AGI vs superintelligence: the difference, explained

The definitions side by side

The chain that connects them

In Situational Awareness, the two are separated by one mechanism: the intelligence explosion. Once AI can do AI research (AGI), hundreds of thousands of automated researchers compress roughly a decade of progress into about a year — and the output of that compression is superintelligence. That's why the scorecard grades them separately: AGI-by-2027 is a live, dated claim; ASI-in-the-2030s is downstream and can't be graded until its preconditions exist.

Why the distinction matters practically

Almost everything contentious about AI risk and geopolitics attaches to ASI, not AGI: decisive military advantage, the case for a government Project, the stakes of the US–China race. Conflating the two makes near-term AI sound apocalyptic and long-term AI sound mundane. The scorecard's verdicts keep them apart — which is also why the current picture (close on capability, not on autonomy) can be true at the same time as "superintelligence remains speculative."

Frequently asked questions

AGI matches skilled humans at general cognitive work; superintelligence (ASI) far exceeds the best humans at essentially everything. AGI is the trigger; ASI is the predicted result of AGI automating AI research.

In every major forecast, yes. The standard sequence is AGI → intelligence explosion (AI automating AI research) → superintelligence. Aschenbrenner dates these roughly 2027 → 2027–29 → early 2030s.

Unknowable with confidence — it depends on AGI and an intelligence explosion that haven't happened. Aschenbrenner forecasts the early 2030s; Altman has suggested 'a few thousand days.' This scorecard grades it Pending.

Neither. Current frontier systems score near skilled humans on scoped tasks (~83% GDPval) but lack the reliable long-horizon autonomy that defines serious AGI bars — and are nowhere near the beyond-all-humans bar of ASI.