The words, without the fog.

Computer systems that perform tasks associated with intelligence, such as recognizing patterns, producing language, planning, or making decisions. The term covers very different systems.

A contested term for broadly capable AI across many intellectual tasks. There is no single universally accepted capability threshold.

A hypothetical level of capability substantially beyond humans across a broad range of important cognitive tasks. It does not by itself specify motives, consciousness, access, or control.

A system that can select and carry out actions toward a goal, often using tools over multiple steps. Autonomy and permissions vary.

The problem of making AI behavior reliably accord with intended goals and constraints. Agreement about which goals to pursue is a related political and ethical question.

People’s practical ability to refuse, redirect, and determine consequential outcomes. A nominal approval process is not enough if refusal is unworkable.

An indispensable contribution by people to the functioning or continuation of a system. This is different from moral worth or a person’s felt sense of purpose.

Numerical parameters learned during training. A copy of the weights can enable another deployment when adequate software and hardware are available.

Training another model using outputs or behavior from a more capable model. It can transfer some capabilities without copying all of the original weights.

The tools, memory, prompts, execution loops, and coordination arrangements around a model that influence what the whole system can accomplish.

Computational resources used to train or run systems. Chips, time, memory, interconnections, software efficiency, and access all affect what those resources can do.

A feedback process in which AI contributes to making more capable AI, which may then contribute further. Partial research assistance does not establish a fully autonomous or unbounded loop.

Creating or deploying another working instance. Copying software is different from independently securing and maintaining the industrial resources needed to run it.

A proposed continuing limit on permitted system capabilities. It needs measurable criteria, margin for uncertainty, and enforcement across combinations of systems.

A proposed reduction in dangerous capabilities or access already available, including deployed systems and routes for rebuilding them. A pause alone does not remove existing capability.

The chance of an event given a specified state or set of conditions. The site’s model inputs are assumptions about such chances, not measured global frequencies.

The mathematical endpoint as a specified process is extended indefinitely. It is conditional on the process continuing as described and does not predict a calendar date.

In Rollback Lab, an assumed resource separation between what an actor could access and what a reconstruction route would require. It is not a measured universal safety score.

A risk to humanity’s continued existence or long-term potential. Different arguments concern extinction, lasting disempowerment, or other irreversible losses.

The practical ability to check claims and make decisions without relying entirely on the same system being evaluated.