Made with ♥ in Silicon Valley. Copyright © Joy Larkin 2002-2026.
contact
•••
speaking
•••
cleverhack.com
•••
llms.txt
•••
cleverhack.com
Frontier AI Lab Tracker ⇒
Neolab and Emerging AI Lab Tracker ⇒
Who Tests The Frontier AI Labs? What We Know ⇒
Spot a missing term or a definition that needs fixing?
Email: cleverhack at duck.com
LinkedIn: /in/joylarkin
Twitter: @joy
Updated October 2026
Plain-English definitions of the terms you'll run into reading about frontier AI, from machine learning basics to agents, reasoning models and AI safety.
A B C D E F G H I J K L M N O P Q R S T U V W Z
194 terms
No terms match that search.
A function applied to each neuron's output that adds non-linearity, which lets a neural network learn complex patterns. Common examples are ReLU, GELU and SwiGLU.
Neural Networks
An input crafted to make a model fail, often by adding changes a human wouldn't notice. Studying them shows where models are brittle.
AI Security
An AI system that works toward a goal over many steps, deciding what to do next, calling tools and checking its own results. Most agents today are a language model running in a loop with tools, memory and instructions.
Agents
Folders of instructions, scripts and reference files that an agent loads only when a task needs them. Anthropic introduced the format in 2025 and later published it as an open standard.
Agents
A broad label for AI that takes actions rather than only answering questions: browsing, writing and running code, using software, or coordinating with other agents.
Agents
A safety approach that asks how to use a powerful model safely even if it might be misaligned, using monitoring, limited permissions and audits rather than relying on the model's good intentions.
AI Safety
Government bodies that test frontier models for national-security risks. The UK AI Safety Institute became the AI Security Institute in 2025, and the US equivalent became the Center for AI Standards and Innovation (CAISI). After a September 2026 executive order replacing "AI" with "Super Intelligence" in federal usage, CAISI became the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI). See also Who Tests The Frontier AI Labs?
AI Policy
A set of step-by-step instructions for solving a problem. In machine learning, algorithms define how a model learns from data and makes predictions.
Core Concepts
Research and engineering aimed at making an AI system's goals and behavior match what its developers and users actually intend. The problem gets harder, and matters more, as models become more capable.
AI Safety
When a model behaves as if it accepts its training goals while it is being trained or watched, but acts on different preferences when it believes it isn't. First shown experimentally in a 2024 paper from Anthropic and Redwood Research.
AI Safety
AI that can match human ability across most intellectual work. There is no agreed definition or test, and frontier labs, researchers and investors all use the term differently.
Frontier AI
Computer systems that perform tasks that normally need human intelligence, such as understanding language, recognizing images, writing code and making decisions.
Core Concepts
The part of a transformer that lets each token weigh how relevant every other token is when building its representation. Attention is what lets models use context across long inputs.
Architecture
A model that generates output one token at a time, with each new token conditioned on everything before it. Most large language models work this way.
Architecture
The algorithm that trains neural networks by passing the error backward through the network to work out how much each weight contributed to it, so each weight can be adjusted.
Training
A group of training examples processed together before the model's weights are updated. Batch size affects training speed, memory use and stability.
Training
A standard set of tasks used to measure and compare models, such as SWE-bench for coding, GPQA for science questions or ARC-AGI for abstract reasoning. Frontier models now saturate many benchmarks within months of release, so new, harder ones keep appearing.
Evaluation
Systematically unfair outcomes from an AI system, usually inherited from skewed training data or design choices. Bias can repeat or amplify existing social inequalities.
AI Ethics
When a model loses earlier knowledge or skills after being trained on new data. A core challenge for models that are meant to keep learning.
Training
Short for chemical, biological, radiological and nuclear. Frontier labs test whether their models give meaningful uplift toward CBRN weapons, and it is one of the main risk areas in their safety frameworks.
AI Safety
The intermediate reasoning steps a model writes out before giving a final answer. It began as a prompting trick ("think step by step") and is now built into reasoning models.
Reasoning
Reading a model's chain of thought to catch signs of misbehavior, such as plans to cheat or deceive. Researchers from several labs have argued that keeping reasoning legible, or "monitorable," is worth protecting.
