Google Exits AI Race, Bets on World Models Instead of Self-Improving Agents

The article argues that Google, despite appearing to lag in the AI race, has actually made a deliberate strategic choice to pursue a different path to artificial general intelligence. The author, Alberto, presents this as an informed hypothesis based on public evidence but notes it lacks official confirmation from Google's executives.

The crux of the disagreement centers on recursive self-improvement (RSI). OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei believe that RSI—where one AI model autonomously creates a better model, which creates an even better one—is the key to achieving AGI. They follow Richard Sutton's "Bitter Lesson," which posits that the best approach is brute-force scaling of algorithms like learning and search, trusting that better models with enough compute will discover necessary innovations themselves. Engineering staff at both companies now manage swarms of coding agents rather than write code directly. Claude Code has written Claude Code and Claude Cowork; GPT-5.6 Sol autonomously post-trained GPT-5.6 Luna. Anthropic co-founder Jack Clark predicted in May that there is a 60% chance RSI will be achieved by 2028, and OpenAI staffers have said they would "press a magic button" for a coordinated global slowdown.

In contrast, Google DeepMind CEO Demis Hassabis—a pioneer of modern AI—appears skeptical that this is the correct approach. Instead, Hassabis is steering Google toward world models: systems that understand and simulate the real world rather than merely predict the next token. This philosophical difference explains why Google, despite its resources and research legacy, appears to be falling behind in headlines and model rankings.

The timeline matters. As of July 21, 2026, Google's Gemini Flash 3.6 ranked 10th on Artificial Analysis's intelligence index, behind every other frontier lab. Meanwhile, Anthropic and OpenAI have shown remarkable acceleration. Anthropic achieved high enterprise LLM API market share by December 2025, surpassing both OpenAI and Google. OpenAI, after recognizing Anthropic's momentum in March 2026, underwent a major restructuring led by CEO of Applications Fidji Simo, who declared the company in "code red." The restructuring cancelled secondary projects like Sora, Atlas, and e-commerce features to concentrate compute and talent on agents and enterprise products.

The shift by OpenAI was not primarily about chasing revenue, the author argues, but about competing for the trajectory toward RSI. Anthropic itself had made a bold claim in April 2023: "We believe that companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles." This concept of "escape velocity" means that once a company reaches a certain performance threshold during exponential improvement, its next model release will be accelerated because the previous model participates in training and post-training the next. As of early 2026, this prediction has proven accurate. Andrej Karpathy's decision to join Anthropic in early 2026 as a member of technical staff underscores the momentum: Karpathy, a renowned AI researcher, had previously built a miniature version of RSI in his autoresearch project, discovering that AI agents could autonomously improve code efficiency by approximately 10%. He joined Anthropic's team focused on teaching Claude to improve the pre-training phase of the next Claude version.

Google, however, is investing in a different direction. The article notes that Google has projects called Gemini, Antigravity, and Omni, but none are competing at the frontier of the code-plus-agents-plus-enterprise race. The author views this not as failure but as strategic withdrawal. Google's bet on world models is a bet that AGI will come from systems that model the physical world, not from scaling text prediction and coding agents. If Hassabis is correct, Google will eventually emerge as the true leader. If he is wrong, Google risks permanent obsolescence—because as Anthropic predicted, the companies training the best models in 2025-26 may indeed be too far ahead for anyone else to catch up in subsequent cycles. The stakes could not be higher.

The article presents an informed speculation that Google's apparent weakness in the AI race is actually a strategic withdrawal rooted in a fundamental philosophical disagreement about the path to AGI. While OpenAI and Anthropic have organized themselves entirely around recursive self-improvement—with engineering teams now managing coding agents rather than writing code—Hassabis and Google's leadership appear skeptical that this approach will succeed. The evidence the author cites includes Google's stagnant model release pace, the departures of key researchers, and an underwhelming I/O event earlier in 2026, all occurring while OpenAI and Anthropic accelerated their releases and revenue growth.

The timing of this shift is critical. Anthropic claimed in April 2023 that companies training the best models in 2025-26 would be "too far ahead for anyone to catch up in subsequent cycles"—a prediction that appears to be materializing. Anthropic and OpenAI now dominate enterprise LLM API market share and are locked in a tight competition around coding agents and autonomous AI development. OpenAI's recent restructuring in March 2026, which CEO Sam Altman used to cancel "side quests" and refocus on RSI, suggests the company views Anthropic's trajectory as a direct threat requiring full organizational commitment.

Google's position is paradoxical. Unlike startups that depend on near-term AI revenue, Google's existing business can subsidize a years-long research bet on world models without immediate return. That buffer is both Google's advantage and its risk: if world models do lead to AGI, Google will have positioned itself correctly; if RSI wins, Google will have ceded the exponential climb to competitors and may not recover.