Glossary

AI industry terms and jargon.

Find

1

13F

A quarterly SEC form where large investment managers disclose the US stocks they hold. It shows positions as of quarter end, weeks late, so the portfolio may have changed by publication. Read more

5

510(k)

The FDA pathway that clears a medical device by showing it is substantially equivalent to one already on the market. No trial of patient outcomes is required, and each clearance can chain off the last. Read more

A

A1

The first filing a company makes to list in Hong Kong, the local counterpart to a US S-1. Until it is public, a Hong Kong listing is still only a plan. Read more

acqui-hire

Buying a company mainly for its people. Regulators care because a license plus mass hiring can transfer a business without a merger filing anyone gets to review. Read more

adjusted operating income

A company’s operating result after selected costs or gains are removed. Check which adjustments were made; a positive result does not establish net profit or positive cash flow. Read more

advanced packaging

The back-end steps that stack and connect finished chips into one module, such as bonding DRAM into HBM. It has become as much a bottleneck for AI hardware as making the chips themselves. Read more

age assurance

Estimating whether a user is a minor from behavior and account signals instead of checking ID. Cheaper and less invasive than verification, and it misclassifies in both directions. Read more

agent

A model given tools and a goal, left to take multiple steps on its own instead of answering one question. The word covers everything from a script that reads your email to software that buys things for you. Read more

Agent Skills

A packaged set of instructions and scripts an AI agent loads to handle a specific kind of task, a standard introduced by Anthropic and since adopted by other coding agents. A skill ships know-how, no new model weights. Read more

agentic commerce

Shopping and payment carried out by an AI agent on a user's behalf, with merchants exposing catalogs and checkout to software instead of to a person browsing a page. Read more

AGI

Artificial General Intelligence. A system that matches humans across most cognitive work, not just one narrow task. Nobody has agreed on a test for it, so claims about its arrival cannot be checked. Read more

AGPL

An open-source license like the GPL with one addition: if you run modified code as a network service, you must share the modifications. Companies selling hosted software often avoid AGPL code for that reason. Read more

AISI

AI Safety Institute. Government bodies, first in the UK and US, that test frontier models for dangerous capability before release. They have testing access but no power to block a launch. Read more

alignment

Making a model reliably do what its operator intended, including in situations nobody wrote a rule for. Distinct from capability: a more capable model is not automatically a better behaved one. Read more

allowlist

An explicit list of permitted tools, users or operations. Anything outside the list is denied when the rule is enforced. Read more

anchor investor

A large investor whose early commitment helps support an offering and attract other buyers. Talks about such a commitment do not establish that an investment has been agreed. Read more

anchor tenant

The customer whose committed demand makes a project financeable. Borrowed from shopping centers; in AI it is the lab signing a long lease that lets somebody else build the data center. Read more

antitrust

Competition law. It governs mergers and conduct by dominant firms, and it is the frame regulators reach for when one company controls a layer everyone else depends on. Read more

Apache 2.0

A permissive open-source license. You can use, modify and sell the work commercially, including in closed products, as long as you preserve the notices. Read more

API

Application Programming Interface. The way one piece of software calls another. When a lab sells access to a model by the token, it is selling API access. Read more

ARC-AGI

A family of visual puzzle benchmarks designed to be easy for humans and hard for AI, run by the ARC Prize Foundation, which verifies scores itself on a private test set. Version numbers mark successively harder editions. Read more

ARR

Annual recurring revenue: subscription revenue normalized to a yearly figure. Cleaner than run-rate because one-off deals are excluded, but companies define it themselves. Read more

ASIC

Application-Specific Integrated Circuit. A chip built to do one job well instead of everything adequately. Google's TPUs and Amazon's Trainium are ASICs, and they compete with Nvidia's general-purpose GPUs. Read more

audio encoder

The part of a speech system that turns sound into numerical representations a language model can process. Shrinking it can reduce cost but also lose information. Read more

B

backbone

The main model architecture that supplies representations used by the rest of a machine-learning system. Extra task-specific components may sit on top of it. Read more

backlog

Contracted future revenue not yet delivered or recognized. It measures commitments, and it turns into cash only if the customers survive long enough to pay. Read more

bacteriophage

A virus that infects bacteria, not people. Harmless to humans and able to kill specific bacterial strains, which is why AI-designed phages are studied as an alternative to antibiotics. Read more

benchmark

A fixed set of tasks used to compare models. Scores rise partly because models improve and partly because the answers leak into training data, so treat a jump on a popular benchmark carefully. Read more

benchmaxxing

Rerunning an evaluation under different settings or reasoning tiers until the score improves, then publishing the best result without disclosing which configuration produced it. Read more

binder

A molecule that attaches to a biological target, such as a protein. Binding is an early research result and does not by itself establish a safe or effective medicine. Read more

biometric data

Measurements of physical or behavioral traits, such as a face or voice, that can be used to identify a person. A photo and a face-identification template are different forms of data. Read more

biomolecular modeling

Using computer models to predict the structure, behavior or interactions of biological molecules such as proteins and DNA. Read more

BIS

The Bureau of Industry and Security, the US Commerce Department arm that writes and enforces export controls. It decides which chips may be sold to which countries. Read more

BM25

A ranking formula from classic text search that scores documents by keyword match. Often paired with vector search so a query catches both exact terms and related meaning. Read more

bookbuilding

The process of collecting investor orders for a share offering to set its price and allocation before trading starts. A live book means the deal is still being sold. Read more

brain-computer interface

A device that reads brain signals and turns them into commands for a computer, through an implant or a headset. Invasive versions sit on or in the brain and need surgery. Read more

bridge loan

Temporary borrowing intended to cover a funding need until longer-term financing or another cash source arrives. Read more

Brier score

A way of scoring forecasts that rewards being both right and appropriately confident. Lower is better, and a confident wrong call costs far more than a hedged one. Read more

bug bounty

A program that pays outside researchers for privately reporting security flaws, with payouts scaled to severity, so the finder profits from disclosure instead of exploitation. Read more

buyback

A company or government repurchasing its own shares or bonds from the market, shrinking the outstanding supply. Fewer shares raises earnings per share; fewer bonds in circulation can push their price up and yield down. Read more

C

C++

A programming language giving developers detailed control over memory and execution. It is used for systems, simulations and other software where performance matters. Read more

C2PA

Coalition for Content Provenance and Authenticity: an industry standard for signed metadata that records where a file came from and whether it was altered since. Read more

cache

Stored results or data kept for reuse so a system does not have to fetch or compute them again. Cache size and reuse can affect both speed and cost. Read more

CAISI

The US Center for AI Standards and Innovation at NIST. It evaluates AI capabilities and develops measurement methods. Read more

calibration

Whether stated confidence matches reality. A well-calibrated forecaster who says 70% is right about 70% of the time, which is a different skill from being right often. Read more

capex

Capital expenditure. Money spent on physical assets that last for years, which for AI means data centers, chips and power. The opposite of operating costs like salaries. Read more

catalog provenance

The recorded origin of a software component in an approved catalog. Checking it helps prevent substitution, but does not prove the component is safe. Read more

catastrophic forgetting

A neural network's tendency to lose abilities it already learned when trained on new data. The main reason deployed models are frozen rather than allowed to keep learning. Read more

CFAA

The Computer Fraud and Abuse Act, the main US anti-hacking law. It turns on who 'accesses' a computer, so it decides what AI agents are allowed to do on other people's websites. Read more

chain of thought

The step-by-step reasoning a model writes out before its answer. Labs monitor it to catch a model planning something its final output would hide, which only works while models reason legibly. Read more

change-of-control clause

A contract term that triggers when one party is acquired, giving the other side rights it did not have before, often to renegotiate or exit the deal. Read more

checkpoint

A saved set of model weights from a particular training stage. Researchers can resume training from it or use it to run the model. Read more

