Google co-founder Sergey Brin announced that the company's AI model, Gemini, is now capable of recursively improving itself, stating that it assists in the development of its next iterations. During a roundtable discussion on the flagship model Gemini 4 Argon, which was made public this week, Brin emphasized the collaborative efforts between developers and AI, citing previous models like Flash as evidence of progress. He hinted that by the time the discussion video was released, the company would already be working on advanced models.
Simultaneously, Google DeepMind released a safety report on Argon, revealing that significant breakthroughs in AI research have not materialized yet. The report, based on surveys from twelve senior executives and specialists, indicates that while over 75% of the working code is generated by AI, the acceleration of AI research remains low. The percentage of code contributions made entirely by AI has increased from 7.5% in April to 11.9% in September, yet around 88% of changes still require human oversight.
Despite the advancements, human intervention remains crucial in directing research priorities, experimental architecture, and diagnosing training challenges. The Argon model addresses only about half of DeepMind's internal test tasks. Furthermore, the internal improvement index for Gemini, based on 92 assessments, shows a slight uptick after April 1, 2026, but the difference is statistically insignificant. The company attributes its performance above the trend to increased investments and computational power rather than to AI-driven acceleration.
Vihan Jain, who leads the RSI and AI for Science division at Google DeepMind, remarked that Argon is integrated into its own evolution. He noted that when the team responsible for the model's fine-tuning begins to request Argon resources instead of raw chips, it signals a new phase of development.
Experts from Anthropic, OpenAI, and Google DeepMind, interviewed for a report released this week, anticipate that true recursive self-improvement in AI could commence within the next 3 to 12 months, potentially sooner if governmental intervention does not impede progress.
The conversation around self-improving AI has reached legislative levels, as U.S. Representative Ro Khanna recently announced a bill aimed at banning such technology, while OpenAI has called for the U.S. to take the lead in establishing global standards for AI development.
This development reflects a significant moment for the market, indicating that while Google is pushing boundaries in AI capabilities, the lack of rapid acceleration in research may offer competitors a chance to catch up in the race for advanced AI technologies.
Informational material. 18+.
