AGI and ASI get used interchangeably in most coverage, but they describe fundamentally different thresholds with fundamentally different implications. The first is a machine that can do what humans do. The second is a machine that can do things humans structurally cannot. And the gap between them — according to the best scenario models available — may be measured in months, not decades.
AGI — artificial general intelligence is an AI system that can perform any cognitive task a human can perform, at human level or better, across all domains. Not specialized expertise in one area. Not pattern-matching at superhuman speed in a narrow domain. General — the kind of flexible, cross-domain reasoning that lets humans adapt to entirely new problems without retraining.
The key word is general. Current AI systems are already superhuman in specific domains — chess, protein folding, certain coding tasks, legal research. AGI is when that reaches across the board: the system that can do your job, then switch to doing your doctor's job, then your architect's, with no additional training and at least human-level reliability.
ASI — artificial superintelligence goes further. Not just matching humans across all domains, but significantly exceeding the best human performance across all domains. The difference isn't quantitative (faster human-level reasoning). It's qualitative — an ASI can reason in frameworks that don't have human analogues, compress research timelines that would take the best human teams decades, and find solutions in complex systems that no human or human team could reach.
| Dimension | AGI | ASI |
|---|---|---|
| Capability ceiling | Matches best human performance across all domains | Exceeds best human performance — by an unknown, potentially large margin |
| Research acceleration | Can do what the best human researchers do | Can discover things human researchers cannot reach at all |
| Self-improvement | Can improve AI systems at human-expert level | Can improve AI systems faster than humans can understand or verify |
| Current AI gap | Missing: cross-domain autonomy, novel problem-solving outside training | Much further — ASI requires AGI first, plus recursive self-improvement |
| Median forecast | ~2027–2031 depending on source | No reliable consensus; could follow AGI by months or by a decade |
Two of the most rigorous public forecasting efforts are worth understanding here, because they model the AGI-to-ASI transition in ways that go beyond vague timelines.
AI 2027, a detailed scenario written by former OpenAI researcher Daniel Kokotajlo, Scott Alexander, and collaborators (informed by ~25 tabletop exercises and feedback from over 100 experts), models what happens after AGI in concrete, quantitative terms. Their core argument: once a system achieves human-expert-level AI research capability, the development of ASI accelerates dramatically. In their scenario, AI agents that can do AI research are themselves being run at scale, compressing years of R&D into months. The transition from AGI to ASI in their model isn't a gradual decade — it's a sudden phase shift driven by recursive improvement.
Their scenario is one of many possible futures — they wrote two endings, one more cautious, one more chaotic — but the underlying mechanism (AI-accelerated AI research) is a point of genuine consensus among researchers who take the timeline seriously.
The AI Futures Model takes a more quantitative approach, tracking the METR task time-horizon benchmark — the length of coding tasks AI can complete at 80% success rate, measured against human completion time. This has been doubling roughly every 7 months since 2020 (and possibly faster since early 2024, with a 4–5 month doubling time). Their model predicts when an "Automated Coder" milestone is reached: an AI collective that fully automates coding work at an AGI research project. That milestone — which the model places in 2026–2027 — is significant precisely because it's the moment when AI research itself gets automated, which is the precondition for the AGI-to-ASI acceleration. In other words: not "when does AGI arrive" but "when does AGI research become self-sustaining" — which may happen before full AGI.
This is the part most people haven't thought through. The AGI-to-ASI transition could happen much faster than the current-AI-to-AGI transition — and the reason is recursive self-improvement.
An AGI system would, by definition, be capable of doing AI research at human-expert level. That means it could work on improving its own architecture, training processes, and algorithms. A system doing AI research at the rate of the best human AI researchers, running continuously on large compute clusters, would accelerate its own development at rates no human team can match. The METR doubling-time data gives a rough sense of the scale: if coding time horizons are doubling every 5–7 months with human-paced research, an AI-driven research cycle could compress that further.
This is the "fast takeoff" scenario — and it's the core thesis of AI 2027. It's distinct from a "slow takeoff" where capabilities increase gradually over years in a way humans can observe and respond to. The honest answer is that nobody knows which scenario is right, because we have no empirical data on recursive AI self-improvement at scale. But the mechanism is not speculative — it's the straightforward consequence of an AGI being applied to the problem of building better AI.
It helps to be precise. The frontier models of late 2026 are doing things that would have been considered AGI-level in 2019 — passing bar exams, exceeding PhD-level performance on science benchmarks, completing multi-hour coding tasks. The goalposts moved because those achievements turned out to be reachable without full general intelligence, just with very good pattern matching at scale.
The genuine remaining gaps to AGI are three things the METR benchmark and the AI 2027 scenario both point to:
These gaps are real. They are also closing on a measurable trajectory — which is why the expert median moved from 2060 to 2031 in four years.
There is no known physical or theoretical law that prevents it. The constraints on human cognition are biological — neuron firing speeds, working memory limits, sequential rather than massively parallel processing. These are engineering constraints, not laws of nature. A system not built on biological hardware isn't bound by them.
The more serious objection is diminishing returns: that intelligence doesn't scale indefinitely, and that real-world problems have physical, social, and political constraints that raw intelligence can't overcome. This is a reasonable argument about degree, not a refutation of ASI's possibility. AlphaFold solved a problem (protein structure prediction) that had resisted 50 years of human research — not because it was smarter in every way, but because it could search solution spaces that were computationally intractable for human-directed methods. That's a preview of what domain-specific superintelligence looks like. General superintelligence extrapolates that across all domains.
AGI is the threshold where most knowledge work becomes automatable and the labor market restructures around that fact. For a profession-by-profession look at what this means in practice, see the full job risk breakdown. The key point: humans are still in the loop post-AGI, still making the decisions that matter — but the economic value of the "execution layer" across most professions collapses.
ASI is the threshold where human civilization stops being the primary driver of scientific and technological progress. Climate modeling, drug discovery, materials science, fundamental physics — the pace of advance in these fields could compress dramatically. This is the optimistic scenario: the same recursive acceleration that worries safety researchers could also compress decades of medical research into years.
The risk at ASI is the alignment problem at a higher-stakes level. A system with goals even slightly misaligned with human welfare, operating at superhuman speed and capability, becomes very hard to correct before damage accumulates. This is why the AGI-to-ASI transition — and in particular how much human oversight exists during it — is the central concern of organizations like Anthropic, DeepMind's safety team, and the Center for AI Safety.
The honest position in 2026 is this: AGI is plausibly close — closer than most institutions are prepared for — and ASI is the harder, longer, and more consequential question that follows it. The gap between them is where the most important decisions about AI development will be made, and most of those decisions are being made right now, by a small number of organizations, largely without public input.