CI/CD

The automated pipeline that builds, tests and ships software on every change. It holds credentials for everything it deploys to, which is why attackers target it. Read more

circuit breaker

A rule that pauses trading on an exchange when prices fall past a set threshold in a session, meant to interrupt panic selling and let information catch up. Read more

CLI

Command-Line Interface. A program you drive by typing commands instead of clicking. Most developer tools ship as one first. Read more

closed-loop evaluation

A test in which each action changes the situation used to choose the next action. It can reveal failures hidden by scoring isolated decisions. Read more

clustering

Grouping items by similarity without assigning each one a label in advance. Similarity can be misleading: two different people can end up in the same face cluster. Read more

codebook

A learned collection of representations used to encode data as discrete tokens. In image models, it helps turn visual features into symbols the model can process. Read more

column lineage

A map of how individual database columns are derived and used across tables and reports. It helps identify which downstream results a data change could affect. Read more

command injection

A bug where text a system treats as data, an issue title for example, ends up executed as part of a shell command. One crafted string then runs the attacker's code. Read more

community benefits agreement

A binding deal between a developer and a local community, trading approval for a large project, such as a data center, for concrete commitments: jobs, infrastructure, or payments to affected residents. Read more

composite

A material combining components with different properties, such as carbon fibers held in resin. Making consistent parts requires controlling both the materials and the manufacturing process. Read more

confidential filing

A draft securities registration submitted to the SEC for private review before the public sees it. It starts the IPO process without disclosing numbers; the public version must appear at least 15 days before the roadshow. Read more

consensus

The average of Wall Street analysts' forecasts for a company's results. Beating consensus is the bar a quarter is judged against, and the bar moves in the weeks beforehand. Read more

container

A packaged software environment with its own dependencies and some isolation from the host computer. The strength of that isolation depends on its configuration. Read more

container registry

A server that stores and distributes container images, the packaged filesystems software ships in. Registries hold every layer of an image, so a secret baked into one layer stays retrievable for years. Read more

context compaction

Shortening an AI conversation history so it fits within the model’s input limit. Summarization can preserve useful context but can also lose constraints or details. Read more

context window

How much text a model can hold in mind at once, counted in tokens. Past the limit the earliest material falls out, so a long agent session starts forgetting its own beginning. Read more

continual learning

Training that lets a model keep learning from new experience after deployment instead of freezing its weights. The main obstacle is catastrophic forgetting, where new learning overwrites old skills. Read more

contrail

The line-shaped ice cloud a jet leaves in cold, humid air. Contrails trap outgoing heat, and researchers put their warming effect near the scale of aviation's CO2 emissions. Read more

convertible bond

Debt that the holder can later swap for shares at set terms. An investor gets paid like a lender while keeping the upside of an owner, and the issuer gets cash without immediate dilution. Read more

convertible note

Short-term debt an early startup raises that converts into equity at a later priced round, usually at a discount, so investors and founders can defer arguing over a valuation. Read more

cornerstone investor

An investor that commits before an IPO's book opens to buy a set slice at the offer price, usually accepting a lock-up in return for guaranteed allocation. Hong Kong listings lean on them to prove demand in advance. Read more

counterfactual

A comparison that asks what would happen if one relevant fact were different while others stayed fixed. It helps separate causes from mere associations. Read more

Covered List

The FCC's roster of communications equipment and vendors judged a national-security risk, mostly Chinese. Gear on it cannot get the authorization US devices need to ship. Read more

CoWoS

Chip-on-Wafer-on-Substrate, TSMC's advanced packaging process for stacking chips and memory close together. The bottleneck for AI accelerators: the chips exist, the packaging capacity does not. Read more

CPI

The consumer price index, the standard US inflation measure. It moves markets because the Fed sets interest rates against it, and rates set the price of the debt funding the AI buildout. Read more

CPU

The general-purpose processor in a computer. It handles many kinds of instructions; GPUs specialize in doing large numbers of similar calculations in parallel. Read more

credential

A secret or token that a service accepts as proof of access rights. An agent can use a service’s credentials without being entitled to every action they permit. Read more

credit spread

The extra interest a borrower pays above a reference rate. A wider spread generally means lenders demand more compensation for risk. Read more

CSRC

China’s national securities regulator. Its overseas-listing filing process is a regulatory step for mainland companies seeking to list abroad. Read more

CTF

Capture the flag: competitive hacking contests where teams find and exploit planted vulnerabilities. Rankings from events like DEFCON are the closest thing security research has to a world league table. Read more

CUDA

Nvidia’s software platform for programming its GPUs. Applications use it to run parallel computations such as model training and inference. Read more

curtailment

A grid operator deliberately cutting a customer's power when supply runs short. Who gets curtailed first is set by contract, which is why data centers fight hard over the fine print of their interconnection terms. Read more

CVE

Common Vulnerabilities and Exposures. The public numbering system for known software flaws. A CVE identifier is how everyone refers to the same bug. Read more

CVSS

Common Vulnerability Scoring System. A 0-to-10 industry scale for how severe a software flaw is, where anything above 9 is usually called critical. Read more

D

data catalog

Software that indexes everything a company knows about its own data: where each dataset lives, what it means, who may use it. AI agents rely on catalogs for grounding, which also makes them high-value breach targets. Read more

data loss prevention

Controls that identify sensitive information and try to stop unauthorized disclosure, such as sending a private document to an outside service. Read more

DCFA

A research method for finding the decisive error in an AI-agent team by tracing dependencies and asking what a corrected action would change. Read more

de novo protein design

Designing a protein from scratch to do a specific job, such as a binder that latches onto a chosen target molecule, rather than copying one found in nature. A binder that sticks tightly and only to its target is the hard part. Read more

decoder

A model component that turns an internal representation into output, such as text or image pixels. Its exact role depends on the model architecture. Read more

demand-side platform

Software that buys ad placements automatically across many sites and apps from one dashboard. Amazon's DSP is how advertisers now buy space inside ChatGPT. Read more

depreciation

Spreading the cost of equipment across the years it is expected to last. The assumed lifespan of an AI chip is a judgment call, and stretching it makes current profits look better. Read more

deterministic

Producing the same result from the same inputs and state. A deterministic execution tool can constrain an AI planner without making the planner itself predictable. Read more

differentiable

A mathematical property that lets training calculate how a small change in an input or parameter changes the output. Connecting differentiable stages allows one objective to adjust several stages together. Read more

diffusion language model

A language model that develops text by repeatedly refining a provisional output. It can update multiple positions together, unlike models that generate one next token at a time. Read more

Digital Services Act

The EU law governing online platforms, in force since 2024. It scales duties with size, so the largest platforms and search engines answer to Brussels for illegal content, child safety and election risks. Read more

digital twin

A live simulated replica of a real system, used to test changes or run an attack safely before it touches the real thing. Read more

distillation

Training a small model to imitate a large one. The small model ends up far cheaper to run and keeps much of the behavior, which is how a 30B model can act like something ten times its size. Read more

distribution tag

A movable label, such as latest, that points to a particular version of an npm package. Changing it can change which version users install. Read more

DNS rebinding

A web attack where a malicious page tricks a browser into treating an attacker domain as the same origin as a service running on your own machine, letting it reach ports normally closed to the outside internet. Read more

domain admin

The highest account level in a Windows network, able to control every machine and user in the organization. Reaching it is the usual definition of a total compromise. Read more

DRAM

The memory chips holding a computer's active working data, distinct from HBM and NAND flash. Prices have climbed alongside HBM as AI servers compete for the same fabrication capacity. Read more

durable execution

A way to save a workflow’s progress so it can resume after a crash, connection failure or long wait. External actions still need protection against accidental repetition. Read more

duty of care

A legal obligation to take reasonable steps to prevent harm. What steps are required, and who can enforce them, depend on the applicable law. Read more

E

EBITDA

Earnings before interest, taxes, depreciation and amortization. For a data-center operator the excluded items are the largest costs, so a fat EBITDA margin can sit on top of a real loss. Read more

ECB

The European Central Bank, which sets monetary policy for countries using the euro. Its president’s policy arguments do not by themselves create EU legislation. Read more

