While OpenAI and Anthropic are focused on recursive self-improvement of algorithms, Google DeepMind, under the leadership of Demis Hassabis, is making a fundamentally different choice. Instead of chasing leadership in benchmarks, the company is investing in "world models" — AI capable of understanding physical laws, not just predicting the next token. This is not just a shift in priorities, but a fundamental pivot that could redefine the entire industry.
Why Google Lags on Paper
Outwardly, the difference is barely noticeable. On July 21, the Gemini 3.6 Flash model was released: fast, cheap, but in the Artificial Analysis ranking, it only took 10th place. The emphasis has shifted to cost — the new version generates 17% fewer tokens than its predecessor. However, in the race for "raw" power, Google is not yet the favorite.
Another alarming signal is the talent drain. In June, two leading researchers left DeepMind, moving to competitors, which caused Alphabet's stock to drop by 6%. Investors are nervous: the company is spending enormous resources but is not demonstrating clear technological superiority.
DeepMind's Strategy: AI That Understands the World
Hassabis is betting on embodied AI and models that analyze gravity, movement, and cause-and-effect relationships. In May, the lab expanded Project Genie, integrating it with Street View, and also introduced SIMA 2 — an agent that learns in virtual 3D worlds.
This is fundamentally different from the approach of competitors. OpenAI and Anthropic focus on recursive self-improvement, where AI independently designs the next generation of algorithms. Already, Anthropic's Claude has written over 80% of the code being deployed in the company's products, and in April alone, their models' performance increased 52-fold. For comparison, an experienced engineer requires 4–8 hours for a fourfold improvement.
Expert Opinion: DeepMind is the "Most Restrained"
Anthropic co-founder Jack Clark estimated the probability that AI could independently manage research by the end of 2028 at 60%. At the same time, he called DeepMind "the most restrained of the big three," citing the lab's own publications on safety issues. Anthropic, on the other hand, acts more boldly: cases have already been recorded where AI created a more advanced AI.
However, it is too early to count Google out. The company holds the lead in the MLE-Bench benchmark (Gemini 3 scored 64.4% in February), and the Gemini app audience has reached 950 million monthly users. This is a colossal reach that provides a huge advantage in data and feedback.
Financial Aspect: Patience is Getting More Expensive
Alphabet's revenue in the second quarter of 2024 reached $119.8 billion (+24% year-over-year), but expenses are growing even faster. Over three months, the company allocated $44.9 billion for data center development — almost double the amount from a year ago. As a result, free cash flow turned negative: minus $5.86 billion. To cover the deficit, Alphabet issued $49.6 billion in shares and raised $20.3 billion in loans. Long-term liabilities grew from $46.5 billion to $98.2 billion over six months.
The AI development expense line resulted in a loss of $5.79 billion (compared to $3.37 billion a year ago). The wait-and-see position is costing the corporation more and more.
What Will Be Decisive in the Next 30 Days
- Whether the release of Gemini 3.5 Pro will take place and what positions the model will achieve in tests.
- Whether DeepMind will present new data on world models for Gemini 4.
- Whether Alphabet's cash flow will return to positive values in September.
- Whether Hassabis will articulate a clear position on self-learning technology.
The launch of Gemini 4 will be a defining moment. If the world model approach proves itself, the current slowdown will turn out to be a precise calculation. If not, Google risks finally ceding leadership in a race where the cost of error is measured in billions.
Cryptalist's Comment: The market is currently underestimating the potential of world models. In the long term, AI that understands the physics of the real world could prove far more valuable than algorithms that simply generate text. However, investors demand quick results, and DeepMind has less and less time to prove its strategy right.