The author expresses skepticism about the effectiveness of Reinforcement Learning (RL), especially the Reformist RL approach. While they appreciate the clarity this perspective offers, they argue that RL is fundamentally inefficient. The core method, policy gradient, often requires immense iterations with the environment to yield results, leading to slow and unreliable outcomes. The author highlights that theoretical results for RL tend to be negative, pointing out that even simple tasks can require an impractical number of samples to optimize, particularly in complex environments like video games. Instead of relying on policy gradient, the author proposes the principle of certainty equivalence as a more efficient alternative. This involves building a model of the environment and optimizing as if that model were accurate. The author cites that this approach has proven to be optimal in various settings, such as multi-armed bandits and MDPs, and allows for faster convergence by leveraging additional signals beyond rewards. They argue that many successful control designs already operate on this principle, making the case that certainty equivalence could also apply to reasoning models, which the author is currently exploring. The author reflects on recent developments in RL as applied to reasoning models, particularly in the context of large language models (LLMs). They note that current methods seem to reduce to a "guess-and-check" approach, where models are fine-tuned based on performance on benchmarks. In their analysis of recent papers, they found that variations in sampling methods could yield results similar to RL, suggesting that substantial improvements in training efficiency might be possible. The author is intrigued by the potential for discovering alternative methods that could drastically shorten training times for reasoning models, speculating that advancements could lead to enhancements in efficiency by factors of 100 or even 1,000.