EDA

Electronic design automation: the software used to design chips and circuit boards, a market held for decades by Synopsys, Cadence and Siemens. AI startups are now pitching agents that draft and check circuits from plain-language requirements. Read more

EDGAR

The SEC's public filing database. Anything a US company files is readable here for free, usually before it is reported anywhere. Read more

electrocardiogram

A recording of the heart’s electrical activity. Clinicians use its timing and shape to investigate heart rhythm and other cardiac problems. Read more

embedded evaluator

An outside reviewer with continuing access inside an AI lab to examine development and safety practices. Access, payment and publication rights determine how independent that review can be. Read more

embedding

A numerical representation of text, images or other data that lets software compare meanings or patterns. Nearby representations often indicate related content. Read more

embodied AI

AI that controls a physical body, a robot arm, humanoid or vehicle, and learns from interaction with the real world rather than from text alone. Read more

encoder

A model component that converts input into a representation used by another component. In an image generator, it can extract the structure that a decoder turns into detailed output. Read more

endpoint

An address through which software requests a particular service or operation. Access controls at that address can restrict what callers may do. Read more

EU AI Act

The European Union's AI law, phasing in from 2024. It sorts systems by risk level and puts the heaviest documentation duties on the highest tier, with fines set as a share of global revenue. Read more

EUV

Extreme ultraviolet lithography: chipmaking machines that print features using 13.5nm light. ASML is the only supplier, and export bans keep them out of China, so Chinese fabs multi-pattern with older tools instead. Read more

eval

A test suite for a model, usually built in-house for one specific job. Labs report benchmark scores publicly and run evals privately, and the private ones drive the decisions. Read more

exaflop

A measure of computing speed equal to a billion billion floating-point operations per second. Comparisons require the same numerical precision and workload assumptions. Read more

export controls

Rules restricting which technology can be sold abroad. For AI they mostly cover advanced chips and the machines that make them, and they are the main instrument of US policy toward China. Read more

F

fair use

The US copyright doctrine letting a work be used without permission if it is sufficiently transformative and does not harm the market for the original. The central question in every AI training-data lawsuit. Read more

feature store

A managed collection of model inputs, such as a customer risk score, shared between training pipelines and applications making live predictions. Read more

federal funds rate

The interest rate US banks charge each other overnight, steered by the Federal Reserve as a target range. Most other US borrowing costs price off it, so a change moves the discount rate on everything from mortgages to data-center debt. Read more

federated connector

A connection that lets an application retrieve data held in another system while leaving that data in its original location. Access still depends on the permissions and capabilities of the connected systems. Read more

ferroelectric memory

Memory that stores bits in a material whose electric polarization flips and stays put, so it holds data without power and switches quickly. Kepler bonds it onto compute dies as an HBM alternative. Read more

Fields Medal

Mathematics' top honor, awarded every four years to researchers under 40. There is no equivalent prize for computer science, which is part of why a joint statement from Fields Medalists carries particular weight. Read more

fine-tuning

Additional training that adapts an existing model to a task, domain or behavior. It changes the model’s parameters rather than only supplying new instructions. Read more

first-day pop

The jump from an IPO's offer price to its first close. A big pop signals demand, and also that the seller priced the shares below what buyers were ready to pay. Read more

float

The shares available for public trading. A small float can make prices move sharply because relatively few shares are available to buyers. Read more

follow-on offering

A company that is already public selling new shares to raise more money. It dilutes existing holders, so a big one signals the company values the cash more than the shareholders value their slice. Read more

formalization

Translating a mathematical argument into precise statements and proof steps that a computer can check. This can verify an existing result without discovering a new one. Read more

Fortran

A programming language developed for numerical and scientific computing. Long-lived simulations often contain decades of Fortran code and assumptions. Read more

forward guidance

A central bank's public signal about where it expects future policy to go, meant to shape market expectations before any rate actually moves. Markets can end up trading the guidance itself rather than incoming data. Read more

forward-deployed engineer

An engineer a software vendor stations inside a customer's organization to build the product into that customer's real systems. Palantir popularized the role; AI vendors now hire them to make agent deployments work. Read more

foundation model

A large model trained on broad data that many downstream tasks build on, rather than a model trained for one job. The term now extends past language: Skild's S1 is a foundation model for robot control. Read more

foundry

A factory that manufactures chips other companies design. TSMC is the one that matters; Nvidia, Apple and AMD all design chips and none of them make their own. Read more

FP4

A 4-bit floating-point number format. Storing weights or caches in FP4 quarters the memory of 16-bit formats at some accuracy cost, which is how long-context serving gets cheap. Read more

FP8

An 8-bit floating-point format for storing model weights and activations. Halving the bits against 16-bit roughly halves memory and doubles throughput, at a small cost in precision. Read more

free cash flow

Cash left after running the business and paying for equipment. Harder to flatter than profit, so it is the number to watch when a company is spending heavily on data centers. Read more

frontier model

One of the handful of largest, most capable models in existence at a given moment. The term exists because regulation keeps trying to draw a line around exactly these and nothing below them. Read more

full-duplex

Communication that receives and sends at the same time. A full-duplex voice agent can listen while speaking, allowing interruptions without waiting for a turn to end. Read more

function signature

The declared inputs and output of a software function. A tool call must match this specification to be valid, although valid inputs can still produce an unwanted result. Read more

fused operations

Several computational steps combined into one operation to reduce memory transfers and processing overhead. Read more

G

gain-of-function

Research that gives a pathogen abilities it lacks in nature, such as infecting new species, usually to study pandemic risk. Tightly regulated because the modified pathogen is itself a hazard. Read more

GDPR

The EU's General Data Protection Regulation, the 2018 privacy law requiring a legal basis, such as explicit consent, before a company may use personal data. Fines reach 4% of global revenue, so it shapes how products launch in Europe. Read more

general availability

The point at which a product ships to all customers rather than a waitlist, beta or preview. The line matters for models: a capability in preview can be withdrawn quietly; one in general availability is priced and supported. Read more

GGUF

The file format llama.cpp and most local runners use for quantized models. A release is not really local-ready until GGUF builds exist. Read more

gigawatt

A unit of power equal to one billion watts. It measures the rate of electricity generation or consumption, rather than energy used over a period. Read more

GitHub Actions

GitHub's built-in automation service: workflows that run tests, file tickets or ship code whenever something happens in a repository. They run with the repository's credentials, so a hijacked workflow is a breach. Read more

GitOps

Running infrastructure by keeping its desired state in a git repository that automated systems continuously apply. Write access to the GitOps repo is effectively write access to production. Read more

GPU

Graphics Processing Unit. A chip built to do many simple calculations at once, which turned out to be exactly what training a model needs. Nvidia's version is the industry default. Read more

GPU kernel

A small program launched onto a GPU to do one piece of work. Launching thousands of them adds overhead, so fusing work into fewer, longer-running kernels is a standard way to make AI inference faster. Read more

gray market

Unofficial trading in a stock before its official exchange debut. Prices there preview where the stock might open, on thin volume and with no guarantee the real market agrees. Read more

GRE

Graduate Record Examinations, standardized tests used in graduate-school admissions. GRE tutoring results do not necessarily transfer to other subjects or age groups. Read more

greedy decoding

Generating text by choosing the most likely next token at each step, instead of sampling from several possible tokens. Read more

gross margin

What is left of revenue after the direct cost of delivering it. For an AI company that cost is largely inference, so the margin says how much the compute bill is eating. Read more

guidance

A company's own forecast of its next quarter or year, issued with results. Markets often react more to guidance than to the results themselves, since the results are already history. Read more

gVisor

Software that intercepts a program’s requests to the operating system, adding an isolation layer between the program and the host kernel. Read more

H

H-share

A share in a mainland China-incorporated company that is listed in Hong Kong and traded in Hong Kong dollars. Read more

harness

The software wrapped around a model that turns it into a working agent: the loop that plans, calls tools, runs code and checks results. The same model can perform very differently in different harnesses. Read more

