Until Now, the Human Was the Operating System
Henry Kim ·
I once thought AGI might be easier to recognize in a game than in a benchmark.
Give an AI a game it has never seen.
No manual.
No walkthrough.
No explanation of the rules.
If it can understand the world, discover the rules, make a plan, and finish the game on its own, perhaps that tells us more than another benchmark score.
GPT-6 Astra has begun rolling out to a limited group. I haven’t tried it yet, so some of today’s excitement may still be launch-day hype.
Still, this release feels different.
OpenAI’s own results show a noticeable jump in computer use. Astra is designed to browse, operate desktop applications, build and test software, and continue working across long sessions. Those results still need broader, independent testing—but they are hard to ignore.
Until now, most AI workflows still had a human in the middle.
AI suggested.
A human checked.
The human corrected it.
Then the human applied it to the real world.
The model was intelligent, but the human was still the operating system.
That is why we were amazed when an agent could search Google and return a clean spreadsheet.
Success was still fragile.
Astra seems to change the question.
Not only:
How smart is it?
But:
Can it keep going until the job is actually done?
If that experience crosses 95% reliability, apps may not disappear. They may simply move into the background.
Instead of opening five applications and moving information between them, we may describe the outcome we want—and let the agent operate the tools.
My prediction is that, before the end of this year, many office workers will begin using computers mainly through conversation.
New devices will probably follow.
Not because keyboards and screens will suddenly disappear, but because intent may become the primary interface.
Maybe AGI will not arrive with a dramatic announcement.
Maybe it arrives when the AI is still working on Day 5—and the human is the one who needs a break.
Honestly, I can’t wait to try the SimCity that Astra is reportedly still building on Day 5.
What would it take for you to trust an AI to use your computer without watching it?
Thoughts from skaia.ca/blog