AGI and superintelligence turn up in the same sentences so often that they’ve started to blur into one idea. They aren’t one idea. One is a threshold that looks a lot like us. The other is everything past that threshold, and nobody knows exactly what shape it takes.
The distinction isn’t academic. It shapes how labs announce models, how regulators write rules, and how much lead time safety researchers believe they have. Blur the two together and you get bad forecasts, followed by worse policy.
Two thresholds, very different distances
Artificial general intelligence, in the usual formulation, describes a system that can handle essentially any intellectual task a human can, at roughly human competence. Learn a new language on Monday, plan a move across the country on Tuesday, debug a codebase on Wednesday, then switch to something completely unrelated without retraining.
Superintelligence goes further. It describes a system that outperforms the best humans across the board, not by a hair but by a margin that makes the comparison awkward. I.J. Good sketched the idea back in 1965 when he described an “intelligence explosion”; Nick Bostrom gave it a book-length treatment in 2014. The core claim is simple. Once a system can improve itself better than humans can improve it, progress stops tracking human patience.
So AGI is a rung. Superintelligence is the ladder above it, and possibly a very long ladder.
What AGI actually means is still contested
Ask five labs to define AGI and you’ll get five answers with different edges. Some talk about economic output, some about generality across tasks, some about autonomy. There’s a useful breakdown of what artificial general intelligence really means and why nobody agrees if you want the full mess of definitions.
The benchmark treadmill
Every few years a test arrives that’s supposed to separate narrow from general. MMLU. ARC-AGI. GDPval. Frontier models saturate each one faster than the previous, and the goalposts move again. That saturation tells you something real: benchmarks measure tasks, and generality is about handling tasks nobody has written down yet.
A model that scores 90% on a graduate-level exam can still fail to notice that a spreadsheet column is misaligned. That gap between test performance and practical competence is exactly where AGI claims get slippery.
Superintelligence is a different kind of claim
Nobody has to argue about benchmarks here, because superintelligence doesn’t exist yet on any mainstream definition. It’s a projection. The interesting part is that it could arrive in more than one form.
Three ways to be superhuman
- Speed. The same quality of thinking as a competent human, but a thousand times faster. A week of research compressed into ten minutes changes what’s possible even without a smarter mind.
- Breadth. Individual specialists still beat it in narrow domains, but nothing beats it across all of them at once. Think of a whole research department folded into a single process.
- Depth. Genuinely better reasoning. New proofs, new materials, arguments no human has managed. This is the version that keeps researchers up at night, and the hardest to forecast.
The speed version is nearly a scaling story. The depth version requires something we don’t yet know how to build.
Where the gap between AGI and superintelligence bites
If AGI and superintelligence arrived on the same day, planning would be simpler. They probably won’t. The interval between them is where most of the real decisions get made.
- Safety lead time. Alignment work that assumes years of warning looks very different if the jump takes months. Systems that are merely human-level can be tested, red-teamed and audited by humans. Systems past that point are harder to evaluate with human judgment alone.
- Deployment economics. A human-level system gets priced against human labour. A superintelligent one gets priced against what, exactly? The business case changes shape entirely.
- Governance. Rules written for AGI tend to assume a slow diffusion curve. Rules written for superintelligence assume an emergency.
Most public argument collapses this interval to zero. Headlines treat “we built AGI” and “we built something we can’t control” as the same headline.
Timelines, and the marketing problem
Labs have a commercial incentive to sound close to something enormous. Recent releases show how the framing works. Coverage of OpenAI’s “generational leap” with GPT-6 Astra leaned on capability jumps rather than definitions. Reports that Meta’s Watermelon ripened into a frontier model followed a similar pattern: striking numbers, fuzzy claims about what the numbers mean.
None of that is dishonest, exactly. It’s just that “frontier model” has become a marketing term, and “AGI” is following it. An announcement that a system “reaches AGI” on an internal benchmark tells you about the benchmark as much as the system.
One discipline helps: translate every claim into a specific task. If a lab says a model plans like a human, ask which plans, over what horizon, at what error rate. If it says a model reasons, ask for the failure cases.
Safety arguments don’t transfer cleanly between the two
The AGI conversation is mostly about misuse, bias and labour displacement. The superintelligence conversation is about control. Different problems, different toolkits.
That split shows up inside the industry too. When more than 1,000 frontier lab staffers signed an open letter asking for an AI brake pedal, the request wasn’t a pause on narrow tools. It was for verifiable slowdown mechanisms that would hold at the point capabilities start compounding faster than oversight can follow.
An eval suite built for AGI-level systems, testing honesty and refusal and task competence, won’t necessarily tell you whether something two rungs up is doing what you asked. Interpretability research is the bridge, and it’s still early.
How to read the next “AGI achieved” headline
Three questions cut through most of the noise.
- What does the claim actually measure, and who designed the test?
- Is the system general across unfamiliar tasks, or just impressive within a family of tasks it was trained on?
- If it is genuinely human-level, what happens in the twelve months after launch, and who is liable when it fails?
Superintelligence claims deserve a fourth: what evidence would change the claimant’s mind? A forecast with no falsifier is a vibe, not a prediction.
The terms will keep getting mixed up, partly because the people building these systems have every reason to let them blur. AGI is a milestone you can announce. Superintelligence is a horizon you can only point at. Knowing which one a given claim is about is the difference between following the field and being sold to by it.