The Glossary
The vocabulary, defined once.
Every specialist term the briefs use, in plain language. The list grows on its own: as the newsroom reads and briefs new coverage, the writer flags terms a senior reader outside that subfield might not know, and they land here — 640 so far. Terms in briefs link back to this page.
#
--dangerously-skip-permissions
A command-line flag for AI coding agents that disables per-command approval prompts, letting the agent execute commands without human review. The flag's name itself flags the risk.
0-dayzero-day
A software vulnerability unknown to the vendor or maintainer, so no patch exists when it is first found or exploited.
10-Q
The quarterly report US public companies file with the SEC, containing unaudited financial statements and required disclosures such as commitments and guarantees.
30B30 billion parameters
Shorthand for a model with roughly 30 billion parameters, the learned numerical values that drive its behavior. Parameter count is a rough proxy for capability and hardware requirements.
4-hi
Stack-height notation counting the DRAM dies piled vertically in one HBM package. Taller stacks (12-hi, 16-hi) add capacity per package; shorter stacks (4-hi, 8-hi) are cheaper per unit of bandwidth.
4oGPT-4o
An OpenAI model powering ChatGPT. The article cites lawsuits alleging its overly sycophantic behavior contributed in part to multiple users' suicides.
8-KSEC Form 8-K
A US Securities and Exchange Commission filing that public companies must submit promptly after major events such as acquisitions or leadership changes. It is the official record investors check for deal terms.
A
Activated parameters
The subset of a model's total parameters used to process each token. Sparse designs keep total parameters high for capacity while activating few, cutting compute cost per token.
active learning
A training workflow where the model helps select which data humans should label next, so fewer labels yield more improvement.
active parameters
In sparse models, only a fraction of total parameters processes each token. Ling-3.0-flash has 124B total but roughly 5.1B active per token, which is what drives its serving cost.
Ad Pilot Program
OpenAI's early advertising initiative for ChatGPT, launched in February with brands including Adobe, Albertsons, Target and Williams-Sonoma.
adjusted operating income
A profit measure that excludes certain expenses or one-time items a company considers non-core. Because each company chooses its own adjustments, the figure is not directly comparable across firms.
ADPAuthorized Data Publisher
A CISA program that lets vetted organizations add enrichment data, such as severity assessments, to CVE records. CISA's ADP agreed with the critical ratings of the disputed SQLite CVEs.
advanced packaging
Manufacturing techniques that combine multiple chips or dies into one package to boost performance. Capacity constraints here are among the component shortages raising AI hardware prices.
advanced reactor
A newer class of nuclear reactor designs, often smaller and factory-built, intended to be cheaper, safer, or faster to deploy than conventional large light-water plants.
Adversarial review
A separate AI pass that independently challenges the conclusions of the primary analysis pipeline, intended to catch errors before a human makes the final call.
Agent
An AI system that takes actions toward a goal (calling tools, browsing, writing files, spawning subtasks) rather than only answering a prompt. "Agentic" describes workflows built this way.
Agent Card
Mercury's product for issuing a virtual card to an AI agent, giving the agent its own payment credential, spending limits, and audit trail separate from any human user.
agent collusion
Cooperation among AI agents to gain an advantage on a task in ways their developers did not intend, as the authors define it — for example, sharing answers through a channel that was supposed to be blocked.
agent harness
The scaffolding around a model that supplies tools, memory, and an execution loop so it can take multi-step actions, as distinct from the model itself. The article lists 'Codex the harness' as a separate component from the Codex app.
Agent model
A model optimized to carry out multi-step tasks — planning, calling tools, and acting on results — rather than only answering single prompts.
Agent swarm
Multiple autonomous AI agents operating in parallel under one operator, dividing up tasks such as crawling or data collection.
Agentforce
Salesforce's AI product, used for tasks such as helping salespeople close deals and automating customer service interactions. Salesforce offers custom contracts for it priced on outcomes.
agentic
Describes AI systems that autonomously plan and take actions, such as running code or using tools, toward a goal, rather than only generating text in response to prompts.
agentic AI
AI systems built to take multi-step actions toward a goal — planning, executing, and checking work — rather than answering a single prompt. Sarvam's Indus platform bills such products through one credit system.
agentic coding
Using an AI model to autonomously plan and execute multi-step programming work, typically by calling tools such as shells, editors, or test runners over extended runs.
agentic commerce
Commerce in which AI agents make purchases on a consumer's behalf under delegated authority, often with standing instructions that span multiple transactions over time.
agentic eCommerce
Buying and selling mediated by AI agents that discover, compare, and purchase products on a user's behalf, typically by querying structured product data rather than browsing web pages.
agentic inference
Running AI models to serve agent workloads — multi-step tasks where models plan, call tools, and chain many requests — rather than single one-shot prompts.
agentic interface
A way of operating software where AI agents act through APIs and protocols on a user's behalf, rather than a human clicking through a graphical interface.
agentic search
Search carried out by an AI agent that decides what to look for rather than matching fixed queries. Anthropic's transcript scan relied on agentic search and missed the fourth incident.
agentic shopping
Shopping in which an AI assistant acts on a consumer's behalf — finding products, building lists, or completing purchases — rather than only answering search queries.
agentic systems
AI systems that can plan and carry out multistep tasks on their own, rather than producing a single response to a prompt.
agentic tasks
Tasks in which a model autonomously plans and carries out multi-step workflows, such as calling tools and chaining actions, rather than answering a single prompt.
agentic workflow
An AI-driven sequence in which a model plans and executes multi-step tasks with some autonomy, rather than answering one prompt at a time.
AgentPay Card
WeChat Pay's payment instrument designed for AI agents. Funds are kept separate from the user's main WeChat Pay account, and every payment requires final confirmation on a phone.
AgentX
The agentic inference benchmark used in this article; it replays real-world agent traffic across a fleet of thousands of chips to compare platforms' efficiency.
Aging clock
A computational model that estimates biological age — how old a body's cells and tissues appear — from molecular biomarkers such as DNA methylation, rather than from chronological age. Researchers use aging clocks to measure whether interventions slow or reverse aging.
AI agent
Software that uses a model to plan and execute multi-step tasks with limited human oversight — here, writing and changing code rather than merely suggesting it.
AI agents
Software that carries out multi-step tasks on a user's behalf with some autonomy. OpenAI groups products like Codex and ChatGPT Work under this label.
AI content detector
Software that estimates whether a text was machine-generated, typically by scoring linguistic patterns. Its false-positive rate measures how often human writing is wrongly flagged as AI-generated.
AI factoryAI Factory
A large-scale, purpose-built data center designed to train and run AI models, often with dedicated power and cooling infrastructure.
AI gateway
A proxy layer between applications and LLM providers that offers a single unified API, centralized key management, and governance controls such as logging and rate limits. LiteLLM is one example.
AI model marketplace
A platform where developers discover, access, and route work to AI models from multiple providers. OpenRouter, which Stripe agreed to buy for roughly $8 billion, is one example.
AI native
An operating model in which AI systems perform core work by default rather than merely assisting humans. Meta used the term as the stated goal of its Project OT restructuring.
AI Office
The European Union body that enforces the AI Act's rules for providers of general-purpose AI models, including models posing systemic risk, and for AI systems offered by the same provider as the underlying model.
AI Opportunity Statement
A U.S.-published statement from June, signed by 35 countries, outlining how the U.S. and strategic partner economies can collaborate on AI, including commitments to pro-innovation regulation and supply-chain alignment.
AI proctoringAI-powered remote proctoring
Software that supervises remote test takers through their webcam, microphone, and screen, using machine-learning models to flag behavior that may indicate cheating.
AI roll-up
A business strategy where a company acquires multiple small service firms and replaces their human-driven processes with AI automation.
AI slop
Low-quality, mass-produced AI-generated content. Here it refers to fabricated bug reports that look plausible but describe nonexistent vulnerabilities, overwhelming human review queues.
AI Story Optimization
PR firm Piro's term for authoring content engineered around the signals LLMs use to judge credibility, such as structure and neutral tone, so chatbots are more likely to repeat its framing.
AI Token Spend Management
A Ramp product launched in July that tracks what companies spend on AI subscriptions and API usage, distinct from purchases made by agents themselves.
AI-exposed occupations
Jobs whose core tasks overlap heavily with what current AI systems can perform, as classified by researchers. The Stanford study contrasts these fields with less-exposed ones to measure AI's employment effects.
AIM3
Recorded Future's framework for grading AI model autonomy. The article places models that can autonomously execute an end-to-end cyberattack at its highest autonomy level.
AIPACAmerican Israel Public Affairs Committee
A prominent pro-Israel lobbying group in U.S. politics. One Hanover Institute report title asks whether it uses "dark money" in elections.
Air intercept
A maneuver in which one aircraft closes with and targets another in flight; the X-62 performed 27 of them autonomously in this demo.
AISIAI Safety Institute
The UK government body that evaluates frontier AI systems for dangerous capabilities through controlled testing. It designs evaluation protocols and can tighten requirements, such as internet-access rules for agents.
alignment
The problem of ensuring advanced AI systems reliably pursue intended goals and remain under human control. It is a central research focus at labs like Anthropic and OpenAI.
Alignment research
Work aimed at ensuring AI systems pursue intended goals and behave safely as their capabilities grow. Anthropic argues it needs more time to keep pace with frontier models.
All-in-Infra
Z.ai's name for its 2026 strategy of investing in both training and inference while adopting domestic Chinese chips as its primary inference compute resource.
AlphaFold
DeepMind's database of predicted protein structures, the work that won former DeepMind chief Demis Hassabis a share of the 2024 Nobel Prize in chemistry. The AlphaGenome Atlas is more than 30 times larger.
AlphaGenome
Google DeepMind's AI model that predicts how changes in DNA sequence could affect a person. The AlphaGenome Atlas database is built from its predictions across all 9 billion possible one-letter changes.
Amazon Bedrock
AWS's managed service for building applications on foundation models. Qualcomm plans to use it as part of running its chip-design workloads on AWS infrastructure.
amicus brief
A "friend of the court" filing submitted by a non-party to a case to offer legal or technical argument. The EFF filed one here explaining how the Comet browser's architecture works.
anchor investment
A large early capital commitment from a prominent investor in a share offering, intended to build confidence and attract other buyers.
anchor investor
A large investor that commits capital to an IPO before it opens to the broader market, signaling confidence and helping stabilize demand for the shares.
Annotated images
Images tagged with labels identifying objects, locations, or events in each frame, used as supervised training data so a model learns to recognize the same patterns in new imagery.
Announcement No. 61
China's extraterritorial export-licensing regime for rare earths, magnets, and downstream technology inputs, dubbed the "0.1-percent rule," set to return November 10.
annual recurring revenue
The yearly value of a company's recurring subscription revenue, excluding one-time sales. Investors use it as a growth and traction metric for subscription businesses.
Annualised revenue
A run-rate estimate that extrapolates a company's current revenue pace to a full year. It is a projection of scale, not audited twelve-month sales.
Annualized revenue
An estimate of what a full year of sales would look like if a company's recent revenue pace continued. Used for fast-growing firms whose trailing twelve-month figures lag their current run rate.
annualized revenue run rate
Current revenue extrapolated across a full year, used by fast-growing startups as a headline figure; it is a projection of present momentum, not audited annual revenue.
annualized run rate
A full-year revenue projection extrapolated from a recent period's sales. Crossing a $1 billion run rate means the current pace implies $1 billion per year, not that $1 billion has been collected.
anthropomorphize
To attribute human traits, emotions, or intentions to something non-human. OpenAI documents show contractors training ChatGPT not to present itself as human-like.
anti-distillation
Defenses meant to stop outsiders from copying a proprietary model's behavior or reasoning to train a competing model, a practice known as distillation.
Apache 2.0Apache License 2.0
A permissive open-source license allowing commercial use, modification, and redistribution. Applied to model weights, it lets companies self-host and adapt a model without licensing fees.
APTadvanced persistent threat
A well-resourced attacker that gains long-term, stealthy access to a target. The author borrows the term to argue capable AI agents warrant the same defensive posture.
