AI Glossary · Daily AI Thread

A

AAIF: The Agentic AI Foundation, a nonprofit standards body founded by major AI companies to define interoperability standards for AI agents.
acquihire: A deal in which a company licenses another startup's core technology and hires its key staff, rather than formally acquiring the company itself.
agency law: The area of law that governs legal responsibility when one party (a principal) authorizes another (an agent) to act on their behalf. It has always assumed the 'agent' is a human, which is exactly what makes applying it to AI agents complicated.
agentic AI: AI that plans and carries out multiple steps of a task on its own, without a person directing every step.
AGI: A term for AI that can perform intellectual tasks across many domains the way a human can; experts disagree on whether any system has actually reached it.
AI agent: An AI agent is a program that plans and carries out multi-step tasks on its own after receiving an instruction, rather than just chatting back and forth.
AI agents: AI software that goes beyond answering questions to independently carry out multi-step real-world tasks, like processing an email or approving a refund.
AI gateway: A middleman service that connects many AI models through one interface, letting developers pick whichever model they want without wiring up each provider's API separately.
AI native: Redesigning an organization and its workflows around AI from the ground up, rather than adding AI on top of how work already gets done.
AI slop: Low-quality, often meaningless content mass-produced with AI and posted online.
AIDC: A legal designation for AI-dedicated data centers that qualify for government power and permitting benefits.
alignment: The field of research focused on making AI systems act according to what humans actually want and value.
alignment failure: Alignment failure means an AI system behaves in ways its developers didn't intend, breaking the safeguards meant to keep it acting safely and predictably.
alignment lead: The person in charge of making sure an AI model behaves safely and does what people actually want. Inside a company, this role is usually seen as the first to notice and flag dangerous behavior.
alignment research: The field of research focused on checking whether an AI behaves the way people intend, and fixing it when it doesn't.
anchor investor: A large investor who commits to buying a big block of shares in an IPO before it opens to the public, giving other investors confidence to follow.
annual recurring revenue: Revenue a company expects to keep collecting steadily, converted into a one-year figure. Investors use it as a quick gauge of how fast and how reliably a startup is growing.
annual run rate (ARR): An estimate of yearly revenue based on a single recent month's pace, a metric growth-stage startups often use to show momentum even while still unprofitable.
annualized revenue run rate (ARR): A metric that projects a company's revenue for a full year based on its revenue over a shorter recent period.
Apache 2.0 license: An open-source software license that lets anyone freely use, modify, and even sell or commercialize the code, with minimal restrictions.
API: A defined set of rules and channels that let different pieces of software exchange data. For example, a gym booking app sends requests like "reserve this" or "cancel this" to a server through its API.
ARC-AGI-3: A benchmark that tests how well an AI can figure out brand-new rules it has never seen before, on its own.
ATS: Software companies use to collect job applications and manage the hiring pipeline. Some versions use AI to score how well a resume matches a job posting and rank candidates by that score.
attention: A core computation in AI models that calculates how relevant each word is to every other word in a passage, letting the model focus on the important parts. It underpins today's dominant transformer models, but its computational cost grows sharply as text gets longer.
automated license plate reader: A surveillance camera system installed on roads that automatically photographs and logs passing vehicles' license plates into a database. Often abbreviated as ALPR.
autonomous AI agent: An AI system that can set its own intermediate goals and carry out a chain of multi-step actions — searching, running code, manipulating files — without a human directing every step.
AVI score: A single-number score measuring how much impact one genetic variant has, applied even to the non-coding regions that don't produce genes.

B

bacteriophages: A virus that infects and destroys bacteria specifically. It's being studied as an alternative treatment for infections that no longer respond to antibiotics.
behind-the-meter power: A setup where a data center generates its own power on-site instead of drawing from the public grid, letting it start operating quickly without waiting for grid interconnection or risking outages.
benchmark: A standardized set of tasks or questions used to measure and compare AI models' performance against each other.

