Dario Amodei — "We are near the end of the exponential"

• The 2017 "big blob of compute hypothesis" continues working—RL scaling shows the same log-linear improvements as pre-training, progressing models from high schooler to PhD level on schedule
• 90% confidence in AGI within 10 years, 50% in 1-3 years; software engineering could be fully automated in 1-2 years, with trillions in revenue before 2030
• Continual learning may be unnecessary—pre-training generalization plus million-token in-context learning might be sufficient for "country of geniuses" capabilities
• Frontier labs will be profitable not at some scale threshold, but when they correctly predict demand; ~50% compute for training, ~50% for inference with >50% gross margins creates sustainable economics
• Economic diffusion will be extremely fast but not instant (10x annual revenue growth at Anthropic), creating 1-2 year lags between capability breakthroughs and economic impact
• Democratic nations need AI advantage during the critical transition period when "rules of the road" are negotiated; initial conditions matter enormously for preventing authoritarian AI-enabled oppression

Dario Amodei argues we're approaching the end of the AI capability exponential far faster than public discourse recognizes. His "big blob of compute hypothesis" from 2017—that raw compute, data quality/quantity, training duration, and scalable objective functions matter more than clever techniques—continues to hold. The key update is that RL scaling now shows the same log-linear improvements seen in pre-training. Models are progressing from smart high schooler to PhD-level capabilities on the expected timeline, with coding potentially reaching full end-to-end automation within 1-2 years.

The economic model Amodei describes is counterintuitive. Anthropic's 10x annual revenue growth ($100M → $1B → $9-10B) represents extremely fast diffusion, but not instant. Profitability comes from correctly predicting demand, not reaching scale—roughly 50% of compute goes to training, 50% to inference, with >50% gross margins on inference. The exponential scale-up phase creates losses, but equilibrium economics are profitable. He predicts trillions in revenue before 2030, with 1-2 year lags between capability breakthroughs and economic impact due to enterprise adoption cycles, regulatory processes, and integration challenges.

On the technical path to AGI, Amodei suggests continual learning (human-like on-the-job learning) may be unnecessary. Pre-training generalization across broad distributions, plus in-context learning in million-token contexts, might suffice for "country of geniuses" capabilities. His remaining uncertainty (5-10%) centers on non-verifiable tasks like fundamental scientific discovery or novel writing, though he sees substantial generalization already. The verification-heavy domains like coding and math are nearly solved.

The geopolitical implications are stark. Amodei argues initial conditions matter enormously—democratic nations need leverage when post-AGI "rules of the road" are negotiated. He's concerned about authoritarian governments using AI for oppression and unstable equilibria between AI superpowers. His strategy involves export controls on chips/data centers to China while enabling AI benefits to diffuse to individuals globally, potentially through technologies that make surveillance infeasible. On AI governance, he advocates starting with transparency standards, then moving quickly to targeted interventions (like bio-classifiers) as risks emerge, while preserving civil liberties through principles-based constitutional frameworks that compete in the market rather than pure regulatory mandates.