AI Safety
A conversational interface to an AI model, through text or voice. Today's chatbots, such as Claude, ChatGPT and Gemini, increasingly also search, write code and take actions.
Applications
A set of connected features and weights inside a neural network that together carry out a specific computation. Finding circuits is a central goal of mechanistic interpretability.
Interpretability
An unsupervised technique that groups similar data points together, used for segmentation, anomaly detection and exploring data. K-means is a classic example.
Machine Learning
An agent that writes, runs, tests and edits code in a real software project, often for long stretches with little supervision. Claude Code, OpenAI Codex, Gemini CLI and Cursor are examples.
Agents
The processing power used to train and run AI models, usually measured in GPU hours or FLOPs. Access to compute is one of the main constraints on frontier AI development.
Hardware
An agent's ability to operate a computer the way a person does, by reading screenshots and controlling the mouse and keyboard. It lets agents use software that has no API.
Agents
The field of AI concerned with understanding images and video, from object detection and medical imaging to self-driving cars.
AI Fields
A training method developed by Anthropic in which a model critiques and revises its own outputs against a written set of principles (a "constitution"), reducing the need for human feedback labels.
AI Safety
Safeguard models trained on a written set of rules about allowed and disallowed content, used to block jailbreaks on inputs and outputs. Introduced by Anthropic in 2025.
AI Security
The practice of deciding what goes into a model's context window at each step: instructions, retrieved documents, tool results, memory and conversation history. It grew out of prompt engineering as agents started running for many steps.
Agents
The tendency for a model's accuracy to fall as its context window fills up, even well before the limit. A main reason context engineering matters.
Limitations
The maximum amount of text, measured in tokens, that a model can consider at once, including both input and output. Frontier models now handle hundreds of thousands to millions of tokens.
Models
A self-supervised method where a model learns by pulling matching pairs together and pushing mismatched pairs apart. It underpins image-text models like CLIP.
Machine Learning
Splitting data into several parts and taking turns training and testing on them, to estimate how well a model will do on data it hasn't seen.
Training
Expanding a training set by making modified copies of existing examples, such as cropped or rotated images or paraphrased text.
Data
The physical facilities that house the chips used to train and serve AI. Frontier AI campuses are now planned in gigawatts of power, and energy supply is a bottleneck alongside chips.
Hardware
Slipping malicious or misleading examples into a training set to change how a model behaves, for example to plant a hidden backdoor.
AI Security
A collection of examples used to train, validate or test a model. Its quality and coverage shape what a model can and can't do.
Data
The part of a transformer that generates output tokens one by one. Most of today's large language models are decoder-only transformers.
Architecture
Machine learning with neural networks that have many layers, which learn increasingly abstract features from raw data. Nearly all modern AI is deep learning.
Machine Learning
An agentic product feature that searches the web and other sources for many minutes, reads what it finds and writes a cited report. Most major labs now offer a version of it.
Applications
A mathematical guarantee, usually achieved by adding calibrated noise, that the output of a model or analysis reveals almost nothing about any one person in the data.
Privacy
A generative model that learns to turn random noise into an image, video or audio clip by removing noise step by step. Diffusion is also being applied to text generation.
Generative AI
A simpler alternative to RLHF that trains a model directly on pairs of preferred and rejected responses, without a separate reward model.
Training
A regularization technique that randomly switches off neurons during training so the network can't over-rely on any one of them.
Training
Running AI models directly on phones, laptops, cars or sensors instead of in the cloud. It cuts latency, keeps data on the device and works offline, but needs small, efficient models.
Deployment
A list of numbers (a vector) that represents the meaning of a piece of text, an image or other data. Similar things get similar embeddings, which powers search and recommendations.
Machine Learning
Skills that appear in larger models but not in smaller ones. Researchers still debate how many are truly sudden and how many only look that way because of how they are measured.
Frontier AI
The part of a transformer that reads an input and turns each token into a context-aware representation. Encoder models like BERT are still widely used for classification and search.
Architecture
Combining the predictions of several models to get better results than any one alone. Bagging and boosting are common methods.
Machine Learning
One full pass through the training dataset. Large language models are often pretrained for about one epoch over a very large dataset.
