Most industry observers are watching the sprint to see who builds the fastest coding agent, but Alberto Romero argues the real story is that Google has quietly exited that specific race entirely. Romero contends that the tech giant isn't losing; it has made a calculated strategic withdrawal because its leadership views the current industry obsession with recursive self-improvement as a fundamental dead end. This is a crucial distinction for busy readers to grasp: we are witnessing a schism in the very definition of intelligence, not just a fluctuation in quarterly model scores.
The Great Divergence
Romero frames the current landscape as a battle between two incompatible theories of how to achieve artificial general intelligence. While competitors like OpenAI and Anthropic double down on the idea that scaling compute and letting AI write its own code is the only path forward, Google DeepMind CEO Demis Hassabis is betting on a completely different architecture. Romero writes, "Google DeepMind CEO Demis Hassabis—a pioneer of AI as we understand it today—doesn't think that automating AI research with coding agents is the correct approach to artificial general intelligence." This is a bold claim, suggesting that the most visible metrics of success—rapid release cycles and autonomous coding agents—are actually misleading indicators of long-term viability.
The author details how the startup ecosystem has embraced what Romero calls "recursive self-improvement," where one model builds a better version of itself in an accelerating loop. He notes that at these labs, "engineering staff... barely write any code anymore. It is swarms of coding agents... that do." This shift has created a sense of urgency, with insiders predicting that by 2028, there is a "60% chance that RSI is achieved." Romero points out that this belief is so strong that even skeptics like Andrej Karpathy have joined Anthropic to help refine these self-improving foundations, having previously demonstrated that AI agents could autonomously improve code efficiency by roughly 10%.
"The AI race is actually a race between two theories of intelligence and also between two kinds of company: startups that need AI to become a business, and an incumbent whose existing business can subsidize AI for years."
Romero's analysis suggests that Google's slower pace and recent underwhelming product launches are not signs of failure, but rather the result of a deliberate pivot away from the "brute force" scaling laws that Richard Sutton famously championed. Instead, the executive branch of Google's research is focusing on "world models"—systems designed to understand and simulate physical reality rather than just predicting the next word in a sentence. This connects to the historical concept of world models, which posits that true reasoning requires an internal simulation of how the world works, a theory that dates back to early cognitive science but has been largely overshadowed by the recent dominance of large language models.
However, this strategic divergence places Google in a precarious position. If the industry consensus proves correct and recursive self-improvement leads to an "escape velocity" that no one can catch, Google risks irrelevance. Romero warns, "He could make Google the absolute leader or an irrelevant laggard." The stakes are existential for the company, as the current trajectory suggests that "companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles."
The Cost of the Pivot
The evidence Romero cites to support his hypothesis is largely circumstantial but compelling. He highlights Google's recent drop in intelligence rankings, noting that their latest model landed "10th on AA's intelligence index, behind every other frontier AI lab." He argues that for a company of Google's resources, this is "an unacceptable score" unless they are intentionally sacrificing short-term performance for a different long-term goal. The author contrasts this with OpenAI's recent "code red" restructuring, where leadership cut side projects to focus entirely on the race for autonomous agents, viewing Anthropic's enterprise gains as a "wake-up call."
Critics might note that Romero's theory relies heavily on the assumption that Google's leadership is unified in this view, which is difficult to verify given the opaque nature of corporate strategy. It is also possible that Google is simply struggling to execute rather than strategically choosing a different path. Yet, Romero's framing of the situation as a philosophical split rather than a competitive failure offers a more nuanced explanation for the observed data than the standard "Google is falling behind" narrative.
Romero emphasizes that the current frenzy is driven by a belief that the future belongs to those who can automate the improvement process. He writes, "OpenAI realized that Anthropic was starting to show the signs of acceleration consistent with proto-RSI. They care about money, sure, but Altman wouldn't make a company-wide restructuring for a short-term trade." This suggests that the race is no longer about immediate revenue, but about securing the foundational architecture of the next era of computing.
"It is out of the race because it withdrew."
The author also touches on the concept of "model collapse," a phenomenon where AI models trained on data generated by other AI models begin to degrade in quality. While Romero does not explicitly detail this in the excerpt, the focus on world models as an alternative to pure token prediction implicitly addresses the risk of collapsing into a feedback loop of synthetic data. By simulating the real world, Google hopes to avoid the stagnation that might plague models relying solely on recursive self-improvement.
The Stakes of the Bet
The core tension Romero identifies is between the speed of the startups and the depth of the incumbent. Startups like Anthropic and OpenAI are betting that the next breakthrough will come from scaling up current methods until they break through to general intelligence. Google, conversely, is betting that these methods hit a ceiling and that a new architectural approach is required. Romero summarizes the high-stakes nature of this gamble: "That said, to write Google off the story would be a serious mistake. I'm going to explain why I think Google has a unique chance here."
The author's argument is strengthened by his observation that the industry is moving so fast that "anything can happen, and what counts is not so much the point as the trajectory." If the trajectory of recursive self-improvement leads to a dead end, Google's bet on world models could be the only viable path forward. Conversely, if the scaling laws hold, Google's hesitation could be fatal. Romero concludes that the current landscape is defined by this uncertainty, with Google essentially saying, "We are not playing your game."
Bottom Line
Romero's most compelling contribution is reframing Google's apparent stagnation as a deliberate, high-stakes philosophical divergence rather than a competitive failure. The argument's greatest vulnerability is its reliance on unconfirmed internal beliefs, yet it offers a necessary counter-narrative to the prevailing hype around autonomous coding agents. Readers should watch whether Google's next major release prioritizes simulation and reasoning over raw generative speed, as that will validate or invalidate this theory of a strategic withdrawal.