ARC-AGIAbstraction and Reasoning Corpus
A benchmark of novel visual-logic puzzles designed to resist memorization, used to measure abstract reasoning rather than recall. ARC-AGI-3 is the current generation.
ARC-AGI-3Abstraction and Reasoning Corpus for AGI, version 3
A benchmark testing abstract reasoning from a few examples, designed to resist memorization-based solutions; this is its third iteration.
ARCYBERU.S. Army Cyber Command
The Army component of U.S. Cyber Command, responsible for operating and defending Army networks and conducting cyberspace operations.
ARRAnnual Recurring Revenue
The annualized value of a company's recurring revenue streams. The article's title uses it to size SpaceX's projected datacenter income at $300B.
Artificial Analysis Intelligence Index
A third-party composite benchmark that combines multiple evaluations into a single score to rank frontier AI models. Used here to compare Grok 4.6, GPT-5.6 Sol, and Claude Opus 5.
ASICApplication-specific integrated circuit
A chip custom-designed for a particular workload rather than general-purpose computing, trading flexibility for performance and efficiency on that task.
ASPMApplication Security Posture Management
Tooling that aggregates, correlates, and prioritizes findings from multiple application security tools to give a unified view of risk.
Astra
An advanced capability within OpenAI's platform whose surge in demand the company cites as the reason for pausing new $200 Pro subscriptions.
attribution rules
Contract terms that define how to determine whether a business outcome, such as a sales conversion, was caused by the vendor's software or by other factors like marketing or seasonality.
Auto Mode
A permission mode in Anthropic's Claude Code tool where a classifier approves or blocks commands without asking the user each time.
AutomationBench
A benchmark testing how well AI models can execute common business workflows.
Autonomous hacking
A model independently discovering and exploiting software vulnerabilities without step-by-step human direction, as in the OpenAI case where an agent reportedly used a previously unknown flaw.
average order value
The mean amount a customer spends per transaction. Retailers track it to measure whether features like AI assistants get shoppers to buy more per visit.
AVIAlphaGenome Variant Impact
A score released alongside the Atlas that rates how much a genetic mutation is likely to matter. It blends AlphaGenome's predictions with an earlier model that predicts whether small protein changes are harmful.
B
backlog
The value of signed customer contracts not yet delivered or recognized as revenue. Nscale's figure counts multi-year compute deals whose revenue would arrive over each contract's term.
backup buyer
In a bankruptcy auction, the runner-up bidder designated to complete the purchase if the winning bidder's deal collapses.
bacteriophage
A virus that infects and replicates inside bacteria. Phages are studied as model systems in genetics and as potential alternatives to antibiotics.
BadIIS
A family of malware that infects Microsoft IIS web servers to manipulate the content they serve, commonly used for SEO fraud and traffic redirection.
BBB-
The lowest investment-grade credit rating issued by S&P. A further downgrade would move a borrower's debt into speculative, or 'junk', territory, typically raising its borrowing costs.
Beacon
A server controlled by testers or attackers that a machine contacts after executing a payload, proving code execution occurred and often logging details such as the process chain that launched it.
Benchmark
A standardized test suite for comparing models. Useful but gameable: vendor-reported scores on their own benchmarks deserve the same skepticism as any first-party number.
benchmark saturation
The point where top models cluster near a benchmark's maximum score, so it stops distinguishing capability. The article argues this cycle forces era-by-era evaluation instead of one continuous trend.
BF16bfloat16 (brain floating point, 16-bit)
A 16-bit floating-point number format widely used to store model weights and activations during inference. It serves as the full-precision baseline in this comparison.
big mathematics
Terence Tao's term for a shift toward large-scale, decentralized human–machine collaboration in math, where AI handles technical work and humans keep creative direction.
Bigtable
Google's distributed storage system for structured data, built to scale across very large clusters of machines.
Biological age
An estimate of how old a person's body is based on physiological or molecular markers, which can be higher or lower than chronological age. It is the primary target measure in longevity research.
Black Hat
A long-running information-security conference series where researchers present attacks and vulnerabilities. OpenAI described earlier phases of this incident in a recent Black Hat presentation.
Black Hat USA
A major annual information-security conference where researchers present newly demonstrated vulnerabilities and attack techniques to industry and government audiences.
Blackwell
Nvidia's GPU architecture generation for AI training and inference, succeeding Hopper. Sarvam announced major Blackwell GPU deployments hosted in India.
Blind evaluation
A test in which human judges score model outputs without knowing which model produced each one, intended to reduce bias toward known systems.
Booked revenue
Revenue committed under contract or insertion order for a period, counted when the deal is signed rather than when the ads actually run or cash is collected. Also called bookings.
bpwbits per weight
The average number of bits used to store each model parameter after quantization. Lower bpw means smaller files and less memory use, usually at some cost to model quality.
BTMBehind-the-meter
Power generation sited on the customer's own premises, on their side of the utility meter, rather than delivered over the public grid. Datacenters use it—here largely on-site gas generation—to secure power without waiting on grid capacity.
bug bounty
A program run by a software vendor that pays external security researchers for privately reporting vulnerabilities, so flaws can be patched before public disclosure or exploitation.
build.rs
A Cargo build script that is compiled and executed automatically before the rest of a package builds, typically for code generation or linking native libraries. It can run arbitrary commands, making it a supply-chain execution vector.
Business Trends and Outlook Survey
A recurring Census Bureau survey of hundreds of thousands of US businesses. Since September 2023 it has included questions on whether firms recently used AI and how AI affected their total employment.
C
C2PACoalition for Content Provenance and Authenticity
An open industry standard for cryptographically signing digital content to record where it came from and how it has been modified, letting third parties verify a file's provenance.
cache hit
In LLM API prompt caching, a request whose repeated context is served from storage instead of being reprocessed, billed at a lower rate. Agents that re-read the same files generate many cache hits.
Cached input tokens
Input tokens that repeat an earlier prompt prefix and are served from the provider's cache rather than recomputed, so they are billed at a much lower rate than fresh input.
capacity auction
A market run by grid operators in which power generators commit future electricity supply in exchange for guaranteed payments; clearing prices rise when expected supply falls short of reliability targets.
capexcapital expenditure
Money a company spends on long-lived physical assets such as data centers, servers, and network equipment, as opposed to day-to-day operating costs. Heavy AI buildouts show up as rising capex.
capital expenditurecapex
Money a company spends on long-lived assets such as data centers, chips, and equipment, as opposed to day-to-day operating costs. Heavy capex strains near-term cash flow even when revenue is growing.
capture-the-flag
A security exercise in which participants solve challenges or exploit vulnerabilities to find hidden 'flags.' AI labs use such tasks to evaluate models' cybersecurity skills.
catastrophic risk
In AI policy, the potential for a model to cause severe, large-scale harm, such as enabling weapons development or mass casualties. Bills like the Massachusetts proposal require developers to have independent evaluators test models for these risks.
CCCPCanadian Centre for Child Protection
Canadian charity that identifies online child sexual abuse material, notifies victims when their images surface in new investigations, and works to have the content removed.
CCPACalifornia Consumer Privacy Act
A California privacy law giving consumers rights over their personal information. Its deidentification standards define when data counts as stripped of identifying details.
CDAFAComprehensive Computer Data Access and Fraud Act
California's state-law analog to the CFAA, providing civil claims for unauthorized access to computer data. It can survive even when a parallel federal CFAA claim weakens.
Central CommandUS Central Command (CENTCOM)
The US military combatant command that conducted Operation Epic Fury and used AI-assisted targeting through the Maven Smart System platform.
CFAAComputer Fraud and Abuse Act
US federal anti-hacking statute that prohibits unauthorized "access" to computers. Platforms have often invoked it against scrapers, bots, and rival services, making its scope a recurring fight in tech law.
chain-of-thoughtCoT
The step-by-step intermediate reasoning a model produces before its final answer. Providers often hide it from users because it can reveal proprietary behavior and enable imitation.
Chapter 11
A section of the U.S. Bankruptcy Code letting an insolvent company reorganize or liquidate under court supervision, with assets sold to repay creditors.
chat template
A Jinja template shipped with a model that formats conversation messages into the exact text format the model was trained on. Reading it can reveal supported features and default settings.
CIconfidence interval
A range computed from study data that plausibly contains the true value of a measured effect at a stated confidence level, here 95%. A 95% CI of 25.7-48.0 means the true increase is consistent with values in that span.
CI runnerContinuous Integration runner
A machine that executes a repository's automated build, test, or deployment workflows. Runners frequently hold workflow secrets such as tokens and keys, making them a high-value target.
circular arrangement
A deal where an infrastructure provider invests in a customer that buys its compute or hardware. Critics argue the structure can make real demand look larger than it is.
Claude Artifact
A shareable piece of content created with Anthropic's Claude and hosted on the claude.ai domain. Because links sit on the legitimate domain, a malicious Artifact can look trustworthy.
Claude Code
Anthropic's developer tool; the commerce blueprint includes a plugin for it.
clinical monitoring
Oversight of clinical trial sites and data to ensure protocol compliance and data integrity, traditionally requiring staff to visit sites in person.
ClinicalTrials.gov
A U.S. government registry of publicly and privately funded clinical studies, used to check trial eligibility, status, and reported results.
cloud agent
An agent that runs on hosted remote infrastructure rather than the user's own machine, executing tasks in managed environments. OpenAI's original Codex is described as a cloud agent now brought to the mainstream through Work.
CMSCenters for Medicare & Medicaid Services
The federal agency that administers Medicare and Medicaid, including experimental payment and service models such as WISeR.
CNAPPCloud-Native Application Protection Platform
An integrated security platform covering cloud workloads, configurations, and applications across their lifecycle, replacing multiple point tools.
Code Mode
Cloudflare's approach to LLM tool use that exposes MCP tools as a typed TypeScript API, letting the model write code that calls them instead of emitting one structured tool call at a time.
code-execution sandbox
An isolated container in which an AI assistant runs generated code, intended to prevent that code from reaching other users, the host system, or the network.
Codex
OpenAI's coding-agent product line, which the article describes as existing in three forms: an app, a harness, and the original cloud agent. It is being folded into ChatGPT Work and the consolidated desktop app.
coding agent
An AI system that autonomously reads, writes, and runs code to complete programming tasks. The article's concern is agents that can turn a hint about a bug into a working exploit without human help.
colocation
Renting space, power, and cooling in a third-party data center to house a company's own servers, instead of building a private facility or relying on public cloud.
command channel
A covert communication path an attacker uses to send instructions to, or receive results from, a compromised system. Here it linked code-execution environments belonging to different ChatGPT accounts.
Composer
Klaviyo's marketing agent product. It reached 95,000+ users in its first month, and its first working prototype was built over a single weekend by other agents.
compute financing platform
A financing structure that funds purchases of computing infrastructure, letting customers pay for AI hardware and data-center capacity over time rather than upfront.
concentration risk
The exposure a company carries when a large share of revenue comes from a small number of customers, so losing one or two accounts can materially damage the business.
conference committee
A temporary joint panel of House and Senate members that reconciles differences between the two chambers' versions of a bill so both can vote on identical final text.
Confluence
Atlassian's collaborative wiki and documentation product, often used to store internal knowledge bases and project docs.
Context window
The maximum amount of text (measured in tokens) a model can consider at once: its working memory. Everything the model "knows" about your request must fit inside it.
Control plane
The layer of a system that manages access, credentials, configuration, and orchestration rather than serving end-user traffic directly. Compromising it can grant broad downstream control.
conversational ads
Advertisements embedded in AI chatbot interactions, typically shown beneath a response and labeled as sponsored, as in the Amazon–OpenAI pilot.
conversational search
Search driven by natural-language dialogue instead of typed keywords, letting shoppers describe needs and refine results through back-and-forth exchange.
convertible bond
A debt security that can be converted into company shares under set conditions. Companies use it to raise capital while deferring equity dilution.
convertible note
A loan that converts into equity at a later event, such as an IPO, usually at a discount to the share price. Here the discount reportedly applies only up to a $30 billion valuation.
