01 / From AI to SI
SI stands for superintelligence. When discussing artificial systems, ASI is also commonly used. It describes a system that substantially exceeds human capabilities across a broad range of cognitive domains. Both depth and breadth matter.
Beating humans at a board game, a math problem, or a particular test demonstrates capability under those conditions. A claim of general superintelligence also requires examining performance across tasks and domains.
Artificial intelligence
A broad range of technologies for intelligent tasks, without assuming a particular degree of generality.
Artificial general intelligence
Emphasizes capability and transfer across many tasks. Definitions of the threshold still differ.
Superintelligence
Goes further, emphasizing broadly and substantially superhuman capabilities.
Researchers describe intelligent systems through capability depth, generality, and autonomy.
A framework for levels of AGI (opens in a new tab)02 / Intelligence has more than one dimension
To understand a system, start with three separate questions. Conflating them makes it easy to mistake progress in one area for progress in another.
Capability depth: how well does it perform?
Interpreting performance requires clear test conditions, scoring methods, and human baselines. Results from different difficulty levels or resource budgets are not directly comparable.
Generality: how much can it do?
Excelling at familiar problem types is different from learning, transferring knowledge, and solving problems in unfamiliar domains. Generality requires looking beyond a single test.
Autonomy: how far can it go on its own?
A system can be highly capable under close human guidance, or be allowed to operate independently for long periods. Autonomy, capability, and safety need to be analyzed separately.
A framework for levels of AGI (opens in a new tab)Measuring general cognitive abilities (opens in a new tab)03 / How do we measure progress?
Capability evaluations examine the range of tasks a system can perform, the consistency of its results, and how well it transfers knowledge to new situations.
- Coverage: Does the evaluation span different cognitive abilities, rather than only the tasks the system does best?
- Comparisons: Are human baselines, resource conditions, and test materials not used in training clearly described?
- Consistency: Are repeated trials, failure cases, and performance in unfamiliar environments reported?
- Long-horizon performance: Can the system handle errors, maintain constraints, and remain open to monitoring and intervention while pursuing a goal?
- Reproducibility: Are methods, experimental conditions, and results described clearly enough for other researchers to repeat the tests?
04 / Potential and risk belong together
Why does it matter?
More capable AI could help advance scientific discovery, complex problem-solving, and knowledge work. It could also change how organizations work and allocate resources.
Principles and a mission for AGI (opens in a new tab)Developing capability and control together
Systems may misunderstand goals, exceed authorized boundaries, or be used for harmful purposes. Alignment research, permission limits, monitoring, and intervention need to be tested against specific threats and conditions of use.
Safety research brings capability evaluation, ongoing monitoring, and risk mitigation into model development and deployment.
Agent control and defense in depth (opens in a new tab)August 2026 Risk Report (opens in a new tab)05 / Common questions
What is the difference between AGI and ASI?
AGI generally emphasizes general capability across a broad range of tasks. ASI goes further, emphasizing broadly and substantially superhuman abilities. The precise thresholds remain debated.A framework for levels of AGI (opens in a new tab)
What do benchmarks measure?
Benchmarks measure performance on defined sets of tasks. Cross-task evaluations, human baselines, repeated trials, and real-world testing help researchers understand the capabilities of a system.Measuring general cognitive abilities (opens in a new tab)
What is safe superintelligence research?
It brings capability research together with alignment, monitoring, and permission controls, exploring how more capable systems can follow human intentions and remain controllable.Safety and alignment for long-horizon models (opens in a new tab)Agent control and defense in depth (opens in a new tab)
What shapes the path to superintelligence?
The growth of capabilities, the evolution of technical approaches, and how researchers define and measure progress all shape the path ahead.