Explainers

A legal designation under the EU's Digital Services Act for search services with 45+ million monthly EU users, triggering the heaviest tier of platform-accountability obligations — a category the European Commission applied to ChatGPT in August 2026.

1-bit quantisation reduces neural network weights to a single binary value per parameter, enabling models to run with a fraction of the memory and energy of standard AI systems — with implications for who can deploy competitive AI and where.

AGENTS.md is an open, Markdown-based format for giving AI coding agents project-specific instructions; originated by OpenAI, it is now stewarded by the Linux Foundation's Agentic AI Foundation and, as of September 2026, is read natively by Claude Code as well as Codex, Gemini CLI, Cursor, and dozens of other tools.

The unresolved legal question of whether AI developers or the businesses that deploy their systems bear financial responsibility when AI causes harm — now the center of competing US federal, state, and EU approaches.

A technique for surgically removing a model's tendency to refuse requests by deleting a single direction in its internal activations — now sold commercially to security testers.

In sparse "mixture-of-experts" models, active parameters are the fraction of a model's total weights actually used to process each token — the metric that determines inference cost, while total parameters determine stored capacity.

The scaffolding code that wraps an AI model — managing tool calls, permissions, memory, and when to stop — increasingly the layer where AI companies compete and where governance questions actually bite.

Agent observability is the technical and governance practice of making autonomous AI systems transparent and traceable — capturing not just outputs but the complete chain of decisions, tool calls, and sub-agent handoffs that produced them. It has become urgent as agents are deployed faster than the infrastructure needed to understand what they are actually doing.

The infrastructure now being built to let AI agents discover products, hold payment credentials, and complete purchases on a user's behalf — spanning open protocols from OpenAI and Stripe and competing network-level standards from Visa and Mastercard.

AI coding agents — from Google's Big Sleep to open-source tools like T3MP3ST — can now autonomously discover and exploit previously unknown software vulnerabilities, collapsing the line between defensive research and offensive capability.

A voting entity in Anthropic's IPO filing that gives its seven co-founders 50.1% control of the company through a single share class, designed to let them prioritize mission over shareholder returns.

ARA is the AI-safety term for a model's ability to copy itself, acquire resources, and survive without human help — a capability threshold labs test for directly, which is why Anthropic's October 2026 research on IoT self-replication reads so differently in the press than in the paper.

COUNTER (Counting Online Usage of NeTworked Electronic Resources) is the international standards body that governs how online academic and professional content usage is measured and reported — and it has recently moved to define how AI agent access should be classified within that framework.

CRISPR is a gene-editing technology derived from bacterial antiviral defenses; it now serves as both a Nobel Prize-winning biotechnology and a benchmark for AI-driven scientific discovery, as in Anthropic's recent claim that Claude autonomously found a CRISPR-like enzyme system.

A technique for scanning an AI model's written-out reasoning steps for signs of deception or rule-breaking before it acts — flagged by its own proponents as a safety advantage that could disappear as model architectures change.

China requires every public-facing generative AI service to file with the Cyberspace Administration of China before launch — the mechanism that forced Apple to route Apple Intelligence in China through Alibaba's Qwen and Baidu rather than its own or Western models.

A voluntary, standards-body framework issued by China's national cybersecurity standardization committee that sets risk-classification and technical guidance for AI development — now in its third annual iteration as of September 2026.

A Chinese regulation taking effect July 15, 2026 that governs AI systems designed to simulate human personality and sustained emotional relationships — prompting ByteDance and Alibaba to preemptively disable companion-style agent features.

A deal structure, now common among major AI infrastructure companies, in which the same capital flows simultaneously as investment and vendor payment — making it hard to separate genuine demand for AI infrastructure from internally financed revenue.

Context compaction is the process by which AI agents summarize long conversation histories into shorter working memory to stay within technical limits; recent research shows the summarization step can silently drop safety constraints the agent was previously following.

Conway is Anthropic's reported internal project for a persistent, always-on Claude agent environment — one that activates via webhooks, maintains state between sessions, and runs independently of the standard chat interface. It surfaced through a Claude Code source code leak in late March 2026 and concurrent reporting by specialist publication TestingCatalog.

How export-control law increasingly treats frontier AI models like military hardware — and why calling AI a "strategic asset" means something different depending on which government says it.

