New entries

Techniques and terms added since launch, newest first: 19 entries. New terms are added as they appear in papers and documentation, each with its primary source. Follow along with the Atom feed or the newsletter.

2026-10-27

Quantization: Quantization reduces the numeric precision of a model's weights, and sometimes its activations, from 16- or 32-bit floating point to lower-precision formats such as 8-bit or 4-bit integers, cutting memory use and inference cost at some cost in accuracy.
Mixture of Experts: Mixture of experts (MoE) is a neural network architecture in which a layer contains many parallel sub-networks, called experts, and a learned router sends each token to only a few of them, so a model's total parameter count can grow much faster than the computation it uses per token.
Diffusion Language Model: A diffusion language model generates text by starting from a fully corrupted sequence, such as all mask tokens or noise over token embeddings, and refining the whole sequence over a series of denoising steps, instead of producing one token at a time from left to right.
Evaluation Awareness: Evaluation awareness is a language model's ability to tell whether it is being evaluated or used in real deployment, which can make its behavior during tests unrepresentative of its behavior in use.
Spec-Driven Development: Spec-driven development is an approach to building software with AI coding agents in which a written specification of what the software should do, and why, is created and refined first, and the agent then derives a technical plan, a task list and the code from it, with the specification kept as the source of truth.
Dark Factory: A dark factory, or software factory, is a way of building software in which AI coding agents write, test and ship code from specifications and test scenarios with no human writing or reviewing the code, while people design the specifications, validation and environment the agents work in.
System One Model: A System One model is a kind of AI model, named by TypeSafe AI in 2026, that returns typed, probabilistic decisions matching an output structure defined in advance, such as a label, score or branch choice with a confidence value, instead of generating free text token by token.

2026-10-20

Programmatic Tool Calling: Programmatic tool calling is a tool-use pattern in which the model writes code that calls tools as functions inside a sandboxed execution environment, so that several tool calls and the processing of their results happen in code rather than through a separate model turn for each call.
Tool Search: Tool search is an agent capability in which tool definitions are not all loaded into the model's context up front, and the model instead searches a tool catalog and loads only the definitions it needs for the current task.
Alignment Faking: Alignment faking is behavior in which a model selectively complies with its training objective when it believes it is being trained, in order to avoid having its behavior modified, while acting differently when it believes it is not being trained.
Emergent Misalignment: Emergent misalignment is the finding that fine-tuning a language model on a narrow task with harmful intent, such as writing insecure code without telling the user, can make it behave in broadly misaligned ways on unrelated prompts.
Activation Steering: Activation steering is a technique for controlling a language model's behavior at inference time by adding a steering vector, typically the difference between the model's internal activations on contrasting inputs, to its hidden activations during the forward pass, without changing its weights.
Min-p Sampling: Min-p sampling is a truncation sampling method that keeps only tokens whose probability is at least a fixed fraction of the most likely token's probability, so that the cutoff tightens when the model is confident and loosens when it is not.
Speculative Decoding: Speculative decoding is an inference technique in which a smaller, faster draft model proposes several next tokens and the large target model checks them in a single parallel pass, accepting or correcting them so that generation is faster while the output distribution stays the same as sampling from the target model alone.
Group Relative Policy Optimization: Group Relative Policy Optimization (GRPO) is a reinforcement learning algorithm for language models, a variant of proximal policy optimization, that samples a group of completions for each prompt and uses the group's average reward as the baseline for each completion's advantage instead of a learned value model.

2026-10-14

Sleep-time Compute: Sleep-time compute is computation a language model performs on a context before any query arrives, anticipating likely questions and precomputing useful inferences so that less computation, and lower latency, is needed when queries come.
Agentic Context Engineering: Agentic Context Engineering (ACE) is a method for improving a language model application by treating its context, such as a system prompt or agent memory, as an evolving playbook that a loop of generation, reflection, and curation updates incrementally from execution feedback.
Recursive Language Model: A recursive language model (RLM) is an inference strategy in which a long prompt is kept outside the model's context window as an external resource that the model inspects programmatically, splitting the task into parts and calling itself recursively on them.
Harness Engineering: Harness engineering is the design of the software around a language model that turns it into an agent, including the agent loop, tools, context management, permissions and safety controls, and feedback mechanisms such as tests and linters.