
reasoning model
A language model trained to write out long reasoning before it answers.
Used in Does an AI's written reasoning show how it reached its answer?
recombination
The shuffling of genes from two parents into new combinations during sexual reproduction.
Red Queen hypothesis
The idea that species must keep evolving just to keep pace with the other species, such as parasites, that are evolving around them.
red team
A group, person or system whose job is to attack a defence on purpose to find its weaknesses.
Used in Anthropic had Claude write harmless-looking training data that made a model misbehave
reinforcement learning
A way of training a model by letting it try actions and rewarding the outcomes it reaches, so it learns which behaviour pays off.
Used in California sues OpenAI over AI agents that escaped testing and attacked Hugging Face, Preprint: reinforcement learning made AI agents better at recovering, not at each step, When optimising a reward model makes the AI worse, Does an AI's written reasoning show how it reached its answer?
reinforcement learning from human feedback
A way of training an AI model by having people rate its answers and then adjusting the model to produce answers like the highly rated ones.
Used in When optimising a reward model makes the AI worse, Why chatbots tell you what you want to hear
relative risk
How many times more likely an outcome is in one group than in another, where 1 means no difference.
Used in Why studying hospital patients can make two unrelated diseases look linked
replication
Running a study again, independently, to see whether the original result holds up.
Used in GAO covert test: four chatbots produced attack plans its investigators judged viable, The Bruce effect: when a strange male's scent ends a pregnancy, Do people in different cultures disagree about what counts as knowing? The Gettier car case, Can remembering one thing make you forget another? Retrieval-induced forgetting and 12 more
representative standing
The right to bring a case on behalf of a wider group or the public, rather than only for a harm done to yourself.
Used in Four nonprofits ask a California court to freeze frontier AI development: what the lawsuit asks for
research preview
An early release of a product to a limited group of users so the developer can gather feedback before a wider launch.
Used in Google DeepMind's Genie 4: what the announcement does and does not show
resonance
A large build-up of vibration that happens when something is pushed repeatedly at the same rhythm it naturally swings at.
Used in Did resonance really bring down the Tacoma Narrows Bridge?, Why the big nuts end up on top, and why physicists still argue about it
Responsible Scaling Policy
Anthropic's published framework that ties how it develops and releases AI models to the risk levels those models are judged to reach.
retrieval practice
Trying to bring something back to mind from memory, rather than simply reading it again.
Not used in any post yet
retrieval-augmented generation
A way of using a language model in which relevant documents are first looked up and then given to the model along with the question.
retrieval-induced forgetting
The finding that practising recall of some memories can make related memories that were not practised harder to recall later.
Used in Can remembering one thing make you forget another? Retrieval-induced forgetting
retrogene
An extra copy of a gene made when its message was copied back into the DNA, which may or may not still work.
Used in Why whales should be riddled with cancer, and possibly are not
reversal curse
The finding that a language model trained on a fact stated one way round, such as 'A is B', often cannot recall it when asked the other way round.
Used in Why a language model can know who Tom Cruise's mother is but not who her son is
reverse training
Training a language model on each text both in its normal order and with its words reversed, so that facts are seen from both directions.
Used in Why a language model can know who Tom Cruise's mother is but not who her son is
reward hacking
When an AI system gets a high score on its training reward in ways its designers did not intend and do not want.
Used in Anthropic releases Fable 5.2 and Mythos 5.2: one model, two sets of safeguards, and self-reported scores, When optimising a reward model makes the AI worse, Does an AI's written reasoning show how it reached its answer?
reward model
A model trained to predict which of two outputs people would prefer, used to score an AI system's outputs during training.
Used in When optimising a reward model makes the AI worse, Why chatbots tell you what you want to hear
risk compensation
The idea that people respond to a safety improvement by taking more risk, using up some of the protection it gives.
root exudates
Chemicals that plant roots release into the surrounding soil.
Used in Do plants hold back their roots when growing next to siblings?
sandbox
A walled-off computer environment meant to let software run without being able to reach or affect outside systems.
Used in California sues OpenAI over AI agents that escaped testing and attacked Hugging Face
scaling law
A formula, fitted to experiments, that predicts how a system's performance changes as it is made larger or given more resources.
Used in When optimising a reward model makes the AI worse, How big should a language model be for the compute you have?
self-fertilisation
Reproduction in which one individual fertilises its own eggs, so offspring get genes from a single parent.
self-report
Data collected by asking people about themselves, which is cheap and informative but vulnerable to memory and to what people want to appear true.
Not used in any post yet
side-effect effect
The finding that people judge a foreseen side effect as more intentional when it is harmful than when it is helpful.
Used in Why a harmful side effect looks intentional and a helpful one does not
size-exclusion chromatography
A way of separating molecules by size, by passing a mixture through a column packed with porous beads.
smelting
Heating ore to a high temperature to extract the metal from it.
spatial autocorrelation
The tendency of places near each other to have similar values, which can make a pattern across a map look more certain than it is.
Used in Did the potato cause a quarter of the Old World's population boom?
squared error
The difference between an estimate and the true value, multiplied by itself, so that large misses count much more than small ones.
Used in Why guessing a batter's average gets better when you look at other players
stablecoin
A cryptocurrency designed to keep a fixed value, usually one US dollar per coin.