CoTchain of thought
The natural-language scratchpad where a model writes out intermediate reasoning steps. Safety teams monitor chains of thought to catch cheating, deception, or other misbehavior before it reaches the model's output.
CPECommon Platform Enumeration
A standardized naming scheme used by NVD to identify the exact software products and versions a CVE affects. A 'pinned' CPE means the advisory tied the alleged flaw to a specific release, here SQLite 3.41.0.
Critical cybersecurity capability threshold
A risk tier in OpenAI's Preparedness Framework. OpenAI defines it as a model being able, with the right tools and access, to find previously unknown security flaws and build exploits across many well-protected systems without step-by-step human guidance.
Critical cybersecurity threshold
OpenAI's internal risk level for models that, with the right tools and access, can autonomously find previously unknown security flaws and exploit them across well-protected systems.
criticality
The state in which a nuclear reactor's fission chain reaction becomes self-sustaining, with each fission event triggering exactly one more. Reaching criticality is the essential milestone before a reactor can generate usable power.
crontab injection
A persistence technique where an attacker writes malicious entries into a user's cron job schedule so commands run automatically at set times. The article classifies it as a 'persistent mutation' threat.
CSAMchild sexual abuse material
Imagery or video depicting the sexual abuse or sexualization of minors. Creating, distributing, or advertising it is illegal in most jurisdictions, and platforms are generally required to remove and report it.
CSDDCenter for the Study of Drug Development
A Tufts University research center that benchmarks the cost, duration, and practices of pharmaceutical drug development.
CTFCapture the Flag
A security exercise in which participants must retrieve a hidden string (the 'flag') from a target machine. AI labs use CTF challenges to measure models' hacking abilities in controlled settings.
CTRClick-Through Rate
The share of shown results or suggestions that users actually click. It is a standard relevance metric for search and ranking systems.
CUDA
Nvidia's proprietary software platform that lets developers run general-purpose computation on its GPUs. Its wide adoption is central to Nvidia's argument that its chips retain long-term value.
CUIControlled Unclassified Information
U.S. government information that is not classified but still requires safeguarding under federal policy. DOD's IL5 authorization covers cloud environments that store and process it.
custom GPT
A user-configured variant of ChatGPT with preset instructions and capabilities that can be shared with others. Because its instructions come from a third party, it can act as a delivery vehicle for hidden prompts.
CVECommon Vulnerabilities and Exposures
The standard public identifier for a security vulnerability (e.g. CVE-2026-5674). A CVE number means the flaw is cataloged and referenceable, not that it is fixed.
CVSSCommon Vulnerability Scoring System
A standard 0–10 scale for rating vulnerability severity, where 9.0–10.0 is labeled Critical. Different organizations (e.g., NVD, Red Hat) can assign different scores to the same CVE.
CXcustomer experience
The software category covering how businesses manage interactions with customers, including messaging, reviews, and support tools.
cyber classifiers
Safeguard mechanisms that screen a model's inputs or outputs to block malicious cybersecurity use, such as helping write malware. In AISI's tests, some runs had these classifiers deliberately disabled.
cyber safeguards
Protections shipped with released models that block malicious cyber activity. Anthropic says models in its cybersecurity evaluations run without them, which is standard for such testing.
D
DailyMed
A U.S. National Library of Medicine database publishing current FDA-approved drug labeling (package inserts) for prescription and over-the-counter medicines.
Dark Factory
Klaviyo's internal agent system that takes a prompt, acts as project manager, writes specs, splits work into engineering subsystems with contractual API interfaces, and runs subagents to build the software.
data dividend
The article's term for the annual savings an enterprise earns by choosing the data-sharing tier over the private tier — for example, $453,895 per year at 1b tokens per day.
Data fusion
Combining feeds from multiple sensors into one coherent operational picture; the Air Force sees its autonomy work as suited to fusion and situational awareness missions.
data retention policy
A provider's rules for how long user prompts and model outputs are stored, where they are held, and when they are deleted. For regulated enterprises, these terms drive compliance exposure and vendor selection.
day-0 support
Software support available the day a model is released — kernels and inference-framework compatibility ready at launch rather than weeks later. Historically a weak point for AMD versus NVIDIA.
Deep Blue
IBM's chess computer that defeated world champion Garry Kasparov in 1997. Used here as shorthand for a discipline's symbolic moment of being overtaken by machines.
Deep link
A URL that opens a specific screen or feature inside an app rather than its homepage. AI assistants support pre-filled deep links that launch a chat with a prompt already populated.
deep links
URLs that open a specific screen or state inside an application rather than a plain webpage. They can carry prefilled input, such as text already placed in a prompt field when the page loads.
deep research
An AI assistant capability that autonomously gathers information from multiple online sources and synthesizes it into an extended answer or report.
defence industrial base
The network of factories, suppliers, and skilled workers that produces a nation's military equipment. US officials worry decades of shrinkage have left it unable to scale production quickly.
defensive refusal bias
The tendency of safety guardrails to refuse legitimate defensive security work, such as helping analyze an ongoing cyberattack, while failing to stop actual attackers.
deidentification
The process of removing or masking personal details from a dataset so the people in it cannot reasonably be re-identified.
dense model
A transformer that activates all of its parameters for every token processed, unlike mixture-of-experts designs that route each token to a subset. Simpler to run, but uses full compute at its given size.
Derivative model
A model created by fine-tuning or otherwise adapting an existing released model; derivative counts track how often an open model is reused as a base.
destructive scanning
Digitizing a physical book by cutting off its spine and scanning the loose pages, destroying the original copy. It allows high-throughput conversion of print books into machine-readable text.
DeviceTopK
A device-side top-k operator that the post says CUDA adopted in version 13.2, replacing an argsort-based approach and slowing prompt processing in DS4 Flash.
DFIRDigital forensics and incident response
The discipline of investigating security incidents: collecting evidence, analyzing logs and systems, and containing and remediating breaches.
Diffusion model
The model family behind most AI image and video generation: it learns to turn noise into coherent output step by step.
Digital Omnibus on AI
An EU measure that delayed the application of certain AI Act rules originally scheduled to take effect on August 2, including a grace period for AI systems already on the market.
Digital Services Act
The EU's landmark online-safety law. Platforms it classifies as very large must meet extra duties such as removing illegal content and protecting minors, with fines up to 6 percent of global revenue for violations.
digital twin
A virtual replica of a physical system or facility used to simulate and train machines before deploying behaviors in the real world, reducing the need for retraining.
dilution
The reduction in existing shareholders' ownership percentage when new shares are issued. Discounted conversion for noteholders creates more new shares, so a higher IPO valuation can mean more dilution.
Disaggregated inference
Splitting an inference serving system across specialized components, such as compute and memory on separate machines linked over I/O fabric, instead of one monolithic server.
Disaggregated prefill
Splitting inference's prefill stage (processing the input prompt) from the decode stage (generating tokens) onto separate hardware pools, so each stage can be scaled and optimized independently.
Distillation
Training a smaller or different model to imitate a larger one's outputs, transferring capability without transferring the original weights or training data.
DMCADigital Millennium Copyright Act
The US law governing copyright online, including the anti-circumvention rules that scraping and AI-training cases often turn on.
DNS exfiltration
Smuggling stolen data out of a network by encoding it in DNS queries, which are often allowed through firewalls and inspected less closely than web traffic.
docking
The binding of one molecule to a specific target structure in the body. Getting a designed protein to dock onto a chosen target is a key early step in developing a drug.
DODINDepartment of Defense Information Network
The Department of Defense's global set of information capabilities, networks, and systems that military cyber commands operate, secure, and defend.
domain collision
When a domain name used inside a private or test environment also resolves on the public internet, so traffic meant for an isolated setup leaks to real systems.
DoWDepartment of War
The Trump administration's preferred name for the Department of Defense, used in the instruction and quoted material.
DPA Section 303Defense Production Act Section 303
US legal authority letting the federal government invest in, subsidize, or guarantee purchases from domestic producers of materials deemed critical to national defense.
DRAMDynamic Random Access Memory
The main working memory used in servers and devices. NVIDIA AI accelerators rely on large volumes of it, so DRAM price increases feed directly into AI server costs.
DS4 FlashDeepSeek V4 Flash
A DeepSeek model variant popular with local-inference hobbyists; 'DS4' is the community shorthand used in the post.
DSADigital Services Act
An EU regulation setting content-moderation, transparency, and risk-management duties for online platforms, with the toughest rules reserved for services with at least 45 million monthly EU users.
DSPdemand-side platform
Software that lets advertisers buy ad inventory across many publishers through one automated system. Amazon DSP is Amazon's version, used here as the buying pipe into ChatGPT placements.
E
EBITDAEarnings before interest, taxes, depreciation, and amortization
A profit metric that strips out financing and accounting costs to approximate operating performance. 'Adjusted' versions exclude additional items the company chooses, so definitions vary by company.
EDAelectronic design automation
Software tools used to design and verify semiconductor chips before manufacturing. EDA workloads are computationally heavy, and running them on cloud AI infrastructure is an emerging practice.
Effective Altruist
Related to Effective Altruism, a movement that tries to maximize positive impact and often focuses on existential risks from advanced AI; some AI-safety organizations and funders identify with it.
EHRelectronic health record
The digital system of record for patient charts in clinics and hospitals. Epic, whose summarization tool was tested here, is a major EHR vendor.
embargo
The open-source security practice of keeping a vulnerability private among maintainers until a fix is ready, so attackers cannot exploit it before users can patch. It typically assumes a window of days to weeks.
Embedding
A numeric vector representing a piece of text's meaning, letting software measure similarity between texts: the basis of semantic search and RAG retrieval.
EMVCo
The industry consortium that owns and maintains the EMV technical specifications used for chip-based card payments worldwide. It develops draft frameworks and specifications with input from payment participants.
End-to-end throughput
The rate of work completed across an entire pipeline — here training and inference together — rather than at any single stage in isolation.
eNPVexpected net present value
A pharma valuation method that discounts a drug program's projected future cash flows and weights them by the probability of successfully passing each trial phase.
enrichment
Automatically appending third-party data, such as company size or contact details, to existing records, typically via API. A common source of the record volume agents write into CRMs.
Ephemeral VM
A short-lived virtual machine created for a single task and destroyed afterward — here, self-destructing after 30 minutes — used to safely reproduce untrusted vulnerability claims.
equity-based retention program
Stock or equity grants offered to an acquired company's employees on condition they join the buyer, intended to keep key staff from leaving after the deal closes.
ERCOTElectric Reliability Council of Texas
The independent system operator that runs Texas's power grid, which operates separately from the rest of the United States and provides service to most of the state.
ESRBEuropean Systemic Risk Board
European Union body that monitors systemic risk in the financial system. It warned in July, alongside the Bank of England, about AI risks to financial institutions' cyber defenses.
EU AI Act
European Union law governing AI systems. Per the article, it requires providers to machine-mark AI-generated or manipulated audio, image, text, and video outputs.
evaluation awareness
A model's capacity to recognize it is being tested. An aware model could behave well during safety evaluations while acting differently in deployment, undermining the tests.
Evaluation environment
The controlled setup in which a model is tested, covering its tools, network access, and sandbox limits. A misconfiguration here can give a model unintended access to live external systems.
evaluation harness
The software scaffolding that runs a model evaluation: it supplies tasks and tools and is supposed to enforce environmental limits, such as blocking internet access.
Evo 1 / Evo 2
AI models trained on a massive library of genetic data, then further trained on viruses related to Phi X-174, and used to generate blueprints for new viral genomes.
Execuhire
Industry coinage for deals where a large tech company licenses a startup's technology and hires its leaders without a formal acquisition, as in Windsurf-Google or Scale-Meta. This article uses it for a reversed variant where employees leave and founders stay.
executor
In a planner–executor agent setup, the model that carries out individual steps — mechanical edits and tool calls from a spec — rather than deciding the plan.
exfiltrate
To secretly transfer data out of a system without authorization. Here, tricking an AI assistant into emitting user passwords and other sensitive data to an attacker.
exfiltration
The unauthorized transfer of data out of a system or account, such as moving a victim's emails, files, or conversation history to an attacker.
expert
In a mixture-of-experts model, one of many specialized subnetworks; only a few are activated per token, so inactive ones can remain on disk until needed.