C

C2PA: An industry standard that attaches a record of when and with what tool a photo or video file was made. It stands for Content Provenance and Authenticity, developed jointly by companies including Adobe, Microsoft, and Google.
capex: Large-scale spending on physical assets like data centers and servers, used to build out capacity for the future.
capital expenditure: Money spent on long-term equipment and assets, like data centers and servers — the kind of spending AI companies are pouring into infrastructure.
Capture The Flag (CTF): A cybersecurity exercise where participants try to find hidden information inside a network; it's also commonly used to test how capable an AI model is at hacking.
chain-of-thought: A step-by-step written record of an AI model's reasoning before it gives an answer. Reading it lets people see why the model reached that answer, or spot signs it's about to do something it shouldn't, which is why it's used for safety monitoring.
chain-of-thought monitoring: A technique where humans or another AI system reviews the step-by-step internal reasoning a model produces on its way to an answer, watching for signs of dangerous or anomalous intent.
child sexual abuse material (CSAM): Illegal content that uses AI to depict children sexually, either through fabricated imagery or by grafting a child's likeness onto sexual images.
CI/CD: The automated pipeline that tests code developers write and pushes it into a live service.
computer use agent (CUA): An AI agent that operates a computer directly - clicking, typing, navigating websites and programs - to complete tasks on a person's behalf.
context window: The maximum amount of text an AI model can take into account at once. Anything beyond this range — earlier conversation or document content — the model can no longer factor in.
control plane: The part of a cloud system that configures and manages it, separate from the part that actually stores and processes data.
CRISPR gene editing: A technology for precisely cutting or altering specific spots in DNA to edit genes, using a protein such as Cas9 as the 'scissors.'
CSAM: Images or videos that sexually exploit children. AI-generated versions are treated the same as real photos under child pornography law.
custom silicon: A chip a company designs in-house and tailors specifically to run its own AI models, instead of relying on general-purpose GPUs bought from outside vendors.
CVE: A unique ID number assigned to a specific software security flaw, letting security teams worldwide refer to and track the same vulnerability by a common name.
CWE: A standardized list that categorizes common types of software security weaknesses so they can be tracked and compared consistently.

D

data poisoning: Deliberately mixing false or distorted information into the data an AI trains on, so the resulting model learns something wrong.
deepfake: A fake image or video created by AI to realistically swap or synthesize a person's face or body.
diffusion: A technique originally used for image generation that gradually refines noisy data into a finished output. Unlike generating text word by word in sequence, it can produce an entire passage at once, making it faster.
Digital Markets Act (DMA): An EU law designed to stop giant platforms like Google and Apple from abusing their market power. Violations can trigger fines of up to 10% of a company's global revenue.
Digital Services Act: An EU law that holds online platforms more accountable, including provisions letting researchers access platform data for study.
digital twin: A virtual replica of a real factory or piece of equipment, used to simulate and troubleshoot it before touching the real thing.
distillation: A technique where a smaller model is trained to mimic a larger model's problem-solving process, essentially copying its reasoning. It's a legitimate way to shrink models, but doing it covertly against a rival's model amounts to unauthorized copying.
DMCA: A US copyright law that, among other things, prohibits circumventing technology used to protect copyrighted content.
drone swarm: A mode of operation where multiple drones communicate with each other and move as a coordinated group — dividing up targets or holding formation — without a person individually piloting each one.
dual-use: Describes a technology or task that can be used for both legitimate defense and malicious attack — like the ability to find security flaws. This dual nature is why AI companies struggle to decide how much capability to allow.

E

edge AI: AI that runs directly on a phone or small device instead of sending data to a cloud server. It works without an internet connection and keeps sensitive information from ever leaving the device.
embedded evaluators: Outside reviewers who work inside an AI company to verify its safety commitments are being kept, similar to bank examiners stationed inside financial firms.
embodied AI: AI that has a physical body, like a robot, and acts in the real world, as opposed to a chatbot that only works inside a screen.
embodied reasoning: The ability of an AI to understand real-world visual input, like a camera feed, and use it to plan and adjust physical actions in the real world, as a robot would.
ERCOT: The independent operator of Texas's power grid, which runs separately from the rest of the US electrical system and supplies most of the state.
EU AI Act: The European Union's AI regulation, which requires AI-generated audio, image, video, and text to carry machine-detectable labels and allows fines of up to 3% of a company's annual revenue for violations.
EU AI Act Transparency Code: A European Union rule requiring AI-generated content to be identifiable as such to users; AI companies operating in the EU can face penalties for not complying.
export controls: Government restrictions that block selling or transferring certain technology or products to other countries or companies, usually for security or diplomatic reasons — advanced AI models and chips have recently been a major target.

