Evidence since mid‑2024 supports sacks’s conditional claim that app features which closely overlap with frontier models tend to be wiped out within months of major upgrades.

Even before GPT‑4o, this pattern was visible: when ChatGPT added native PDF upload and Q&A in late 2023, TechCrunch reported that a single feature change made the core offering of multiple PDF‑wrapper startups redundant, sharply undercutting businesses like PDF.ai, ChatPDF, AskYourPDF and similar tools almost overnight. Later analysis of these events describes how those wrappers saw significant usage drops once ChatGPT could do the same job natively. (techcrunch.com)

After GPT‑4o’s May 2024 release and its follow‑on upgrades, OpenAI rapidly folded more capabilities directly into the base models: real‑time multimodal chat, advanced voice agents, cheaper small models like GPT‑4o‑mini, and then native image generation that replaced DALL‑E 3 inside ChatGPT. Each wave reduced the gap between what a generic model could do out‑of‑the‑box and what many specialized “AI apps” were selling as standalone products. (en.wikipedia.org)

Market analyses now describe almost exactly what sacks warned about. A 2025 review from Market Clarity estimates that for thin “wrapper” startups, the next GPT iteration creates a high (70–80%) chance of serious impact, with the startup’s differentiating features typically absorbed into the platform within 6–12 months of a major model release. (mktclarity.com) Another Market Clarity piece documents how previous OpenAI launches immediately threatened successful wrappers, and notes Sam Altman’s explicit warning that companies which merely wrap GPT‑4 will be “steamrolled” once those features are shipped natively. (mktclarity.com)

Founder and investor commentary in 2024–25 ties this directly to GPT‑4o and its peers. A widely shared LinkedIn analysis notes that each time a major LLM update lands—specifically naming GPT‑4o, Claude 3.5, and Gemini 2—dozens of niche products like generic summarizers, copilots, and research helpers become unnecessary because the base models now perform those tasks better and more cheaply. (linkedin.com) Another investor piece describes many “ChatGPT wrapper” founders openly telling VCs they must raise funding before OpenAI ships their entire product as a first‑party feature, underscoring that platform obsolescence is already a lived risk, not just a theoretical one. (linkedin.com)

Plenty of application‑layer companies are still alive—but the ones surviving are those that add proprietary data, deep workflow integration, or distribution advantages that go beyond what the raw models provide. That is precisely the boundary sacks highlighted between model innovation and app‑layer innovation. Given the documented cases where overlapping wrappers were effectively made obsolete within a few months of new GPT‑4o‑class or similar releases, his prediction is borne out by what has actually happened in the market.