NOTE-1269: Shaping Humanity's Longterm Trajectory

Contribution

A formal tool for comparing longtermist interventions. Ord says its aim is “not to argue for longtermism, but to develop a technique for comparing different ways of shaping the longterm future” (p. 7). He represents the whole future as one curve of value over time and derives, for four idealised changes and for existential risk reduction, how each one’s value scales with the future’s height and its length. The results are mostly qualitative: which interventions grow with the future’s duration and which do not.

Key insight

Shifting the future earlier and multiplying it are different in kind. An advancement replaces a sliver of the present with a sliver of the end, so its value is (v_τ − v_0)δt however long the future lasts. Existential risk reduction and enhancements multiply the whole area, v̄τ. So, the longer and better the future, the less advancing progress matters relative to securing it, unless the advance also changes the shape of the curve, for instance by cutting risk or locking in better values (pp. 26–29).

Assumptions

• Value is roughly quantitative and separable across time (pp. 3–4), a ratio-scale flow that can be integrated. This admits the total view, the critical-level view, an average view integrated over time, “some person-affecting views”, and views with diminishing value of population at a time. It excludes views that are not time-separable, such as the average wellbeing of everyone who ever lives.
• Only value. “If there are important parts of morality that are unrelated to value (such as personal freedoms or unbreakable rules) the framework will remain silent about those” (p. 4).
• No pure time preference (p. 9). Discounting money for growth is set aside because the curve is value itself. Discounting for risk is set aside because outcomes are evaluated ex post; risk would enter through uncertainty over trajectories, not a fixed hazard rate (p. 10).
• Finite duration (p. 10). Infinite sums are avoided by treating τ as a parameter and asking how values scale with it.
• Marginal, idealised, permanent changes from t = 0 (pp. 11–12), with diagrams deliberately not to scale (pp. 7–8).

Key results

• The size of the future (pp. 5–7). The species has lasted about 300,000 years; a typical species lasts about a million; Earth stays habitable for about a billion more years, and other stars could extend that “at least a thousand times”. The future’s value relative to this century is (v̄/v_0) · (τ/100): how much better the average century is, times how many centuries there are. “Each of these factors might be truly vast — quite possibly a billion or more” (p. 7). Duration alone “is more than enough to get longtermism off the ground” and is “the most widely accepted part” (p. 7).
• Advancements (pp. 12–17): the curve shifted left by δt.
  • Exogenous end time: δV ≈ (v_τ − v_0)δt. It does not depend on τ or on the curve’s shape (p. 15). If value changes by under 5% a year the approximation is within 5% for advancements up to a year; a year’s advancement “may require something comparable to the entire effort of all currently existing humans working for a year” (p. 15).
  • Endogenous end time (the end comes δt sooner too): δV ≈ −v_0δt. Unless the present has negative value, the advancement makes the future worse (p. 16).
  • A temporary advancement that lasts until time t_e is worth about (v_e − v_0)δt, best if it lasts to the peak and no further (p. 25). Ord’s footnote: Bostrom 2003 “was focused on an advancement of this kind” (fn. 11).
• Speed-ups (pp. 17–20): the curve compressed by γ_t. Exogenous end: δV ≈ (v_τ − v̄)(1 − 1/γ_t)τ, positive only if the end is better than the average. Endogenous end: δV = −v̄(1 − 1/γ_t)τ, positive only if the future’s total value is negative. A possible speed-up is likely overdetermined, and then it is only an advancement (p. 20). The agricultural revolution arose independently in at least five places.
• Gains (pp. 20–22): a fixed rise δv; δV = δv τ. Examples: a fixed rise in everyone’s wellbeing in a stable population; saving an ecosystem, species or work of art, where its value lasts as long as it does. A gain must be contingent; saving “something irreplaceable from permanent destruction” resists the objection (p. 22).
• Enhancements (pp. 22–23): a proportional rise γ_v; δV = (γ_v − 1)v̄τ. Examples: better health or prosperity for all, proportionally more people (on some views), better moral understanding or motivation. Contingency is again the challenge; the best case is values improved before a lock-in (p. 23).
• Comparisons (pp. 26–29).
  • Advancements scale with v_τ − v_0; speed-ups with (v_τ − v̄)τ; gains with τ; enhancements with v̄τ (p. 26).
  • An advancement beats a gain if δt/τ > δv/(v_τ − v_0). “Advancing progress by a millionth of humanity’s lifespan is ten times as good as permanently gaining a ten-millionth of the gap” (p. 27).
  • An enhancement beats a gain if (γ_v − 1) > δv/v̄; τ cancels (p. 27).
  • Existential risk. Model near-term risk as occurring at t = 0 and a catastrophe as leaving about zero value. Raising survival by a factor γ_p gives δV = (γ_p − 1)v̄τ. Example: 20 points of risk, survival raised one point, γ_p = 81/80 = 1.0125 (p. 28). This is “almost identical” to an enhancement, and irreversibility, built into existential risk, makes contingency easier to meet.
  • An advancement beats risk reduction only if δt/τ > (γ_p − 1) · v̄/(v_τ − v_0) (p. 28). “On a million-year lifespan (that of a typical species) a one-year advancement would be roughly as important as a one-in-a-million improvement in nearterm survival probability — but the latter seems much more achievable” (p. 29).
• Conclusion (pp. 29–31). For advancing progress to beat risk reduction, it must do something more, such as advance defensive technology, wisdom or moral progress, effects that can also scale with τ. Proposed extensions: uncertainty over trajectories or hazard curves, non-marginal changes, and idealised shapes (a plateau, exponential or cubic growth; fn. 14 suggests quadratic if new locations are fixed resources).

