AGI is already here. Now the question is superintelligence

For years we asked “when will AGI arrive?”. Today I think the question is outdated. In my view we have already reached AGI, and I’m not alone: a growing number of researchers, entrepreneurs and daily users now think about the next step, superintelligence.

What I mean by AGI

AGI stands for Artificial General Intelligence. There is no single definition, which is why the debate is so confusing. I use a functional one: systems able to handle most cognitive work, move between language, code and documents, understand a goal and use tools to reach it.

With that definition, I look at what happens every day:

an agent like Claude Code takes a goal and completes a whole project, testing and fixing on its own;
Grok Bot, Claude Duet and the other 24/7 agents work for days on a cloud computer, coordinating with each other;
models solve olympiad maths problems, write software, analyse contracts, research with dozens of sources and speak naturally.

These aren’t “narrow” abilities like a chess program’s. They are general: they apply to almost any job done at a computer. To me, that is AGI.

What remains open

It would be dishonest not to say what’s missing. Models learn little from experience once training ends: memory and skills help, but they aren’t continuous learning like ours. Surprising mistakes remain, such as an invented court ruling next to a flawless proof. And the physical world is still largely out of reach.

People who say “it’s not AGI yet” usually point to these limits, or use stricter definitions, such as “better than humans at most economically valuable work” (OpenAI’s Charter definition). It’s a legitimate disagreement. My judgement is that these limits don’t cancel the generality of the abilities we already have.

The direction: superintelligence

If AGI is here, the interesting question becomes superintelligence (ASI): abilities far beyond human ones in almost everything.

Several signals point there. Sam Altman, as early as 2025, wrote that he knew how to build AGI and shifted the ambition to superintelligence; Elon Musk at Davos predicted AI smarter than any single human by the end of 2026. In the US an executive order of 29 September 2026 even asked federal documents to use “Super Intelligence” instead of “Artificial Intelligence”. These are predictions and communication choices, not proof; but they show where the conversation is moving.

The signal that matters most to me is self-improvement: AI helping to build the next AI. Anthropic says it uses its own models more and more to speed up development and experiments, while distinguishing this from a fully autonomous loop. I think we are getting close to that point, and that labs have capabilities beyond what we see. That’s my hypothesis, not an established fact.

What changes for us

If I’m right, waiting for “AGI to arrive” before changing how you work is a mistake: now is the time to learn to delegate to agents, with clear goals, context, permissions and checks. Those who do it today will have a huge advantage when systems become even more capable.

And if the direction is superintelligence, the non-technical questions become urgent: who controls these systems, with what checks, within which limits and for whose benefit. I discuss this in the opinion on how worried we should be.

In one line

For me AGI isn’t a future milestone: it’s the tool we already use every day. The real game now is superintelligence, and the best way to prepare is to learn to work with agents right away.

This is a personal opinion: other experts use different definitions and reach different conclusions.

Sources

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.