In late July 2026, Varun Singh, pre-training lead at Arcee AI, said on the AI Engineer podcast that the share of training data from the open web has collapsed from roughly 85% to 15% in five years. When OpenAI trained GPT-3 in 2020, web text from Common Crawl and Wikipedia dominated the data mix, with reinforcement learning described as "just a cherry on top." Singh argued that releases such as OpenAI's o1 in late 2024, DeepSeek-R1 in January 2025, and Claude Code demonstrated that RL can dramatically improve model performance on various tasks, making it the primary capability engine rather than a garnish. He posed the question: "is your standard base model still the best prior for the large-scale reinforcement learning phase that agentic models now use?" He suggested the answer is no, meaning pre-training data strategy must be rebuilt to focus on code, structured reasoning problems, and sequences needed for RL. Singh walked through training reports from MiniMax-M1, NVIDIA's Nemotron 3 Ultra, Kimi K2, and Arcee's Trinity Lodge, showing the pattern of declining web text share and replacement by synthetic data.