Is AI Starting to 'Build AI'? Google Raises $25 Billion in Debt, Jeff Dean Exits to Launch Startup — Experts: All-in on RSI

Amid growing market skepticism over the return on AI capital expenditures, Alphabet, Google's parent company, is stockpiling resources through a record-breaking bond issuance to fuel its next technological gamble.

Alphabet recently announced plans to issue $25 billion in bonds, attracting over $115 billion in subscriptions—an oversubscription ratio of 4.6 times—making it the third-largest investment-grade bond transaction globally this year. This massive capital influx points to a clear strategic direction: AI self-evolution through Recursive Self-Improvement (RSI).

Jasjeet Sekhon, Chief Strategy Officer at Google DeepMind, acknowledged that AI-generated revenue currently cannot cover the enormous capital outlays. However, he emphasized that the risk of not betting on a technological inflection point is even greater.

He stated that RSI is the core logic behind Google’s massive capital spending. RSI enables AI to autonomously review and optimize its own algorithms, architecture, and training processes with minimal human intervention, thereby spawning more powerful next-generation models.

Sekhon likened this to "using a steam engine to build a more advanced steam engine." While the technology is still immature, early signs of model-assisted design are already emerging.

This strategic move has triggered a chain reaction within Google. Jeff Dean, Google’s Chief Scientist and AI pioneer who served for 27 years, recently departed, taking three top-tier talents to launch a new company, Discovery Loop, aimed squarely at automating scientific research and advancing RSI.

Before leaving, Dean predicted in an interview that by 2027, machine learning systems will achieve significantly higher automation, enabling self-improvement through massive experimentation.

RSI has become a new focal point in the race among tech giants. OpenAI introduced the "RSI Index" in its GPT-5.6 release notes to measure a model’s self-evolution capability and has brought back renowned researcher Li翁 to lead the related team.

Anthropic disclosed that, with AI-assisted development, engineers’ quarterly code delivery volume increased eightfold between 2021 and 2025. The company predicts that, given sufficient computing power, the era of AI fully autonomously developing its next versions may arrive sooner than expected.

A recent research report by Industrial Securities defines RSI as the new主线 for large model iteration, categorizing it into three stages: "Qualified," "Parity," and "Surpassing."

Although the technology remains in its infancy—currently limited to code generation and narrow self-optimization—Anthropic notes that the next-generation model’s focus will be on building a "self-iterative loop."

With massive capital flows and top talent converging, the tipping point for AI toward autonomous evolution may be closer than imagined.