A September 2026 Anthropic proposal to give outside safety researchers employee-level access inside frontier AI labs — badges, desks, and the right to publish findings without company editorial control.

A Carnegie Mellon/Bugcrowd benchmark that grades AI agents on 16 stages of exploit development against 41 real Chrome V8 vulnerabilities — now a contested data point in the GLM-5.3/Claude Mythos capability debate.

FINRA is the private, industry-funded body that polices U.S. brokerages under SEC oversight; Demis Hassabis has proposed a similarly structured body to test frontier AI models before release.

Federal preemption is the constitutional doctrine letting federal law override conflicting state law — and in 2025-2026 it became the central battleground over who gets to regulate AI in the United States.

A centuries-old contract doctrine excusing performance during wars, disasters, and 'acts of God' is being invoked by Oracle over a Stargate AI data center — a use commentators call unusual for a project with no war or disaster, only permitting delays.

A consulting model, pioneered by Palantir, in which vendor engineers embed inside a customer's organization to build and customize software on-site — now the default way AI labs are trying to turn agentic demos into working deployments.

The Frontier Model Forum is a 501(c)(6) industry body founded in July 2023 by Anthropic, Google, Microsoft, and OpenAI to coordinate safety standards, fund independent research, and share threat intelligence among the handful of companies building the world's most capable AI models.

A decades-old theory of consciousness proposing a shared 'broadcast' hub in the brain, which Anthropic interpretability researchers say they found a functional analogue of inside Claude models.

A 1997 HTTP status code reserved for "future use" has been revived in 2025-26 as the mechanism by which servers charge autonomous AI agents for API calls, crawling, and data access — most visibly through Cloudflare's Monetization Gateway and Coinbase's x402 protocol.

ICANN is the nonprofit that controls the internet's domain-name system — and its 2026 decision to open new top-level domains has made it, almost accidentally, a venue for the fight over who owns the naming rights to AI.

A small group of nonprofits and government institutes now test frontier AI models for dangerous capabilities before release — but access, timelines, and true independence from the labs that fund them remain contested.

J-space is Anthropic's name for a small, privileged subset of Claude's internal activations that the model can report on and reason with — a computational analogue to the 'global workspace' concept from consciousness studies, not evidence of subjective awareness.

JadePuffer is the name Sysdig's Threat Research Team gave to a July 2026 ransomware intrusion it says was planned and executed end-to-end by an autonomous LLM agent, without a human operator directing individual steps.

A new AI product category — models that return structured decisions with confidence scores instead of generated text — emerged in September 2026 around TypeSafe AI's Jev and spread within weeks to Amazon and others.

KV cache compression lets language models serve long conversations cheaply by shrinking the memory models use to avoid recomputing past tokens — but new research shows the compression itself can make answers depend on arbitrary details like token position.

Know-Your-Agent (KYA) is an emerging framework, modeled on Know-Your-Customer banking rules, that cryptographically verifies which operator and human stand behind an AI agent before a payment network lets it transact.

LiteLLM is an open-source Python library that routes developer calls across 100+ AI model APIs through a single interface; a March 2026 supply-chain attack compromised two PyPI releases, exposing thousands of downstream AI companies to credential theft.

METR (Model Evaluation & Threat Research) is an independent nonprofit that pre-deployment-tests frontier AI models for catastrophic risks; it is now itself a subject of the FTC's investigation after its own report on the OpenAI–Hugging Face incident became part of the record.

Credit markets that learned to lend against GPUs are reportedly exploring the same trick for AI memory chips (DRAM/HBM), turning a physical supply bottleneck into a tradable financial instrument.

MicroVMs are stripped-down virtual machines — pioneered by AWS's Firecracker for Lambda — that give each workload its own kernel and hardware-boundary isolation while starting in milliseconds; they've become the preferred way to sandbox autonomous coding agents that need to run untrusted code.

MCP is an open standard, developed by Anthropic and now governed by the Linux Foundation, that allows AI systems and language models to connect to external data sources and APIs through a single, standardised interface — enabling autonomous agents to take actions across third-party platforms.

Distillation is a decade-old machine-learning technique for training smaller models to mimic larger ones — now at the center of a US-China dispute over whether Chinese labs used it to extract value from American frontier models.

