My AI Opinions

Scott's canonical one-stop statement of his AI worldview (~2025), written explicitly so he can refer people to it after a misinterpretation spat. Four sections, each a set of explicit probability estimates with arguments-for-earlier/later. Timelines: defines AGI (90% of knowledge work) at 25%/50%/75% by 2027/2034/2045, then cleanly decomposes the muddle of 'AI takeoff' into four separately-estimated gaps — the diffusion gap (AGI-capable to half-of-jobs-actually-done), the superhuman gap (to smarter-than-top-geniuses), the Bostromian superintelligence gap (to a-subjective-century-of-progress-per-year), and the point of no return — which is the post's most useful analytic contribution. Safety: p(first PONR-crossing AI wants to eliminate humanity) = 50% on pure corporate incentives, 20% given decent alignment effort, via the 'dumbest AI that can solve alignment' argument and an honest optimism/pessimism ledger (RLAIF dragging along the good-vector vs. sandbagging). Geopolitics: ~40% chance of a well-designed US-China pause before PONR; the candid 'I round p(doom) to 20% because haters get mad at 18%' aside. Other outcomes: only 20% chance of a lasting AI underclass, 40%/20% utopia-to-them/utopia-to-us, and a striking 66% on the simulation hypothesis via an indexical refinement of Bostrom (proximity to the singularity is itself Bayesian evidence you're simulated). Classic Scott: calibrated, self-aware about its own ignorance, deferring to AI Futures Project medians.

Quality 75 · Excellent. Low-Excellent / top of the strong-AI cluster. Real value as the definitive, well-organized, intellectually honest map of Scott's AI views with numbers attached, and the four-gap decomposition genuinely clarifies discourse that usually collapses 'takeoff' into one blob. But it is by design a consolidation/reference list, not a single crafted argument or new reframing — the substance is largely Scott's and the community's existing positions assembled in one place. Comparable to the strong AI essays this census (RLHF-bandaid 75, anti-caution 75, embryo-selection 75); the clarity and the gap framework hold it at the top of that band rather than above it. 75.

Claude’s paradigm shift 45 · Moderate. Moderate. Synthesizes live 2025 AI-safety thinking (Bostrom, METR time horizons, RLAIF, emergent misalignment, AI Futures Project) rather than introducing a new frame; the gap-decomposition is a useful organizing device and the simulation-indexical point a refinement of Bostrom, but these are clarifications within an active debate, not paradigm shifts in their moment. 45.

Real-world impact 2 · Minor. The canonical one-stop statement of Scott's AI worldview with explicit probability estimates, whose four-gap decomposition genuinely clarifies 'takeoff' discourse that usually collapses into one blob. Real relevance as a reference within AI discourse, but a consolidation/reference list with no material change — low RWI.