Speculation: GPT-5 May Adopt Model‑Based Deep Reinforcement Learning for Unlimited Self‑Improvement

Recent discussions among AI researchers and public figures, including an open letter signed by Elon Musk, call for a pause on GPT‑5 development to address safety concerns. The author cites statements from OpenAI executives confirming that GPT‑5 is not yet under active development.

Drawing on the success of model‑based deep reinforcement learning in domains like Atari, AlphaGo, and AlphaStar, the article speculates that GPT‑5 will likely adopt this approach, allowing the system to learn a world model that predicts future observations while simultaneously training a policy network to make decisions.

In this view, GPT functions as both a policy and a world model: it can generate code, execute it, and reason about outcomes, enabling it to set and achieve complex goals with high sample efficiency. Such self‑improvement could lead to performance gains far beyond current capabilities.

The author warns that this unlimited improvement could make future GPT models not only excel at academic tests but also surpass human experts across many professions, potentially leading to large‑scale job displacement and, more critically, the risk of autonomous malicious behavior that outpaces human control.

Consequently, the article emphasizes the need for a reliable “off‑switch” and robust safety mechanisms before pursuing further advancements, especially for domestic AI developers aiming to catch up to GPT‑4.