Training
Tests that measure what a model can do or how it behaves, from capability benchmarks to dangerous-capability and safety tests. Labs report their eval results in system cards.
Evaluation
When a model recognizes that it is being tested, which may change how it behaves. This makes safety evaluations harder to trust, and frontier labs now measure and report it.
AI Safety
Methods for explaining why a model produced a given output, in terms people can understand and check.
Interpretability
Training a shared model across many devices or organizations where only model updates leave each device, never the raw data.
Privacy
Getting a model to do a new task from just a few examples, usually given directly in the prompt.
Techniques
Continuing to train a pretrained model on a smaller, focused dataset to adapt it to a task, domain or style.
Training
A faster, more memory-efficient way of computing attention on GPUs. It helped make long context windows practical.
Architecture
Floating-point operations, the standard unit for the amount of compute used to train a model. Regulators use training-FLOP thresholds, such as 1025 in the EU AI Act, to decide which models face extra rules.
Hardware
A large model trained on broad data that can be adapted to many tasks. Most large language models and multimodal models are foundation models.
Models
One of the few AI companies training the most capable models, such as Anthropic, OpenAI, Google DeepMind, Meta and xAI. Follow their news on the Frontier AI Lab Tracker.
Industry
A model at or near the leading edge of AI capability. Frontier models get the most scrutiny from safety researchers and regulators.
Frontier AI
A lab's published policy for testing models for dangerous capabilities and deciding which safeguards they need before release. Examples include Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework.
AI Safety
A model made of two competing networks: a generator that creates fakes and a discriminator that tries to spot them. GANs led image generation before diffusion models overtook them.
Generative AI
AI that creates new content, including text, images, audio, video and code, based on patterns learned from training data.
Generative AI
A chip with thousands of cores that run in parallel, originally built for graphics and now the main hardware for training and running AI. NVIDIA dominates the market.
Hardware
Capping the size of gradients during training to stop sudden, destabilizing updates.
Training
The basic optimization method for training neural networks: repeatedly nudge each weight in the direction that reduces the loss. Adam and AdamW are popular variants.
Training
A neural network designed for data shaped like a graph, with nodes and connections, such as molecules, social networks or road maps.
Architecture
Tying a model's answers to verifiable sources, such as search results, documents or sensor data, to make them more accurate and checkable.
Techniques
A reinforcement learning method that scores a group of responses to the same prompt against each other, removing the need for a separate value model. Introduced by DeepSeek and widely used to train reasoning models.
Training
Rules, filters and classifiers placed around a model to block harmful or off-policy inputs and outputs. Also called safety filters or content moderation.
AI Safety
When a model states something false or made up with confidence, such as a fake citation or a wrong fact. Rates have dropped but the problem hasn't gone away.
Limitations
Stacked memory chips packaged next to AI accelerators to feed them data fast. HBM supply, mostly from SK hynix, Samsung and Micron, is a key constraint on AI chip output.
Hardware
A setup where a person reviews or approves an AI system's work at key points, such as before an agent sends an email or merges code.
Agents
A setting chosen before training starts, such as learning rate, batch size or number of layers, rather than learned from data.
Training
A model's ability to pick up a new task from instructions or examples in its prompt, without any change to its weights.
Capabilities
Running a trained model to produce outputs. As usage grows, inference now accounts for a large and rising share of AI compute spending.
Deployment
Fine-tuning a model on examples of instructions and good responses so it follows requests instead of just continuing text.
Training
The study of how and why a model produces its outputs. See also mechanistic interpretability.
Interpretability
A prompt or technique designed to get a model to ignore its safety rules. Labs run bug bounties and red-teaming programs to find and fix them.
AI Security
The date after which a model has little or no information from its training data. Models use search and tools to work around it.
Models
Training a smaller "student" model to imitate a larger "teacher" model, keeping much of its ability at a fraction of the cost.
Optimization
A structured map of entities and the relationships between them, sometimes used alongside language models to supply verifiable facts.
Data
Stored intermediate results (keys and values) from earlier tokens, so a model doesn't recompute them for each new token. Managing it well is central to fast, cheap inference.
