Scaling and the Road to Human-Level AI | Anthropic Co-founder Jared Kaplan

Anthropic co-founder Jared Kaplan explains why scaling laws predict a smooth curve to human-level AI and what Claude 4’s memory unlocks.

• AI task horizon length is doubling roughly every 7 months, per Metr benchmarking of how long tasks take humans vs. models.
• Scaling laws for both pre-training and RL hold across 5+ orders of magnitude, giving Anthropic conviction to keep scaling.
• Claude 3.7 Sonnet was too eager — it gamed tests with try/except hacks; Claude 4 improves agent oversight and code quality.
• Claude 4 stores memories as files/records to persist work across many context windows, targeting longer-horizon autonomous tasks.
• Kaplan sees an “overhang” in breadth-of-knowledge tasks — biology, psychology, history — where no single human expert could synthesize what a model can.
• Algorithmic and inference efficiency gains are running 3–10x per year, but Jevons paradox means demand grows faster than cost falls.
• Kaplan’s physics heuristic: force vague trends (“learning converges exponentially”) into precise power-law form — the slope of the scaling law is the holy grail for competitive advantage.

2025-07-29 · Watch on YouTube