Claims

id	claim	strength	support
C1	Under time-separable, ratio-scale value with no pure time preference, the values of four idealised changes depend only on v_0, v_τ, v̄, τ and the size of the change, not on the curve’s shape	strong within the model: derived, with stated approximations	pp. 12–23
C2	Advancing progress does not grow in value with the future’s duration, and if the end time moves with it, it is a loss	strong within the model; the real question is whether real advances behave so	pp. 15–16
C3	Reducing near-term existential risk multiplies the future’s whole value, like an enhancement, and is more surely contingent	moderate: the multiplier is derived; contingency is argued	p. 28
C4	Pure advancements and gains rarely rival existential risk reduction; a year’s advancement on a million-year future matches a one-in-a-million gain in survival	moderate: derived for a plateau-like case; “much more achievable” is judgement	pp. 28–29
C5	People are roughly indifferent to the timing of benefits to others, such as their children or future generations	weak here: one citation (Frederick 2003), not in the chapter’s bibliography, given in a sentence	p. 9
C6	Duration alone suffices for longtermism, and the future’s value could exceed this century’s by “more than a billion billion”	weak: illustrative orders of magnitude	p. 7

Concepts

• longterm trajectory: humanity’s instantaneous value v(t) over time, from t = 0 to the end τ.
• instantaneous value of humanity: value per unit time, under whatever theory of value; on classical utilitarianism, happiness minus suffering at that time (p. 3).
• advancement / delay: the future curve shifted earlier / later.
• speed-up / slow-down: the curve compressed / stretched in time.
• gain / loss: a fixed rise / fall in value at all times.
• enhancement / diminution: a proportional rise / fall.
• exogenous versus endogenous end time: an end fixed by nature (the Sun) versus one that moves with our development (a dangerous technology).
• contingency: whether a change would have happened anyway; an overdetermined speed-up is only an advancement.
• lasting change: one that persists for at least a tenth of the future (p. 25).

Connections

• The Precipice (LIT-1678), unread. See section 3.
• Bostrom 2003 (NOTE-1314). Bostrom compared the cost of delay with existential risk reduction and found a percentage point of risk worth over 10 million years of delay. Ord’s framework generalises that comparison and says Bostrom’s case was a temporary advancement (fn. 11). It adds what Bostrom’s does not: if the end time moves with progress, advancing it is a loss.
• Greaves and MacAskill (NOTE-1302) call a zero pure rate of time preference “near-essential”. Ord adopts it with the same lineage (Pigou, Ramsey, Harrod; Ramsey is LIT-1733) and adds the empirical claim that people do not discount benefits to others.
• Tarsney (NOTE-1247) and Thorstad (NOTE-1308). Ord sets risk aside by evaluating outcomes ex post and putting all risk at t = 0. Their objections live exactly where he leaves the framework for later: a hazard that persists after this century. Inference: once risk after t = 0 is added, the τ in (γ_p − 1)v̄τ becomes an expected duration, and Thorstad’s point is that this expectation may be short.
• Hershfield (NOTE-1259). Ord notes that apparent pure time preference for oneself is hard to tell apart from weakness of will (p. 9). Hershfield ties it to felt discontinuity with one’s future self. See Bearing on the record.
• MacAskill (LIT-1747), unavailable, is cited on value lock-in (2022: 78–9).

