In June, fourteen researchers — most of them at Google DeepMind, including one of its co-founders — published a careful 57-page study of what happens to artificial intelligence after it matches us. They mapped every technical route forward and every obstacle in the way. Then they arrived at the question of what all this does to ordinary human lives, wrote about a hundred and thirty words, and stopped.
That stop is the most interesting thing in the paper. This post is about why.
First, what the paper actually says
The paper is called From AGI to ASI, and it's worth being precise about it, because the coverage hasn't been. It is not a timeline. It names no year for human-level AI and no year for anything beyond it. It deliberately brackets the whole question of when, saying its goal is to map the terrain "independent of when humanity hits this milestone." Anyone telling you this paper predicts a date is telling you about a headline, not the paper.
What it does instead is assume human-level AI arrives, then ask two questions with real rigor: by what routes could machine capability keep climbing past that point, and what could slow it down or stop it? Four routes get mapped — building bigger, inventing genuinely new methods, systems improving themselves, and intelligence emerging from large collectives of AI agents working together. Then come the brakes, and this is the half nobody quotes: the world may run out of high-quality training data; the energy, chips, minerals, and money required may simply stop being affordable; today's methods may hit a ceiling; research itself gets harder as fields mature. Two of the brakes are especially grounding. Models trained on human concepts may be fenced in by human concepts. And any new idea about the physical world still has to be tested in the physical world, at the speed the physical world runs. As they put it, that alone "could potentially limit the rate of intelligence growth to the rate of empirical science rather than the rate of computational scaling."
Their own conclusion is offered, in their words, "with a lot of uncertainty (and thus low confidence)." That hedging isn't weakness. It's what serious people sound like when they're describing something they can't see the end of — and it's a good reason to trust the one place where they're unambiguous.
The paragraph that stops
Deep in the section on why societies might deliberately slow all this down, the paper turns, briefly, to people. Here is essentially the whole of it:
"It is unclear what a post-AGI labor force would look like, if many cognitive tasks could be automated and offered cheaper through AI services than through human workers."
They go on: at sufficient scale, this "would require a rethinking of fundamental economic mechanisms" — specifically, they name "the shift from labor to capital as the main economic resource, and its potential impact on the social contract." Many more aspects of how humans live together, they add, need examining for how well they hold up.
And then the paragraph ends like this: "Giving answers to these important questions is beyond the scope of this report."
That's it. In a document that spends pages on training-data supply, memory bandwidth, and the theoretical limits of intelligence, the question of how people will eat gets a hundred and thirty words and a polite handoff. Search the whole paper for inequality, for who captures the gains, for how the plenty gets distributed, and you will not find it. Not because the authors are careless — because they drew a boundary, and said so out loud.
Why they stopped, and why that's fair
It would be easy and wrong to read that sentence as a dodge. It isn't. It's a category boundary, and an honest one.
Everything else in that paper is an engineering question. Will this architecture scale? Does the data hold out? How fast can experiments run? Those questions have right answers that evidence can eventually settle, and these are among the best-qualified people alive to chase them.
"How should the benefits of all this reach people?" is not that kind of question. No amount of compute resolves it. It's a design question — a question about how we choose to arrange things, answerable only by proposing an arrangement and arguing for it in the open. Different training entirely. And to their credit, one of the paper's own authors is working the problem elsewhere: the text points readers to a companion analysis of what they call post-labor prosperity, rather than pretending this paper had covered it.
So the blank isn't negligence. It's a vacancy. Someone has to fill it, and it won't be filled by the people building the engines.
They mapped every route the machines might take. The one thing they didn't map is the one thing that decides whether any of it is good news.
What an answer looks like
Copiosis is one candidate answer — a specific, checkable proposal for exactly the gap that paper leaves open.
Consider what "the shift from labor to capital as the main economic resource" actually threatens. It's alarming because of a single hidden assumption in the system we happen to use: that a person's access to food, shelter, and care flows through selling their labor. Break that link and the shift is a catastrophe. Never make that link in the first place and the shift is just… machines doing more of the work.
That's the structural move. In Copiosis, necessities are simply provided — food, housing, healthcare, education, with no bill attached. People earn Net Benefit Rewards for making life better for others and the planet, and those rewards buy life's luxuries. Nothing in that design requires a human to perform the routine labor. If a machine grows the food, the food still reaches people, because access was never routed through a wage.
Which means the labor-to-capital shift doesn't have to be fought, slowed, or reversed. It has to be landed somewhere that was built to receive it. We made the fuller version of this case in AI is building the on-ramp; what's new here is that the people building the technology have now put the same question on the table themselves, in a peer-reviewed venue, in their own words.
Two more lines worth sitting with
The paper contains two sentences that deserve more attention than they've gotten.
The first, on whether the world could coordinate its way to a safe pace: given the difficulty of international coordination and the historical rarity of global regulatory frameworks, they write, aligned multilateral governance "remains an elusive, perhaps unrealistic, target."
The second is sharper. Building on an established framework about how technology spreads under geopolitical rivalry, they conclude that sustained competition between groups "systematically favours the development and adoption of competitiveness-enhancing technologies, irrespective of their implications for human welfare."
Read that again, because it isn't a complaint about technology. It's a description of an incentive structure — one in which "does this make people's lives better?" is not among the questions being asked, because nothing in the machinery asks it. That is precisely the diagnosis Copiosis starts from, and precisely what the algorithm is built to correct: a system where the thing rewarded is the benefit produced, so that the question can't be skipped.
And then, generously, the paper offers a standard both systems can be judged against. Progress, the authors write, means developments that "preserve or enhance individual and collective autonomy, promote human flourishing and dignity, and are broadly recognized as beneficial and useful by societies." We'd sign that sentence without changing a word. The honest follow-up is simply: which arrangement is more likely to produce it — one that routes prosperity through money, or one built to route it toward benefit directly?
Honesty section, as always
A few things this post is not claiming, stated plainly so nobody has to guess.
- No one at DeepMind endorsed anything like Copiosis. They named a question. We're offering an answer. Those are different acts, and conflating them would be exactly the kind of overreach we try not to commit.
- The paper gives no dates. If you see this post cited as "Google says AGI by year X," the citation is wrong. It says the possibility of getting well past human level "within the next decade or two" can't easily be dismissed — a double negative about a possibility, and a careful one.
- The paper assumes the safety problem gets solved. The authors flag this themselves: it is "by no means a given, nor is it a light assumption." Everything downstream, including our argument, inherits that assumption.
Fill in the blank
Here's what stays with us. Fourteen people with as clear a view of this technology as anyone on Earth sat down to map where it goes. They filled in the routes, the obstacles, the physics, the open research questions — and left one region unlabeled, marked only important, and not ours to answer.
The blank is an invitation. It gets filled either deliberately, by people who thought hard about how prosperity should reach human beings, or by default, by whatever the existing incentives grind out while nobody's specifying. One of those produces something we chose. The other produces whatever we get.
That's why this project exists — so that when the question finally gets asked in earnest, there's a fully worked answer already on the table. Start with how it works, test it against the transition and the hard objections, and then tell us where it breaks. Skepticism welcome. Filling in a blank this important should be hard.