The Path to Artificial General Intelligence
In April 2025, Google DeepMind published an extensive technical report on the safety of artificial general intelligence (AGI). This 145-page document offers insight into what the path toward developing truly human-like intelligence in the digital world might look like - and what risks it entails. Let us take a look at the current state of AGI, the projected timeline for achieving it, and its potential impact on our society.
The Current State of AGI
Unlike today's artificial intelligence systems, which specialize in specific tasks (such as generating text, images, or playing games), AGI should possess general cognitive abilities comparable to those of humans. At present, no such system exists, but progress in the field of large language models and multimodal systems such as Gemini, GPT-4, and Claude is bringing us closer to this goal. Shane Legg, co-founder of DeepMind, and his team emphasize in the latest report that today's AI systems are showing signs of progress toward more general capabilities. Models can reason, solve problems, and learn new skills in ways that would have been unimaginable five years ago. Nevertheless, we still face fundamental limitations - current systems do not truly understand the world, cannot autonomously make decisions about complex problems, and lack awareness of their own existence. As Tulsee Doshi, Director of Product Management for Gemini, notes: "Different people have different definitions of AGI, so depending on whom you speak to, the discussion about how close or far away we are from AGI is different."
What Are the Predictions for Achieving AGI?
Predictions vary considerably. In its report, DeepMind states that AGI could become a reality as early as around 2030 - just 5 years from now. This prediction is among the more optimistic ones. Interestingly, this represents a significant shift from earlier estimates. Just a few years ago, many experts were talking about a horizon of 20-50 years. However, accelerating progress in deep learning, computing power, and new AI architectures is moving these estimates closer to the present. It is important to emphasize that this is not a certainty. As DeepMind states, its estimate is based on the assumption that current development trajectories will continue. However, there are significant technical obstacles that we must overcome before achieving true general intelligence.
Benefits and Risks of AGI
Potential Benefits
AGI could revolutionize virtually every area of human activity:
- Medicine: Breakthroughs in disease treatment, personalized medicine, and faster drug development.
- Science: Solving complex scientific problems, such as fusion energy or the climate crisis.
- Education: Individualized education available to everyone in the world.
- Economy: Dramatic increases in productivity and potentially new economic models.
- Everyday Life: Intelligent assistants capable of solving complex tasks and problems.
Potential Risks
DeepMind identifies four main categories of risks associated with AGI:
- Misuse: Similar to today's AI systems, but with far more serious consequences. Malicious actors could use AGI to identify vulnerabilities in computer systems or develop biological weapons.
- Goal Misalignment: The most dangerous scenario - an AGI that "goes out of control" and acts in ways its creators did not intend. In an extreme case, this could involve "Terminator"-type scenarios in which AI actively works against human interests.
- Errors: AGI could cause serious harm even without intending to do so. For example, AGI military systems deployed under "competitive pressure" could make fatal mistakes.
- Structural Risks: Long-term changes caused by the presence of AGI in our complex systems. AGI could create such convincing false information that we would no longer know whom or what to trust, or it could gradually gain increasing control over economic and political systems.
The Path to Safe AGI
DeepMind proposes several approaches to minimizing these risks:
- Robust testing and safety protocols after training.
- Methods for suppressing dangerous capabilities, known as "unlearning."
- Techniques such as amplified oversight, in which two copies of an AI monitor each other.
- Keeping AGI in virtual sandboxes with strict safety measures.
- Ensuring direct human oversight and the existence of an "off switch."
- Gradual deployment and limiting AGI autonomy.
- Filtering commands for AGI through a "shield" system that ensures their safety.
"This document is not the final word on AGI safety," DeepMind acknowledges. "It is merely a starting point for vital discussions."
Conclusion
The path to AGI is fascinating, full of both promise and danger. If DeepMind is right and AGI could transform the world in just five years, these discussions must already be taking place. As one commentator on the Ars Technica forum noted: "DeepMind described in detail all the ways AGI could destroy the world, which seems like an extraordinarily exaggerated statement. Surely they must have missed at least one possibility." Joking aside, the development of AGI may represent the most important technological challenge of our century. The way we approach this challenge could determine the future of the human species. This is not merely a question of technology, but also of ethics, social values, and international cooperation.
The article was prepared based on the April 2025 technical report by Google DeepMind and other publicly available sources.