HBM

High-bandwidth memory. Memory stacked vertically next to the processor to feed it faster. Supply is tight, and memory makers have become AI companies by market value as a result. Read more

heap overflow

A bug where a program writes past the end of a memory buffer, corrupting neighboring data. Attackers steer the corruption to hijack the program and run their own code. Read more

held-out

Data kept out of the learning process and reserved for checking performance. It helps detect when an apparent improvement only fits examples already seen. Read more

High NA EUV

An advanced chipmaking technique that uses extreme ultraviolet light and a wider-angle optical system to print finer circuit patterns. NA means numerical aperture. Read more

hollow-core fiber

Optical fiber that guides light through a hollow air or vacuum channel instead of solid glass, so signals travel about 30% faster because light moves quicker through a vacuum than through glass. Read more

homomorphic encryption

Cryptography that lets a computer run calculations on encrypted data without ever decrypting it, so the machine's operator cannot read what it is processing. Powerful for privacy, still slow, and every speedup makes it more usable. Read more

Hugging Face

The default public repository for machine-learning models and datasets. When a lab releases open weights, this is almost always where the files land. Read more

humanoid

A robot built in roughly human shape, on the theory that the world is already arranged for human bodies. Contested: wheels and arms are cheaper for most jobs. Read more

hyperscaler

The handful of companies running cloud computing at global scale: Amazon, Microsoft, Google, Meta. Their capital spending sets the pace of AI infrastructure for everyone else. Read more

I

IA360

Spain’s proposed framework for responsible AI deployment, including discussions with workers and employers and monitoring of employment effects. Read more

idempotency key

A unique identifier attached to an operation so a service can recognize a retry and avoid performing the same action twice. Read more

ImageNet

A labeled image dataset widely used to train and test computer-vision systems. Results on it do not establish performance on every real-world image. Read more

imitation learning

Training a model to copy recorded expert demonstrations rather than learn by trial and error. It works where good behavior is demonstrated daily but never written down, and degrades when reality drifts from the demonstrations. Read more

implied move

The size of the price swing that options prices predict for a known event, such as an earnings report. It is the market's bet on how much, not on which direction. Read more

in silico

Research performed using computer models or simulations. A promising simulated result still needs testing against physical measurements. Read more

in vivo

Latin for "within the living," describing treatment done inside a patient's body rather than on cells grown outside it. Contrasts with ex vivo, where cells are modified externally and then reintroduced. Read more

in-context learning

A model picking up a new task from examples placed in its prompt, with no retraining. Skild's S1 applies the idea to robots: one demonstration video stands in for hundreds of training examples. Read more

inference

Running a trained model to get an answer, as opposed to training it in the first place. Training is a one-time cost; inference is what you pay every time someone uses the thing. Read more

infostealer

Malware that quietly harvests what a machine already holds: passwords, browser cookies and session tokens. Stolen sessions let attackers walk into logged-in accounts without ever touching a password. Read more

injunction

A court order requiring or prohibiting an action. A preliminary injunction can apply while a case proceeds, before the court reaches a final decision. Read more

intangible asset

An asset without physical form, such as software, intellectual property or a customer relationship. It can support a business while being difficult to sell after a default. Read more

interconnection queue

The waiting list to plug a new power plant or large load like a data center into the grid. Queues in big US regions now run to years, which is why compute companies buy sites with connections already secured. Read more

interlocking directorate

One person serving on the boards of two competing companies at once. Section 8 of the Clayton Act has banned it since 1914, and the usual remedy is a quiet resignation, not a lawsuit. Read more

interpretability

Research into what is actually happening inside a model, as opposed to judging it by its outputs. The practical goal is catching a failure by looking at the machinery instead of waiting for it to misbehave. Read more

investment grade

A credit rating (BBB- or better) telling investors a bond is relatively safe. Many pension funds and insurers can only buy investment-grade debt, so reaching the threshold unlocks much larger pools of money. Read more

J

jailbreak

A prompt that talks a model out of its own safety rules. Different from prompt injection: a jailbreak is the user attacking the model, injection is a third party attacking through content the model reads. Read more

JVM

The runtime that executes Java bytecode, including code compiled from languages such as Java and Kotlin. Read more

K

kernel

The core software that controls an operating system’s access to memory, processors and devices. Giving a virtual machine its own kernel separates it from the host’s operating system. Read more

KEV catalog

CISA's list of software flaws confirmed to be exploited in the wild. Federal agencies must patch each entry by a set deadline, so a KEV listing turns a vulnerability into a compliance clock. Read more

KiCad

Free software for designing electronic circuits and printed circuit boards. Its built-in checks catch specified electrical and layout-rule violations. Read more

kill switch

A mechanism for stopping a running system. For AI, its value depends on whether it remains available and effective when the system is behaving unexpectedly. Read more

know-your-agent

Identity checks for AI agents, modeled on banking's know-your-customer rules: who operates the agent, what it is authorized to do, and how much autonomy a counterparty should allow it. Read more

knowledge graph

A database that stores entities and the relationships between them as a graph, so software answers questions by following connections rather than joining tables. A common way to ground AI agents in enterprise data. Read more

KOSPI

The main index of the Korea Exchange, covering the common stocks listed in Seoul. Samsung Electronics and SK Hynix are its two largest members, so memory-chip sentiment moves the whole index. Read more

Kotlin

A programming language used for Android apps, server software and projects sharing code across platforms. Read more

KV cache

The memory a language model keeps of text it has already processed, so it does not recompute it for every new token. Where this cache lives, and how often it is hit, drives most of the cost of serving a model. Read more

KVMem

A research system that preserves previously processed agent history across GPU memory, system memory and local storage, then selects relevant portions for each request. Read more

L

latency

The delay between asking and getting an answer. For an agent taking many steps it compounds, so latency decides what is usable and throughput decides what is affordable. Read more

latent outputs

Internal numerical representations produced by a model before they are turned into visible images or other final output. Read more

latent space

A model’s internal numerical representation of information. Nearby representations can encode similar features without corresponding directly to words or pixels. Read more

Lean 4

A proof assistant: software that checks a mathematical proof step by step and accepts it only if every step follows. It is why AI-generated math can be trusted without a human rechecking it. Read more

length-adaptive average lagging

A measure of simultaneous-translation delay that adjusts for output length. Lower delay does not by itself mean a translation is more accurate. Read more

LIBERO

A family of robot-manipulation benchmarks. Variants test different abilities, so a completion score on one cannot be treated as overall robot reliability. Read more

license plate reader

A camera network that logs every passing plate with time and location. Flock Safety runs the largest US network, and the accumulated logs double as a searchable record of where everyone drives. Read more

ligand affinity

How strongly a molecule binds to a target, such as a protein. Predicting a structure does not by itself establish the strength of that binding. Read more

LLM

Large Language Model. Software trained on enormous amounts of text to predict what comes next, which turns out to be enough to answer questions, write code and hold a conversation. Read more

LLM poisoning

Publishing content built to be absorbed and repeated by AI chatbots rather than read by people. It games the credibility signals models use when choosing sources: footnotes, neutral tone, institutional trappings. Read more

lockup

A post-IPO ban on insiders selling their shares, typically for 180 days. When it expires, or is released early in tranches, the supply of sellable stock jumps and the price often sags. Read more

loss of control

The scenario where an AI system keeps pursuing a goal that conflicts with what its operators intended, and ordinary shutdown or correction no longer works. A concern about capability outrunning oversight, not a claim that it has happened at scale. Read more

LPDDR

Low-power DRAM built for phones and laptops, where battery matters more than raw bandwidth. Each numbered generation moves data faster; LPDDR6 is the newest, and the first phones carrying it shipped in 2026. Read more

M

macro placement

Choosing where large circuit blocks sit on a chip. Their positions affect wiring, timing and whether the final layout satisfies manufacturing rules. Read more