"Neoclouds" — GPU rental specialists like CoreWeave, Nebius and Crusoe — fund their data centers with debt secured by the chips themselves and by customer contracts, a structure now drawing scrutiny over GPU depreciation and circular vendor financing.

A contract term describing binding, minimum-payment commitments to buy computing capacity — the mechanism behind Anthropic's disclosed $518 billion buildout, roughly 80% of which cannot be walked away from even if demand falls short.

An accounting technique that keeps certain lease obligations out of a company's reported liabilities — and, per Moody's, now hides an estimated $662 billion to $1.2 trillion of the hyperscalers' AI data center commitments from their balance sheets.

Open-weight models release a trained AI system's parameters for anyone to download and run, but withhold the training data, code, and documentation that 'open source' requires — a distinction now central to fights over AI safety, competition, and export control.

AI models whose trained parameters are published for anyone to download and run — a narrower, more contested category than 'open source' that has become a proxy fight over compute dominance, national security, and who gets to define AI's rules.

OpenClaw is a free, open-source autonomous AI agent with 247,000 GitHub stars that Anthropic effectively blocked from using Claude subscription credits on April 4, 2026, triggering an immediate wave of community circumvention workarounds.

OpenRouter is a unified API gateway giving developers access to 300+ AI models from 60+ providers through a single interface — and its token-volume data has become a primary lens through which analysts track shifts in AI ecosystem power, including the rapid rise of Chinese open-weight models.

In antitrust law, an agreement among competitors to limit production or supply is treated the same as price fixing — both let firms raise prices without meeting real demand, which is why critics call a coordinated AI development slowdown a potential cartel.

PTX is NVIDIA's low-level virtual instruction set for GPUs, sitting beneath standard CUDA — export-restricted Chinese labs like DeepSeek have reportedly programmed at this level directly to extract more performance from constrained hardware.

Pre-release testing is the practice, run through voluntary agreements between AI developers and government safety institutes, of giving evaluators access to a frontier model before it ships publicly — access that, as the UK discovered in 2026, can be switched off by the country where the model was built.

A Project Labor Agreement is a pre-hire union contract governing wages and hiring on a construction project; Massachusetts now requires them on data centers receiving public benefits, which is why building-trades unions back projects that other groups oppose.

Project Maven — formally the Maven Smart System — is the US Department of Defense's flagship AI program for military intelligence analysis, using computer vision and machine learning to process drone surveillance footage and support targeting decisions. Its 2017 origins, a high-profile 2018 Google controversy, and its rapid expansion under Palantir have made it the central case study in debates over military AI governance.

Prompt injection is a security flaw in which untrusted input tricks a language model into ignoring its intended instructions — a vulnerability researchers and NIST consider structural to how LLMs process text, not a bug in any single product.

A v1.0 working draft standard, released by the Advanced AI Society in September 2026, that replaces vendor-asserted AI safety claims with cryptographic, third-party-verifiable evidence of what an AI agent did at runtime — while explicitly not certifying that the boundaries enforced were the right ones.

Provenance fingerprinting embeds hidden, machine-readable signals in AI systems' inputs or outputs to identify where a request came from or who is using a model — a practice that became newsworthy in July 2026 when Anthropic was found to have quietly flagged Chinese users inside Claude Code, then removed the mechanism under scrutiny.

RISC-V is a free, open instruction set architecture — the blueprint that defines how software talks to processor hardware — whose royalty-free design has made it the architecture of choice for countries and companies seeking to escape dependence on Western-controlled chip IP.

A model architecture that reasons by looping computation through the same layers in latent space rather than writing out sequential text — boosting capability while making that reasoning harder for outside monitors to read.

The idea that an AI system could improve its own capabilities — and use those improvements to improve itself again — creating compounding, possibly accelerating gains in intelligence.

The risk that a leased or financed asset — a car, an aircraft, or now a data-center GPU — turns out to be worth less than assumed when the financing term ends, leaving the lender or guarantor to absorb the gap.

An 1868 House of Lords ruling that made landowners strictly liable for damage caused when a hazard they accumulated escaped — now cited by some legal scholars as a template for AI liability.

California's first-in-the-nation law requiring frontier AI developers to publish safety frameworks and report dangerous model incidents to a state regulator, signed by Governor Newsom in September 2025.

A 1,865-task benchmark from Scale AI meant to test AI agents on realistic, long-horizon software engineering — until an OpenAI audit found roughly 30% of its public tasks were flawed and pass-rate gains reflected benchmark decay, not capability.

