June 30, 2024
A year and a half ago, ChatGPT burst onto the scene, sparking an intense race among tech giants like OpenAI, Microsoft, and Facebook to develop the most advanced AI models. Despite initial rapid advancements, the AI field has hit a plateau, prompting a reevaluation of how we push the boundaries of artificial intelligence.
The AI Plateau: When More Data Yields Less Progress
The initial models were trained on roughly 20% of the total text on the Internet. Subsequent models accessed more substantial segments of the web, with the latest iterations trained on 15 trillion tokens, covering a vast expanse of the entire history of written content produced by humanity.
The anticipation was that the more data the models were trained on, the better their capabilities would become. Expectations were high that training on the entirety of human-written language might lead to the emergence of superhuman intelligence. Predictions included groundbreaking achievements such as discovering new laws of physics, developing innovative drugs, and solving long-standing mathematical puzzles.
However, these expectations were not fully realized as the incremental improvements in AI capabilities began to plateau. For instance, the leap in performance from GPT-2 to GPT-3 was significant, but subsequent improvements, such as from GPT-3 to GPT-4 and then to GPT-4o, demonstrated diminishing returns. This suggested that simply amassing more data might not necessarily translate into exponential growth in AI intelligence.
AI’s New Paradigm: The Path to Mastery Through Self-Play
In contrast, the development of AlphaZero by DeepMind showcased a different model of AI excellence. AlphaZero, an AI program designed to master the games of chess, Go, and others, did not rely solely on human expertise. Instead, after an initial phase of learning from human game data, it transitioned to a phase of self-exploration and self-play.
This second phase involved the AI playing against itself billions of times, creating and exploring numerous new game scenarios never seen before in human play. This extensive experimentation allowed the AI to develop a profoundly deep and innovative understanding of game dynamics, leading it to discover strategies and techniques previously unknown or underappreciated by human grandmasters.
From Imitation to Innovation: AI’s Next Leap Forward
This approach underscores a potential pathway for AI to achieve superhuman capabilities in fields beyond games. By adopting a similar two-phase approach—initial learning from existing data followed by a phase of creative self-exploration—there is a possibility for AI to achieve superhuman intelligence and make significant contributions to fields such as physics, medicine, or mathematics. This method could enable AI to not just mimic human thought processes but to diverge and possibly innovate beyond them.
By leveraging the strategies evolved through countless self-play sessions, new frontiers in AI development are being opened. AI systems can independently generate knowledge that hasn’t been explicitly programmed or previously conceived by humans. This independent creative capability could lead to breakthroughs in numerous domains, offering new solutions that are currently beyond our grasp.
Drug Discovery, Novel Materials, and Sustainable Energy: Embrace the Excitement, Not the Fear
Looking ahead, these capabilities could extend far beyond traditional game boards. AI systems could revolutionize drug discovery by predicting molecular behaviors and interactions, rapidly developing new medicines. They could also aid in formulating a unified theory of physics, unraveling cosmic mysteries. In materials science, AI might invent novel materials, enhancing sustainable energy solutions and driving technological advancements.
While these technologies may take years to mature, their potential societal benefits should inspire excitement rather than fear. As these tools evolve, they promise to solve some of humanity’s most pressing challenges, turning distant possibilities into achievable realities.
Michael Zolotov, 33, is an AI expert with a master’s degree in electrical engineering. He is a Co-Founder and CTO of a group of leading AI companies, including the publicly traded Razor Labs, which develops artificial intelligence technologies for asset-intensive industries; Axon Vision, which develops AI solutions for the defense market; and more. Michael also co-founded the Future Learning school, the first Deep Learning training academy for AI engineers in Israel.