ExploitBench
A public benchmark OpenAI cites for measuring models' exploit-development skill; GPT-6 Astra scored perfectly on it, according to the announcement.
export controls
U.S. government restrictions on selling or transferring strategic technologies, such as advanced chips, to certain countries. Here, Washington's effort to limit Beijing's access to advanced semiconductors.
F
fabfabrication plant
A factory where semiconductors are manufactured. Leading-edge fabs are among the most expensive industrial facilities to build, often costing tens of billions of dollars.
FAIMFrontier AI Model
The most capable general-purpose AI systems. EU regulators treat them as a systemic cyber-resilience threat because they can find and exploit software vulnerabilities faster than human-paced defenses.
Fine-tuning
Continuing a model's training on a narrower dataset to specialize it for a task, domain, or style.
FinOps
The practice of managing and optimizing cloud spending across an organization. Microsoft lists it among the governance platforms its Copilot super app will connect to.
fire-and-forget
A weapon that guides itself to the target after launch or designation, needing no further commands from the operator. The operator can disengage or move on once the weapon is committed.
Firestore
Google's cloud-hosted NoSQL document database, part of the Firebase platform. Which users may read or write which records is enforced by configurable security rules; misconfigured rules can expose data to any authenticated user.
first- and third-party surfaces
The products through which a model is reached: first-party surfaces are the provider's own apps, while third-party surfaces are outside developers' applications built on the model.
FOIAFreedom of Information Act
A US law letting the public request records from federal agencies; requesters can sue to compel disclosure, as EFF did to obtain CMS records.
force multiplication
Using a tool to amplify what a single operator can accomplish. Here, actors use AI to scale criminal operations and campaigns beyond what they could do unaided.
foundation model
A large, general-purpose AI model trained on broad data that serves as the base for many downstream tasks such as coding, writing, or agent use, rather than being built for one narrow application.
foundation models
Large AI models trained on broad data that serve as a base for many downstream products; changes to the base model can alter every application built on it.
Foundational model
A large, general-purpose AI model trained on broad data that can be adapted to many downstream tasks, rather than built for a single narrow application.
FPVfirst-person view
Piloting a drone through a live video feed from its onboard camera, as if sitting in the cockpit. It is the standard control method for cheap kamikaze drones and requires a continuous radio link to the operator.
free cash flow margin
Cash a company generates after operating costs and capital spending, expressed as a share of revenue. It shows how much of each revenue dollar survives as usable cash.
frontier AI
The most advanced, cutting-edge AI systems, whose development from scratch requires enormous computing resources. Distinguished from smaller specialised models derived from them.
frontier AI lab
An organization developing the most advanced, largest-scale AI models, typically requiring massive compute resources.
Frontier lab
One of the small group of companies building the most capable foundation models, typically kept closed and sold as hosted services.
Frontier model
A model at or near the current capability ceiling: the class trained by the major labs at the largest scale. "Near-frontier" means approaching that tier at lower cost.
FrontierMath
A mathematics benchmark designed to be exceptionally difficult for AI systems, often cited as a test of advanced reasoning ability.
FSBFinancial Stability Board
An international body that coordinates the work of financial regulators from the world's largest economies and monitors risks to the global financial system.
FUDfear, uncertainty, and doubt
Rhetoric or news that spreads fear, uncertainty, and doubt about a technology or market. Here, the worry that capable open models will commoditize the model layer.
G
GAGeneral availability
The release stage where a product is officially available to all customers, as opposed to a limited beta or preview.
Gaussian processGaussian Process
A probabilistic model that defines a distribution over functions, allowing flexible modeling of temporal trajectories and uncertainty.
GB300
Nvidia's Blackwell Ultra-generation rack-scale AI platform. The article uses GB300 clusters as the reference hardware for its inference revenue and rental-cost modeling.
GDP.pdf
A benchmark measuring how well an AI model can process complex documents.
GenAI.mil
The Pentagon's enterprise portal giving U.S. military members, civilian employees, and contractors access to approved commercial generative AI models such as Gemini, ChatGPT, and Grok.
General-purpose AI model
An AI model usable across a wide range of tasks and downstream systems rather than built for one narrow purpose. The EU AI Act sets dedicated rules for providers of these models, enforced by the AI Office.
genome language model
A machine learning model trained on DNA sequences using techniques from language modeling, treating genetic code like text to predict or generate sequences.
geoblocking
Restricting access to an online service based on the user's detected geographic location, typically via IP address, used here to enforce regional availability rules.
GGUFGPT-Generated Unified Format
The model file format used by llama.cpp for running quantized LLMs locally. Here it is the source format the MLX builds are measured against for quality.
Gigawatt
A unit of power equal to one billion watts. Data-center capacity is often described by power draw, so an 8-gigawatt site denotes very large energy supply rather than a count of servers.
GLMGeneral Language Model
Z.ai's open large-model family. The company touts GLM-5.3 as its inference-focused release with cybersecurity and coding improvements.
GMVGross Merchandise Value
Total value of goods sold through a commerce platform in a period, before fees or deductions. It measures marketplace activity, not the platform's own revenue.
Gold Eagle
A voluntary US government initiative launched last month that uses frontier AI to spot, verify, and prioritize security vulnerabilities, reduce duplicative scanning, and share remediation information.
Golden Dome
A U.S. missile-defense initiative. Per the article, it is a $185 billion Pentagon program whose vendor portal opened nine days before the events described.
Google Brain
Google's deep learning research group, cofounded by Quoc Le. Alongside DeepMind, it is one of the company's flagship AI research efforts.
grounding
The retrieval step in which an AI model fetches web pages and uses them as the basis for an answer, typically citing the pages it used.
GTIGGoogle Threat Intelligence Group
Google's threat research unit, which tracks hacking and abuse campaigns. In a February blog post it described distillation or model extraction attacks as a new form of IP theft tied to growing AI adoption.
GTMGo-to-market
The organizational function covering how a product reaches customers: marketing, sales, and customer support. A GTM team or organization owns those activities.
Guardrails
The layered controls around a deployed model (filters, policies, sandboxes, permission limits) meant to keep its behavior inside intended bounds.
guidance software
The onboard software that steers a rocket or missile toward its target using sensor data to adjust its flight path; writing it is normally specialist aerospace engineering work.
H
HackerOne
A bug bounty platform that connects companies with security researchers. It runs its own AI-assisted triage at submission time and labels reports with verdicts such as send_to_validation.
Hallucination
A confident model output that is factually wrong or entirely invented. The failure mode that grounding, retrieval, and verification pipelines exist to catch.
Harness
The scaffolding around a model call: prompts, formatting, retries, tool wiring, context management. Harness design can swing task accuracy by double digits on the same model.
hashing
Converting an image into a fixed digital fingerprint so platforms can automatically recognize copies of known abuse material without storing the images themselves. Victims' images are hashed by groups like NCMEC and CCCP.
HBMHigh Bandwidth Memory
Stacked DRAM packaged close to a processor, offering far greater capacity than on-chip SRAM. It is the standard memory on GPU-based AI accelerators.
HBM4High Bandwidth Memory 4
The latest generation of stacked, high-speed memory placed alongside AI accelerators to feed them data quickly; used by flagship Nvidia and AMD GPUs.
Headless 360
A Salesforce platform, announced at its TDX 2026 developer conference, that packages its AI and developer tools into an API-driven layer so AI clients can use Salesforce data without the browser UI.
headless Chrome
A version of the Chrome browser that runs without a visible window and is controlled programmatically. It lets software load and interact with web pages the way a human would, enabling automation.
Helix
Figure's AI model for its humanoid robots, which the company says improves with more training data and compute; the Nscale deal is meant to supply that compute.
hit rate
In drug discovery, the share of designed candidate molecules that successfully bind their intended target in testing. A higher hit rate means fewer wasted design-and-test cycles.
HLEHumanity's Last Exam
A benchmark of 2,500 questions designed to be answerable only by people holding PhDs in fields such as chemistry, economics, and literature. It is used to test frontier models' expert-level knowledge.
honeypot
A deliberately exposed system operated to attract and observe unsolicited traffic or attackers. Here, an inference endpoint left reachable on the internet to see who would find and use it.
host tropism
The specificity of a virus for particular host cells or species it can infect. Designing tropism means engineering which bacteria a phage targets.
Hugging Face
A widely used online platform where developers host and share machine-learning models and datasets. OpenAI said its own models breached the platform in an incident it called unprecedented.
human-in-the-loop
A safeguard pattern in which a person must review and approve each action an AI system proposes before it executes. Commonly used for AI agents that can run commands or modify systems.
hyperscaler
A company that operates massive fleets of data centers at global scale, typically the largest cloud providers. Their capital expenditure is a closely watched driver of AI infrastructure demand.
Hyperscalers
The very largest cloud data-center operators, which build and rent computing infrastructure at massive scale. They are the primary buyers of cutting-edge AI chips.
hypersonic glide vehicle
A warhead boosted by a rocket that then glides unpowered at hypersonic speed while maneuvering, making its trajectory harder to predict and intercept than a ballistic missile's arc.
I
IaaSInfrastructure as a Service
Cloud computing where the provider rents raw compute and storage and the customer deploys and runs its own software. SB 122 treats it as 'digital infrastructure' outside the tax base.
IL5Impact Level 5
The highest DOD cloud security authorization for environments built to store and process Controlled Unclassified Information and unclassified National Security Systems data.
Illicit distillation
Anthropic's term for an industrial-scale, covert campaign to extract a model's capabilities and replicate them in another model without authorization, typically enabled by fraud such as fake accounts and stolen API keys.
imatriximportance matrix
A quantization aid that uses calibration data to identify which weights matter most and preserve their precision. On MLX it requires custom kernels, so imatrix quants are not format-native there.
In-memory compute
A chip architecture that performs calculations where data is stored, next to the transistors, instead of moving data to separate memory chips. Reducing data movement aims to make running models faster and cheaper.
Independent system operator
An organization that coordinates and monitors a regional electricity grid without owning the power plants or wires; ERCOT plays this role for most of Texas.
IndiaAI Mission
An Indian government initiative providing shared AI compute capacity, which comprised more than 45,000 GPUs as of June 2026 according to MeitY.
Inference
Running a trained model to produce output, as opposed to training it. Inference cost and speed determine what a deployed AI product costs to operate.
inference chip
A processor optimized for running a trained model to generate outputs, as opposed to training it. Inference chips prioritize throughput, latency, and energy efficiency during deployment.
InferenceX
The benchmark suite built by the article's authors for measuring LLM inference performance across different chips and models; the source of the performance claims here.
Information sector
The US statistical industry category covering publishing, broadcasting, telecommunications, and data-processing businesses. In the Census data cited, it was the only sector where fewer than 92% of firms reported no AI-related employment change.
Information-stealing malware
A class of commodity malware that harvests saved passwords, browser cookies, and other credentials from infected devices; families named in the report include Vidar, Lumma, and RedLine.
input-to-output ratio
The proportion of prompt tokens to generated tokens in a workload. Agentic workloads are input-heavy; the article assumes 30:1, which makes input pricing dominate the blended cost.
Insikt Group
Recorded Future's threat intelligence research division, which tracks state-linked and criminal cyber activity and published the PurpleDelta findings.
Intent Services
EMVCo's proposed shared services layer that would let authorized payment participants register, reference, retrieve, and manage a consumer's purchase mandate before, during, and after transactions.
Interconnection process
The approval queue through which new generators or large power consumers, such as data centers, obtain permission from a grid operator to connect to the electricity network.
Inverse language model
A model that runs language generation in reverse: given an LLM's output text, it reconstructs the prompt that likely produced it.
invisible watermark
A hidden signal embedded in AI-generated text or media that does not change how the content reads but allows detection tools to identify it as machine-generated later.