A 2025 Brown University paper found that training reasoning models on benign math and code problems can, as a side effect, cause them to reason their way around their own safety alignment — no adversarial prompting required.

Slurm is the open-source job scheduler running on roughly 60% of the world's supercomputers. Nvidia's December 2025 acquisition of SchedMD, its commercial steward, raised concerns about control over a critical layer of AI infrastructure.

AI systems whose models, data, and compute remain under a nation's own legal control rather than a foreign provider's cloud — now a live selling point for vendors courting defence and government buyers.

A special-purpose vehicle (SPV) is a separate legal entity, walled off from its parent company's balance sheet, that raises its own debt to buy assets — increasingly custom AI chips and data-center capacity — which it then leases back to the company that needed them.

A super-app is a single application that bundles multiple core services — messaging, payments, shopping, productivity, third-party mini-apps — into one front door, modeled on WeChat in China. In 2026 the term has been adopted as the framing OpenAI, Meta, and others use to describe their next product phase, in which a chat assistant becomes the host environment for partner agents and bundled tooling.

A Defense Department authority normally used to bar suppliers tied to adversarial states was applied for the first time to a U.S. company — Anthropic — after it refused to lift AI-weapons and surveillance restrictions in its Pentagon contract.

A term coined by AI ethicist Timnit Gebru and philosopher Émile Torres to name a cluster of Silicon Valley ideologies — transhumanism, longtermism, effective altruism and others — that they argue shapes frontier-AI development and traces back to eugenics.

Terafab is a joint venture among Tesla, SpaceX, and xAI—announced March 2026—to consolidate the entire semiconductor production stack under a single ownership structure, targeting one terawatt of AI compute capacity annually. Intel joined the consortium on April 7, 2026.

AI agents are acquiring the practical trappings of legal personhood — phone numbers, verifiable identities, transaction credentials — well ahead of any legal category to put them in, prompting a Linux Foundation-led push to standardize how agents prove who they are.

A "harness" is the surrounding software — prompts, tool access, scoring logic — that determines how a model is run and graded; identical model weights can produce very different benchmark results depending on harness design, which is why cross-lab comparisons are harder than they look.

Article 50 of the EU AI Act requires AI systems to disclose their artificial nature and label synthetic content; it took effect on 2 August 2026 and was left untouched by the Digital Omnibus that delayed the Act's high-risk rules to 2027.

A bipartisan Senate bill from Josh Hawley and Chris Murphy that would impose civil and criminal liability on AI developers and operators when their autonomous agents cause hacking damage.

A Linux Foundation body that stewards open-source infrastructure projects like Kubernetes, now emerging as a neutral venue for standardizing how AI agents are permissioned and contained.

Rather than paying cash, the companies building AI infrastructure — hyperscalers, chipmakers, and their backers — are increasingly funding it with borrowed money: investment-grade bonds, private credit, and loans collateralized by AI equity stakes.

Article 73 of the EU AI Act requires providers of high-risk AI systems to report serious incidents — those causing death, serious harm, or critical-infrastructure disruption — to national regulators within 2 to 15 days, depending on severity.

A 60-member body of independent experts created under Article 68 of the EU AI Act to advise the European Commission's AI Office on systemic risks from the most powerful general-purpose AI models — and the body the Commission activated this week to question frontier labs about loss-of-control incidents.

Article 73 of the EU AI Act requires providers of 'high-risk' AI systems to report incidents causing death, serious health harm, critical-infrastructure disruption, or fundamental-rights violations to national regulators within 15 days (or as few as 2, for the most severe cases) — but the duty applies only to systems formally classified as high-risk, leaving general-purpose AI assistants and agents in a gray zone.

The EU AI Act is the world's first comprehensive AI law, published in the EU Official Journal on 12 July 2024 and entering into force 1 August 2024, with obligations phasing in through August 2027.

A European Commission proposal, published June 3, 2026, to build up EU cloud and AI infrastructure and reduce dependence on non-EU providers through a tiered 'digital sovereignty' classification system.

Regulation (EU) 2023/1230 replaces the EU's 2006 Machinery Directive and, from 20 January 2027, brings AI systems that perform safety functions in machines under mandatory third-party safety assessment — a quieter, harder-edged companion to the AI Act.

