AI vs AGI vs SI vs ASI: Definitions and Differences

#The four terms

The terms answer three different questions, and mixing them is the main source of confusion:

• How capable is the system? (depth of performance)
• How general is it? (breadth across tasks and domains)
• How autonomous is it? (how much it does without human direction)

AI to AGI is mainly a jump in breadth. AGI to SI is mainly a jump in depth. Autonomy is a separate axis. Consciousness is a fourth, separate question: none of these definitions require a system to have subjective experience.

#SI and ASI are not two rungs on one ladder

Bostrom's definition is about capability and says nothing about what the intellect is made of. A superintelligence could in principle be artificial, a biologically enhanced human, or a coordinated collective. ASI is the artificial case. So ASI is a kind of SI, not a stronger level above it. In everyday use, "superintelligence" almost always means the artificial kind, so the two terms are often used interchangeably.

#Why AGI has no single definition

The definition above is a working one, not a consensus. Stanford HAI notes that the term is controversial because people mean different things by "human-level intelligence", and that there is no universally accepted test, so claims are hard to verify. Three approaches recur:

1. Capability-based: does the system perform a sufficiently broad range of cognitive tasks?
2. Human-equivalence-based: does it match human performance across cognitive domains?
3. Learning-based: can it acquire new skills and solve unfamiliar problems efficiently?

A system can satisfy one and fail another. Google DeepMind researchers proposed splitting the question rather than answering it with one label. Their paper, Levels of AGI (Morris et al., 2023; ICML 2024), classifies systems by the depth (performance) and breadth (generality) of their capabilities and ties the levels to deployment questions such as autonomy and risk. The practical lesson: ask "how good, at what, and how independently?" instead of "is it AGI?"

Common proposed ingredients of general intelligence include reasoning, learning new skills, transferring knowledge across domains, long-horizon planning, memory, adapting to unfamiliar situations, and checking and correcting one's own errors. Researchers disagree on which are required. Continual learning, for instance, is an important proposed criterion, not an agreed prerequisite.

#What would make an intelligence superhuman?

Superintelligence is usually described along several dimensions that need not move together:

• Quality: better reasoning and judgement than any human expert.
• Speed: the same work done much faster.
• Scale: many copies working in parallel.
• Integration: combining knowledge across fields that no single person can hold.
• Creativity and strategy: producing new science, technology and plans.

Bostrom's often-cited speed argument is that neurons fire at a peak of roughly 200 Hz while processors run at gigahertz rates. That compares one narrow physical quantity. Firing rate is not computation speed, and the brain gets much of its capability from massive parallelism, so treat it as an illustration of hardware headroom, not a measure of intelligence.

#How AGI might become ASI

This is hypothesis, not established mechanism. A June 2026 Google DeepMind report, From AGI to ASI (Genewein et al., arXiv, 10 June 2026), names four candidate pathways and says the transition may be less a single step change than a series of changes across many fields:

1. Scaling AGI: more compute, data and model size.
2. AI paradigm shifts: new architectures or methods.
3. Recursive improvement: AI contributing to its own development.
4. Multi-agent collectives: large numbers of coordinating agents producing superhuman capability as a group.

The abstract also points to frictions and bottlenecks along each path and says their size is an open research question. Secondary coverage lists hardware supply, energy and data limits; I have not verified that list against the full paper. The intelligence explosion idea, from I.J. Good's 1965 argument, is the recursive path in its strongest form: an ultra-intelligent machine designs better machines. It is a thought experiment. Whether each cycle would actually be faster, or would hit diminishing returns, is unsettled. Improving one's own research output is also different from rewriting one's own code.

#Examples

Today's general-purpose models do not fit the old "narrow AI" label well, since one model handles many kinds of tasks. Whether that breadth amounts to AGI is exactly what disagreement over the definition is about.