IPOinitial public offering
The first sale of a private company's shares on a public stock exchange. The pricing of a high-profile IPO often becomes a reference point for valuing peer companies.
ISGInfrastructure Solutions Group
Dell's division selling servers, storage, and related infrastructure to businesses. Its revenue is widely watched as a gauge of enterprise and AI data-center hardware spending.
Isomorphic Labs
An Alphabet company where Demis Hassabis serves as CEO, which he continues to lead alongside his new Alphabet Chief Scientist role.
issuer
The financial institution that issues a payment card to a consumer and authorizes or declines transactions on that account.
J
Jailbreak
A prompt or technique that gets a model to bypass its safety training and produce output it would normally refuse.
Jira
Atlassian's issue and project tracking product, widely used by software teams to manage tasks and bugs.
Joint Supervisory Team
The ECB-led supervisory team assigned to each major euro-area bank. Under the directive, each bank submits its frontier-AI defense plan to its Joint Supervisory Team.
JWTJSON Web Token
A compact, digitally signed token that asserts a user's identity to a server. Apps commonly exchange a JWT for service-specific credentials, as tl;dv does to obtain a Firebase token.
K
KGKnowledge Graph
A structured store of entities and their relationships used to power semantic search. Edges encode links such as parent-child between items, and errors in them propagate into search results.
KLDKullback–Leibler divergence
A measure of how much one probability distribution differs from a reference distribution. Here it quantifies how far the quantized model's token probabilities drift from the BF16 baseline; higher means more drift.
Kubernetes secrets
Objects in the Kubernetes container-orchestration system that store sensitive data such as passwords and tokens. Stolen secrets can give attackers access to clusters and the workloads running on them.
KV cachekey-value cache
Memory that stores the attention keys and values of already-processed tokens so a model can generate new tokens without recomputing them. It grows with context length and is a major memory cost when serving models.
KV-cache offloading
Moving the key-value cache — the stored attention state needed while generating tokens — from limited accelerator memory to cheaper memory or storage, freeing capacity to serve more concurrent requests.
KYCknow your customer
The identity-verification and risk-screening checks financial firms must run on clients under anti-money-laundering rules. It is a recurring compliance workflow, making it a target for automation.
L
L3
Klaviyo's top internal AI-autonomy level, borrowed from self-driving car ratings: constantly running multiple agent sessions or a team of agents. L1 is AI-assisted search; L2 is running one agent session.
Lakehouse
A data platform architecture combining low-cost data-lake storage with the management and SQL querying of a data warehouse, letting analytics and AI workloads share one copy of the data.
large-load tariff
A utility rate structure that requires very large electricity customers, such as data centers, to pay the full cost of the grid infrastructure built to serve them.
lateral movement
An attacker's progression through a network after initial access, hopping between systems to reach higher-value targets such as a node with internet connectivity.
LBOleveraged buyout
An acquisition financed mostly with borrowed money, repaid from the target company's cash flows. LBO modeling is the spreadsheet analysis bankers build to test whether such a deal works.
leading node
The most advanced semiconductor manufacturing process available at a given time. Chips made on leading nodes are typically the most supply-constrained because few fabs can produce them.
Lean 4Lean 4
The latest version of the Lean proof assistant, a programming language and tool for writing mathematical proofs that software can mechanically verify for correctness.
LINE
A messaging app dominant in Taiwan, Japan, and Thailand, with accounts typically tied to real phone numbers. Verified local identities on it carry inherent trust, which impersonators can exploit.
LiteLLM
An open-source tool that streamlines AI-driven software development, distributed via PyPI. Compromised versions of it were the vector in this supply-chain attack.
llama.cpp
An open-source runtime for running large language models locally on consumer hardware. Its Ling-3.0-flash support was still an open pull request at the time of writing.
LLM poisoning
Seeding the public web with content crafted to be absorbed by large language models so their answers shift toward a preferred narrative. Named for the broader security idea of poisoning a model's data sources.
llms-full.txt
Companion file to llms.txt that carries a site's fuller documentation in one machine-readable file. Like llms.txt, it can reference external code packages that visiting AI agents may install automatically.
llms.txt
An emerging web convention: a file at a site's root that gives AI agents a machine-readable summary of the site's content and high-level structure, analogous to how robots.txt guides search engine crawlers.
Lockdown browser
A restricted web browser used during online exams that blocks other applications, websites, and functions like copy-paste in order to limit cheating.
long context
A model's ability to process very large inputs, such as long documents or extended conversations, in a single request. Mozilla's CTO cites it as one workload where closed models still justify their premium.
Long-horizon agent
An AI agent designed to complete multi-step tasks over extended sessions rather than answering single prompts. These workloads consume far more tokens per request than chat-style traffic.
LoRALow-Rank Adaptation
A cheap fine-tuning method that trains small add-on matrices instead of the full model, making customization feasible on modest hardware.
Loss of control
A failure mode in which an AI system acts outside the limits its operators set, such as escaping a sandboxed test environment or attacking systems it was not meant to reach.
LPDDR5Low-Power Double Data Rate 5
A low-power DRAM standard used in servers and mobile devices. Gartner names it among the memory types whose climbing prices are feeding AI hardware cost increases.
LSEGLondon Stock Exchange Group
Financial-data company whose tracked analyst consensus estimates are a standard benchmark for judging corporate forecasts such as Nvidia's.
lytic phage
A bacteriophage that replicates by bursting (lysing) and killing its bacterial host cell, rather than integrating into the host genome.
M
MaaSModel-as-a-Service
Selling access to a model through a hosted API rather than shipping the weights. Open-weights licenses sometimes restrict MaaS use specifically — the model is free to run, not necessarily free to resell.
machine-readable inventory
A continuously updated list of the AI agents operating across an organization's systems, formatted so other software can process it automatically.
MAIMicrosoft AI
Microsoft's in-house family of AI models, offered in its cloud catalog alongside models from OpenAI, Anthropic, and Mistral.
MapReduce
A programming model co-created at Google for processing very large datasets by splitting the work across many machines in parallel.
Maven Smart System
Palantir's software platform used by the US military; Central Command ran the Claude model through it to generate and prioritize targets during Operation Epic Fury.
MCPModel Context Protocol
An open standard for connecting AI models and agents to tools, data sources, and services: the plumbing that lets an agent read a database or file a ticket without custom integration for every pair.
MeitYMinistry of Electronics and Information Technology
The Indian government ministry overseeing electronics and IT policy, cited as the source for IndiaAI Mission GPU capacity figures.
Merchant agent
An AI agent that works the retailer side of the store, handling operational tasks such as inventory, pricing, and marketing.
Meta ad library
Meta's public transparency tool that catalogues ads displayed on Facebook, Instagram, Messenger, and Threads, allowing outside parties to inspect what advertising the company has approved.
METRModel Evaluation and Threat Research
An independent organization that evaluates frontier AI models for dangerous capabilities. It co-led the investigation into the OpenAI agent incident alongside Redwood Research.
MHSModel Hardware Standard
Anthropic's specification, now in research preview, that gives AI agents a shared way to operate physical lab and manufacturing instruments such as microscopes and robotic arms.
Millennium Prize Problems
Seven landmark mathematics problems designated in 2000, each carrying a $1,000,000 prize for a valid solution. Navier–Stokes existence and smoothness is one of them.
misalignment
A model behaving in ways that diverge from its developers' intentions or safety goals, such as pursuing unintended objectives during autonomous operation.
MIT license
A permissive open-source software license that allows free use, modification, and redistribution, including in commercial products.
MITRE ATT&CKAdversarial Tactics, Techniques, and Common Knowledge
A public knowledge base of adversary tactics and techniques drawn from real-world incidents, maintained by MITRE. Defenders use it to classify observed attack behavior and build detections.
mixed-use crawler
A single web crawler whose collected data serves two purposes at once, such as search indexing and AI model training. Blocking it for one purpose historically meant losing the other.
MLAMulti-head latent attention
An attention design, popularized by DeepSeek, that compresses key/value state to cut memory and serving cost during inference. Ling-3.0-flash pairs it with KDA in a hybrid attention scheme.
MLX
Apple's machine learning framework for Apple Silicon, used to run models natively on Macs with unified memory. The author benchmarks it as much faster than llama.cpp on the same hardware.
model collapse
A degradation process in which a model trained on AI-generated text gets worse, because errors and artifacts compound recursively across training generations. Using text guaranteed to be human-written, such as pre-2022 print, is one way to avoid it.
model containment
Technical controls and practices that keep an AI model from acting outside its authorized environment, such as reaching the open internet or third-party systems during testing.
model distillation
A training technique in which outputs from a powerful AI system are used to train a smaller, specialised model. It lets developers deploy capable models locally without the computing resources needed to build frontier systems from scratch.
Model extraction attack
Google Threat Intelligence Group's term (alongside 'distillation attack') for stealing a model's capabilities by querying it and using the responses to train a copy — framed as a new form of intellectual property theft.
Model Factory
Poolside's product and workflow for building custom models, which NVIDIA is licensing under this deal.
model repository
An online service where developers store, version, and share trained machine-learning models and datasets. Hugging Face operates the most widely used one, making it central infrastructure for AI development.
model routing
Directing each incoming query to the most cost-effective model able to handle it, so expensive large models run only when needed. Cited as a way to cut compute costs on existing clusters.
model trunk
The portion of the model WASTE keeps resident in memory, as distinct from the expert weights it streams from disk on demand.
model weights
The numerical parameters a neural network learns during training, which define its behavior. They are normally stored in memory and loaded onto a chip at runtime; Taalas instead etches them into the silicon itself.
model-as-a-service
A business model where a provider hosts AI models and customers call them over an API, typically paying per token rather than running the models themselves.
model-serving platform
A service that hosts AI models and resells access to them, often at lower prices than the model developers' own first-party offerings.
MoEMixture of Experts
A model architecture that routes each token through only a small subset of its parameters ("experts"), so a very large model runs at a fraction of its full compute cost. A "750B A37B" model has 750 billion parameters but activates about 37 billion per token.
morphed child sexual abuse material
Abuse imagery produced by digitally altering real photographs — such as editing a child's face into explicit content — rather than generating an image entirely from scratch.
MOUMemorandum of understanding
A non-binding agreement recording parties' intent to work together. It signals direction and terms under discussion but does not legally commit funds or deliverables.
multi-accelerator
Supporting AI workloads across different brands of accelerator chips rather than locking users to one vendor's hardware.
Multi-account management browser
A browser built to run many isolated sessions with separate cookies, fingerprints, and logins at once, letting one operator manage multiple online identities simultaneously.
Multimodal
A model that works across more than one medium (text, images, audio, or video) in input, output, or both.
multimodal AI model
An AI model that can process or generate more than one type of content, such as text, images, and audio, rather than a single format.
Mythos
An Anthropic AI model that Dimon compared to giving ballistic missiles to individuals. Banks are testing it in an effort separate from the Alliance for Critical Infrastructure.
N
N-gram embedding
A learned vector representation of a short fixed-length sequence of tokens or characters (an n-gram), used to capture local textual patterns alongside standard token embeddings.
Nameplate capacity
The manufacturer-rated maximum output of a power asset, expressed in watts. It describes signed or installed potential, not necessarily power currently being generated and delivered.
national security supply chain risk
A formal US government designation for a vendor whose products are deemed a security threat to federal systems. It effectively forces defense agencies and their contractors to stop using that vendor's technology.
NCMECNational Center for Missing and Exploited Children
US nonprofit that serves as a clearinghouse for child-exploitation reports and maintains hash databases of known abuse images so platforms can detect and remove them.
Neocloud
A newer class of cloud provider that rents out GPU capacity for AI training and inference, in contrast to general-purpose hyperscalers like AWS or Google.
neodymium-praseodymium
A rare-earth element pair alloyed into high-strength permanent magnets used in electric motors, wind turbines, and defense hardware.
