An analysis from Wall Street News points out new developments in the AI industry: frontier labs are shifting their competitive focus towards "Recursive Self-Improvement" (RSI), where AI participates in more stages of its own research and development. This trend is supported by multiple signals:

Google released its Gemini 3.8 Flash model on September 2nd. DeepMind researcher Shunyu Yao called it a "huge leap" for RSI, as the model is designed to evaluate and optimize its own results through agentic loops.

OpenAI CEO Sam Altman recently stated publicly that due to the "awe-inspiring" speed of model capability improvement, OpenAI has for the first time delayed a frontier reinforcement learning (RL) training run to ensure safe follow-up, hinting at the proximity of RSI.

SSI (Safe Superintelligence Inc.), founded by OpenAI co-founder Ilya Sutskever, recently received a strategic investment from NVIDIA and announced that its hashrate will expand tenfold in the next 12 months, emphasizing that "our research is already worth scaling up massively."

The analysis suggests that RSI could change AI training models from "one-shot" to "never-ending," where hashrate advantage directly translates into research and model advantage. Renowned tech investor Gavin Baker warns that top labs might actively sacrifice short-term revenue for long-term technological advantage, allocating more hashrate to training rather than inference. This would pose a significant risk to AI stock valuations.