Artificial general intelligence and artificial superintelligence are often discussed as if they are simply two stages of increasingly powerful AI. The difference is more important than that.
In its From AGI to ASI report, Google DeepMind uses AGI to describe AI with roughly human-level general intelligence. ASI sets a much higher bar: intelligence that broadly exceeds humans and can outperform even large groups of experts across virtually all important domains.
That changes the question. Reaching AGI would mean machines have crossed a human benchmark. Reaching ASI would mean that human intelligence may no longer be the useful benchmark at all.
AGI and ASI Define Two Separate Levels of Intelligence
The simplest way to understand the distinction is to look at who or what the AI is being compared against.
| AGI | ASI | |
|---|---|---|
| Basic idea | Human-level general intelligence | Broadly superhuman general intelligence |
| Reference point | Roughly an individual human | Large groups of human experts |
| Capability | Performs across many cognitive tasks | Exceeds humans across virtually all important domains |
| Can it be superhuman at some tasks? | Yes | Yes, but much more broadly |
| Must it be a single AI? | Not necessarily | No |
| Does it know everything? | No | No |
This distinction also explains why systems such as AlphaGo or AlphaFold are not ASI under DeepMind’s framing. They can outperform humans in specific domains, but superhuman performance in one area is different from superhuman general intelligence.
Understanding What AGI Means
AGI does not mean an AI must be better than every person at every task.
DeepMind uses a practical working definition: a system performing at roughly median human level across most cognitive tasks. A future AGI could still have an uneven capability profile. It might already outperform humans dramatically in areas such as memory, calculation or information processing while remaining closer to human ability elsewhere.
A simple way to picture it is a general-purpose digital worker.
Instead of being designed for one narrow task, it could potentially move between research, analysis, planning, coding, communication and other cognitive work with broad human-level competence.
Current AI systems can already exceed humans in selected areas, but the report argues that they are not yet general enough to meet this AGI definition.
Understanding the ASI Benchmark
Now change the comparison. Instead of asking whether an AI can perform like one capable person, imagine asking whether it can outperform an entire research institution, engineering organization or highly coordinated group of experts.
That is much closer to the report’s ASI threshold.
DeepMind describes ASI as broadly superhuman intelligence capable of exceeding the performance of large human-expert collectives, not simply an AI that scores slightly higher than a person on existing benchmarks.
The difference can therefore be compressed into one idea:
AGI approaches the capability of a human. ASI moves beyond the capability of human organizations.
But this raises a bigger question: why would AI continue advancing after reaching human-level intelligence?
How Digital Systems Expand Intelligence at Scale
The answer is not simply that future models might become “smarter.”
Digital intelligence has properties biological intelligence does not easily have.
DeepMind highlights several differences that can become increasingly important as computing resources grow:
- Speed: digital systems can process and exchange information at much higher bandwidth.
- Memory: machine working memory and storage can scale far beyond biological limits.
- Replication: software and its memory state can potentially be copied.
- Hardware flexibility: an AI can potentially move to newer or more powerful computing infrastructure.
- Parallelism: many instances can operate at the same time.
- Knowledge sharing: digital experiences and learning can be stored, replayed and shared between systems.
Consider an excellent human researcher.
You cannot instantly create 10,000 identical copies of that person, preserve all of their memories, let the copies work simultaneously and then transfer what one learned directly to the others.
With digital intelligence, versions of those operations become technically conceivable.
That is why the jump from AGI to ASI may involve more than improving the intelligence of one model.
How Millions of AGIs Could Build a Superintelligent System
This is one of the most interesting ideas in the report.
ASI does not necessarily need to be one extraordinary machine.
Imagine human-level AGIs being replicated into thousands or millions of instances. Some specialize in mathematics, others in engineering, biology, software, planning or experimentation. They work simultaneously, communicate digitally and divide enormous problems into smaller tasks.
The result could resemble a digital research organization or company operating at a scale and speed that human institutions cannot match.
DeepMind explores both centrally coordinated AI organizations and more decentralized systems resembling markets or networks. It also raises an unresolved research question: whether collective AI capability could follow its own multi-agent scaling laws, where intelligence grows as more agents and stronger coordination are added.
This leads to an important shift in how ASI should be imagined.
It may not look like:
one machine → becomes unimaginably intelligent.
It could look more like:
human-level AGI → many AGIs → specialization and coordination → superhuman collective capability.
Understanding the Boundaries of ASI
Superintelligence should not be confused with unlimited intelligence.
The report explicitly argues that ASI would be neither omniscient nor omnipotent.
Even an extremely capable system would operate inside the physical world. Information cannot travel infinitely fast. Experiments take time. Manufacturing requires materials and energy. Some information cannot be perfectly observed. Computational complexity and logical limits do not disappear simply because a system becomes more intelligent.
A superintelligent system might solve problems far beyond human ability while still being unable to instantly solve every scientific, economic or physical problem.
That distinction matters because discussions about ASI often move too quickly from “far more intelligent than humans” to “capable of anything.”
Those are not the same claim.
Why This Intelligence Gap Matters
AGI would be a historic threshold because machines would have reached broadly human-level cognitive capability.
ASI would represent something different: the possibility that intelligence itself could begin operating beyond the scale of individual humans and even beyond the organizations humans use to combine expertise.
That is why the AGI-versus-ASI distinction matters more than terminology.
AGI matches the human benchmark. ASI changes the benchmark itself.
And if AGI is eventually reached, the next question will not simply be whether AI becomes smarter.
It will be how digital intelligence could move from human-level capability to something humans have never had before: broadly superhuman intelligence at organizational scale.