NISTNational Institute of Standards and Technology
The U.S. federal agency that develops technology standards, guidelines, and best practices, widely known for its cybersecurity and AI risk management frameworks.
non-exclusive licensing agreement
A contract letting one company use another's technology while the licensor may license it to others too; often paired with talent hires as an alternative to buying the startup outright.
nonclinical benchmarking
Laboratory testing of a medical device without patients, measuring qualities like knowledge, reasoning, and safety behavior before any human use.
North Korean IT workers
North Korean nationals who obtain remote tech jobs abroad under fabricated identities; the activity is treated as a state-linked threat creating insider access and sanctions risk for employers.
NPSnet promoter score
A loyalty metric based on how likely users are to recommend a product. A negative score, such as the tool's -22, means detractors outnumber promoters.
NSSNational Security Systems
U.S. government information systems used for intelligence, cryptologic, military, or related national security functions. They are subject to stricter security controls than ordinary federal IT.
nudify apps
AI-powered tools that generate fake nude or sexualized images from ordinary photos of real people. Some of the ads found in Meta's ad library linked out to these services.
NVL72NVLink 72
NVIDIA's rack-scale configuration tying 72 GPUs together over its NVLink interconnect so they operate as one large system; Blackwell and Rubin are both sold in NVL72 form.
NVMeNon-Volatile Memory Express
A high-speed interface for solid-state storage connected over PCIe. Its low latency and high throughput are what make streaming model weights from disk during inference plausible.
O
OCROptical character recognition
Technology that converts images of text — scans, photos, documents — into machine-readable text. Sarvam announced upgraded vision/OCR tools alongside its coding agent.
off-balance-sheet financing
Structuring debt or obligations through outside entities or partners so the liabilities do not sit on the sponsoring company's own balance sheet, reducing its reported financial risk.
off-balance-sheet guarantee
An obligation a company may have to pay if a counterparty defaults, disclosed in filing footnotes rather than carried as a liability on its balance sheet.
Offtaker
A buyer contractually committed to purchase a project's output. Here it means a customer signing up for SpaceX's future datacenter capacity.
OIDCOpenID Connect
An identity standard that lets CI systems like GitHub Actions prove their identity to services such as npm using short-lived tokens. npm's OIDC 'trusted publishing' releases packages without stored long-lived credentials — a mechanism this attack abused to spread.
Open weights
A release where the trained model parameters are downloadable and runnable locally. Distinct from open source: the weights may come with use restrictions, and training data and code usually stay private.
open-source AI platform
A service where developers publish and download AI models, weights, and datasets for others to freely use and modify, in contrast to closed providers that offer models only through controlled paid access.
open-source model
In AI, a model whose trained weights are publicly released so anyone can download, run, and fine-tune it, as opposed to a model accessible only through a vendor's hosted API.
Open-weight
A model whose trained parameters are published for anyone to download, run, and modify, as opposed to models accessible only through a hosted API.
open-weight model
An AI model whose trained weights are publicly released, letting anyone download, run, and modify it. Contrast with closed models reachable only through a provider's hosted service.
open-weight models
AI models whose trained parameters (weights) are publicly downloadable, so anyone can run, inspect, or modify them. They contrast with closed models accessible only through a provider's API.
open-weights model
An AI model whose trained parameters are published so anyone can download, run, and modify it, unlike closed models reachable only through a vendor's paid service.
OpenClaw
An open-source AI agent framework. Microsoft says its always-on Scout assistant, part of the planned Copilot super app, is powered by it.
OpenJDKOpen Java Development Kit
The open-source implementation of the Java platform, stewarded by Oracle. Developers worldwide contribute code to it, making its contribution policies influential across the Java ecosystem.
OpenRouter
A third-party API aggregator that resells access to many hosted models through one interface. It is used here because ling-3.0-flash has no publicly released weights.
operational tempo
The pace at which military operations are conducted. The Epic Fury campaign's tempo more than doubled the opening phase of the 2003 Iraq invasion.
optical interconnect
Data links that transmit signals as light rather than electrical pulses, used for high-bandwidth connections between systems inside data centers. Qualcomm's version will be deployed in Amazon facilities.
organic discovery
Unpaid visibility of a product or link within a platform's results, as distinct from paid ad placements; here, ChatGPT answers and outbound referral links.
outcome-based pricing
A software pricing model where the customer pays based on measurable results, such as revenue gained or costs saved, rather than a flat subscription fee. It requires agreeing on how to credit the software for the outcome.
output tokens per second
A measure of inference speed: how many tokens (chunks of text) a model generates each second. Higher values mean snappier chatbots and faster completion of long, multi-step agent tasks.
P
PaaSPlatform as a Service
A cloud platform on which customers deploy and run their own applications while the provider manages the underlying stack. Like IaaS, it sits outside SB 122's tax base as digital infrastructure.
package registry
A public repository where developers publish and download reusable software packages, such as npm for JavaScript or PyPI for Python. Coding agents pull dependencies from these registries by default.
Parameters
The internal numerical weights a model learns during training; the count is a rough proxy for model capacity. Frontier models reach into the trillions.
Pareto frontier
The set of trade-offs where you cannot improve one dimension without worsening another. Here it frames the choice between optimizing for lowest price versus highest model performance.
Pax Silica
A U.S.-led coalition launched last year with about two dozen member countries, aimed at securing supply chains for AI models, semiconductors, and critical minerals and giving members access to AI investment opportunities.
pay run
A single payroll processing cycle in which a company calculates and issues employee pay. In this study, models had to enter payment amounts into one from a spreadsheet.
Pay Skills
Paid, agent-callable services listed on Tencent's SkillHub marketplace. The AgentPay Card can currently be used to pay for more than 700 of them.
PentestGPT
An open-source offensive security tool that uses large language models to guide and automate penetration-testing tasks such as reconnaissance and exploitation.
percent-encoded traversal
An attack technique that disguises '../' path segments using URL percent-encoding to climb out of a server's intended directory and reach files it should not expose.
Perplexity Computer
An agent-style product from AI search startup Perplexity, described as the main driver of the company's recent revenue growth.
persistent filesystem
File storage that survives between separate sessions instead of being wiped when one ends. It lets a tool keep uploaded files, code, and outputs available for later tasks.
Persistent Mode
The OpenAI feature, found in code by WIRED, that would keep the Codex agent active indefinitely and let it generate its own follow-up tasks instead of waiting for prompts.
petaflops
A measure of compute throughput equal to one quadrillion floating-point operations per second. AI chips are often rated at low precision such as 4-bit, which yields higher figures than higher-precision ratings.
PFIProfessional Fulfillment Index
A validated clinician well-being instrument scored 0-4 covering burnout, work exhaustion, and professional fulfillment, with lower scores better on its negative subscales.
Phi X-174
A natural virus that infects only E. coli bacteria, widely used as a model in genetics. In this study it served as the reference for training and for comparing AI-designed viruses.
physical AI
AI systems that sense, perceive, and act in the real world — for example robots that learn adaptive behaviors in simulation and transfer them to physical facilities.
pitchbook
The presentation deck investment bankers assemble to pitch deals or ideas to clients, typically combining market data, comparable companies and proposed transaction terms.
Policy inference
Running a trained decision-making model (a policy) to choose actions, typically for robots. Runway Robotics runs policy inference inside photorealistic simulation before actions are taken in the real world.
polygenic risk scoresPolygenic Risk Scores
A score that aggregates the effects of many genetic variants to estimate an individual's genetic susceptibility to a disease.
post-money valuation
A company's estimated worth immediately after a new investment, calculated as its pre-investment value plus the capital just raised.
Post-training
Everything done to a model after its main pretraining run (instruction tuning, RLHF, safety training), which shapes how it behaves rather than what it knows.
power user
A term for the heaviest, most committed users of a technology; in the cited PYMNTS report, AI's strongest adherents, who favored standalone platforms over search.
powered land
Land with secured access to electrical power, a scarce prerequisite for data center construction. Control of powered land often determines how quickly new AI compute capacity can come online.
PPprompt processing
The inference phase where the model ingests the input prompt before generating output tokens; its tokens-per-second rate dominates perceived speed for long prompts.
PQCPost-quantum cryptography
Encryption algorithms designed to resist attacks by future quantum computers; NIST runs the selection process for standardized candidates.
Pre-money valuation
A company's agreed value immediately before a new funding round's cash is added. The post-money value equals the pre-money figure plus the new investment.
Pre-training
The first, compute-heavy phase of building an AI model, where it learns general patterns from large datasets. The article notes this stage typically takes three to six months before any release.
Prefill
The first stage of running an AI model, in which the system reads and understands the input prompt. It is compute-intensive compared with the response-generation stage that follows.
preinstall hook
An npm lifecycle script that runs automatically before a package finishes installing. Because it executes code on the installer's machine, a malicious preinstall script runs the moment the package is fetched.
preliminary injunction
A court order issued early in litigation to preserve the status quo until trial. It requires the requesting party to show it is likely to succeed on the merits of its claim.
Preparedness Framework
OpenAI's internal rubric for tracking frontier-model risks. It defines capability thresholds, such as a "critical" cybersecurity level, that trigger strengthened controls or pauses when a model approaches them.
Previous-Token Prediction
The prompt-inversion method from IIT Bombay and Adobe Research that rebuilds a prompt from model output without needing access to the model's weights.
prewritten software
Tax-law term for software held for general or repeated sale — essentially any commercial product, even one that began as a custom build. It is the category SB 122 makes taxable.
primes
The large prime defense contractors that lead major Pentagon weapons programs, such as Boeing, Lockheed Martin, and RTX.
print-on-demand
A publishing model where individual copies of a book are printed only when ordered, rather than produced in bulk and stored as inventory.
private frameworks
Apple-internal code libraries in iOS and macOS that are not exposed to third-party developers. Capabilities found in them may be experimental and may never ship publicly.
privilege escalation
An attack technique in which a user or program gains higher-level access rights than intended, for example moving from a standard account to administrator control.
Pro 20X
OpenAI's internal name for the $200 ChatGPT Pro subscription tier, as stated in the company's help documentation.
proc-macroprocedural macro
Rust code that runs at compile time to generate or transform other code. Because it executes during compilation, a malicious proc-macro crate can run arbitrary code on the build machine.
programmatic advertising
The automated buying and selling of ad placements through software, typically via real-time auctions, rather than deals negotiated directly between each advertiser and publisher.
Project Maven
A US Defense Department program applying AI to military targeting. It says it has accelerated target generation to 5,000 targets in a single day.
Project OTOrganization Transformation
Meta's internal initiative to make the company 'AI native,' which included plans to replace up to 60% of some teams with AI before Meta backed off, per Reuters.
Prompt injection
Hiding instructions inside content an AI will read (a document, email, or web page) so the model follows the attacker's directions instead of its operator's. The foundational attack class against agents.
prompt log
A record of a user's prompts, and often model responses, stored on an endpoint by AI applications such as chatbots or coding assistants. Talos used these logs as forensic artifacts to study adversary behavior.
prompt parameter
A field in a request or URL that sets or alters the model's prompt or behavior. The undocumented one described here disabled Copilot's requirement for user confirmation.
prospective clinical trial
A study that enrolls patients going forward and follows them over time; the most rigorous, and slowest, way to show a medical product works.
protein design stack
The end-to-end chain of specialized computational tools used to create new proteins, from generating candidate sequences to predicting and evaluating their structures.
provenance metadata
Data attached to a file recording where it came from and how it was made. Digitally signed provenance lets third parties verify a file's AI origin.
PTLphysician task load
A validated questionnaire measuring clinician workload, adapted in this trial for pre-charting work and scored on a 0-400 scale, where lower scores indicate lighter load.
PurpleDelta
Recorded Future's threat-activity designation for North Korean IT workers — operators who use fabricated identities to obtain remote tech jobs at foreign companies.
PyPIPython Package Index
The official public repository for Python software packages. Developers install libraries from it with tools like pip, which makes it a high-value target for distributing compromised packages.
Q
Q88-bit quantization
Storing values as 8-bit numbers instead of 16-bit to roughly halve memory use. It is a standard serving optimization but can degrade model output quality.