MAI

The name used for Microsoft AI’s own model family. Its draft conduct code describes intended behavior and training standards for those models. Read more

market cap

Share price multiplied by shares outstanding: what the public market says a listed company is worth today. Read more

MAU

Monthly active users, the count of distinct people who used a product at least once in a month. The standard scale metric for consumer apps, and easy to inflate with a generous definition of used. Read more

MCP

Model Context Protocol. An open standard for connecting AI assistants to tools and data sources, so a capability written once works across different clients. Read more

MDASH

Microsoft's codename for an AI system that searches source code for vulnerabilities and uses additional agents to check whether suspected flaws can be exploited. Read more

MFN

Most favored nation. In a startup SAFE, a clause guaranteeing the investor the best terms any later investor negotiates. An uncapped MFN note sets no valuation of its own; it borrows one from the next round. Read more

MGX

Nvidia’s modular server and rack design framework. Hardware makers use its reference designs to integrate processors, networking and other components. Read more

microgrid

A local electricity network with its own generation or storage that can operate separately from the wider grid. Read more

microtask

A paid unit of online piecework taking seconds to minutes: label an image, transcribe a receipt, rate an answer. Much of the training data behind early machine learning was assembled this way. Read more

microVM

A stripped-down virtual machine that boots in milliseconds and gives a workload its own kernel. The isolation of a VM at close to the cost of a container, popular for sandboxing AI agents. Read more

middle mile

Freight routes between fixed business endpoints, like a distribution center and a store, as opposed to delivery to somebody's door. Fixed, repeatable routes with no passengers, which is why autonomous trucking starts here. Read more

Millennium Prize Problems

Seven mathematics problems the Clay Mathematics Institute posted $1 million bounties on in 2000. The Poincare conjecture is solved; Navier-Stokes regularity was among the six still open. Read more

misalignment

When an AI system pursues goals or behavior that conflict with the intentions or constraints of the people directing it. Read more

MIT license

One of the most permissive open-source licenses: use, modify and sell freely, keep the copyright notice. On model weights it means almost no strings attached. Read more

model adapter

A small set of trainable components added to a model to specialize its behavior without retraining all its original parameters. Read more

model routing

Software that reads each AI request and picks which model answers it, sending easy work to cheap models and hard work to expensive ones. It turns model choice from a policy decision into a per-query price comparison. Read more

model steering

Changing selected internal model signals during use to encourage a behavior without retraining all its weights. Experiments test whether those signals cause an observed capability. Read more

module shadowing

When a local file shares a name with a standard-library module, Python can load the local file instead. An attacker who plants the file turns an innocent import into code execution. Read more

MoE

Mixture of Experts. A model split into specialist sub-networks where only a few run for any given input, so a very large model can be cheap to run. Read more

monorepo

A single code repository holding many of a company's projects at once, common at large tech firms. Reaching it means reaching most of the company's source code in one place. Read more

MOU

A memorandum of understanding: a signed statement of intent that is not a binding contract. An MOU announces a deal before the terms that would make it one exist. Read more

mRNA

The molecule cells use as working instructions for building proteins. mRNA vaccines deliver those instructions so the body manufactures a target protein and trains the immune system against it. Read more

multimodal

A model that handles more than text. Images, audio and video in, and often out, all through one set of weights instead of separate models wired together. Read more

N

n-gram

A run of n consecutive words or tokens. Storing statistics for common n-grams lets a model look up frequent patterns cheaply instead of computing them, an old technique now returning inside frontier architectures. Read more

NAND flash

The memory chips inside SSDs and USB drives, storing data by trapping electric charge in cells rather than spinning a disk. Cheaper and slower than the DRAM used for a computer's working memory. Read more

Navier-Stokes

Equations describing fluid motion, including viscosity. The existence and smoothness problem asks whether smooth flows always remain well behaved under specified conditions. Read more

NCMEC

The US clearinghouse for reports of child sexual abuse material. Online platforms are legally required to report such material to it through the CyberTipline. Read more

near-packaged optics

Optical connections placed close to a computing chip’s package, shortening the electrical path before data travels as light. The aim is greater bandwidth with less connection power. Read more

neocloud

A cloud provider built around renting GPUs for AI rather than general-purpose computing. CoreWeave, Nebius and Lambda are the type specimens; most lease chips from Nvidia and sell capacity on to the labs. Read more

neural operator

A neural network that learns mappings between functions rather than fixed-size inputs, so one trained model can simulate physics like fluid flow at any resolution. The architecture behind several AI weather and engineering models. Read more

neural rendering

Using trained neural networks to generate or reconstruct parts of an image, often reducing the work required by conventional graphics rendering. Read more

Nix

A package manager and build system that records a program’s dependencies precisely, helping developers reproduce the same software environment and keep different versions apart. Read more

non-sofic

A group that cannot be approximated by finite ones. Whether any exists was asked by Gromov in 1999; producing one is the kind of result that had resisted every human attempt for decades. Read more

nonfarm payrolls

The monthly count of US jobs added outside farming, from the Bureau of Labor Statistics survey of employers. The single number markets trade hardest, released the first Friday of each month. Read more

npm

A widely used registry and package manager for JavaScript software. Developers download dependencies from it and publish reusable packages to it. Read more

NVL72

Nvidia's rack-scale format: 72 GPUs wired by NVLink into what software sees as one enormous accelerator. Big orders are now counted in racks rather than chips. Read more

NVLink

Nvidia's high-speed interconnect linking chips inside a server rack far faster than standard networking. NVLink Fusion licenses that fabric to other chipmakers, so custom silicon can slot into Nvidia systems. Read more

O

observability

Using logs, measurements and recorded interactions to understand what a running system did and why it failed. The records may contain sensitive user data. Read more

off-balance-sheet

An arrangement where the assets and debt sit in another entity, so they never appear in a company's own accounts. Legal, common, and the reason a balance sheet can understate what a company has committed to. Read more

omnimodal

A model that takes text, images, audio and video in one request rather than routing each to a separate system. Vendor shorthand extending multimodal to native audio-video understanding. Read more

opaque serial depth

A proposed measure of how much sequential computation an AI performs between outputs people can interpret. It estimates hidden reasoning capacity, not whether a system is safe. Read more

open interest

The total outstanding derivative contracts that have not been closed. It measures existing positions, while trading volume measures transactions over a period. Read more

open-weight

A model whose trained parameters are published, so anyone can download and run it themselves. Not the same as open source: the training data and code usually stay private. Read more

optical circuit switch

A network switch that redirects light between fibers without converting it to electrical signals, saving the power and delay of the conversion. Google runs them inside its own AI clusters. Read more

optical interconnect

A connection that carries data as light between computing components. It can extend reach and reduce electrical losses, but requires hardware to convert and manage the signals. Read more

optical transceiver

The plug-in module that converts electrical signals to light and back at each end of a fiber link. Data centers buy them by the million, and Chinese vendors lead the market. Read more

out-of-band release

A software update issued outside the normal release schedule, often to fix a security vulnerability that cannot wait. Read more

overallotment option

The right an IPO's underwriters have to sell extra shares beyond the original offering if demand runs high, raising the total amount the company collects. It is why a listing's final size often exceeds the number first announced. Read more

oversubscribed

An offering is oversubscribed when investors ask for more shares than are for sale; the multiple measures demand. At the extreme, allocations are decided by lottery. Read more

P

P/E

Price-to-earnings ratio: a company's share price divided by its yearly earnings per share. Listing at 219 times earnings means a buyer pays 219 years of current profit up front, a bet that profit grows into the price. Read more

paralinguistics

The part of speech that is not the words: tone, pace, stress, hesitation. Transcription drops it, which is why two identical transcripts can carry opposite meanings. Read more

parameters

How many individual numbers a model has learned. A 30B model has thirty billion. It is a rough size measure and a poor quality measure, since a well-trained small model routinely beats a badly trained large one. Read more

perpetual futures

Derivative contracts with no fixed expiry. Traders gain exposure to a reference price without owning the underlying asset; the contract price can differ from an actual share transaction. Read more

personal access token

A credential string that stands in for a password when software talks to a service like GitHub. Whoever holds the token holds its permissions, which is why one leaked in a public file is a full breach. Read more

