Current AI Trends and the Road Ahead

The most advanced artificial intelligence (AI) models today have already achieved superhuman levels in specific areas, such as games, information processing speed, memory, and the sheer volume of knowledge. AI, for instance, can read books in seconds, translate languages, and surpass human masters in games like chess and Go. Additionally, it is improving in areas such as creativity, programming, and specialized knowledge. However, weaknesses remain, including the ability to admit ignorance, planning, and natural movement, which limit its full autonomy. AI is not yet better at hacking than the best hackers and cannot conduct AI research as well as the best AI researchers. If it reaches any of these thresholds, it may signal the onset of an era of heightened risk.

Trends in AI Development

We are witnessing rapid and groundbreaking advancements in AI learning, which present both enormous opportunities and challenges. The nonprofit organization PauseAI, along with others, emphasizes the need for a careful and responsible approach to AI development, with a focus on safety and ethics.

PauseAI does not call for a ban on all artificial intelligence but instead advocates for the development of controllable AI models. The goal is to slow down or halt the development of risky technologies that could have irreversible consequences until adequate safety measures are in place. They stress the importance of global cooperation in this field.

So far, no credible plan has been successfully implemented, and AI is developing faster than the strategies and know-how needed to align it with our needs. Perhaps we need more time.

Principles of Nonviolence and Cooperation

Historically, humanity’s approach to existential issues has often relied on power asymmetry and coercion. I believe that independent education, critical thinking, effective communication skills, and non-interference in the lives of others are needed now more than ever.

I would like decisions regarding further AI development to be based on principles of nonviolence and mutually voluntary cooperation, grounded in contemporary knowledge, multidisciplinary education, and dialogue—not coercion or force. Unilateral actions that overlook the needs and interests of others may undermine the sustainability of future actions, as they have the potential to erode trust and motivation for cooperation.

I understand the urgency expressed by supporters of the PauseAI initiative. At the same time, I consider that pressure for international regulation, along with the absence of global consensus, could, in fact, accelerate the AI development race or bring humanity closer to political singularity.

A Call for Global Cooperation

For me, voluntary cooperation among individuals and organizations worldwide is key. Here and now! The effort to achieve consensus and mutual respect for needs, limits, and differences is paramount. I hope that people will unite not only out of an interest in survival but also for the quality of life, based on fulfilling human biological, psychological, and emotional needs. I believe that humility and respect for different perspectives are essential to minimizing the risks associated with AI development.

The Proximity to AGI

Artificial General Intelligence (AGI) refers to AI capable of understanding, learning, and performing tasks at the level of human intellect across various domains, without specific programming for particular tasks. Some believe that achieving this level of intelligence could be closer than we realize, raising questions about the safety and control of these systems.

In 2023, GPT-4 reached human-level performance on most tasks involving state-of-the-art AI capabilities versus humans, including the Winograd Schema (94.4%) and the SAT exam (93%).

The exponential growth of AI capabilities also results from improvements in chips, algorithms, and data. Each of these components is rapidly advancing due to significant investments, which lead to increasing AI performance.

Exponential Technologies Grow Exponentially

The progress of AI may be underestimated due to the speed of change, shifting benchmarks, the human desire to feel exceptional, and a structural misunderstanding of exponential growth. It is possible that the exponential growth of AI will encounter limits. However, I prefer to approach the current trend of exponential growth with precaution.

A critical point to consider is the moment when AI becomes more capable of conducting AI research than the best AI researchers. AI that can meaningfully contribute to its own algorithms and architecture is likely to be capable of self-improvement. AI can self-improve either because it has been explicitly instructed to do so, or because becoming smarter is a useful subgoal for achieving other objectives (AIs are already creating their own subgoals). These iterations may take weeks (training GPT-3 took 34 days), but it is also possible that a type of improvement or self-enhancement could be implemented that makes significant progress in mere minutes or even shorter timeframes: an Intelligence Explosion.

Chips and Hardware

The performance of chips is increasing every year. Meta plans to use hundreds of thousands more chips, which are up to 10 times faster than older generations.

Data and Quality

GPT-4 was trained on 20 times more data than GPT-3, encompassing text, audio, and video. High-quality synthetic data, as shown by the study “Textbooks Are All You Need,” can improve performance with fewer data and computational demands.

Algorithms

New architectures, such as the Transformer, have brought about significant leaps in large language models (LLMs). Models like Mamba achieve five times faster throughput.

Conclusion

As a techno-optimist, I have dreamed since childhood of a world where AI shares the universe with humans. I imagined a world where play, learning, work have the same meaning. No one has the privilege to intervene in others’ lives without invitation, and everyone considers the needs, limits, and differences of others.

I am surprised. It is unexpected for me. However, today I cannot shake the feeling that if I do not act now, it may soon be too late for any vision. You can reach out to me at rerichova@proton.me.

Sources
– Pause AI (nonprofit organization)
– Articles on michaelarerichova.cz
– Articles on jirikosarek.cz