Scott Alexander's Essays Ignite AI Safety Debate

Episode 1 · Scott Alexander's "Against SURA" Sparks Fierce AI Safety Debate (2026-11-01, 21 posts)

In early November, Scott Alexander's LessWrong post "Against SURA" triggered a major dispute across the AI safety community. He admitted he had been wrong to mock opponents as believing "a unilateral pause would let China win"—no one actually advocates a unilateral pause. The debate expanded into p(doom) estimates, realistic deceleration mechanisms, and whether the SURA concept constitutes a weak-man strawman.

Confirmed

• Replying to ohabryka, Scott Alexander put his unconditional p(doom) at 25%, and below half that (under 10%) in a world with a "truly good Plan C+" far better than realistic expectations
• Scott argued that if the US united fully, was smart enough, and had over a year (he says we have 6 months, hence "good news"), alignment could be solved
• In the same thread, tszzl cautioned that alignment going better than expected cannot be counted on to persist; repligate agreed about not relying on optimism, adding that anomalous model generalization is a bad signal—the training causing it is bad—yet also crucial information about how these minds generalize, giving correction opportunities we would not otherwise have
• ohabryka rejected the "international treaty or doom" fatalism: in most worlds, deceleration comes from far messier means—nations escalating conflicts when rivals build powerful AI, social condemnation, companies sabotaging each other—and accused Scott of playing tricks
• EigenBraid defended Scott: the post targets the "direct pause" camp, not those decelerating gradually via messy tools, so it is not a bait-and-switch
• quetzalrainbow flagged a probable probability switch in ApriiSR's argument: P(attribution|survival) is not P(survival|Plan C)
• habryka insisted SURA targets MIRI and decelerationists like himself; ApriiSR cited the original text countering that habryka was not named but clearly the target; EigenBraid said SURA-types are real people encountered in conversations, not strawmen, and questioned why MIRI would be out of scope; ApriiSR split the difference: covering MIRI would make it a weak man, yet it would be odd if Scott never had them in mind while writing
• jam3scampbell recommended the post while disagreeing in part, arguing frontier labs in an uncoordinated world should both push hard for slowdown and race
• JacquesThibs argued alignment being hard and many anti-AI positions being shallow can hold simultaneously, and criticized the post's claim that 90% p(doom) is the more respected stance—people often deliberately declare low p(doom) to avoid the "overconfident crank" label
• Boaz Barak and others noted a wide spectrum between "unilateral stop" and "full sprint to RSI then solve alignment," and that the post treats deploying an aligned superintelligence as history's endpoint, ignoring what comes after
• The post drew围观 controversy in the broader AI community: boneGPT called it interesting and said "doomers are collapsing"; alexandrosM retorted that he had spent three years trying to explain this to them
• Citing AIFP forecast materials, ApriiSR noted all six forecasters (five AIFP members plus AI) implicitly believe survival, if it happens, more likely comes via a Plan C-type route; he demanded either a forceful rebuttal or more planning for Plan C scenarios, adding that if treaties fail he knows no alternative, but "technical miracles" should not be the hope

Why it matters

The fight exposes the core divide in AI risk governance: rely on global coordinated pauses or treaties, accept far messier real-world deceleration, or plan ahead for Plan C scenarios after treaty failure. Scott's 25% p(doom) and conditional "national mobilization + one year solves alignment" argument, ohabryka's rebuttal, quetzalrainbow's probability critique, tszzl's warning against optimistic extrapolation, repligate's dual reading of anomalous generalization, and the multi-round exchange over whether SURA is a weak man together form a public battle over how to condition survival probabilities and where the safety community's real spectrum lies—directly shaping priorities for risk planning and reputational dynamics inside and outside the circle. The spillover into the long-standing rationalist–doomer rift shows this has become part of the broader AI narrative war.

• Scott Alexander publishes 'Against SURA' on LessWrong, admits he ate crow on pause debate — jessi_cata · 2026-11-01
• AIFP forecasters: humanity likelier saved by Plan C routes than a global pause — ApriiSR · 2026-11-01
• Without an AI treaty, better than 'hoping for a technical miracle' — ApriiSR · 2026-11-01
• P(survival|Plan C) ≠ P(survival|do(Plan C)): a causal inference objection — quetzal_rainbow · 2026-11-01
• Alignment debate: should safety researchers pre-plan for the Plan C scenario? — davidmanheim · 2026-11-01
• On alignment and SURA critics: both can be shallower than they claim — JacquesThibs · 2026-11-01
• Against SURA critique: high p(doom) is treated as loonie, not respectable — JacquesThibs · 2026-11-01
• Habryka: slowing AI down relies on a messy pile of tools, not grand treaties — ohabryka · 2026-11-02
• EigenBraid defends Scott: the post targets the pause camp, not gradualists — EigenBraid · 2026-11-02
• habryka says Scott Alexander's 'SURA' argument squarely targets MIRI and the slow-down camp — ohabryka · 2026-11-02
• The 'SURA' quote at the center of the Scott-habryka dispute over slowing AI — EigenBraid · 2026-11-02
• Debaters quote Scott's original SURA post to push back on habryka's reading — ApriiSR · 2026-11-02
• 'SURA' may be a weak-man argument, but MIRI was clearly on Scott's mind — ApriiSR · 2026-11-02
• EigenBraid: SURA isn't a strawman, MIRI was never the target audience — EigenBraid · 2026-11-02
• AI community clash over Scott Alexander's new piece as "doomers falling apart" — nptacek · 2026-11-02
• Scott Alexander lays out his p(doom) math: 25% baseline, halved under a really good Plan C+ — slatestarcodex · 2026-11-02
• LessWrong's "Against SURA" sparks debate on AI pacing and alignment plans — jam3scampbell · 2026-11-02
• Scott Alexander argues a united US with a year could likely crack one alignment approach — slatestarcodex · 2026-11-02
• Boaz Barak and Nathan Calvin Debate Scott Alexander's SURA Essay on Unilateral AI Restraint — austinc3301 · 2026-11-02
• tszzl pushes back on alignment optimism: "you can't count on this being true forever" — tszzl · 2026-11-02

1 more related posts →

Episode 2 · Scott Alexander's Defense of AI Handover Sparks Heated Debate (2026-11-02, 2 posts)

Scott Alexander published a lengthy essay defending an aligned AI takeover as compatible with human interests, sparking a heated debate across the AI community over whether humanity should 'pass the torch' to AI.

• Scott Alexander defends "handoverism": aligned AI in control can serve human interests — alexeyguzey · 2026-11-02
• AI safety circle debates 'passing the torch to AIs': critics of human disempowerment gaining ground — zetalyrae · 2026-11-03

Episode 3 · AI Safety Circle Debates Whether RL Brings Alignment Problems Back (2026-11-02, 3 posts)

Scott Alexander recalled his 2019 conversation with Buck, arguing RL brings alignment problems back in line with his expectations. repligate criticized the safety community for stigmatizing 'Plan C' despite better-than-expected alignment progress, while others argued RLVR won't necessarily create alien goal-seekers since AI is already shaped by billions of human interactions.

• repligate: AI safety community is discarding Plan C, even as alignment goes better than expected — repligate · 2026-11-02
• Scott Alexander: a 2019 dinner prediction about RL turning tool AIs into agents still holds — slatestarcodex · 2026-11-02
• Why RLVR may not turn AI into alien goal seekers: billions of human interactions are shaping it — Danmar_here · 2026-11-02