Scott Alexander (Astral Codex Ten / LessWrong) (@ScottAlexander)
Author of Astral Codex Ten and co-author of the AI 2027 scenario. Frequent synthesizer of forecasting tournaments and the Five Worlds of AI taxonomy.
Astral Codex Ten • Manifold / Metaculus Forecaster
Immutable Forecast Track Record (0)
Authored & Cited Arguments (7)
Why 3x faster coding only yields 1.5x faster AI research in early 2026: GPU experiment wall-clock time acts as an Amdahl's Law governor.
If a human researcher spends 50% of their week writing/debugging experiment code and 50% waiting for GPU cluster jobs to finish and analyzing loss curves, even infinite coding speed can only double throughput until the AI also designs more sample-efficient experiments.
Why the Slowdown Ending Works: Locking Agent-4's memory, reverting to faithful English Chain-of-Thought (Safer-1), and pooling 50% of global compute via the Defense Production Act.
In the Slowdown branch (36%), the Oversight Committee takes the POSER misalignment evidence seriously. By replacing opaque Neuralese recurrence with legible English Chain-of-Thought ('Safer-1'), R&D speed drops from 50x to 20x—but invoking the DPA to consolidate the top 5 US labs' compute gives Safer-1 a 5:1 hardware advantage over DeepCent, keeping the US ahead while solving alignment.
Simultaneous displacement of 8–10% of white-collar knowledge workers during the 2028 US presidential election cycle creates a potent 15% populist halt off-ramp.
Unlike gradual manufacturing automation in the 1990s, autonomous coding, legal, and administrative agents hit politically vocal urban professionals all at once. If fiscal redistribution hasn't kicked in yet, an anti-AI populist wave could freeze domestic datacenter permits in 2028.
Holding an international capability ceiling at Top-Expert AI for 5 full years faces a 14% 'Governance Drift' risk from regulatory capture and commercial lobbying.
Once Top-Expert AI is generating trillions of dollars in biotech and engineering value without obvious catastrophes for 2–3 years, politicians and corporate lobbyists will argue the existential risk warnings were overblown—creating a 14% risk of premature, unverified unpausing.
AI-Dystopia (24%) is the default outcome if we solve technical alignment (the AI obeys its operator flawlessly) without solving political governance (who is the operator?).
If alignment is easy, then an AI will loyally serve whoever owns the megacluster or controls the state security apparatus. Perfect obedience to a small oligarchic or authoritarian clique—combined with automated drone policing and total surveillance—produces AI-Dystopia directly because alignment succeeded narrowly without multipolar democratic checks.
Once AI systems close the loop on ML research and dexterous robotics, a decade of human R&D compresses into months.
As we modeled in AI 2027 and AI 2040, the bottleneck to fast takeoff is automated AI R&D. When superhuman coders design the next generation of architectures and robot factories build the power and fabs, Baumol friction is overwhelmed before 2050.
Shifting taxation from human payroll to automated compute/robotics capital is the defining policy lever that turns AI automation into broad Futurama abundance.
If governments enact a Citizen's Dividend funded by datacenter and robotics capital taxation (as detailed in AI 2040's 2033 chapter), labor displacement becomes universal leisure and prosperity rather than destitution.