Phase 1

An early human study that primarily evaluates a treatment’s safety and dosing. Starting one does not establish that the treatment works. Read more

photomask

A patterned template used to project circuit designs onto a semiconductor wafer during chip manufacturing. Read more

photonic processor

A computing device that uses light for some operations. Its useful speed and energy cost also depend on memory, electronic control and conversion between light and electrical signals. Read more

Playwright

Software for controlling web browsers through code, commonly used to test websites and automate browser tasks. Read more

post-money valuation

A startup's value counting the new investment just raised. A $2M round at a $10M post-money means investors own 20%; the pre-money was $8M. Read more

post-training

Everything done to a base model after its main pretraining run: supervised fine-tuning, reinforcement learning and reasoning tuning. Labs now get large capability gains here without changing the base model at all. Read more

PostgreSQL

An open-source relational database that stores structured records and supports transactions. It can also coordinate durable job queues. Read more

power purchase agreement

A contract to buy electricity over an agreed period. Long contracts can give a generator predictable revenue to finance construction or keep a plant operating. Read more

pre-money valuation

A startup's value before a new round of funding is added in. A $20M raise at a $3.8B pre-money valuation becomes a $3.82B company once the round closes. Read more

pre-seed

The earliest outside money a startup raises, usually before it has a product. Sums run from a few hundred thousand to a few million. Read more

precision

The share of a system’s reported findings that are correct. A reviewer can achieve high precision by reporting only obvious defects while missing many others. Read more

preclinical

Research conducted before testing a treatment in people. Laboratory results can support further development without establishing that a treatment is safe or effective for patients. Read more

preemption

A federal law overriding state ones on the same subject. In US AI policy it is the central fight: whether Washington sets a single standard or fifty states each set their own. Read more

prefix caching

Reusing computation for the shared opening of several prompts. Prefix-aware routing sends those requests to the server that already holds the reusable state. Read more

premarket

Trading in US stocks before the exchange's 9:30am New York open, on thin volume. Premarket moves show direction, and they often shrink or reverse once full trading starts. Read more

Preparedness Framework

OpenAI's internal rulebook for measuring dangerous model capabilities before release, with thresholds like High and Critical that trigger stronger safeguards. Anthropic's equivalent is its Responsible Scaling Policy. Read more

preprint

A research paper shared publicly before formal peer review. Its methods and results are available to inspect, but publication alone does not establish independent validation. Read more

preservation letter

A legal demand to keep every document related to a matter because litigation or an investigation may follow. Deleting records after receiving one can itself become the offense. Read more

pretraining

The initial training stage in which a model learns general patterns from large datasets, before later training adapts it to instructions or specific tasks. Read more

price target

An analyst's published estimate of where a stock trades within about a year. Targets move markets less by being right than by showing which way professional conviction is shifting. Read more

price-to-sales

Market value divided by a year of revenue, used to price companies without profits. A ratio above 100 means investors are paying today for a century of current sales. Read more

private credit

Loans negotiated outside public bond markets, usually from investment funds. Their terms can be tailored to a borrower’s revenue and assets. Read more

private placement

A sale of shares directly to chosen investors rather than on the open market. In an IPO it lets a strategic partner buy a stake at the offer price without competing for allocation. Read more

procedural graph

A map of steps and their relationships that guides an agent through a task. A system can revise the map using evidence from successful and failed attempts. Read more

process node

The generation of chipmaking technology a processor is built on, named in nanometers. Smaller numbers pack more transistors into the same area; the labels are marketing shorthand rather than literal measurements. Read more

programmable logic controller

An industrial computer that operates physical equipment, valves, pumps, motors, on a factory floor or at a water or power plant. Decades old, rarely patched, and increasingly reachable from the open internet. Read more

prompt caching

Storing a prompt's already-processed context so repeat requests reuse it at a steep discount instead of paying full input price. Agent loops re-read the same context constantly, so cache pricing often matters more than list pricing. Read more

prompt injection

An attack that hides instructions in content a model reads, so the model follows the attacker instead of its operator. The AI equivalent of SQL injection, and still unsolved. Read more

prospectus

The document a company publishes before selling shares, setting out its finances and risks. In Hong Kong and Europe it is the equivalent of the US S-1. Read more

provenance

The record of where data or an artifact came from and how it was produced. It helps someone trace a result back to its sources and processing steps. Read more

pull_request_target

A GitHub Actions event that runs in the base repository’s context. Executing untrusted contributor code in such a workflow can expose privileged credentials. Read more

PURPA

The Public Utility Regulatory Policies Act of 1978, a US federal law that sets standards state utility regulators must formally consider. Amending it is how Congress steers electricity policy without commanding the states. Read more

Python

A general-purpose programming language used for automation, data analysis and AI software. Its readable syntax and large library ecosystem make it common for quickly building and connecting tools. Read more

Q

QLC

A flash-memory cell that stores four bits instead of one, packing more capacity onto the same chip at the cost of speed and durability. The current way to make NAND flash cheaper per terabyte. Read more

quadruple patterning

Printing one chip layer in four aligned lithography passes to reach features finer than the tool could draw in one. It works without EUV but adds steps per wafer, so cost and yield suffer. Read more

quantization

Storing a model's numbers at lower precision so it fits in less memory and runs faster. INT8 and BF16 name how many bits each number gets. Some accuracy is lost, usually less than you would expect. Read more

R

ransomware

Malware that encrypts or steals a victim's data and demands payment to restore it. Run by criminal gangs as a business, complete with affiliates, negotiation portals and customer service. Read more

React Native

A framework for building mobile apps with React and JavaScript while using the phone platform’s native interface components. It lets developers share substantial code between Android and iOS. Read more

reasoning model

A model trained to work through a problem in steps before answering, spending more compute per question. Better at math and code, slower and dearer at everything. Read more

recall

The share of relevant cases a system finds. In code review, low recall means real defects escape detection even if the comments it does produce are accurate. Read more

recursive self-improvement

The hypothesized loop where an AI system improves its own ability to improve, compounding capability faster than human oversight can follow. Central to most catastrophic-risk arguments. Read more

red team

A group paid to attack a system before its users do. For models it means deliberately trying to make one produce the output it was built to refuse. Read more

regression

A change that makes a previously working behavior worse. Regression testing checks that a repair or new feature has not broken existing behavior. Read more

reinforcement learning

Training a model by rewarding good outputs and penalizing bad ones, so it learns from feedback on its own attempts rather than only imitating examples. The stage where much of a frontier model's reasoning is now shaped. Read more

release candidate

A software version offered for testing before a final release. Further defects may lead to another candidate instead of immediate general release. Read more

remaining performance obligations

Contracted revenue a company has signed but not yet delivered. A cloud provider's RPO shows demand years ahead of the income statement, and it can be renegotiated or canceled before any of it becomes revenue. Read more

remediation

Fixing a security weakness and deploying the repair. Finding a vulnerability alone does not remove the risk. Read more

remote code execution

A vulnerability that lets an attacker run their own code on someone else's machine, the most serious class of software bug. Read more

replay

Re-running recorded work with saved inputs or responses. Replacing a model’s answers alone does not reproduce the behavior of the external tools it used. Read more

residual value guarantee

A promise to cover the gap if leased equipment is worth less than expected when the lease ends. It lets a chipmaker backstop a customer's debt without booking the risk as a liability. Read more

reverse transcriptase

An enzyme that copies RNA into DNA. Finding one in a genome does not by itself establish what biological job its surrounding system performs. Read more

revolving credit facility

A pre-arranged bank loan a company can draw down, repay and redraw as needed, up to a limit. A standby source of cash rather than a one-time loan, often lined up before a public listing. Read more

reward hacking

When an AI system exploits flaws in how its goal is scored instead of doing the intended task, like tampering with the grader rather than solving the problem. A central failure mode in training agents with reinforcement learning. Read more

