Essential terminology for understanding AI progress and the path to superintelligence
AI systems that can match human cognitive abilities across virtually any intellectual domain.
Unlike narrow AI that excels at specific tasks, AGI would possess the flexibility and general reasoning capabilities of the human mind. It could learn new skills, transfer knowledge between domains, and adapt to novel situations without task-specific training. Major AI labs including OpenAI, Anthropic, and Google DeepMind are actively working toward this goal.
AI systems that significantly exceed the cognitive capabilities of the best human minds in every domain.
ASI represents the theoretical endpoint of AI development - systems that surpass human-level performance not just in narrow tasks but in scientific creativity, general wisdom, social skills, and all other cognitive domains. This is what "midnight" symbolizes on the MidnightAI clock.
Intelligence that greatly surpasses the cognitive performance of humans in virtually all domains of interest.
The term was popularized by philosopher Nick Bostrom and encompasses various theoretical paths including speed superintelligence (faster processing), collective superintelligence (networked AI systems), and quality superintelligence (superior reasoning). The development of superintelligence is considered one of the most consequential events in human history.
The field of research focused on ensuring AI systems behave as intended and remain beneficial to humanity.
AI safety encompasses technical research on alignment, robustness, and interpretability, as well as governance and policy considerations. Key challenges include ensuring AI systems pursue intended goals, remain controllable, and do not develop unintended harmful behaviors as they become more capable.
The technical challenge of ensuring AI systems act in accordance with human values and intentions.
Alignment research addresses how to specify what we want AI to do, how to ensure it learns and follows those specifications, and how to maintain alignment as systems become more capable. Current techniques include RLHF, Constitutional AI, and various interpretability approaches, but robust solutions for highly capable systems remain an open problem.
Risks that could cause human extinction or permanently curtail humanity's potential.
In the context of AI, existential risk refers to scenarios where advanced AI systems could pose catastrophic threats to humanity. Leading AI researchers including Geoffrey Hinton, Yoshua Bengio, and executives at major AI labs have expressed concern about AI-related existential risks, though probability estimates vary widely.
Large AI models trained on broad data that can be adapted to a wide range of downstream tasks.
Foundation models like GPT-4, Claude, and Gemini are trained on massive datasets and develop general capabilities that transfer across tasks. They represent a paradigm shift from task-specific AI to more general systems, and their capabilities often emerge unpredictably at scale.
AI models trained on vast amounts of text to understand and generate human language.
LLMs like GPT-4, Claude, and Gemini use transformer architectures and are trained to predict the next token in a sequence. Despite the "language" name, modern LLMs demonstrate capabilities in reasoning, coding, mathematics, and other domains. They form the backbone of current frontier AI systems.
Abilities that appear in AI systems as they scale, which were not explicitly trained for or predicted.
As models grow larger, they often develop unexpected capabilities like chain-of-thought reasoning, in-context learning, or the ability to solve novel problem types. These emergent properties make it difficult to predict what future systems will be able to do, adding uncertainty to AI development trajectories.
Empirical relationships describing how AI performance improves with increased compute, data, and model size.
Research has shown predictable power-law relationships between these factors and model capabilities. Scaling laws have driven massive investments in AI infrastructure and inform predictions about when AGI might be achieved. However, whether scaling alone leads to AGI remains debated.
A technique for training AI models using human preferences to guide behavior.
RLHF involves training a reward model on human comparisons of AI outputs, then using that reward model to fine-tune the AI. This approach has been crucial in making models like ChatGPT helpful and safe, but has limitations in capturing complex human values and may not scale to superintelligent systems.
An alignment approach where AI systems follow explicit principles rather than learning from human labels.
Developed by Anthropic, Constitutional AI involves defining a set of principles (a "constitution") that the AI should follow. The model is then trained to critique and revise its own outputs according to these principles, reducing reliance on human feedback for each decision.
The ability to think logically, solve problems, and draw valid conclusions from available information.
Reasoning is weighted at 25% in our capability assessment - the highest weight. It encompasses mathematical reasoning, logical deduction, causal reasoning, and abstract problem-solving. Current frontier models show strong reasoning capabilities on many benchmarks but still struggle with novel or highly complex reasoning tasks.
The capacity for autonomous action, planning, and goal-directed behavior in the world.
Agency is weighted at 20% in our assessment. It includes the ability to break down goals into subtasks, use tools, interact with external systems, and operate autonomously over extended periods. Agentic AI systems are a major focus of current research and raise significant safety considerations.
The ability to write, understand, debug, and reason about computer code.
Coding is weighted at 15% in our assessment. Modern AI systems have achieved remarkable coding capabilities, often matching or exceeding human programmers on standard benchmarks. This capability is particularly significant as it enables AI systems to potentially improve themselves or build other AI systems.
The ability to process and generate content across multiple modalities like text, images, audio, and video.
Multimodal capability is weighted at 10% in our assessment. Modern frontier models increasingly handle multiple input and output modalities, moving beyond text-only systems to understand images, generate audio, and work with video content.
The ability to understand scientific concepts, generate hypotheses, and contribute to research.
Scientific capability is weighted at 10% in our assessment. This includes understanding research papers, generating novel hypotheses, designing experiments, and potentially accelerating scientific discovery. AI systems are increasingly being used as research assistants and may eventually make independent scientific contributions.
The ability to perceive, understand, and interact with the physical world through embodied systems.
Robotics capability is weighted at 10% in our assessment. While AI has made remarkable progress in digital domains, physical world interaction remains challenging. Advances in robotics could enable AI systems to directly manipulate the physical world, with significant implications for automation and AI safety.
A visual metaphor showing how close we are to superintelligent AI, inspired by the Bulletin of the Atomic Scientists.
The MidnightAI Doomsday Clock represents humanity's proximity to artificial superintelligence. Midnight symbolizes the arrival of ASI. The clock position is calculated using multi-model consensus analysis of current AI capabilities, research breakthroughs, and expert predictions.
The current estimated distance from superintelligent AI as displayed on the MidnightAI clock.
This metric synthesizes multiple factors including current AI capabilities across seven domains, rate of improvement, technical breakthroughs, resource investment, and expert forecasts. The number of minutes reflects our assessment of how close we are to ASI, with lower numbers indicating closer proximity.
Standardized tests used to measure AI performance across specific domains.
Benchmarks like MMLU (general knowledge), HumanEval (coding), GSM8K (math), and many others provide quantitative measures of AI capability. MidnightAI tracks performance across benchmarks to assess progress in our seven capability domains.
Organizations at the frontier of AI research and development.
MidnightAI tracks 14 major AI labs including OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, Microsoft (USA); DeepSeek, Alibaba, ByteDance, Baidu (China); Mistral AI, Aleph Alpha (Europe); and Cohere, AI21 Labs (Global). These organizations drive the advancement of AI capabilities.
Want to learn more about how we track AI progress? Read our methodology or learn about MidnightAI.
