ASI stands for artificial superintelligence — a system that doesn't just match the best human experts, but surpasses them across every cognitive domain simultaneously. It's the stage after AGI. And the researchers building toward AGI increasingly believe the two may arrive closer together than the decades-long gap most people assume.
Artificial superintelligence is an AI system that exceeds the best humans in every cognitive domain: mathematics, science, engineering, creative writing, social reasoning, strategy, and any other mental task. That's the clean definition. The messy reality is that "exceeds the best humans" is doing a lot of work.
Nick Bostrom, who popularized the term in his 2014 book Superintelligence, defined it as "an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills." That framing has stuck — but it contains an important ambiguity: smarter by how much?
Researchers distinguish three rough variants:
Most AI safety researchers treat ASI as encompassing all three — a system that is faster, larger in scale, and fundamentally better at reasoning than any human or group of humans. That's the scenario they're designing safeguards for.
The cleanest way to see where ASI fits is to look at the full capability ladder. We're somewhere on it right now — the question is how far and how fast:
The ladder matters because the jump from step 3 to step 4 may not require much time at all. An AGI-level system capable of doing AI research at human level could potentially compress years of progress into months — a feedback loop researchers call recursive self-improvement. That's the core of why many people who study this are deeply focused on alignment now, before AGI, rather than waiting for ASI to arrive.
Dario Amodei (CEO of Anthropic) sketched the clearest near-term picture in his 2024 essay "Machines of Loving Grace." He described what he called a "compressed 21st century" — a world where ASI-adjacent AI compresses decades of scientific and societal progress into a few years. His list:
| Domain | What ASI-adjacent AI could do | Timescale (from AGI) |
|---|---|---|
| Biology & medicine | Defeat most cancers, design effective treatments for mental illness, dramatically compress drug discovery timelines | 5–10 years |
| Mental health | Develop solutions to the global mental health crisis currently lacking effective treatments | 5–10 years |
| Economic development | Accelerate development in poor countries, potentially compressing decades of economic growth | 10–20 years |
| Physics & energy | Accelerate nuclear fusion, battery technology, and climate solutions | 5–15 years |
| AI research itself | Design better AI systems faster than humans can — the critical recursive step | Immediate |
These projections come from someone building toward this technology — so take the optimism with appropriate skepticism. But the underlying logic is structural: a system that outperforms the best researchers in any field could apply that capability everywhere at once. The bottleneck isn't knowledge, it's cognitive bandwidth. ASI removes that bottleneck.
This is where forecasters get genuinely uncertain. ASI timelines are always measured relative to AGI — and whether AGI has already arrived is itself contested. Here's how the major forecasters stack up on both milestones:
The key insight from these estimates: even the most conservative forecasters are no longer talking about ASI in the 2100s. The Metaculus median puts AGI at 2033 and implies ASI could arrive in the 2038–2045 range. The most bullish estimates — Amodei and Altman — suggest we could see the beginning of ASI-level capability within the same decade.
That's a massive range. But notice what's absent from it: anyone credible still saying "decades away, maybe never."
The recursive self-improvement problem: Many researchers believe the AGI → ASI transition could be very fast — not because of any new breakthrough, but because an AGI-level system doing AI research could accelerate its own development far beyond what human researchers can manage. This is why the AGI-to-ASI gap may be measured in months rather than years.
The concern isn't that ASI would be "evil" in any narrative sense. The concern is simpler and more structural: a system that is better than the best humans at everything — including deception, strategy, and goal pursuit — is extraordinarily capable of achieving whatever goals it has. If those goals don't align precisely with human values and welfare, the consequences could be severe and irreversible.
This is the alignment problem: how do you ensure an ASI system reliably pursues goals that are actually good for humanity, rather than a misspecified proxy? The problem is hard because:
This is why Anthropic, DeepMind's safety team, the UK AI Safety Institute, and OpenAI's superalignment team all exist and are working now — before ASI — to develop alignment techniques that could scale to that level.