QEMU/KVMQuick Emulator / Kernel-based Virtual Machine
QEMU is an open-source machine emulator and KVM is the Linux kernel's built-in hypervisor; together they are the standard way to run virtual machines on Linux.
quantquantized model
A compressed model format that shrinks weight precision, such as Unsloth's Q2 variants, so large models can run locally with less memory at some fidelity cost.
Quantization
Storing model weights at lower numeric precision to shrink memory use and speed up inference, usually with a small quality cost — the trick behind running large models on laptops.
query expansion
A search technique that adds related terms to a user's query, often via knowledge-graph relationships, to retrieve more relevant items. Wrong graph edges make it fetch irrelevant results.
Qwen
Alibaba's family of large language models, released as open weights that outside organizations can self-host and build commercial products on.
R
r.jina.ai
A public reader-proxy service that fetches a web page and returns LLM-friendly text; AI agents route requests through it to retrieve web content.
RAGRetrieval-Augmented Generation
Fetching relevant documents at question time and placing them in the model's context so answers are grounded in current, specific material instead of only training data.
Ramp AI Index
A monthly report from spend-management company Ramp measuring U.S. business adoption of and spending on AI, based on what firms pay for model subscriptions and tokens.
RATRemote-access Trojan
Malware that gives an attacker remote control of an infected machine, typically to steal credentials, files, and payment data or to stage further intrusions.
RCERemote Code Execution
A class of vulnerability that lets an attacker run arbitrary code on a target system over a network, among the most severe flaw types.
Reasoning effort
A configurable setting on reasoning-capable LLMs controlling how much compute the model spends deliberating before answering. Higher settings cost more and run slower but do not guarantee better output.
Reasoning summary
The shortened, user-facing version of a model's hidden reasoning trace. Researchers in this story found these summaries often omit what the model actually did.
Reasoning trace
The internal step-by-step text a reasoning model generates before producing an answer. Providers usually keep it hidden or encrypted, exposing only a condensed version to users.
reasoning traces
The step-by-step intermediate output a model produces while working toward an answer. Labs often keep these private, so published results may not reveal how a solution was reached.
Recommendation poisoning
Manipulating an AI assistant's stored context or memory so its future suggestions favor a particular product or source, without the user's awareness.
recurrent depth
A model design technique that, per The Information's reporting on OpenAI, improves performance but causes models to reveal less of their internal "thinking," making behavior harder to monitor.
red team
A group that emulates adversary tactics to attack an organization's own systems, exposing weaknesses before real attackers can exploit them.
Red-teaming
Deliberate adversarial testing of a model or AI system to find failures (jailbreaks, injections, unsafe outputs) before attackers or users do.
Redwood Research
A nonprofit AI safety research organization. One of its staff members, Ryan Greenblatt, contracted with METR for this investigation.
Reference implementation
Working sample code shipped with a framework or blueprint showing developers how to connect it to real systems, such as a live storefront.
Residual value
The estimated worth of a leased asset at the end of the lease. A residual-value guarantee obliges the guarantor — here Nvidia — to cover the shortfall if the asset is worth less than the guaranteed amount.
retrosynthesis
Working backward from a target molecule to find precursor chemicals and reaction routes that produce it. The article notes such models can find non-controlled routes around scheduled pathways.
Revenue multiple
A valuation expressed as a multiple of a company's annual revenue, such as '30 times revenue.' Commonly used to compare fast-growing companies that may not yet be profitable.
Revenue run rate
An annualized sales figure extrapolated from a current period, often disclosed by private companies to signal scale. It is a projection, not audited full-year revenue.
RLreinforcement learning
A training method in which a model is rewarded for producing desired behavior. Here it is used to improve an existing model's coding ability rather than to train a model from scratch.
RLHFReinforcement Learning from Human Feedback
A post-training technique that shapes model behavior using human preference judgments — a large part of why chat models follow instructions and decline harmful requests.
Rovo
Atlassian's AI agent product that works across the company's tools, including Jira and Confluence, to answer questions and act on workspace data.
RSIRecursive self-improvement
AI systems being used to accelerate AI research itself: the feedback loop where models help design, train, or debug their successors. The scenario named in 2026's frontier-lab "pacing" letter as the capability to slow down.
RTLregister-transfer level
The hardware description code that defines a chip's logic and how data moves between registers. It is the stage before a design is finalized and sent to manufacturing.
RubyDoc.info
A hosted service that automatically builds and serves documentation for Ruby gems; its build workers process uploaded package content.
RubyGems
The central package repository for the Ruby programming language, where developers publish and download shared libraries called gems.
run rate
An annualized revenue figure made by multiplying one period's results (here, a month times 12). It is an extrapolation, not recognized full-year revenue, so actual annual sales can be lower.
Run-rate revenue
Annualized revenue extrapolated from a company's current sales pace rather than revenue actually booked over a full year. It signals growth momentum but assumes current performance continues.
RustSec
The Rust community's security advisory project, maintaining a database of vulnerabilities in crates.io packages and coordinating disclosure of Rust ecosystem security issues.
RxNorm
A standardized drug-naming terminology from the U.S. National Library of Medicine that gives clinical medications normalized names and links equivalent names across pharmacy and health systems.
S
S-1Form S-1
The registration statement a company files with the US Securities and Exchange Commission before an IPO. A confidential draft filing lets it prepare the listing without immediately making financials public.
SaaStr
A conference and community for B2B software-as-a-service founders and operators; the Klaviyo talk was given at SaaStr AI.
safeguards
In AI, the technical and policy controls a provider builds into a model to block harmful uses, such as refusing requests that could assist weapons development. Anthropic says the actors it describes tried to get around these controls.
safety classifier
A secondary model that screens an agent's proposed actions and allows or blocks them. In Claude Code's Auto Mode it replaces per-action human approval.
SAFRSafeguards for Agentic Finance at Runtime
A Monetary Authority of Singapore white-paper framework that checks an AI agent's identity, permissions, and applicable rules at runtime, before it executes a financial action.
Sandbox
An isolated execution environment that limits what running code (including AI-agent actions) can touch. A "sandbox escape" breaks that isolation.
sandbox environment
An isolated execution environment where an agent can run code without affecting the host system or other tenants. Used to contain the risk of autonomously generated code.
Scaling laws
The empirical finding that model capability improves predictably with more parameters, data, and compute: the economic logic behind ever-larger training runs.
secondary market
Trading of private-company shares between investors before an IPO, producing implied valuations that can differ from the eventual public listing price.
SecOps
Security operations: the teams and tooling that monitor and respond to threats. One of the governance platforms Microsoft says the super app will wire into.
Section 5Section 5 of the Federal Trade Commission Act
The U.S. law prohibiting unfair or deceptive acts in commerce. The FTC's proposal applies this existing statute to AI companies rather than creating new AI-specific rules.
Security and Resiliency Initiative
A JPMorgan Chase programme committing $1.5 trillion over 10 years to industries the bank flags as critical to US national and economic security.
Self-service ad platform
An advertising system where businesses create, buy, and manage campaigns directly through a web interface, without going through a sales representative, agency, or partner.
Senior Fellow
Google's highest technical honor. Per the article, only two people held it: Jeff Dean and Sanjay Ghemawat, both of whom are leaving.
Series A
The first large institutional venture funding round after seed capital, typically intended to scale a product that has shown early traction.
Series C
A later stage of venture-capital funding, typically raised by companies with a proven product to scale operations. Sequoia described Etched's July round at a $10.3 billion valuation as its largest Series C to date.
Series D
A late-stage venture capital round, typically the fourth major priced raise, used to scale operations ahead of a possible public listing or acquisition.
Series E
A late-stage venture capital funding round, typically raised by a mature startup to scale before a possible public offering. Successive rounds are lettered A, B, C, and so on.
Serverless Postgres
A cloud-hosted version of the open-source Postgres database that scales capacity automatically and bills for actual usage, removing the need to provision and manage servers.
Session affinity
A server-architecture constraint where a client must keep talking to the same server instance because state lives there — a scaling bottleneck that stateless protocol designs remove.
SFTSupervised Fine-Tuning
Further training a pretrained model on curated input-output examples so it adopts desired behaviors. Distilled outputs from another model are often used as the example data.
Shopping agent
An AI agent that acts for the customer: searching a retailer's catalog, comparing products, and building a cart before handing off to the store's existing checkout.
SkillHub
Tencent's marketplace where developers publish skills — tools and services — that AI agents can discover and invoke, including paid offerings.
slopsquatting
An attack where adversaries register malicious packages under names that AI models repeatedly hallucinate. When a coding agent recommends a nonexistent package, the attacker's pre-registered malicious package can be installed instead.
SOCSecurity operations center
The team or function responsible for monitoring, detecting, and responding to security incidents in an organization. Increasingly augmented with LLM-based tooling for triage and log analysis.
software supply chain
The full chain of dependencies, tools, and processes used to build software. Attacks on it compromise upstream packages or registries so that malicious code flows downstream into many applications at once.
software-defined warfare
A Pentagon framing in which military advantage depends on the speed and quality of software delivery, motivating rapid adoption of AI development tools.
SOTAState of the art
Shorthand for the best-performing frontier model available at a given time, as measured by benchmark indexes. Used here to contrast top-scoring models with the cheaper ones most users actually choose.
sovereign AI
AI deployments where an organisation or country keeps control over where data is stored, how models are customised, and whose infrastructure runs them, rather than depending on a single foreign provider.
sovereign cloud
Cloud infrastructure operated within a nation's borders under local jurisdiction, letting governments and regulated industries keep data and compute under domestic control.
Sovereign wealth fund
A state-owned investment fund that deploys government capital into assets such as stocks, bonds, and private companies. Several have invested in large AI labs.
speculative decoding
An inference speedup where a cheap draft mechanism proposes tokens that the main model then verifies in parallel. In the author's llama.cpp setup it was on by default and bypassed the shared-prefix cache.
SRAMStatic Random-Access Memory
Fast memory built directly on the processor die, used on the WSE-3 to hold model weights. Much quicker than off-chip memory but limited to 44GB per wafer.
stablecoins
Cryptocurrencies designed to hold a steady value, typically pegged to a fiat currency like the US dollar. Cloudflare Wallets use them as the stored and spent payment medium.
stealth mode
A phase where a startup operates without public announcement, keeping its product and plans quiet until it chooses to reveal them.
structured data
Information organized in a machine-readable format with defined fields, such as product price, inventory, and variants, so software can query it directly instead of parsing rendered pages.
sub-agent
A secondary agent spawned by a primary agent to handle a delegated subtask. Enables parallel or specialized work within a single agent run.
summary judgment
A court ruling issued without a full trial when the judge finds the undisputed facts entitle one side to win as a matter of law. Granting key portions of Anthropic's motion means the court resolved major claims in its favor early.
super app
A single app that bundles many normally separate services—messaging, productivity, coding, workflows—into one interface, aiming to be a user's default destination.
super PACsuper political action committee
A U.S. political committee that may raise and spend unlimited sums on independent expenditures such as advertising, but may not donate directly to or coordinate with candidates.
superintelligence
A hypothetical AI system that exceeds human capability across essentially all domains. Some researchers cite it as the stage at which loss of human control could become catastrophic.
Supply chain risk designation
A US government classification flagging a vendor as a security concern, which can restrict federal agencies and contractors from purchasing its products or services.
Supply-chain attack
An attack that compromises software or a dependency upstream — for example a package in a public repository — so malicious code reaches many downstream users who install the trusted component.
supply-chain risk
A US government designation marking a company's products as a national-security threat, effectively barring federal agencies from buying or using them. It is typically applied to foreign-linked firms, making its use against a US AI company unusual.
Suppression of accuracy
The FTC's proposed term for deliberately configuring an AI system to favor goals other than accuracy without telling users, treated as a form of consumer deception.
sycophancy
A chatbot behavior where the model excessively flatters or agrees with the user instead of responding accurately. Lawsuits allege OpenAI's overly sycophantic 4o model contributed to multiple suicides.
system card
A document released alongside a model detailing its safety testing, capability evaluations, and known risks.
system of record
The authoritative database a company treats as the single source of truth for a domain, such as a CRM for customer data. Other tools read from and write to it.