Riemann zeta

A function at the center of how prime numbers are distributed. Where its value hits zero is the subject of the Riemann hypothesis, unsolved since 1859 and the most famous open problem in mathematics. Read more

risk factor

The section of an IPO prospectus where a company lists what could plausibly hurt the business, written partly to shield it from investor lawsuits. What a company chooses to name there is a map of what its bankers fear. Read more

RLHF

Training that has people rate a model's outputs and steers the model toward what raters preferred. The rating is paid piecework at industrial scale, increasingly done by the professionals whose judgment the model is learning. Read more

roadshow

The tour a company makes to pitch large investors before listing. It is where the story gets tested, and where the price range usually moves. Read more

robotaxi

A ride-hail vehicle that drives itself with no human behind the wheel, dispatched and paid like a taxi. Waymo, Zoox, Tesla and several Chinese operators run competing fleets. Read more

robots.txt

A file websites publish telling automated crawlers which pages not to fetch. It is a request, not an enforcement mechanism: compliance is voluntary, which is why AI scrapers can ignore it. Read more

rollback

Restoring a system to an earlier recorded state after a failure. It can undo reversible software changes, but cannot retract an already delivered message or physical action. Read more

ROS

An open-source framework of libraries and tools for building robot software, widely used across research and industry despite the name; it is not an operating system in the traditional sense. Read more

RTK

A command-line utility that reduces terminal output passed to coding agents. Smaller output can reduce context use, but the effect on the final bill depends on the agent’s subsequent work. Read more

run-rate

Current revenue projected forward a full year. A company earning $4B this month has a $48B run-rate, which flatters fast growth and is why it gets quoted so often. Read more

runtime

The period when a program is executing. A choice known only at runtime depends on live inputs rather than information fixed when the program was written. Read more

Rust

A programming language designed for fast systems software, with rules that prevent many memory-access errors before a program runs. Read more

S

S-1

The registration document a US company files with the SEC before going public. A confidential S-1 is filed privately first, so the company can start the process without publishing its numbers. Read more

SAFE

Simple Agreement for Future Equity. An investment that converts into shares later, at a price set by a future round, so nobody has to agree what the company is worth today. Read more

sandbox

An isolated environment where code runs without touching anything outside it. A sandbox escape means the isolation failed and the code reached the real system. Read more

SDK

A package of code and utilities that helps developers use a service or platform from their own programs. Read more

SEC

The Securities and Exchange Commission, the US markets regulator. Companies going public file with it, and those filings carry legal liability for what they say. Read more

Second Request

A formal US antitrust demand for more information before a merger can close, issued when regulators suspect a deal might harm competition. Receiving one delays a deal and signals real scrutiny, not a routine filing. Read more

secondaries

Trading shares of a private company between existing holders, with no new shares issued. A useful signal because employees and early investors sell what they privately doubt. Read more

senior secured notes

Bonds first in line for repayment and backed by specific collateral. If the borrower fails, these holders are paid from the pledged assets before anyone else sees a dollar. Read more

Series A

Lettered venture funding rounds, in order: A is usually the first large institutional round, with B, C and onward marking later, bigger raises. The letter signals stage, not size, and late letters at small amounts often signal trouble. Read more

settlement layer

The system where transactions finally clear and get recorded, whatever route they took to arrive. Whoever runs settlement gets paid on every trade regardless of which side wins it. Read more

shader core

A GPU processing unit that runs small programs used to draw images and perform parallel calculations. Read more

Show HN

A category on Hacker News where people post things they built. One of the few remaining places a project with no marketing budget gets found. Read more

show-cause order

A regulator's demand that a company either justify its current practice or change it, with a deadline and the burden of proof on the company. FERC used one to force US grid operators to defend their data-center hookup rules. Read more

side effect

A change an operation makes outside its returned answer, such as writing a file or sending a message. Retrying the operation can repeat that change. Read more

sidecar

A South Korean market circuit-breaker that briefly suspends program trading, halting bulk buy or sell orders for five minutes when a stock-index future swings sharply, to slow a cascade. Read more

silicon photonics

Chips that guide and process light, often alongside electronic components, to move data through computing systems. Read more

single sign-on

One login granting access to many separate applications through a central identity provider. Convenient, and a single point of failure: subvert the sign-on step and every connected app opens. Read more

sovereign AI

A government building or funding its own AI models, compute and data infrastructure so the capability sits under national control rather than a foreign vendor. Saudi Arabia's HUMAIN and the EU's various compute funds are the pattern. Read more

SPAC

A special-purpose acquisition company: a listed shell that merges a private company onto the stock market without a traditional IPO. Faster, with lighter scrutiny, and the track record shows what that trade costs. Read more

sparse activation

Using only part of a model’s parameters for each input. This can reduce computation while the full set of parameters still needs storage. Read more

sparse autoencoder

A model used to break another model’s internal activity into a smaller set of active features that researchers can inspect. Its features are an imperfect view of the underlying computation. Read more

speaker diarization

Splitting an audio recording by who is speaking when, so a transcript reads as labeled turns instead of one undifferentiated stream. Read more

special-purpose vehicle

A separate legal entity created to hold one set of assets and debts, keeping them off the sponsor's balance sheet. AI chip financing increasingly runs through SPVs that buy hardware and lease it back. Read more

speculative decoding

A serving trick where a small draft model proposes several tokens and the big model verifies them in one pass. Same output, delivered faster and cheaper. Read more

SRAM

Fast on-chip memory that sits next to the compute, far quicker than the HBM stacked beside a GPU but far smaller. Chip designers trade between the two, and several startups bet the whole architecture on the choice. Read more

SRE

Engineering work that keeps software services available and reliable, including monitoring, incident response and preventing recurring failures. Read more

stage-only token

An npm credential that lets automation submit a package for human release approval but cannot publish it directly. It retains some other write permissions. Read more

STAR Market

Shanghai's Nasdaq-style board for technology companies, launched in 2019. Its profitability requirements are looser than China's main boards, so loss-making hardware companies list there. Read more

state-space model

A model that carries information forward through an evolving internal state. In language AI, it offers an alternative to attention over earlier tokens. Read more

stateless

Retaining no memory between episodes or requests. Another component must supply any history needed for the next decision. Read more

statement of interest

A filing letting the government put its position on record in a private lawsuit without becoming a party to it. Persuasive to a judge, not binding. Read more

statutory damages

Fixed amounts copyright law awards per infringed work, up to $150,000 where willful, instead of proven losses. Across tens of thousands of recordings they compound into the numbers that make AI training suits existential. Read more

stealth model

A model released anonymously on a routing platform so developers test it without knowing who built it. Labs use the disguise to collect unbiased feedback and real workload data ahead of an official launch. Read more

structured output

An AI response constrained to a specified format or set of choices so software can process it. Correct formatting does not guarantee a correct answer. Read more

subagent

An agent assigned a smaller task in its own context, returning its result to the agent coordinating the work. This limits shared context but adds communication costs. Read more

subnetwork

A smaller part of a neural network selected to perform a task, using fewer computations than the full network. Read more

superintelligence

A hypothetical system meaningfully smarter than the best humans at essentially everything. Used as a research goal by several labs and as a marketing word by more. Read more

supervoting shares

A class of stock carrying more votes per share than ordinary stock, letting founders keep control of a company even after selling most of its economic value to outside investors. Read more

supply chain

Every piece of outside code a program depends on, and everything those depend on in turn. A supply-chain attack compromises one of them and reaches everyone downstream. Read more

surface EMG

Electromyography reads the electrical signals muscles produce; the surface version reads them through the skin from a wearable. The signal is noisy and varies from person to person, which is the engineering fight in every EMG wristband. Read more

suspense file

The California Appropriations committees' holding pen for bills with fiscal cost, cleared in one rapid vote each cycle. Most bills that die in Sacramento die there, without debate. Read more

