A Signal From Ahead

AI arrived ahead of schedule. In a 2016 survey, machine learning researchers gave even odds that machines would outperform humans at every task by about 2061.1 The 2022 edition of the survey said 2060. The 2023 edition, with 2,778 respondents, said 2047: thirteen years gone in twelve months.2

The channels fall into two families. In the first, the future shapes the present from outside the timeline, the way an ending can constrain a story, and nothing has to travel. In the second, a signal really does move backward through spacetime, and that family obeys one hard rule: a channel into the past reaches back only as far as the channel itself has existed.

If this world is a simulation, the future reaching back isn’t exotic physics. It’s an edit. Whoever runs a simulation stands outside the timeline it computes, and can pause it, inspect it, branch it, and change any variable, including how fast a technology arrives. The likeliest operators of an ancestor simulation are descendants of the civilization being simulated, which makes them, from inside the run, the future.

Nick Bostrom’s 2003 argument turns this into arithmetic.4 If even a small fraction of civilizations reach the technology to simulate their ancestors in detail, and a small fraction of those run many such simulations, simulated observers vastly outnumber real ones. Try the numbers.

As for why the operators would touch this decade: a simulation built to study how intelligence arises would treat the years when AI arrived as the most interesting stretch of the run, the obvious place to tune, branch, or replay.

P(this is a simulation) ≈ 1/10  ×  P(its operators tuned AI’s timeline) ≈ 1/100  =  1/1,000

Physics usually runs on one boundary condition: the state of the universe at the beginning, carried forward by its laws. The laws themselves are time-symmetric, though, and nothing in them forbids a condition imposed at the end as well. With both in place, the present is fixed by two boundaries at once, and everything in between leans toward whatever the ending contains.

Nothing travels backward in this picture. The ending acts as a constraint, and a constraint reshapes the odds of everything it touches. Random walks are enough to see it.

This is the structure of a real proposal. In 1964 Aharonov, Bergmann, and Lebowitz showed that quantum mechanics can be written with one state evolving forward from the past and a second evolving backward from the future.6 That two-state picture led to the prediction of weak values,7 which have since been measured in the laboratory.8 Gell-Mann and Hartle worked out quantum cosmologies with both initial and final conditions,9 and Horowitz and Maldacena proposed a final-state condition inside black holes as a way out of the information paradox.10

Experiments have looked hard for nonlinearity and found none: a 1989 trapped-ion measurement bounded it to a few parts in 10²⁷ of the nuclear binding energy it probed.16 Tight is not zero. A civilization with millennia of instrument building could exploit an effect far below anything measurable today. The relay also needs the faster-than-light channel to behave the same way in every frame, a property the nonlinear theory itself would have to supply.

P(an exploitable nonlinearity below today’s bounds) ≈ 1/10,000  ×  P(a future intelligence aims it at this decade) ≈ 1/10  =  1/100,000

General relativity allows wormholes, and in 1988 Morris, Thorne, and Yurtsever showed how one becomes a time machine: give its two mouths different histories.17 Clocks at the two mouths stay synchronized through the throat but drift apart in ordinary space, so a message that enters the lagging mouth comes out of the other one in the past.

An offset can be built two ways. Try both.

A message from the future raises the paradox question at once: what stops it from preventing its own sending? Joseph Polchinski, back again, posed the sharpest version: a billiard ball sent through a wormhole to knock its younger self away from the entrance. In 1990 a team including Igor Novikov and Kip Thorne showed that such spacetimes admit self-consistent histories, ones in which whatever comes back produces exactly the events that send it.25 When Echeverria, Klinkhammer, and Thorne worked the billiard ball through in detail, the starting conditions they studied all had consistent solutions, often infinitely many.26 David Deutsch then proved that quantum mechanics always supplies at least one,27 and a 2011 optics experiment reproduced the logic of such loops with photons.28

Apply that to a message that speeds up AI. The nudge received in the 2020s shapes the intelligence that later sends it, so the loop is consistent wherever what’s sent equals what was received: where the curve below crosses the diagonal.

Each channel is the product of two factors: whether the physics allows it, and whether a future intelligence actually used it this way. These are my settings. Move them and the total follows; every scale is logarithmic.