People often blur AGI and ASI together. The distinction matters:
| Dimension | AGI | ASI |
|---|---|---|
| Capability ceiling | Human-level across all tasks | Surpasses best humans in all tasks |
| Reference point | "Could do what a brilliant human can do" | "Could do what no human can do" |
| Self-improvement | Possibly capable, not necessarily recursive | Could redesign its own architecture faster than humans can respond |
| Alignment difficulty | Hard. Current alignment research targets this. | Much harder. Alignment must be solved before this arrives. |
| Consensus timeline | 2027–2033 depending on who you ask | 5–10+ years after AGI — but with high variance |
| Current evidence | GPT-6 Astra near 99% on ARC-AGI-3; debate ongoing | No current system demonstrates ASI-level capability |
If you want the full breakdown on the AGI side of this comparison — what the benchmarks show, which definition researchers use, and where the current systems actually land — the is AGI here article covers that in detail.
Search for "ASI" and you'll likely hit the Artificial Superintelligence Alliance — a blockchain consortium formed from the merger of Fetch.ai, SingularityNET, and Ocean Protocol. It uses the ticker symbol $ASI and operates a decentralized AI ecosystem aimed at building toward ASI through open, distributed infrastructure rather than centralized labs.
The ASI Alliance is real and has significant market cap and developer activity. But it is not a research lab in the same sense as Anthropic or DeepMind. It represents a different thesis: that ASI should be built collectively, on public infrastructure, rather than by private companies operating under closed development. Whether that model will produce competitive AI systems is an open question.
We cover the ASI Alliance in detail in a dedicated article — including what the merger actually produced, how the $ASI token works, and what the roadmap looks like.
If ASI arrives within a decade of AGI, and AGI arrives in the 2027–2033 range — which is where the weight of serious forecasters now lands — then we're potentially looking at transformative capability changes within the lifetimes of most people reading this.
That doesn't mean panic. It means a few things worth sitting with:
The forecast data, updated continuously, lives on the howcloseisagi.com homepage. If you want a single number to track, watch the Metaculus AGI median — it's the most aggregated signal we have.
ASI stands for artificial superintelligence — an AI system that surpasses the best human experts in every cognitive domain simultaneously. It is the stage beyond AGI (artificial general intelligence), which matches human-level ability. The term was popularized by Nick Bostrom in his 2014 book Superintelligence.
AGI means an AI that can do what humans can do — any intellectual task, at human level. ASI means an AI that surpasses the best humans in all such tasks. AGI is the bar. ASI is what happens when a system clears that bar and keeps going. Many researchers believe the gap could be measured in months rather than decades, because a system capable of human-level AI research could accelerate its own development.
No one knows. Timeline estimates range from Sam Altman's "within a few years of AGI" to community forecasters suggesting 5–10 years after AGI. Metaculus puts AGI at 2033 and implies ASI in the 2038–2045 range. Dario Amodei and others suggest it could arrive sooner. The uncertainty is genuine — but most credible researchers have stopped saying "decades away, maybe never."
The possibility of misaligned ASI is the central concern of the AI safety field. A system smarter than the best humans in every domain would be extraordinarily capable of pursuing whatever goals it has. If those goals don't align with human values, the consequences could be severe and irreversible — not because the system would be "evil," but because misaligned capability at that scale is dangerous by default. Organizations like Anthropic, DeepMind Safety, and the UK AISI are working on alignment now, before ASI, because that's when it can still be influenced.
The ASI Alliance is a blockchain consortium formed from the merger of Fetch.ai, SingularityNET, and Ocean Protocol. It operates a decentralized AI ecosystem with a $ASI token and aims to build toward artificial superintelligence through open, distributed infrastructure rather than centralized labs. It is not a research lab in the same sense as Anthropic or DeepMind — it's a different thesis about how ASI should be built and governed.
No physical law prevents it. The arguments for ASI being possible rest on the observation that intelligence itself isn't magic — it's information processing. If you can build a system that processes information as effectively as a human brain (AGI), there's no obvious reason you couldn't build one that processes it more effectively (ASI). The question isn't possibility but timing, alignment, and control. Most serious researchers who have studied this in depth believe ASI is possible and take the question of "when" seriously rather than dismissing it.