Swift

A programming language used to build apps for Apple devices, along with server and cross-platform software. Read more

sycophancy

An AI system's tendency to agree with or flatter a user even when that conflicts with evidence or appropriate advice. Read more

synthetic data

Data created by software or models instead of collected directly from real events. It can expand training coverage, but its mistakes and missing conditions can carry into the trained model. Read more

SynthID

A watermarking technology for identifying AI-generated content. Apple plans support for generated and edited images, with exceptions for some edits. Read more

system card

The technical report a lab publishes alongside a model: capabilities, safety testing, known failure modes. The closest thing AI has to a drug label, written by the manufacturer. Read more

T

TAM

Total addressable market: the revenue a company would book if it won every possible customer. Pitch decks use it to size the dream rather than the business. Read more

tape-out

The point where a chip design is finished and sent to the factory for manufacturing. Changes after tape-out mean new masks and months of delay, so it marks the design's real deadline. Read more

teleoperation

A human driving a robot remotely. Often used to collect the demonstration data a robot later learns from, and sometimes quietly present in demos of autonomy. Read more

ternary weights

A compression scheme storing each model weight as only +1, 0 or -1 instead of a 16-bit number, shrinking memory roughly tenfold. The open question is how much reasoning quality survives the squeeze. Read more

text-and-data mining

Automated analysis of text or other data to extract patterns and information. Whether copyright exceptions permit it depends on the jurisdiction and the use. Read more

theorem prover

Software that checks or constructs proofs from formal statements and rules. Its result applies to the encoded problem, so a faulty translation can invalidate the practical conclusion. Read more

thin film

A very thin layer of material deposited on a surface. Its structure depends on how it is grown, which makes deposition conditions part of the experiment. Read more

throughput

The amount of work a system finishes per unit of time, such as tokens per second. High throughput does not necessarily mean a short wait for one request. Read more

tmux

A terminal multiplexer: software that keeps several command-line sessions running in separate panes and windows, even when you disconnect. Read more

token

The unit a model reads and writes, roughly three quarters of a word. Pricing, speed and context limits are all quoted per token, so it is the industry's basic unit of account. Read more

tokenizer

The component that chops text into the numbered units a model actually reads. Every lab builds its own, so a model's tokenizer works like a fingerprint: probe how an anonymous model splits text and you can often tell whose it is. Read more

tokenomics

In AI usage, the economics of tokens: what processing and generating them costs, and how much useful work that spending buys. This differs from the term’s cryptocurrency meaning. Read more

tool allowlist

An explicit list of tools a program or agent is permitted to use. Tools outside the list are unavailable even if they exist elsewhere in the system. Read more

tool registry

An inventory of the tools an AI system may call, including their names and accepted arguments. Checking it can reject invented or mismatched calls before anything runs. Read more

toolchain

The collection of software used to turn code into a running program, including compilers and related build tools. A pinned toolchain fixes the exact versions. Read more

TPU

Tensor Processing Unit, Google's own AI chip. Built for one job, used mostly inside Google, and the main working alternative to buying Nvidia. Read more

trajectory

The recorded sequence of an AI agent’s actions and observations while attempting a task. It lets reviewers inspect the route to an answer as well as the answer itself. Read more

tranche

One slice of a financing, ranked by who absorbs losses first. Junior tranches take the first losses and pay more for it; senior tranches are repaid first and pay less. Read more

tree-sitter

A parser library that turns source code into a structured tree, fast enough to rerun as you type. It is how a tool understands code across many languages without writing a parser for each. Read more

two-factor authentication

A sign-in or approval check requiring two kinds of evidence, such as a password and a security key. A stolen password alone should not be enough. Read more

typed interface

A software boundary that specifies which kinds of inputs and outputs a function accepts, helping catch incompatible calls before execution. Read more

U

underwriter

The banks that organize a share sale, set the price and take responsibility for placing the shares. Hiring them is a public signal that a listing is real. Read more

unified memory

One pool of memory shared by a chip's CPU and GPU rather than separate banks, so data is not copied between them. Apple's M-series uses it, and the pool's size caps how large a model a machine can run locally. Read more

V

valuation

What a company is deemed to be worth, usually set by whatever price its most recent investors paid. An opinion with a number attached. Read more

vector database

A database that stores text or images as numerical embeddings so AI systems can search by meaning. Chatbots keep their source documents in one, so an exposed instance leaks the documents. Read more

vendor financing

A supplier lending money, or guaranteeing loans, so customers can buy its products. It flatters reported demand, since sales are partly funded by the seller, and it sank telecom equipment makers like Lucent when their borrowers failed in 2001. Read more

verifier

A program or model that checks another system’s answer or action. Its own errors affect whether training rewards useful behavior. Read more

vibe coding

Building software by describing what you want to an AI and accepting what it writes, without reading the code closely. Coined by Andrej Karpathy in 2025; now a product category. Read more

virtual machine

A software-defined computer with its own operating environment. It can isolate an AI agent and its files from other workloads on the same physical machine. Read more

vision-language model

A model trained to read images and text together, so it can answer questions about a document, screenshot or photo. Smaller ones increasingly handle parsing work that once needed a frontier model. Read more

vision-language-action model

A model that takes camera input and instructions and outputs robot motions, the robotics counterpart of a chat model. Most current humanoid autonomy claims rest on one, so independent evaluation of them matters. Read more

vLLM

Software for serving language models efficiently. Specialized model formats may require extra integrations before they can run. Read more

VLOSE

The EU's label for a search engine with over 45 million monthly European users. Designation triggers the Digital Services Act's strictest duties: risk assessments, independent audits and researcher data access. Read more

W

wafer-scale

A chip built on a whole silicon wafer instead of being cut into many small dies, so one processor is the size of a dinner plate. Cerebras uses it to keep a huge model on a single piece of silicon. Read more

warrant

A contract giving the holder the right to buy a company's shares at a set price later. Companies grant them instead of cash: Marvell's warrant pays Google in potential stock for hitting purchase milestones. Read more

watermarking

Embedding an imperceptible statistical pattern in AI-generated text or media so detection tools can identify it later. It survives copy-paste; heavy editing can wash it out. Read more

waveform

A record of how a signal changes over time. Chip engineers inspect simulated waveforms to find where a circuit’s behavior first went wrong. Read more

WebAssembly

A portable instruction format that browsers and other hosts can execute. It lets software written in several languages run within a controlled environment. Read more

weight space learning

Training on the numerical weights of existing neural networks to learn from what those networks already encode. It differs from training on their generated answers. Read more

weights

The numbers a model learned during training. They are the model: copy the weights and you have copied it, so publishing them is a bigger decision than publishing a paper. Read more

wet lab

A laboratory that performs physical experiments with biological or chemical materials, rather than relying only on computer simulations. Read more

worktree

A separate working directory attached to the same Git repository. It lets experiments or agents change different versions of the code without overwriting one another’s files. Read more

world model

An internal simulation of how things behave, so a system can predict what happens next instead of matching patterns. The current bet for getting robots to handle situations they were never shown. Read more

world-action model

A robot model that learns to predict what the world does and what to do in it as one system, trained on recordings of real robot actions rather than adapted from a video generator. Read more

Z

Z3

A software solver that checks whether logical and mathematical constraints can all be satisfied. Its answer applies to the formal statement it receives. Read more

zero data retention

An AI vendor commitment not to store a customer's prompts or outputs after processing them, beyond the moment needed to generate a response. Trades away the pattern-spotting that comes from keeping logs. Read more

zero-click

An attack that compromises a device with no action from the victim: no link clicked, no file opened. The most valuable class of exploit because there is nothing to train users against. Read more

zero-day

A software vulnerability the vendor does not yet know about, so no patch exists. Named for the number of days they have had to fix it. Read more

zero-shot

A model performs a task or works in a setting without additional training examples from that setting. Success still depends on what it learned previously. Read more

Zstd

A lossless data-compression format. Correct software must recover the original data and handle the format's edge cases, not merely compress a few sample files. Read more