Bearing on the record

• QUESTION-299. It sharpens the impersonal ground without defending it. Ord’s argument needs an evaluative premise, value summed over time with no pure time preference, and no claim about anyone’s self. Two results matter for the question.
  • Duration alone carries the case (p. 7). What the living owe, on this view, grows with how long humanity lasts even if life never gets better.
  • Besides survival, the only other route to the whole future is to change what will be valued before it locks in (p. 23). Inference: this makes the present generation’s values a long-run lever, which brings the legacy question (QUESTION-301) into the normative one. To shape the future’s values is to leave a legacy of the living’s values, whatever the impersonal justification.
• The owner’s thesis (CLAIM-491). Ord says nothing about motivation; his agent is an “altruist” (p. 20). Two points bear on it. Both are inferences.
  • Concern for others’ futures is not discounted, on Ord’s report. People “don’t have” pure time preference “regarding benefits to other people, such as one’s children or future generations”; they are “roughly indifferent to the timing” (Frederick 2003, p. 9). Hershfield (NOTE-1259) finds people discount their own future more the less continuous they feel with it. If concern for future generations were concern for a self extended through time, it should fade with distance the same way, since felt continuity with people further off must be weaker. Indifference to timing does not fit that. It fits concern for future people as interchangeable members of a category (“humanity”, “future generations”), not as parts of a self continuous over time (TERM-096). The thesis can absorb this only with a category-level self whose extent is not graded by time. That is the self-categorization reading (THEORY-1114), and CLAIM-492 asks what fixes its extent apart from caring.
  • Caveats. The result is reported in one sentence, from one study, and Frederick 2003 is not in the chapter’s bibliography. It is about judgements of when benefits should come, not felt concern or action. And Schubert et al. (NOTE-1270) find that lay judgement reaches the far future only when prompted. Timing-neutral judgement and weak unprompted concern can both be true.
  • “Humanity” as one life. The framework treats humanity as one subject with a “lifespan” (p. 27) and compares its survival risk with an individual’s at eighty or ninety (p. 10). The analogy is rhetorical; the value summed is impersonal. It is not evidence that anyone identifies with humanity.
• A dated note is proposed for QUESTION-299 (section 4), not for CLAIM-491. The discounting point bears on the thesis’s test, but it rests on a secondhand report the record should read first (see Open questions).
• No THEORY filed. A valuation framework, not an account of what is true.
• No ML instruction.

Limitations

• Author’s version, April 2023, not compared with the 2025 book text; pages differ.
• Idealised and marginal. Ord grants real interventions are “much messier” (p. 12). No uncertainty, so no hazard after t = 0.
• Silent on what is not value, by its own statement (p. 4), and so on rights, special obligations and what the past is owed.
• The empirical aside on time preference (C5) is thin; see above.
• Typos and slips.
  • “We can all such a change a gain” for “call” (p. 20); “had in not started” for “it” (p. 20); “speedup” beside “speed-up” (p. 20); “achieved later the default trajectory”, missing “on” (p. 22); “preferrable” (p. 26); a missing full stop after “who will ever live” (p. 4).
  • “Ord 2022: 153–8” (p. 23) for Ord 2020: 153–8, as cited on p. 2.
  • Bibliography: “Collision, P.” for Collison (p. 32); “That Is Not Dead Which Eternal Lie” for “Which Can Eternal Lie” (p. 32); “Long- Term” (p. 31). Cited but not listed: Frederick 2003, MacAskill 2022, Pigou 1920, Ramsey 1928, Harrod 1948, Koopmans 1960, Ng 1991.

Open questions

• Does Frederick 2003 show what Ord says? If people weigh lives saved later the same as lives saved now, but discount their own future with felt discontinuity, then concern for others’ futures and for one’s own are structured differently. That is a direct test of whether concern for future generations is an extended self’s. The record does not hold the paper. Crossref’s likely match, given the chapter does not list it: Frederick, S. (2003), “Measuring Intergenerational Time Preference: Are Future Lives Valued Less?”, Journal of Risk and Uncertainty 26, doi:10.1023/A:1022298223127.
• What happens to the comparison of advancement and risk reduction when risk persists after this century, as Thorstad argues? Ord names the extension but does not make it.