AI Hurtles Ahead

Howard Marks argued that AI crossed a threshold in early 2026 from a productivity tool into a labor substitute, framing this shift as the difference between a faster horse and the automobile. Drawing on a nine-module tutorial built around his own intellectual frameworks and a widely-read blog post by OthersideAI CEO Matt Shumer, Marks described three capability levels: Level 1 (chat), Level 2 (tool use), and Level 3 (autonomous agents that set their own sub-tasks, self-test, and deliver finished work). He noted AI was at Level 1 in 2023, Level 2 in 2024, and reached Level 3 in early 2026 — a pace he said has no historical parallel. On the bubble question, Marks maintained his December 2025 position: the technology is real and demand is growing faster than supply, making underestimation more likely than overestimation. He drew a distinction between hyperscalers such as Microsoft, Amazon, and Alphabet — whose diversified profits make ruinous overvaluation unlikely — and early-stage AI startups he described as lottery tickets. He flagged that some AI revenue remains circular, flowing between AI companies rather than from end users paying for genuine economic value, and that pipeline infrastructure spending may yet outrun demand. Marks reserved his sharpest concern for labor displacement, arguing AI's speed of adoption will outstrip society's ability to retrain workers — faster and across more job categories than offshoring did to manufacturing. He closed by reiterating his December advice: neither all-in nor all-out, but a moderate position applied with selectivity and prudence.