Google hasn't lost the AI race; it has deliberately stepped out of it. While OpenAI and Anthropic are racing toward recursive self-improvement (RSI), a process where AI systems autonomously build better versions of themselves, Google DeepMind's leadership is pursuing an entirely different path: world models that understand and simulate the real world rather than simply predicting the next word.

Why Is Google Abandoning the Recursive Self-Improvement Race?

DeepMind CEO Demis Hassabis, a pioneer in artificial intelligence, has steered Google away from the scaling-focused approach that dominates OpenAI and Anthropic's strategies. According to analysis of public statements and strategic decisions, Hassabis views the recursive self-improvement path as either a dead end or merely a temporary off-ramp, not the true route to artificial general intelligence (AGI), the kind of AI that performs as well as humans across all tasks.

The contrast is stark. OpenAI and Anthropic have embraced what's known as the "Bitter Lesson," a principle suggesting that brute-force scaling of compute and algorithms, rather than clever architectural innovations, drives AI progress. Both companies are betting heavily on coding agents, AI systems that can write and improve code better than humans, to accelerate their path to RSI.

"RSI is coming," stated researchers at Anthropic, with some predicting a 60% chance of achieving recursive self-improvement by 2028.

Anthropic researchers, as reported in company communications

At OpenAI and Anthropic, the shift is already visible in how engineers work. Staff members report barely writing code anymore; instead, they manage swarms of coding agents like Claude Code and Codex that handle the programming. Claude Code has even written Claude Cowork, and GPT-5.6 Sol "autonomously post-trained" GPT-5.6 Luna, demonstrating AI systems improving other AI systems.

What Is Google Betting On Instead?

Hassabis is steering Google toward world models, AI systems designed to understand and simulate the real world rather than simply predict text tokens. This represents a fundamentally different theory of how intelligence works and how to build AGI. Rather than scaling datacenters and letting compute solve all problems, this approach focuses on building models that can reason about physical reality.

This strategic divergence matters because it reflects a deeper disagreement about the nature of intelligence itself. OpenAI and Anthropic believe that with enough compute and the right algorithms, AI systems will autonomously improve themselves toward AGI. Google DeepMind's leadership believes that understanding the world, not just predicting sequences, is the missing piece.

How Google's Unique Position Shapes the Outcome

Google faces a paradoxical situation. Unlike OpenAI and Anthropic, which are startups that must prove AI is a viable business to survive, Google is an incumbent with massive existing revenue from search and advertising. This gives Google a critical advantage: it can subsidize AI research for years without needing immediate returns.

However, this same position creates existential risk. If Hassabis is wrong about world models and right about the importance of RSI, Google could become irrelevant while competitors race ahead. Conversely, if world models are indeed the path to AGI, Google's willingness to pursue a longer-term, less flashy strategy could position it as the ultimate leader.

The evidence of Google's withdrawal from the traditional AI race is visible in several ways:

• Slower Release Pace: Google has released models at a slower cadence than OpenAI and Anthropic, signaling a shift away from the rapid-iteration strategy both competitors are pursuing.
• Leadership Departures: Several heavyweight researchers have left Google for Anthropic, including Andrej Karpathy, who joined as a member of technical staff to work on making Claude improve its own pre-training process.
• Underwhelming Public Events: Google's I/O conference earlier in 2026 failed to generate the excitement or demonstrate the momentum that OpenAI and Anthropic have achieved with their recent releases.

Karpathy's departure is particularly telling. The AI researcher had built a miniature version of recursive self-improvement in his autoresearch project, discovering that AI agents could autonomously improve code efficiency by roughly 10%. He recognized the potential of RSI and joined Anthropic to scale that approach, signaling that he believed the recursive self-improvement path was viable.

What Does This Mean for the AI Industry?

The AI race is now fundamentally a race between two competing theories of intelligence and two different types of organizations. Startups like OpenAI and Anthropic are betting everything on recursive self-improvement and the scaling laws derived from the Bitter Lesson. They have no choice; they need AI to become a transformative business to justify their valuations and funding.

Google, by contrast, can afford to be patient. If world models prove to be the correct path to AGI, Google's long-term bet could pay off spectacularly. If recursive self-improvement is the answer, Google risks becoming a laggard in an industry it helped pioneer.

The stakes are enormous. Anthropic and OpenAI have achieved near-total dominance of the AI startup sector, with the two companies accounting for roughly 90% of annualized revenue in the AI startup space. OpenAI has achieved this despite a major weakness: only 2.2% of US households are willing to pay for an AI subscription, forcing the company to rely heavily on enterprise and API revenue.

Yet neither company is ahead of the other in any truly relevant metric. They are essentially tied on revenue, reach, and ambition. OpenAI has recently fixed its main weakness by cutting secondary projects and focusing on core capabilities, while Anthropic has maintained an unbreakable focus on coding agents and enterprise customers.

How to Understand Google's Strategic Gamble

• Recognize the Philosophical Divide: The AI industry is split between those who believe scaling compute solves all problems and those who believe understanding the world is essential. Google has chosen the latter path, which is a fundamental departure from industry consensus.
• Monitor World Model Progress: Watch for announcements from Google DeepMind about advances in world models and simulation. These will be the key indicators of whether Hassabis's bet is paying off or falling behind.
• Track Leadership and Talent Flow: Pay attention to where top AI researchers are choosing to work. Departures from Google to Anthropic suggest confidence in the recursive self-improvement approach, while future hires at Google could signal renewed commitment to world models.

The outcome of this strategic divergence will shape the entire AI industry for years to come. If Google is right, the company's patience and resources could position it as the ultimate leader in AGI. If it is wrong, Google risks becoming a footnote in a story written by OpenAI and Anthropic. For now, the only certainty is that Google has made a deliberate choice to pursue a different path, and the results of that choice will not be known for years.