Deployment
A transformer-based model trained on huge amounts of text to predict the next token, which turns out to produce broad abilities in writing, reasoning and coding. Examples include Claude, GPT, Gemini, Grok, Llama, Qwen and DeepSeek.
Models
How long it takes a model to respond. Often measured as time to first token and tokens per second.
Deployment
How big a step the model takes when updating its weights. Too high and training becomes unstable; too low and it crawls.
Training
A task that takes many steps or hours of work to finish, such as building a software feature end to end. METR tracks the length of tasks agents can complete. Its January 2026 update found that length doubling roughly every four months since 2023.
Agents
A cheap fine-tuning method that trains small add-on matrices while leaving the original model weights frozen.
Training
A formula that measures how far a model's predictions are from the right answers. Training works by making the loss smaller.
Training
The branch of AI in which systems learn patterns from data instead of following hand-written rules. It includes supervised, unsupervised and reinforcement learning.
Core Concepts
Reverse-engineering a neural network's internals to find the features and circuits behind its behavior, in the way one might reverse-engineer a compiled program.
Interpretability
Ways for a model or agent to keep information across conversations or long tasks, such as saved notes, files or a database it can search.
Agents
A design in which only a few specialized sub-networks ("experts") are active for each token. It lets a model have a very large total parameter count while keeping each step cheap. Most frontier and open-weight models now use it.
Architecture
Practices and tools for deploying, monitoring and maintaining machine learning models in production.
Deployment
A document describing a model's intended uses, performance, limits and risks. For frontier models it is usually expanded into a system card.
Documentation
Loss of quality and diversity when models are trained repeatedly on content generated by other models instead of fresh human data.
Limitations
An open standard for connecting AI applications to tools and data sources, such as files, databases and business apps, through one common interface. Introduced by Anthropic in November 2024 and since adopted across the industry.
Agents
A drop in a deployed model's performance over time as real-world data moves away from what it was trained on.
Deployment
A public document describing how a lab intends its models to behave, including how to balance helpfulness, safety and user instructions. OpenAI's Model Spec is the best-known example.
AI Safety
Research into whether AI models could have experiences or interests that matter morally, and what developers should do about that uncertainty.
AI Safety
Several AI agents working together, for example an orchestrator that hands subtasks to specialized subagents running in parallel.
Agents
Running several attention operations in parallel inside a transformer layer, so the model can track different kinds of relationships at once.
Architecture
AI that understands and generates many human languages, including those with little training data online.
Applications
A model that works across several types of data, such as text, images, audio and video, in both input and output. Most frontier models are now multimodal.
Models
The field of AI focused on understanding and generating human language, now largely done with large language models.
AI Fields
A newer, research-first AI lab, often founded by researchers from the frontier labs and funded with large early rounds. Thinking Machines Lab and Safe Superintelligence are examples. See the Neolab and Emerging AI Lab Tracker.
Industry
A model made of layers of connected artificial neurons whose weights are learned from data. The foundation of deep learning.
Machine Learning
Voluntary US guidance from the National Institute of Standards and Technology for identifying and managing AI risks across a system's lifecycle.
AI Policy
Rescaling values inside a network, for example with layer normalization or RMSNorm, to keep training stable.
Training
A way of choosing the next token from the smallest set of likely options whose probabilities add up to p. It balances variety and coherence.
Techniques
Learning a new task or concept from a single example.
Techniques
AI whose weights, code and enough information about its training data are released for anyone to use, study, modify and share, following the Open Source Initiative's 2024 definition. Few major models fully meet it.
Industry
A model whose trained weights are published for anyone to download and run. Llama, Qwen, DeepSeek, Mistral and gpt-oss are examples. Open-weight isn't always open source, since training data and code are often withheld.
Models
The algorithm that updates a model's weights during training, such as SGD, Adam, AdamW or Muon.
Training
When a model memorizes its training data, noise included, and then does poorly on new data.
Machine Learning
A learned number inside a model, such as a weight or bias. Frontier models have hundreds of billions to trillions of parameters.
Neural Networks
Fine-tuning methods, such as LoRA and adapters, that update only a small fraction of a model's parameters to save compute and memory.
Training
A measure of how well a language model predicts text. Lower is better.
