The AI Read

Glossary

The Jargon, Defined

79
Terms
79
With a source

Terms are marked in the articles themselves the first time each one appears. Everything the publication uses that a reader could reasonably not know should end up here.

Find

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

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

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

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

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

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

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

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

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

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

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

CLI

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

market cap

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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