#Misconceptions

• "Current AI is just statistical pattern matching." Models are trained by optimisation on data, but that description undersells what the learned systems can do and says little about their limits. Better to name the limits directly: reliability, grounding, and long-horizon autonomy.
• "Model weights never change." Deployed models usually have fixed weights, but they can adapt in context, use external memory and tools, and be retrained later. These are different mechanisms from learning continually.
• "Scale alone gets us to AGI." One of several live hypotheses. DeepMind lists scaling beside paradigm shifts, recursion and collectives.
• "AGI or ASI needs consciousness." Intelligence is functional capability. Consciousness is a separate, unresolved question.
• "ASI would invent new physics." A hypothetical capability, not a guaranteed outcome.

#Risks and safety

Risk is tied to capability and autonomy, not to the label. Alignment (getting systems to pursue intended goals), control (retaining oversight of capable systems) and governance are active research and policy areas. The claims that capable agents would resist shutdown or seek resources (instrumental convergence) are hypotheses, not demonstrated facts. See the encyclopedia's safety chapters for the evidence, including ASI control and alignment and existential risk and long-term futures.

#History in brief

• 1950: Alan Turing publishes Computing Machinery and Intelligence and proposes the imitation game.
• 1956: The Dartmouth workshop coins the term "artificial intelligence".
• 1965: I.J. Good describes an "intelligence explosion".
• 1997–98: Bostrom writes "How Long Before Superintelligence?" (published 1998). Mark Gubrud uses "artificial general intelligence" in 1997.
• 2000s: Marcus Hutter's AIXI gives a formal model of universal AI. Shane Legg and Ben Goertzel popularise "AGI" around 2002.
• 2014: Bostrom publishes Superintelligence: Paths, Dangers, Strategies.
• 2023–2026: General-purpose models make AGI a corporate and policy topic; DeepMind publishes Levels of AGI (2023) and From AGI to ASI (2026).

#Frequently asked questions

#What is the difference between AI, AGI, SI and ASI?

AI is the whole field. AGI is hypothetical AI with general, human-level ability. SI is any intellect far beyond the best humans, and ASI is the machine version. The key difference is breadth and level of capability. No agreed test marks AGI, and no ASI exists.

#Is AGI the same as superintelligence?

No. AGI is defined at roughly human level and across many domains. Superintelligence goes well beyond the best humans. A system could reach AGI without becoming superintelligent.

#Is ASI more advanced than AGI?

By definition, yes: ASI exceeds human performance in practically every field, while AGI is pegged near human level. Whether AGI leads to ASI, and how fast, is unresolved.

#Does AGI exist today?

There is no consensus that it does. Current models are broad but unreliable on long, open-ended tasks, and with no agreed test, any AGI claim depends on the definition used.

#How could AGI become ASI?

DeepMind's 2026 report names four possible paths: scaling, new paradigms, recursive improvement and multi-agent collectives. They are not mutually exclusive, and bottlenecks such as compute, energy and data could slow any of them.

#What is an intelligence explosion?

A hypothetical feedback loop in which an AI improves the process that produces AI, so each generation arrives faster. I.J. Good described it in 1965. It is a hypothesis, not an observed phenomenon.

#What are the risks of artificial superintelligence?

The main concerns are misaligned goals, loss of human control, and concentration of power. These are studied as hypotheses and scenarios, since no ASI exists to test them on.

#Sources

• Stanford HAI: What is AGI? - Working definition, disagreement over "human-level", and the absence of an agreed test.
• Morris et al., Levels of AGI for Operationalizing Progress on the Path to AGI - Google DeepMind; performance and generality levels (arXiv:2311.02462, ICML 2024).
• Genewein et al., From AGI to ASI - Google DeepMind; four pathways and bottlenecks (arXiv:2606.12683, June 2026).
• Bostrom, How Long Before Superintelligence? - Implementation-independent definition of superintelligence (1998).
• Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
• Good, I.J. (1966). "Speculations Concerning the First Ultraintelligent Machine." Advances in Computers 6.
• Turing, A. (1950). "Computing Machinery and Intelligence." Mind 59.

#Bottom line

Treat the four terms as answers to different questions: AI is the field, AGI is about breadth, SI is about level, and ASI is SI built from machines. When someone says "AGI", ask which definition they mean and how it would be tested.