Optimal Timing for Superintelligence
New paper by Nick Bostrom here.
Claude Opus 4.6 on online reactions/commentary:
This paper just dropped three days ago (February 12, 2026), so the commentary is still very fresh and developing. Here's a synthesis of the paper's core argument and the reactions so far.
The Paper's Core Argument
Bostrom frames the superintelligence question not as Russian roulette but as risky surgery for a terminal condition. The paper adopts a person-affecting ethical stance — focusing on people alive today — and argues that models incorporating safety progress, temporal discounting, quality-of-life differentials, and concave QALY utilities suggest that even high catastrophe probabilities are often worth accepting. The key move is reframing the status quo (no superintelligence) as itself deadly: 170,000 people die every day of disease, aging, and other tragedies, while existential risk from other technologies increases regardless. Against the Yudkowsky/Soares position (articulated in their recent book If Anyone Builds It, Everyone Dies) that advocates a global ban, Bostrom counters that if superintelligence could deliver radical life extension, the expected-value calculus shifts dramatically. His models suggest the threshold is startlingly high — under conservative assumptions, developing superintelligence increases remaining life expectancy as long as the probability of AI-induced annihilation is below about 97%.
The paper's recommended strategy is captured in the memorable phrase: swift to harbor, slow to berth — move quickly to AGI capability, then pause briefly before full deployment. But Bostrom also cautions that poorly implemented pauses could do more harm than good. He explicitly says he doesn't think this work implies specific policy prescriptions, given its many simplifying assumptions, and that his main purpose is highlighting key considerations and tradeoffs.
The Major Lines of Criticism
1. The assumptions do enormous hidden work. One of the sharpest critiques, raised on Hacker News, is that Bostrom's models depend heavily on what the AI can do — reducing mortality to 20-year-old levels, yielding ~1,400-year life expectancy, curing all diseases, and developing rejuvenation therapies. If "AGI" meant something much less capable — systems that are economically transformative but can't solve aging within a relevant timeframe — the upside shrinks dramatically and the case for tolerating high catastrophe risk weakens considerably. Bostrom's choice not to impose precise definitions of AGI/superintelligence was seen by some as conveniently sidestepping the fact that the model's conclusions are deeply sensitive to capability assumptions.
2. The person-affecting framing is a controversial choice. On LessWrong, WilliamKiely pointed out that an impersonal (x-risk-minimization) perspective seems much more plausible than a person-affecting one, and questioned whether the analysis from a person-affecting view is useful even if person-affecting views are wrong. Bostrom himself acknowledged he had originally planned to analyze optimal timing from an impersonal perspective as well but cut it for length — a decision some see as leaving the most important analysis undone.
3. Missing considerations: s-risks and species-level values. The LessWrong linkpost by Julian Bradshaw identified what it called probably the biggest weaknesses of Bostrom's treatment: (a) discounting how much people care about the continuation of the human species, separate from their own lives or those of loved ones; and (b) ignoring the possibility of s-risks (suffering risks) worse than extinction. The commenter argued these aren't even that "arcane" — permanent victory for one's ideological enemies, or authoritarian AI-enabled takeover, are concrete bad outcomes people can easily imagine and would weigh heavily against rushing.
4. Skepticism about superintelligence-as-cure-all. A detailed comment on Hacker News by user "jibal" challenged nearly every premise: that all diseases are curable, that anti-aging is physically feasible, that superintelligence functions as a kind of genie or magic wand. The commenter argued Bostrom treats unproven technological possibilities as near-certainties, comparing it to assuming an ASI could find the largest prime or solve the halting problem. The broader charge was that the paper is "apologetics, not an intellectually honest search for truth."
5. Asymmetry between destructive and constructive capability. Multiple HN commenters noted a disturbing asymmetry in Bostrom's framework. User "wmf" observed that it sounds much easier to wipe out ~90% of humanity than to cure all diseases, which could create a "valley of doom" where the downsides of AI arrive well before the upsides. User "cameldrv" reinforced this by pointing out that humans have driven countless species to extinction without even trying, and that an AI with substantial resources could disrupt humanity's long, fragile supply chains — without needing anywhere near the capability required to cure all diseases.
6. The gerontocracy problem. The LessWrong discussion raised the politically uncomfortable point that if AI begins delivering promising biomedical results before full AGI, the gerontocratic political class in Western democracies will be extra-incentivized to rush toward superintelligence for personal longevity reasons — essentially that Bostrom's framework, even if not intended as policy, could become a rationalization for powerful people who have the most to gain personally from speed.
7. The "everyone dies anyway" rhetorical move. Bostrom's framing that "if nobody builds it, everyone dies" drew pushback from multiple directions. HN user "jibal" called this a red herring, noting that everyone is going to die regardless, and that the paper misframes the issue by treating natural mortality and existential catastrophe as morally equivalent. The surgery analogy was seen by some as doing rhetorical rather than analytical work — death from aging over decades is qualitatively different from the sudden annihilation of the entire species.
8. Risk aversion and irreversibility. User "wmf" quoted the observation that "death is so final" — if we lose the game, we don't get to play again, and whether anyone should accept the deal depends on their risk aversion, which the expected-value framing systematically underweights. Another commenter noted the irony that people with 1,400-year life expectancies would themselves become extremely risk-averse about anything threatening their lives — a second-order effect the model ignores.
More Sympathetic Takes
Not everyone was critical. One commenter noted that regenerative medicine has a massive manpower issue, so even sub-ASI robotic lab work should appreciably move the needle on aging research. Tyler Cowen's Marginal Revolution post signaled the paper without heavy editorial comment, treating it as a serious contribution. Jaime Sevilla of Epoch AI called it "hardcore." And some on HN appreciated Bostrom for at least modeling the tradeoffs explicitly rather than defaulting to either pure doomerism or unbounded techno-optimism — even if the model's parameters are debatable, the structure of the argument forces people to articulate which parameters they disagree about.
The Broader Context
The paper is landing in a particular moment. It's partly a direct response to the Yudkowsky/Soares book advocating a permanent ban, and partly a continuation of Bostrom's evolving position — notably different in emphasis from his 2014 Superintelligence, which was primarily focused on danger. The shift toward explicitly weighing the opportunity cost of delay (measured in daily deaths from disease and aging) represents either a maturation of the EA/x-risk discourse toward cost-benefit realism, or — depending on your priors — a drift toward rationalizing the status quo of breakneck AI development. As one HN commenter put it: there's far too much money in AGI for people to actually take a step back, and the sort of analysis Bostrom provides may end up giving intellectual cover to what's happening regardless.