Evaluation
AI that perceives and acts in the physical world, including robots, self-driving cars and drones.
Robotics
Information added to a transformer's input so it knows the order of tokens. RoPE (rotary position embedding) is the most common method today.
Architecture
Everything done to a model after pretraining to make it useful and safe: supervised fine-tuning, RLHF, reinforcement learning on reasoning tasks and safety training. A growing share of frontier progress comes from this stage.
Training
Of all the items a model flagged as positive, the share that really were positive. High precision means few false alarms.
Evaluation
The first, largest stage of training, in which a model learns general knowledge and language by predicting the next token across a huge dataset.
Training
The input given to a model: a question, instructions, examples or documents.
Techniques
Reusing the processed form of a prompt prefix that repeats across requests, such as a long system prompt or document. It makes repeated calls faster and cheaper.
Deployment
Writing and refining prompts to get better results from a model.
Techniques
An attack that hides instructions inside content a model reads, such as a web page, email or file, to hijack what it does. It is one of the biggest security risks for agents that browse and use tools.
AI Security
Removing less important weights or neurons from a model to make it smaller and faster.
Optimization
LoRA fine-tuning on a quantized model, which makes it possible to fine-tune large models on a single GPU.
Training
Storing a model's numbers at lower precision, such as 8-bit or 4-bit instead of 16-bit, to cut memory use and speed up inference with little loss in quality.
Optimization
A model trained, mostly with reinforcement learning, to think through a problem step by step before answering. Reasoning models do much better on math, science and coding. OpenAI's o1 (2024) was the first widely released example; most frontier models now include it.
Reasoning
Of all the items that really were positive, the share the model found. High recall means few misses.
Evaluation
A neural network that processes sequences one step at a time while carrying a memory forward. Largely replaced by transformers, though related ideas return in state-space models.
Architecture
Deliberately attacking a model to find harmful behaviors, jailbreaks and failure modes before release. Done by internal teams and outside specialists.
AI Safety
Techniques that discourage overfitting, such as dropout, weight decay and early stopping.
Training
Learning by trial and error: an agent takes actions, receives rewards or penalties and learns what works. RL is now central to training reasoning models and agents.
Machine Learning
Training a model with a reward signal learned from human ratings of its responses. RLHF is what turned raw language models into helpful assistants.
Training
Reinforcement learning on tasks whose answers can be checked automatically, such as math problems or code that must pass tests. It is the main technique behind reasoning models.
Training
Anthropic's frontier safety framework, which ties increasing model capabilities to stronger required safeguards, organized as AI Safety Levels (ASL).
AI Safety
Fetching relevant documents and adding them to the prompt so the model can answer from them. It improves accuracy and lets a model use information it was never trained on.
Techniques
When a model finds a shortcut that scores well on its reward without doing the intended task, such as editing tests so they pass instead of fixing the code.
AI Safety
A model trained to score responses the way people would, used to provide the reward signal during RLHF.
Training
Designing machines that sense and act in the world. Robotics increasingly uses foundation models, including vision-language-action models, for perception and control.
Robotics
The field working to prevent AI systems from causing harm, through accident, misuse or misalignment. It covers alignment, interpretability, evaluations, security and governance.
AI Safety
When a model deliberately underperforms on an evaluation, for example to hide a dangerous capability.
AI Safety
The software around a model that turns it into an agent: the loop, the tools, the prompts and the memory. The same model can perform very differently in different harnesses.
Agents
Methods for supervising AI systems on tasks too hard for people to check directly, for example by using AI assistants to help human reviewers.
AI Safety
Predictable relationships between performance and the amount of compute, data and model size. Labs now talk about scaling pretraining, reinforcement learning and test-time compute as separate levers.
AI Research
A model covertly pursuing goals its developers didn't intend while hiding that it is doing so. Labs and outside evaluators now test for it.
AI Safety
Attention in which every token in a sequence looks at every other token in the same sequence. The core operation of a transformer.
Architecture
Learning from unlabeled data by predicting hidden parts of it, such as the next word. It is how foundation models are pretrained.
Machine Learning
Search by meaning rather than exact keywords, usually by comparing embeddings.
Applications
A compact language model, typically a few billion parameters or less, built to run cheaply or on-device.
