Plate VII
AGI and the Future: What It Is, When It Comes, What Changes
AGI
30 OCTOBER 20267 Min Read
Everyone is arguing about when AGI arrives. Almost nobody agrees on what it is. Here's how I think about the term, where AI actually stands, and what changes on the way there.
Ask five people in AI when AGI arrives and you'll get five dates. Ask them what AGI is and you'll get six definitions.
That's the first thing worth saying about artificial general intelligence: the argument about timing is mostly an argument about meaning. A system that writes better code than most engineers can still fail to count the objects in a photo. Is that general? Depends who you ask.
I'm not going to pretend to settle that. But I build with these systems every day, and I think there's a more useful way to look at it than waiting for a single moment when someone declares the finish line crossed.
What People Mean by AGI
Strip away the hype and there are roughly three definitions in circulation.
The capability definition. AGI is a system that can do most economically valuable cognitive work as well as a skilled human. This is the one most labs care about. It's measurable in principle: hand it the job, see if it does the job.
The learning definition. AGI is a system that can pick up a new skill as efficiently as a person can. The ARC Prize uses a version of this. It's a higher bar. Being good at a thousand things you were trained on isn't the same as learning the thousand-and-first from a few examples.
The "you'll know it when you see it" definition. A system that feels like talking to a capable colleague across any topic. This is the one the public uses, and the one that keeps moving: every time AI clears a bar, the bar gets redefined as "not really intelligence".
None of these is wrong. They just measure different things, which is why people looking at the same model can honestly disagree about whether it's close.
Where AI Actually Stands
The honest description of today's AI is jagged.
On some axes it's already superhuman. In 2025, AI systems from Google DeepMind and OpenAI reached gold-medal level at the International Mathematical Olympiad, one of the hardest reasoning tests there is for humans. Frontier models write production code, read and summarise faster than any person and work across dozens of languages.
On other axes it's surprisingly weak. Models still make basic mistakes a child wouldn't. They struggle to remember and build on what happened last week. They're poor at learning a genuinely new task from a handful of attempts, which is exactly what benchmarks like ARC-AGI-3 were built to test. And put them in a body, as I wrote about in my posts on physical AI, and the gap gets wider: folding laundry is still harder than passing a law exam.
So the shape isn't a single line creeping toward "human level". It's a spiky outline that's well past human on some sides and well short of it on others, and it's growing unevenly.
The Trend That Matters Most
If I had to pick one measurement to watch, it wouldn't be a benchmark score. It would be how long a task AI can complete on its own.
METR, an independent research group, has been tracking exactly this: the length of tasks, measured in how long they take a skilled human, that AI agents can finish with reasonable reliability. Over the past six years that length has doubled roughly every seven months, and on recent models the pace looks faster than that. Systems that could manage a few minutes of autonomous work in 2023 now handle tasks that take a person many hours.
That's the number that maps onto real work. A tool that can do a 5-minute task is an assistant. A system that can reliably do a week-long project is a colleague.
That's a big if. Trends like this can bend. But it's the most concrete evidence anyone has, and it points in one direction.
Soon, or Not for Decades?
The timelines split cleanly by who's talking.
The labs say soon. Leaders at Anthropic, OpenAI and Google DeepMind have publicly suggested systems with broadly human-level capability could arrive somewhere between the next couple of years and around 2030. Their argument is simple: the curves have held for a decade, and every predicted wall has failed to appear.
Researchers as a group say later. Large surveys of AI researchers have put the 50% estimate for human-level machine intelligence closer to the 2040s. Their caution isn't anti-AI. It's that scaling might hit real limits: running out of high-quality data, no easy way to check whether an AI did ambiguous work well, and the physical cost of ever-bigger data centres.
My own view sits in between, and is mostly about shape rather than date. I don't think there will be an "AGI day". I think we'll look back and realise the line was crossed gradually, domain by domain, with coding and analysis first and physical and social skills last. By the time everyone agrees it's here, it'll have been here for a while.
What Changes
Assume capability keeps climbing, whatever we end up calling it. Here's what I expect to shift.
1. Work gets reorganised around judgment. When the cost of doing a task falls toward zero, the scarce thing becomes deciding which tasks matter and checking the results. The people who thrive will be the ones who can define problems precisely, delegate to AI systems, and catch when the output is subtly wrong. That's already true for software engineers. It will be true for analysts, lawyers, designers and researchers.
2. Science speeds up. The most exciting case for AGI isn't chatbots. It's research: systems that read the entire literature, propose hypotheses, design experiments and run them in simulation. AlphaFold already compressed years of lab work on protein structures into a database anyone can search. Multiply that across materials, medicine and energy.
3. The physical world catches up last, then fast. General intelligence without a body still can't fix a pipe or care for a patient. That's why physical AI matters so much: it's the bridge between AGI as software and AGI that changes how things are made, moved and maintained.
4. Power concentrates, unless we choose otherwise. Training frontier models takes resources only a handful of companies and countries have. If AGI-level systems arrive, who controls them, who can access them and who benefits are political questions as much as technical ones. Open models, public compute and real competition all matter here.
5. Safety stops being a side topic. A system that can complete week-long projects on its own can also make week-long mistakes on its own. Making sure these systems reliably do what we intend, and can't be easily turned to harm, is the hard engineering problem of the next decade. It's not a reason to stop building. It's a reason to build carefully.
What I'm Doing About It
I don't find "AGI is coming" scary or exciting on its own. It's a forecast. What matters is how you prepare for it.
For me that means three things. Build with these systems every day, so I understand their failure modes rather than reading about them. Go deeper on the parts AI is slowest to absorb: systems thinking, hardware, the physical world. And treat reliability and safety as core skills, not afterthoughts, because the bottleneck in every AI system I've shipped has been trust, not capability.
What Comes Next
The most likely future isn't a sudden leap to a machine god or a sudden collapse of the hype. It's a steady climb where AI does a bit more of the work every few months, unevenly, until one day the question "is it AGI yet?" stops being interesting because the answer has stopped mattering.
The useful question was never when. It's what we build with it, and who it's built for.
That part is still up to us.
Sources
Written & engraved by Mayank Bairagi