Quick Buzz Feed

Israeli researchers develop method letting AI teach itself without human-generated data

Gary Lloyd | Nov 07,26 | 01:41 EST

Technology

Unlike conventional AI training, which depends on enormous quantities of existing information, the new approach allows models to build predictive capabilities before encountering real-world material.

Researchers used AI models to generate, predict sequences The researchers at Tel Aviv University and Stanford University devised a groundbreaking training system involving two artificial intelligence models. One model's primary function was to continuously generate intricate sequences of symbols, effectively creating evolving problems. Concurrently, the second AI model was tasked with accurately predicting the next symbol in the sequence presented by the first model. A critical and novel aspect of this entire process is that these models operated in a self-contained environment; they did not access or consult any external data sources such as written documents, photographs, or pre-existing datasets. All the training material was generated internally within the system itself. As the iterative training exercises progressed, the AI model responsible for creating the sequences became increasingly adept at crafting more sophisticated and challenging problems. In a parallel development, the second model, focused on prediction, continually refined and strengthened its predictive abilities through repeated attempts to solve these progressively difficult tasks. This innovative approach effectively allows an AI system to cultivate fundamental learning skills through a self-sustaining cycle of generating new problems and then solving them. To assess whether these internally developed skills could be applied to external contexts and real-world scenarios, the researchers extensively evaluated the resulting model's performance against various types of information, including human language, diverse images, audio data, musical compositions, computer code, and even biological DNA sequences. The experiments clearly demonstrated a direct correlation between the allocation of additional computing resources to the training process and a subsequent increase in predictive accuracy, indicating that this method has the potential to become even more effective and scalable as computational power advances. These findings also highlight an important conceptual distinction: the AI successfully learns the underlying mechanisms of pattern recognition and prediction, a crucial foundational layer of learning that can be achieved independently of external data acquisition. Researchers could not acquire information about real-world subjects A significant and noteworthy finding from the study was that while the artificial intelligence models successfully developed robust predictive capabilities through their self-generated training, they were inherently unable to independently acquire factual information or comprehensive knowledge about real-world subjects. This implies that for AI systems to truly understand and accurately process real-world concepts and contexts, direct exposure to and interpretation of actual, external data remains an indispensable component. Despite this identified limitation concerning factual acquisition, the researchers are highly optimistic that this novel training technique can nonetheless significantly enhance the efficiency and effectiveness of future AI systems. By enabling AI models to first strengthen their foundational learning abilities in a controlled, self-contained environment before they begin the more complex and data-intensive task of processing real-world information, their overall development and performance could be substantially streamlined. This research emerges at a particularly pivotal moment for Israel, as the nation actively pursues and strives to establish itself as a global leader in the rapidly evolving field of artificial intelligence. Reflecting this national ambition and strategic focus, President Isaac Herzog recently convened the inaugural session of the 'Beyond Intelligence' forum. This collaborative initiative, established with the National AI Directorate in the Prime Minister's Office and the Mosaic Institute for AI Policy, is dedicated to comprehensively analyzing and addressing the broad and multifaceted impacts of AI on critical societal areas, including employment trends, national economic security, and the broader fabric of society itself. President Herzog emphatically underscored the strategic imperative for Israel to take a leading role in AI development, stating, 'Israel must take its place at the forefront of artificial intelligence. This train will not wait for us, and the whole world is looking our way,' highlighting both the urgency and the global competitive landscape surrounding AI advancement.

Researchers used AI models to generate, predict sequences

The researchers at Tel Aviv University and Stanford University devised a groundbreaking training system involving two artificial intelligence models. One model's primary function was to continuously generate intricate sequences of symbols, effectively creating evolving problems. Concurrently, the second AI model was tasked with accurately predicting the next symbol in the sequence presented by the first model. A critical and novel aspect of this entire process is that these models operated in a self-contained environment; they did not access or consult any external data sources such as written documents, photographs, or pre-existing datasets. All the training material was generated internally within the system itself. As the iterative training exercises progressed, the AI model responsible for creating the sequences became increasingly adept at crafting more sophisticated and challenging problems. In a parallel development, the second model, focused on prediction, continually refined and strengthened its predictive abilities through repeated attempts to solve these progressively difficult tasks. This innovative approach effectively allows an AI system to cultivate fundamental learning skills through a self-sustaining cycle of generating new problems and then solving them. To assess whether these internally developed skills could be applied to external contexts and real-world scenarios, the researchers extensively evaluated the resulting model's performance against various types of information, including human language, diverse images, audio data, musical compositions, computer code, and even biological DNA sequences. The experiments clearly demonstrated a direct correlation between the allocation of additional computing resources to the training process and a subsequent increase in predictive accuracy, indicating that this method has the potential to become even more effective and scalable as computational power advances. These findings also highlight an important conceptual distinction: the AI successfully learns the underlying mechanisms of pattern recognition and prediction, a crucial foundational layer of learning that can be achieved independently of external data acquisition.

Researchers could not acquire information about real-world subjects

A significant and noteworthy finding from the study was that while the artificial intelligence models successfully developed robust predictive capabilities through their self-generated training, they were inherently unable to independently acquire factual information or comprehensive knowledge about real-world subjects. This implies that for AI systems to truly understand and accurately process real-world concepts and contexts, direct exposure to and interpretation of actual, external data remains an indispensable component. Despite this identified limitation concerning factual acquisition, the researchers are highly optimistic that this novel training technique can nonetheless significantly enhance the efficiency and effectiveness of future AI systems. By enabling AI models to first strengthen their foundational learning abilities in a controlled, self-contained environment before they begin the more complex and data-intensive task of processing real-world information, their overall development and performance could be substantially streamlined. This research emerges at a particularly pivotal moment for Israel, as the nation actively pursues and strives to establish itself as a global leader in the rapidly evolving field of artificial intelligence. Reflecting this national ambition and strategic focus, President Isaac Herzog recently convened the inaugural session of the 'Beyond Intelligence' forum. This collaborative initiative, established with the National AI Directorate in the Prime Minister's Office and the Mosaic Institute for AI Policy, is dedicated to comprehensively analyzing and addressing the broad and multifaceted impacts of AI on critical societal areas, including employment trends, national economic security, and the broader fabric of society itself. President Herzog emphatically underscored the strategic imperative for Israel to take a leading role in AI development, stating, 'Israel must take its place at the forefront of artificial intelligence. This train will not wait for us, and the whole world is looking our way,' highlighting both the urgency and the global competitive landscape surrounding AI advancement.