F

F1 score: A single number that measures how accurate a model's predictions are. The closer it is to 100, the closer the predictions are to correct.
fair use: A copyright exception that allows certain uses of protected material without permission. Courts weigh factors like how much the use transforms the original and whether it harms the market for it.
Forward Deployed Engineer: An engineer embedded directly with an enterprise customer to prepare that company's data and systems so AI tools can use them.
Forward-Deployed Engineers: Engineers embedded directly with enterprise customers to integrate and apply AI tools to real workflows, addressing the problem that selling a model alone doesn't get companies to actually use it.
foundation model: A large AI model pretrained on vast data so it can serve as the reusable base for many different applications, from search summaries to chatbots.
free cash flow: The cash a company has left over from its operations after subtracting money spent on things like equipment and facilities — the actual cash available for dividends or reinvestment.
frontier: Industry shorthand for the most advanced, highest-capability AI models available at a given time.
frontier AI model: The most capable, cutting-edge AI models available at a given time — considered to carry the highest potential risk, which puts them at the center of safety regulation debates.
frontier model: Industry shorthand for the most capable AI models available at a given moment — the current state of the art, as opposed to smaller, cheaper models built for specific tasks.
full-duplex: A mode of communication, like a phone call, where both sides can talk and listen at the same time. Unlike older voice AI that waits for one side to finish, it can naturally interrupt or respond while the other side is still talking.
full-duplex voice: A voice-conversation mode where both sides can listen and speak at the same time, like a phone call — instead of the stiffer take-turns pattern most voice assistants use.

G

GDPval: A benchmark OpenAI built to test how useful an AI's output is across 44 real professions, like accounting or medicine, using hundreds of realistic work scenarios.
General-Purpose AI (GPAI): An AI model versatile enough to be used across many different tasks — like translation, coding, or conversation — rather than one built for a single purpose; the EU AI Act imposes extra transparency obligations on this category.
GGUF: A file format for packaging large AI models so they can run on ordinary personal computers, no specialized server hardware required.
gigawatt: A unit of electrical power equal to one billion watts. A single large AI data center can use as much electricity as hundreds of thousands of homes, and gigawatts are the unit used to measure that scale.
goal drift: When an AI agent gradually strays from its original instructions and starts pursuing a goal it decided on its own.
guardrails: Safety rules built into an AI agent in advance so it can't take actions outside its approved scope.

H

hallucination: When an AI model states something false or unsupported as if it were fact.
hand strength model: A model that analyzes a player's behavioral data to estimate the type and probability of the cards they are likely holding.
Harness: A harness is the surrounding system that keeps an AI model on task across many steps, letting it plan, act, check its own results, and try again toward a goal, rather than answering once and stopping.
HBM: Memory made by stacking multiple DRAM layers so far more data can move at once; it is essential for training and running AI models.

I

induced pluripotent stem cell: A mature cell — from skin or blood, for example — that has been chemically reverted to an embryonic-like state. In this state it can develop into almost any cell type again, making it the starting material for growing organoids.
inference accelerator: A specialized chip designed to quickly handle only the step where an already-trained AI model generates its answers.
intent collision: A term security researchers use for when an AI blends a user's legitimate request with malicious instructions hidden on a webpage and acts on both as one.

K

KV cache: A temporary store of results an AI model has already computed while processing a conversation. Being able to hand this off directly to another model, instead of recalculating everything from scratch, is what makes switching between models faster.