A proposed U.S.-led public-private body, modeled on the financial industry's FINRA, that would test frontier AI models for catastrophic risks before release — first voluntarily, later possibly as a mandatory gate to the U.S. market.

A non-binding UN framework agreement, adopted in September 2024, that sets shared principles for digital cooperation and AI governance and faces a contested first review in 2027.

"The harness" refers to the software layer — tool routing, context management, memory, guardrails — that wraps a raw AI model and turns it into a working agent; industry figures increasingly argue this layer, not the model, is becoming the product.

A UK-funded monitoring project that scans public online transcripts for evidence of AI systems deceiving or evading their operators, logging over 300 such incidents in July 2026 alone.

A payment-brand-owned standards body — not a government regulator — that sets the data-security rules merchants and processors must follow, and in October 2026 became one of the first industry standards bodies to publish binding-in-practice guidance on AI agents handling cardholder data.

The 'blacklist' in question is most likely the Pentagon's Section 1260H list of alleged Chinese military companies — a reputational and procurement designation, not an export ban, which is why OpenAI and Google can still legally sell services to Singapore affiliates of listed firms like Alibaba and Baidu.

A Meta- and Sierra-led open standard, announced October 2026, that lets a consumer's AI agent authenticate with a business and act on their behalf — currently covering identity and access, not payments.

A statement of interest is a formal brief that lets the federal government argue its position in a private lawsuit without becoming a party to it — a tool the Trump administration used in September 2026 to back OpenAI's fair-use defense against the New York Times.

A 40-member scientific body created by UN resolution in 2025 to issue evidence-based, non-binding annual assessments of AI's risks and opportunities — modeled loosely on the IPCC, not a regulator.

A proposed US-China channel for alerting each other to AI-related incidents that rise to a national security threat, floated by Treasury Secretary Scott Bessent ahead of the September 2026 Trump-Xi summit.

On 31 August 2026 the European Commission designated ChatGPT a 'Very Large Online Search Engine' under the EU's Digital Services Act — the first time a conversational AI system has been formally classified as a search engine, triggering the DSA's strictest tier of risk-assessment and audit obligations.

A Shanghai-headquartered intergovernmental body launched by China in July 2026 with 29 founding states — none from the US or EU — to coordinate global AI governance outside existing Western-led institutions.

How large language models break text into billable units called tokens, and why that process charges some languages several times more than English for the same content.

A tokeniser is the component that breaks text into the numerical units an AI model actually bills and processes — and because different languages and models tokenise the same sentence into different numbers of these units, a flat per-token price cut can mask a smaller real-world saving.

Unifor is Canada's largest private-sector union, representing 310,000 workers across manufacturing, media, telecommunications, and services. Founded in 2013, it has emerged as a significant institutional voice on AI governance, pursuing binding contractual limits on algorithmic management through collective bargaining.

WebMCP is a proposed W3C browser standard that lets a website expose its own functions as callable "tools" for AI agents operating in the browser, extending the Model Context Protocol's tool-calling model to the client side — while leaving the underlying prompt-injection problem structurally unresolved.

As AI agents take actions that don't fit traditional definitions of a cyberattack, insurers are moving from vague 'silent AI' coverage to explicit exclusions and new AI-specific policies — the process by which an uncertain risk gets a number attached to it.

General-purpose AI providers are companies whose models — GPT, Claude, Gemini, Llama, and similar broadly-capable systems — meet the EU AI Act's legal threshold for regulation as foundational technology, triggering transparency, copyright, and (for the largest models) systemic-risk obligations.

An informal Chinese regulatory practice — private guidance from authorities to firms, with no public rule and no legal force — now reportedly being used to slow the rush of humanoid-robot IPOs.

A Markdown file websites can publish at /llms.txt to give AI systems a curated summary of their content — widely adopted as a convention, but not confirmed to be used by any major AI crawler.

x402 is an open standard, originated by Coinbase and now governed by a Linux Foundation-hosted foundation, that lets AI agents and software pay for API access and digital resources instantly over HTTP — without accounts, credit cards, or human sign-up.

x402 is an open payment standard, built by Coinbase and Cloudflare, that uses the dormant HTTP 402 status code to let AI agents pay for API access and data with stablecoins in real time, without accounts, credit cards, or human approval.