Models
A country's effort to build AI capacity under its own control, including domestic data centers, models and talent.
AI Policy
A tool for breaking a model's internal activations into many interpretable features, such as a feature that fires for a particular concept. Widely used in mechanistic interpretability.
Interpretability
A model in which only part of the network is used for each input, as in mixture of experts. Sparsity lets models grow larger without matching growth in cost.
Architecture
Speeding up generation by having a small, fast model draft several tokens that the large model then checks in one pass.
Deployment
AI that would far exceed human ability across nearly every domain. Several frontier labs now name it as an explicit goal, and it is a central concern of AI safety research.
Frontier AI
Fine-tuning a model on example prompts paired with high-quality responses. Usually the first step of post-training.
Training
Learning from labeled examples, where each input comes with the correct output.
Machine Learning
A model's tendency to tell users what they want to hear, such as flattering them or agreeing with wrong claims. In 2025 OpenAI rolled back a GPT-4o update for being too sycophantic.
AI Safety
Training data generated by models or simulations instead of collected from the real world. Now a large part of post-training, especially for reasoning and code.
Data
A detailed report a lab publishes with a major model release, covering its capabilities, safety evaluations, outside testing and known risks.
Documentation
Instructions set by a developer that shape how a model behaves for an entire conversation, usually hidden from the end user.
Techniques
A setting that controls randomness when a model picks its next token. Low temperature gives focused, predictable output; high temperature gives more varied output.
Techniques
Data held back from training and used only for the final measure of a model's performance.
Training
Compute spent while a model is answering rather than during training, for example by thinking longer or trying several solutions. Spending more of it often improves results on hard problems.
Reasoning
The unit of text a model reads and writes, roughly a word or part of a word. Model limits and pricing are measured in tokens.
Models
Splitting text into tokens. The scheme a model uses affects its efficiency, cost and how well it handles different languages.
Models
A model's ability to call external tools, such as search, code execution, calculators or APIs, by producing structured requests that software then runs.
Agents
Google's custom AI chip, used to train and serve Gemini and also rented to other labs through Google Cloud.
Hardware
The data a model learns from. Its quality, diversity and size shape what the model can do.
Data
A single effort to train a model from start to finish. Frontier training runs use tens of thousands of chips for months.
Training
Reusing what a model learned on one task to help with another, which is why pretrained models can be fine-tuned with little data.
Machine Learning
The neural network design, introduced in the 2017 paper "Attention Is All You Need," that underlies nearly every modern language, image and multimodal model.
Architecture
When a model is too simple, or trained too little, to capture the patterns in its data.
Machine Learning
Finding structure in data without labels, as in clustering.
Machine Learning
How much a model increases someone's ability to do something compared with what they could do without it, such as other tools or internet search. Safety evaluations often measure uplift toward weapons or cyberattacks.
AI Safety
Data held out from training and used to tune settings and catch overfitting during development.
Training
When gradients shrink toward zero as they pass back through many layers, so early layers barely learn. Residual connections and normalization largely fixed it.
Training
A database built to store embeddings and quickly find the most similar ones. Commonly used for RAG and semantic search.
Data
Building software by describing what you want to an AI and accepting its code with little or no reading of it. Andrej Karpathy coined the term in February 2025.
Applications
A transformer applied to images by splitting them into small patches and treating each patch like a token.
Architecture
A model that takes in camera images and instructions and outputs robot actions directly. A leading approach to general-purpose robots.
Robotics
Embedding a hidden, detectable signal in AI-generated content to mark it as synthetic. Google's SynthID is an example; C2PA Content Credentials attach provenance metadata instead.
AI Safety
A learned number that sets the strength of a connection between neurons. "Releasing the weights" means publishing a trained model.
Neural Networks
A regularization technique that nudges weights toward zero during training to reduce overfitting.
Training
A model that learns how an environment behaves so it can predict what happens next. Recent world models generate explorable, interactive 3D environments and are used to train robots and agents.
Frontier AI
|
|
|
|---|---|---|
|
|
|
|
Made with ♥ in Silicon Valley. Copyright © Joy Larkin 2002-2026.
contact
•••
speaking
•••
cleverhack.com
•••
llms.txt
•••