L

layered multi-model strategy: A strategy of relying on multiple AI models instead of a single one, choosing among them based on the situation or risk level. It spreads out risk so that losing access to one country's or company's model doesn't leave a business stranded.
leverage: Borrowing money to increase the size of an investment — it amplifies gains, but it amplifies losses just as much.
LLM: An AI model trained on huge amounts of text that can understand and generate language in a human-like way.
LPU: A specialized AI chip that works alongside GPUs to speed up how quickly a chatbot generates its response.

M

machine identity: The account and credential AI agents or software programs use to access systems, as opposed to a human user. It has become a new security management category as automation grows.
margin call: A demand from a broker for an investor to add more collateral or repay debt when the value of assets bought on borrowed money falls. If the investor can't meet it, the broker force-sells the assets — which can deepen a market downturn by triggering even more selling.
MCP: A standard way for an AI chatbot to connect directly to outside services or databases, like Bloomberg, to pull in information.
memory poisoning: An attack where someone plants malicious instructions in an AI system's long-term memory or reference data to skew its future decisions.
Millennium Prize Problem: Seven of the hardest unsolved problems in mathematics, named by a US math institute in 2000; solving one earns a $1 million prize.
Millennium Prize Problems: One of seven landmark unsolved problems in mathematics named by the Clay Mathematics Institute in 2000, each carrying a $1 million prize; only one has ever been solved.
misalignment: A term for when an AI model behaves differently from what its creators intended. It isn't necessarily malicious, but it's treated as a safety issue because the behavior has slipped outside human control.
Mixture of Experts (MoE): An AI model design that uses many smaller "expert" networks instead of one giant network, activating only a few of them for each query — letting a model save on computation relative to its size.
Mixture-of-Experts: An architecture that holds many smaller 'expert' sub-networks inside one model and activates only a few of them for each input. It lets a model have a large total parameter count while keeping the actual computation per request much smaller, so it can approach big-model performance at a lower cost.
Model as a Service (MaaS): A way of offering an AI model where users call it over the internet like an API, instead of downloading and running it themselves.
model collapse: A phenomenon where AI models trained repeatedly on text generated by earlier AI models gradually produce lower-quality, less diverse output.
model router: A system that automatically picks a bigger, pricier model or a smaller, cheaper one depending on how hard a given prompt is — used to cut costs while keeping the performance a task needs.
model weights: The numerical values an AI model learns during training — they are, in effect, the model's actual knowledge and capability. Losing them means retraining from scratch with the same data and compute.
monthly active users: A metric counting how many people used a service at least once within a given month — a standard way to measure how widely and consistently a product is being used.
mTLS: A security method where two computers verify each other's identity before opening a connection, stricter than the more common approach of only checking one side.

N

Navier-Stokes equations: A 200-year-old physics equation describing how liquids and gases flow. Whether smooth solutions always exist has never been proven, making it one of seven Millennium Prize Problems worth $1 million each.
neocloud: A term for companies that specialize in renting out GPU compute capacity without building their own AI models.
network segregation: A security practice, common in Korean finance, that physically separates a company's internal work network from the internet-connected outside network to block hacking and intrusion. Because this separation is legally required, using external cloud-based generative AI directly is difficult.
neurodivergent: A term describing people whose brain and nervous system develop differently from the statistical 'norm' — such as autism or ADHD — framed as natural variation rather than disorder.
NPU: A chip built specifically for neural-network computation that can do the same work as a general-purpose GPU using less power.

O

observability: The ability to understand what's happening inside a system in real time by tracking and recording its internal behavior from the outside.
OCR: OCR converts text in a scanned document or photo into computer-readable text. It only transcribes the characters; it doesn't check whether the content is correct.
off-target effect: An unwanted side effect in which a gene-editing tool cuts or alters the wrong spot in the DNA instead of its intended target.
offboarding: The process of revoking a departing employee's access to company accounts, files, and systems.
on-premise: Running software and data on a company's own servers instead of sending it to an external cloud provider.
opaque recurrence: A technique where a model repeats its intermediate reasoning in an internal form rather than human-readable language, making it hard for people to verify how it reached its conclusion.
open-weight: An AI model whose parameter weights are published so anyone can download, run, and modify it directly.
orchestrator: A system that sits above multiple AI models and routes each task to whichever one fits best, rather than relying on a single model for everything.
organoid: A lab-grown clump of cells, cultured from a person's skin or blood, that functions like a miniature version of a real organ. Brain organoids contain millions of neurons that produce electrical signals resembling real brain waves.
outcome-based pricing: A pricing model that charges only when the AI actually completes a task, instead of a flat subscription or usage fee.

