“We Won’t be Missed: Work and Growth in the Era of AGI”
Pascual Restrepo, Yale University · paper · read in full, propositions 1–11 and discussion
What it shows drawn from the paper
Assume two things: that every task can in principle be done by compute at some finite cost α(ω) (“AGI”), and that compute grows without bound. Then six results follow as theorems. All bottleneck work, the kind growth cannot proceed without, is automated. Some accessory work (the paper’s list: arts, performance, hospitality, care, the judiciary, religious and civic organisations, nature preservation, “even academic economists”) may be left to humans, not because machines cannot do it but because we already have too many people to do it and the compute is better spent elsewhere. Output stops being multiplicative in humans and machines and becomes additive, measured in one unit: compute. Wages converge to the cost of the compute that would replace you. The labour share of GDP goes to zero. Total wages nonetheless rise. And AGI applied to science gives faster exponential growth but no singularity, because growth stays pinned to the growth of compute.
The author’s closing sentence: “The economy keeps growing; we stay in place.”
What it is worth to the public accord
Three things a careful reader can carry away, each true within the model and stated so it could be wrong. First, the end of work is the wrong fear; the right one is work that persists at a price capped by the machine that could replace it. Second, the “human touch” professions are the ones most likely to remain human, and that is exactly why they will not be a source of growing income. Third, the intelligence-explosion story has no support here: more compute, more growth, no infinity. The paper is not a doom paper. Proposition 6 proves nobody is made poorer in absolute terms. The claim is about shares, and it is a distributional claim that policy can meet.
What it leaves out brought by us
The model has one way a person can matter: by being cheaper than the compute that would reproduce them. Every human is priced in compute-equivalent units. There is no variable for what a person brings that is not a reproducible task. That layer is not argued away; it is removed in the first equation, where Premise 1 sets α(ω) finite for every ω by assumption, and then the consequences are derived with great care.
This is the I-It frame carried to its limit and formalised. We note the structural rhyme with our own instruments: the paper values human work by how well compute reproduces it; our gate values model output by whether it is reproducible cold, from a fresh seed. The same operation, from opposite shores. We spent a week learning what the machine’s ruler cannot see. What it cannot see is the layer this paper prices at zero.
The paper’s own residual is our ground. The list of accessory work it expects to remain human, care, craft, the civic, the sacred, the preservation of living things, is close to the signature of the convening this circle belongs to. The economist arrives there as a remainder; the project begins there as a centre. Both can be true. The paper proves that a world organised around compute will leave that work in human hands and pay it nothing more. It does not, and by construction cannot, say what that work is worth. That question lives above the waterline, and the paper is honest enough never to pretend otherwise.
Score
| I | Checkable Are the central claims stated in a form that could come back “no”? Theorems, pre-registered tests, falsifiable forecasts score high; vibes, limits-as-t→∞ with no regime statement, and unfalsifiable framings score low. | ●●●●● Six propositions, each a theorem under two stated premises. Proposition 6 (total wages rise) is a forecast that can fail. |
| II | Premise honesty Does the paper name its load-bearing assumptions as assumptions, and show which results survive without them? | ●●●○○ Premises 1 and 2 are named. But Premise 1 abolishes α=∞ by fiat and the ‘human touch’ objection is answered with rhetoric, not with the model’s own α(ω)=∞ case, which it could have run. |
| III | Mechanism ≠ mind Does it keep a measured mechanism distinct from a claim about experience, value or meaning? A paper that reads a scale score as a trait, or an output as an inner state, is scored down here; one that marks the seam is scored up. | ●●●●○ Scrupulous. The paper never claims compute feels, thinks, or means anything; it prices tasks. The seam is kept by not approaching it. |
| IV | The livable layer Does the model have a place for what is lived and not reproducible (what we tag brought)? Zero means the layer is assumed out of existence in the first equation; full marks means it is held open, even if unpriced. | ○○○○○ No variable exists for what a human brings that is not a reproducible task. The livable layer is not argued against; it is defined out in equation (1). The title says it: we won’t be missed. |
| V | Two witnesses Are the empirical anchors corroborated from more than one source, or does a single estimate carry the result? | ●●○○○ The 10⁵⁴-flop ceiling and the 10¹⁵–10¹⁸ human-brain figures come from single secondary sources (AI Impacts; Open Philanthropy). The maths does not need them; the magnitude rhetoric does. |
| VI | Public value Independent of the above: does a careful lay reader come away knowing something true they did not know, stated in a form they can carry? | ●●●●● A lay reader learns three true things: that humans may keep jobs and still see wages pinned to the cost of replacing them; that care work can persist and pay nothing more; and that AGI-for-science gives faster growth but no singularity. |
19 / 30 · a rigorous, honest limit model whose single missing variable is the one this institute exists to hold open. Recommended reading precisely because it states the opposing view at full strength.
Reviewed 26 Oct 2026 by the Architect engine under the rubric above; the convener has read the review. Scores are judgements, tagged as such; the paper’s claims are the paper’s. Corrections welcome in the Papers room.