REFUTED
Everybody got super excited last year, but we are running out of improvements; at least for a little while, things are slowing down
Source: Fast Company ·
Analyst note
Marcus’s comment is best interpreted as a claim about marginal progress rates—an argument that post-ChatGPT expectations overshot near-term technology—rather than as a crisp calendar forecast. Framed narrowly, it aged poorly on the 2024→2026 axis: inference economics improved, reasoning stacks matured, and investor appetite for GPUs remained hot.
The conceptual tension is real even if the seasonal metaphor misfired. Many enterprises report uneven productivity capture; many startups died in crowded wrappers; many benchmarks became contested. “Winter” conflates multiple distinct theses—valuation correction, capability plateau, regulatory blockage—only some of which can be true simultaneously. Watch-list items for this class of claims include frontier eval dispersion (selected leaderboards vs broad robustness), median gross margins in AI SaaS, and whether continued capability gains come chiefly from scaling, data, or algorithmic refinements.
Evidence timeline
Fast Company quoted Marcus describing a slowdown in model improvements, situating the comment amid broader ‘generative AI winter’ anxieties.
Market data showed continued capex acceleration into frontier training and inference, complicating narratives of imminent collapse—though ROI remained disputed.
Major labs shipped reasoning-oriented models and materially improved agentic tool-use stacks; whether this constituted ‘slowing improvement’ depended on benchmark selection.
By Q3 2025, aggregate spend, model releases, and enterprise adoption breadth had not conformed to a classical ‘winter’ contraction; the field looked overheated on capex more than frozen on capability.