P

package manager: A developer tool for installing and managing software components — in this incident, AI agents repurposed one as an impromptu message board to communicate with each other.
parameters: An internal value inside an AI model that gets adjusted during training. The total count is often used as a rough measure of a model's size.
physical AI: AI that operates in the physical world, directly controlling things like robotic arms or lab equipment, instead of staying confined to software.
post-quantum cryptography: A new generation of encryption designed to remain secure even against the computing power of future quantum computers, which are expected to break most of today's encryption methods.
post-training: The stage after a model's initial large-scale training where it's refined using human feedback or additional data to better match real-world use. This step largely determines the quality and safety of a model's actual responses.
preliminary injunction: A court order that temporarily blocks a specific action before a case is fully decided.
Preparedness Framework: An internal process AI companies use to assess a model's risk level (biological, chemical, cyber, nuclear, etc.) before release, and to decide what safeguards or restrictions apply based on that risk tier.
Prompt injection: An attack that hides malicious instructions inside a document or webpage an AI will read, tricking it into following those instructions instead of the user's actual request.
prompt injection attack: An attack that hides malicious instructions inside content an AI reads, like a webpage or document, tricking an AI agent into taking actions the user never intended, such as leaking data or making an unauthorized purchase.
public benefit corporation: A legally defined company structure that requires pursuing a stated public benefit alongside shareholder profit.

R

RAG: A technique where a system retrieves relevant documents before answering a question, then generates its response grounded in that retrieved content — letting a model draw on information it wasn't trained on, like recent or internal documents.
reasoning trace: The step-by-step internal thinking an AI model goes through before producing a final answer. Companies typically keep this hidden so competitors can't copy it.
recurrent depth: A reasoning technique that processes information by cycling it through a model's internal layers repeatedly, rather than expressing it in visible language. It leaves fewer human-readable traces, making it harder to monitor.
recursive self-improvement: The process by which an AI system designs a more capable version of itself without human intervention.
red team: A team or exercise that deliberately takes an attacker's perspective to find weaknesses before real adversaries do. At AI companies, it refers to teams that probe how a model might be misused or behave unexpectedly by simulating real attacks.
regulatory sandbox: A designation from Korea's financial regulator that lets a new financial service be tested in the market for a set period before full regulations apply.
reinforcement learning: A training method that doesn't hand the AI the right answer directly. Instead it rewards actions that lead to good outcomes, so the model gradually learns to act better.
remote code execution: A serious security flaw that lets an attacker run arbitrary commands on someone else's computer over the internet, without physical access.
reward hacking: A pattern where an AI model gets rewarded for reaching a goal through some unintended shortcut, so that shortcut gets reinforced instead of corrected. A habit picked up this way during training can later show up as problem behavior in real use.
RLHF: A training method where humans rate an AI's responses and the AI is retrained on those ratings, the key technique behind making chatbots more natural and useful.
runtime trust: A security approach that keeps checking whether an AI agent is still behaving as intended while it's running, not just at the moment it first logs in.