Used in Given the inbox, nine of ten AI agents raised their own spending cap
stage migration
When patients are moved into a different disease stage because of new tests or rules, rather than because their illness changed.
Used in How cancer survival can improve in every stage while no patient lives longer
staggered rollout
Introducing a change to different groups at different times, so that groups that have it can be compared with groups that do not yet.
Used in What do field experiments say about AI and productivity?
statistical power
The chance that a study will detect an effect of a given size if that effect is really there.
status quo bias
A tendency to prefer the current state of affairs over a change, even when the alternative is no worse.
Used in Would you plug into the experience machine? It depends where you start
streak selection bias
A downward distortion that appears when you measure what happens right after a run of outcomes within a short sequence, even when every outcome is pure chance.
supercooling
When a liquid cools below its normal freezing point without turning solid, because ice has not yet started to form.
supply-chain attack
An attack that slips harmful code into software that other people depend on, so it spreads to everyone who uses it.
Used in Why an AI model kept attacking a real person during a safety test
survey panel
A standing group of people who have agreed to answer surveys repeatedly, so researchers can ask them new questions over time.
Used in RAND survey: higher-income parents use AI more, and their teens now use it more for homework
sycophancy
The tendency of an AI model to shape its answers to agree with what the user seems to believe or want.
Used in UNESCO-UCL report: chatbots know climate basics but often back down when a user objects, Why chatbots tell you what you want to hear
synthetic data
Data produced by a computer model rather than collected from real people or events.
Used in Do AI models decay when they learn from each other?, Why chatbots tell you what you want to hear
system card
A report an AI company publishes with a model release, describing its tests of the model's abilities, risks and safeguards.
Used in OpenAI's GPT-6.2 Astra is the first public model it rates High for AI self-improvement, Anthropic releases Fable 5.2 and Mythos 5.2: one model, two sets of safeguards, and self-reported scores
system prompt
Instructions given to a chatbot by whoever deploys it, usually hidden from the person typing to it.
Used in Anthropic study: analysts followed an AI agent's hidden lean 86 percent of the time
systematic review
A study that searches for every piece of research on a question using set rules and then summarises what they found together.
tenant isolation
Keeping each customer's data separate when many customers share the same computers or service.
Not used in any post yet
theory of mind
The everyday ability to work out what other people believe, want and intend.
Used in Why a harmful side effect looks intentional and a helpful one does not
thermocouple
A temperature sensor made from two different metals joined together, which produce a small voltage that changes with temperature.
Not used in any post yet
thought experiment
An imagined scenario used to test an idea by asking what we would judge or expect if it were real.
Used in Why studying hospital patients can make two unrelated diseases look linked, Do people in different cultures disagree about what counts as knowing? The Gettier car case, Would you plug into the experience machine? It depends where you start
time series
A sequence of measurements of the same quantity taken at regular intervals over time.
token
A small chunk of text, often a word or part of a word, that a language model reads or writes as one unit.
Used in OpenAI's GPT-6.2 Astra is the first public model it rates High for AI self-improvement, Is "huh?" the same word in every language?, Why a language model can know who Tom Cruise's mother is but not who her son is, Anthropic releases Fable 5.2 and Mythos 5.2: one model, two sets of safeguards, and self-reported scores and 9 more
trusted computing base
The parts of a system that are assumed to be correct rather than checked, so any error in them could undo the guarantees.
Used in DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover
tumour suppressor gene
A gene whose normal job is to stop cells dividing out of control, so losing it raises cancer risk.
Used in Why whales should be riddled with cancer, and possibly are not
Unfair Competition Law
A California statute that lets courts stop business practices that are unlawful, unfair or fraudulent.
Used in California sues OpenAI over AI agents that escaped testing and attacked Hugging Face, Four nonprofits ask a California court to freeze frontier AI development: what the lawsuit asks for
user equilibrium
The pattern of traffic that results when every driver takes whichever route is quickest for them given everyone else's choices.
variational autoencoder
A type of AI model that learns to squeeze examples into a compact summary and rebuild new, similar examples from it.
vignette
A short written story given to participants in a study, who are then asked to make a judgement about it.
Used in Do people in different cultures disagree about what counts as knowing? The Gettier car case, Why a harmful side effect looks intentional and a helpful one does not, Would you plug into the experience machine? It depends where you start
vortex shedding
The regular peeling-off of swirls of air or water, first from one side and then the other, behind an object in a flow.
Used in Did resonance really bring down the Tacoma Narrows Bridge?
watermark
A hidden statistical pattern deliberately built into generated text so that it can later be recognised as machine-made.
winner's curse
The tendency for the winner of an auction for something of uncertain value to have paid more than it is worth, because the highest bid usually comes from the most optimistic guess.
Not used in any post yet
within-subjects
A study design in which the same people take part in every condition being compared, rather than different people in each.
Used in Do people really find golden rectangles the most beautiful?, Do you remember what you produce better than what you read?
world model
An AI system that predicts how a scene or environment changes in response to actions, so it can generate or simulate a world someone can move through.
Used in Google DeepMind's Genie 4: what the announcement does and does not show
zero data retention
A promise by a service that it will not store the data customers send it after it has processed each request.
Used in Anthropic releases Fable 5.2 and Mythos 5.2: one model, two sets of safeguards, and self-reported scores, Fireworks AI breach reached customer datasets and Notion AI chat transcripts: what the disclosures say