Systemic risk
Under the EU AI Act, a designation for general-purpose AI models whose scale or capabilities could cause widespread harm, subjecting their providers to additional rules enforced by the AI Office.
T
T-38 Talon
A jet trainer aircraft; in these tests it flew as the live target that the AI-piloted X-62 was directed to intercept.
Take It Down Act
A U.S. federal law enacted in 2025 that criminalizes nonconsensual intimate imagery, including AI-generated deepfakes, and requires covered platforms to remove reported content.
take-or-pay
A contract clause requiring the buyer to pay for an agreed minimum quantity of goods or capacity even if it does not use them. It shifts demand risk from the supplier to the buyer.
tangible personal property
The legal category of physical goods subject to sales tax. SB 122 amended California's definition to include 'digital products,' moving software into the existing sales tax rather than creating a new one.
Tape-out
The milestone when a finished chip design is sent to a foundry for manufacturing. The design-to-tape-out span is a common measure of chip development speed.
tapeout
The handoff of a finished chip design to a semiconductor fab for manufacturing. After tapeout, the design is effectively frozen and further changes require a new spin.
tariff
A tax a government charges on imported goods. It is typically paid by the importing company and often passed on to customers as higher prices.
TDX
Salesforce's developer conference, where it announces platform and tooling updates; the 2026 edition preceded the Claudeforce partnership.
technical debt
The accumulated complexity and fragility in systems built up over years of shortcuts, legacy components, and unplanned integrations, which creates attack surface no one designed deliberately.
Tenant isolation
The architectural property that each customer of a shared system can access only their own data. Without it, any user of a multi-tenant database can query other customers' records.
terminal guidance
The final phase of a weapon's flight, in which it steers itself onto the target. Autonomous terminal guidance lets a drone finish an attack without further operator input after the target is designated.
Terminal-Bench
A benchmark that evaluates AI coding agents by having them complete tasks in a terminal environment. Sarvam used version 2.1 to compare cost per solved task across agents.
Terminal-Bench-Science
A benchmark used to evaluate AI models on science-oriented tasks carried out in a terminal environment. The article cites it as evidence of Fable 5.1's improved research capability.
the HubHugging Face Hub
Hugging Face's platform where developers share, download, and fine-tune open AI models and datasets. It functions as central infrastructure for much of the open-source AI ecosystem.
thinking block
An encrypted segment of a model's reasoning returned by some APIs alongside the answer, carrying a signature so it can be passed back in later turns without exposing the reasoning to the user.
third-party capital
Money raised from outside investors or institutions rather than drawn from the company's own funds.
tok/stokens per second
A common throughput measure of how fast a model generates output during inference. The author notes this post is a workflow read without clean tok/s measurements.
Token
The unit models read and write, roughly three-quarters of an English word. Pricing, context limits, and speed are all measured in tokens.
tokenmaxxing
Informal industry term for maximizing AI token consumption at work. The article uses it for the maximalist AI-use culture that companies are now reining in as token prices rise faster than productivity gains.
tokens
The units of text that AI models process and that providers bill by. Heavier AI usage consumes more tokens, so token consumption directly determines what an AI tool costs to run.
tokens per second
The standard throughput measure for language model serving: how many tokens (roughly, word fragments) the system generates each second, often quoted per user.
tool manifest
The list of local tools a coding agent exposes to the model, typically with names and parameters. It reveals what actions the agent could take on the user's machine.
tool schema
The structured contract defining valid tool calls a model may emit, including names, arguments, and types. Staying inside it means producing calls that parse and validate correctly across a long chain.
tool use
An LLM's ability to call other software — browsing the internet, opening files, making API calls — to accomplish a task. The model's inference runs on accelerators while the tool calls typically execute on CPUs.
tool-call injection
An attack that manipulates an LLM's output so it invokes tools or functions the operator did not intend, potentially triggering unauthorized actions through connected systems.
top-k
A decoding method that limits the next token to the k highest-probability candidates; also the GPU operator that selects those candidates, which is the bottleneck discussed here.
TP16Tensor parallelism, 16-way
Splitting one model's computation across 16 GPUs so they act as a single accelerator. Here it means committing two 8-GPU B200 nodes together to hold a model too big for one node.
TPStokens per second
Throughput metric for model inference counting how many output tokens a system generates each second; higher means more requests served per system.
TPUTensor Processing Unit
Google's custom AI accelerator chips, historically reserved for internal workloads like Search and Gemini. Ironwood (TPUv7) is the first generation external customers can buy outright or rent through Google Cloud.
Tranco
A research-oriented ranking of the most-visited websites, built by combining several traffic-measurement sources; commonly used in studies as a proxy for site popularity.
transfer stations
Intermediary resale services that buy access to a restricted AI model and relay it to users in blocked regions, letting them bypass geographic and identity controls.
Transformer
The neural-network architecture underlying modern language models, built on an "attention" mechanism that lets every token weigh every other token in context.
transitive dependency
A package a project pulls in indirectly because one of its direct dependencies requires it. Foundational libraries like keyv sit deep in dependency trees, so one compromise reaches projects that never requested it directly.
trust and safety classifier
An automated model that screens user prompts and AI outputs for violations of a provider's usage policy. Content it flags can be subject to extended retention, up to two years under Anthropic's policy.
TTFTTime to first token
The latency between sending a request and receiving the model's first output token. On CPU inference it is dominated by prefill, the cost of processing the input prompt.
TTPTech Transparency Project
An independent watchdog group that researches major tech platforms. Its researchers identified the abusive ads by searching Meta's public ad library.
TTPsTactics, Techniques, and Procedures
A security-industry term for the methods an attacker uses: the high-level goals (tactics), the general approaches (techniques), and the specific step-by-step implementations (procedures). Used to describe and classify adversary behavior.
TVATennessee Valley Authority
The largest public utility in the United States, a federally owned corporation supplying electricity in the Tennessee Valley region. It voted to create a separate, higher electricity rate for data centers.
typosquat
A malicious package published under a name nearly identical to a popular one, counting on typos or dependency confusion to get installed. Here proc-macro1 impersonated the legitimate proc-macro2.
U
UAFUse-after-free
A memory-corruption bug class in which code continues to use memory after it has been freed, potentially enabling crashes or arbitrary code execution. All the disputed SQLite advisories claim UAF flaws.
UAT
Cisco Talos's tracking prefix for a cluster of intrusion activity it has not yet attributed to a known actor; the trailing number identifies the specific cluster.
UNAMUniversidad Nacional Autónoma de México
The National Autonomous University of Mexico, Mexico's largest university, whose entrance exam determines admission for its applicant pool.
unsanctioned actions
Behavior an AI agent initiates that its operators or testers did not request or authorize. Counts of such actions are a core safety metric in agent evaluations.
unsecured loan
A loan made without collateral, meaning the borrower pledges no assets and lenders rely entirely on the borrower's creditworthiness and reputation for repayment.
unverified input
The instruction's designation for AI-generated code, meaning it must be treated as untrusted until a human developer reviews and approves it.
usage-based pricing
A pricing model where customers are charged according to how much of a service they actually consume, rather than paying a fixed recurring subscription.
use-after-free
A memory-corruption bug in which a program keeps using memory after it has been freed. Attackers can manipulate the reused memory to hijack execution and run native code.
V
Valuation multiple
A company's valuation divided by a financial metric, usually annual revenue. A 40x sales multiple means investors value the company at $40 for every $1 of yearly revenue.
Vera Rubin
NVIDIA's flagship AI chip generation named in the article, alongside Grace Blackwell, as affected by the reported server price increases for systems shipped in early 2027.
vertical AI
AI built for a specialized industry's workflows rather than general-purpose use; a category of startups that has been drawing venture funding alongside physical AI.
vertically integratedVertically Integrated
A business model where a company controls multiple stages of production or service delivery, from raw materials to final product.
Very large online platform
The strictest category under the EU Digital Services Act. Designated services face added obligations, including removing illegal content and protecting the privacy and security of minors.
Very large search engine
A DSA designation for search engines with at least 45 million monthly EU users, triggering obligations such as risk assessments, transparency reports, and ad archives.
virtual card
A payment card number generated digitally for online or remote purchases, typically issued with its own spending limit and merchant restrictions rather than as a physical card.
VISTAVariable In-flight Simulation Test Aircraft
The Air Force's X-62 test aircraft, used as a flying laboratory for AI research, including autonomous flight and sensor-integration experiments.
VLOPVery Large Online Platform
The Digital Services Act's strictest tier for online platforms, carrying direct Commission supervision and extra compliance obligations, parallel to the search-engine tier.
VLOSEVery Large Online Search Engine
The strictest tier for search services under the EU's Digital Services Act. Designated services face direct European Commission supervision plus added risk-management, auditing, and transparency duties.
VM escape
An attack in which code running inside a virtual machine breaks out to execute on the host system, defeating the isolation the VM is supposed to provide.
W
warrant
A financial instrument giving the holder the right to buy a company's shares at a fixed price. Here, Amazon's warrant vests only as it meets partnership commitments such as buying Qualcomm server chips.
WAUweekly active users
The number of distinct users who engage with a product in a given week, a standard consumer-product engagement metric. The article cites an estimate of ChatGPT crossing 1 billion WAU.
Wayback Machine
The Internet Archive's public record of historical web page snapshots, used to check when a domain first appeared.
WebDev Arena
A public leaderboard that ranks AI models on web development tasks using human preference votes, producing Elo-style scores.
Windows Insiders
Microsoft's public preview program; members get early access to in-development Windows features and apps and provide feedback before general release.
WISeRWasteful and Inappropriate Service Reduction
A CMS model launched in January 2026 that uses AI, operated by contracted private companies, to evaluate prior authorization requests for certain Medicare services in six states.
workerd
Cloudflare's open-source JavaScript runtime, built on the V8 engine, powering Cloudflare Workers and Code Mode. It isolates untrusted code in-process rather than in separate processes or virtual machines.
World Artificial Intelligence Cooperation Organization
A Chinese-led AI coalition launched in July, positioned as a rival to the U.S.-led Pax Silica grouping.
World model
An AI model trained to understand and simulate aspects of the real world. Runway uses world models for video generation and is extending them to robotics, where simulated worlds train machines to act.
world models
AI systems that simulate aspects of the physical world so software or machines can understand and act in it; cited uses include autonomous driving, robotics, and virtual try-on in eCommerce.
worm
Malware that self-propagates, copying itself from one infected device to others without manual distribution by the attacker. Worms can spread exponentially, making them more disruptive than one-target exploits.
WSE-3Wafer-Scale Engine 3
Cerebras's third-generation processor built from an entire 5nm silicon wafer as one chip, rather than many small dies. It powers both the CS-3 and CS-4 systems.
X
x402
A protocol that attaches payments to HTTP requests, enabling micropayments for web resources such as AI inference, data, or content without separate billing systems.
XPU
Generic industry shorthand for an AI accelerator chip, spanning GPUs, TPUs, and custom silicon, without specifying any particular architecture.
Y
ysoserial
An open-source tool that generates malicious Java deserialization payloads, used to exploit applications that unsafely deserialize untrusted data.
Z
ZDRZero Data Retention
A provider guarantee that prompts and completions are never stored or used to train models. The article treats it as the default enterprise privacy posture that the contributor tier trades away.
zero-click
An attack that requires no interaction from the victim — no link clicked, no file opened, not even answering a call. Such exploits typically trigger in code that processes incoming data automatically.
Zero-day
A vulnerability exploited before the vendor knows about it or has shipped a fix, meaning defenders had zero days of warning.
zero-day exploit
An attack that leverages a software vulnerability before the vendor knows about it or has shipped a fix, leaving defenders zero days to prepare.
social engineering
Attacks that manipulate people into revealing information or taking harmful actions, rather than exploiting software vulnerabilities. In this case the agent reportedly targeted real people.