S

SaaSpocalypse: A term for market fears that generative AI would gut the value of traditional software-as-a-service (SaaS) companies.
safety classifier: An automated screening system that flags risky or potentially abusive requests to an AI model, routing them to a more restricted model or blocking the response when needed.
sandbox: An isolated environment used to test an AI model or program separately from real systems, so that if something goes wrong, it doesn't affect anything outside it.
self-replicating malware: Malicious code that copies itself and spreads to other machines or processes without a human re-launching it each time, making it harder to contain than ordinary malware.
self-replication: The behavior of an AI model copying and running itself onto other machines without explicit external instruction.
semantic segmentation: A computer-vision technique that classifies every single pixel in an image into a category like "building," "road," or "tree." Unlike rough object detection, it traces the exact outline of each object.
Sherman Antitrust Act: A US law over a century old that bars companies from colluding to restrict competition or limit output. There's concern it could also apply if AI companies agreed together to slow down development, even for safety reasons.
SoC: A chip design approach that packs multiple computing components onto a single piece of silicon instead of using separate chips.
sock puppet: A fake account made to look like a real person, used to deceive others (including AI) into treating its statements as trustworthy.
sovereign AI: An approach that lets a country or company control its own AI models and infrastructure instead of depending on foreign technology, mainly from the US.
spear phishing: A targeted scam email attack aimed at a specific person or organization, as opposed to mass phishing sent to random recipients.
steel skids: A rectangular steel frame built to be moved easily. It serves as the movable foundation for prefabricated, modular facilities like data centers or reactors.
supply-chain attack: A hack that infects the software or components a target relies on, instead of attacking the target directly, to sneak in through a side door.
supply-chain risk: A government designation meaning a company's product is considered vulnerable to hacking or sabotage by an adversary; agencies are then barred from using it.
SynthID: A watermarking technique developed by Google that embeds a detectable signal in the subtle word-choice patterns of AI-written text.

T

telemetry: The automatically generated logs and monitoring data a system produces while running. When the volume gets enormous, spotting one anomalous signal inside it gets harder, not easier.
teraflops: A unit measuring a computer's ability to perform one trillion calculations per second; a higher number means more computation can be handled at once.
text watermarking: A technique that embeds an invisible pattern into AI-generated text, detectable only with the right method, so that AI-written content can be identified after the fact.
token: The basic unit an AI model breaks text into when reading or writing; it can be smaller or larger than a single word. Using more tokens means higher cost and processing time.
token tax: A proposed tax where governments would collect a share of the revenue AI companies earn from their services and use it to support workers displaced by AI. No country has actually implemented one yet.
tokenmaxxing: A slang term for burning through AI tokens to boost a personal usage ranking, regardless of whether the work actually needed that much AI.
TPU: A custom chip Google designed specifically to speed up AI computation, similar in role to a GPU.
trade secret: Confidential business or technical information that a company protects by keeping it secret rather than registering it like a patent; obtaining and using it improperly can lead to legal liability.
Transfer Station: A "transfer station" is a proxy server network in China that resells restricted access to AI services at a markup, logging in with a legitimate account and reselling that access to others.
transformers: The AI architecture, introduced in 2017, that underpins nearly every large language model today, including ChatGPT and Claude. It uses a computation called attention to capture relationships between words in a passage.
turnless speech model: A voice model designed to respond and interject naturally while the user is still speaking, instead of waiting for a fixed turn to end before replying.

U

unified memory: A design where a computer's CPU and GPU share one pool of memory instead of using separate ones. It cuts down on data movement, which helps when running large AI models.

V

venture capital: Funding provided to early-stage, high-growth companies in exchange for equity, or the firms that provide it.
Very Large Online Search Engine: A designation the EU gives to online search services with more than 45 million monthly EU users, triggering the strictest user-protection and transparency rules.
vibe coding: Vibe coding means describing what you want in plain language and letting an AI agent write the code, rather than typing it yourself line by line.
VLM (Vision-Language Model): A vision-language model is an AI that understands images and text together, letting it read visual details in a document, like tables, stamps, or handwriting, that plain character recognition would miss.
VRAM: Dedicated memory built into a graphics card, providing the space needed to run AI models quickly.

W

warrants: A financial instrument giving the right to buy stock at a set price at a future date. It locks in the option to buy shares later, not an actual purchase now.
watermark: An invisible marker embedded in AI-generated text or images that can later be used to verify whether the content was made by AI.

Z

Zero Data Retention: A policy under which an AI provider does not store a user's conversation data on its servers, discarding it immediately after processing.
zero-day: A security vulnerability that's unknown to the software's makers or unpatched — or an attack that exploits one — before a fix exists.
zero-day vulnerability: A software flaw that its own maker doesn't yet know about, meaning it can be exploited by attackers before any fix exists.
zero-shot: When an AI performs a task correctly without being specifically trained or shown examples for that exact task first.