The AI treadmill will really become unstoppable when AI systems begin improving themselves, and there are initial signs that this has begun happening.
This tantalizing prospect was recently given more weight by a researcher at the forefront of artificial intelligence development. Matej Balog, a Research Scientist at Google DeepMind, has suggested that we are witnessing the nascent stages of AI self-improvement. His statement provides a cautiously optimistic yet grounded perspective on a development that could fundamentally alter the trajectory of AI.
Balog, whose work includes contributions to projects that use AI to discover more efficient algorithms, contextualizes these early signs. When discussing the current state of AI’s ability to enhance itself, he is careful to specify the nature of these advancements. “I would also agree that we are maybe seeing the first sign of self-improvement, but one also needs to be very specific about what we have shown so far,” Balog stated. He pointed to a concrete example from his team’s work: “It’s the speeding up the training of the next generation of the Gemini model.” This is a significant, tangible outcome where an AI has contributed to making a future version of itself more efficient. However, Balog also injects a dose of reality into the timeline of this feedback loop, noting, “So the feedback loop is fairly long at least currently, maybe on the order of months.”
Despite the current extended timeframe for these improvements to manifest, Balog affirms that the process can indeed be categorized as a form of self-betterment. “But there is, you can call it self-improvement for sure,” he asserts. The crucial question, as he frames it, is one of extrapolation and the future trajectory of this capability.
He outlines several potential scenarios for how AI self-improvement could evolve, highlighting the uncertainty that even top researchers in the field grapple with. “Maybe the big question that many people are curious about is how does this extrapolate into the future? And you can have different types of self-improvement,” Balog explained. He describes a one-off benefit as a possibility: “One is where you get maybe just a one-off benefit, like the model improves itself once, and that’s it.” Another, more dynamic possibility involves continuous but diminishing returns: “Another one is, okay, the model keeps improving itself continuously, but maybe the improvements get, uh, marginally smaller and smaller and smaller, and you converge to some limit.” The most transformative, and debated, scenario is one of accelerating progress: “or maybe the, the improvements will, will keep accumulating, uh, up and up and up. And that’s a big open question that we don’t have an answer to today.”
The implications of true recursive self-improvement are profound, and the landscape of AI is already showing trends that align with this potential future. For instance, Google DeepMind’s AlphaEvolve, which Balog has worked on, has been used to discover new, more efficient algorithms for fundamental computational tasks, which in turn can accelerate AI research and development. This creates a positive feedback loop where better AI helps create better tools for building even more capable AI. This concept is not confined to Google. Across the industry, there’s a concerted effort to use AI to automate and optimize various stages of AI development, from data cleaning and augmentation to model architecture search and hyperparameter tuning. Each of these represents a small, yet significant, step towards more autonomous and self-improving systems. While Balog and his colleagues remain measured in their assessment, the initial signs they are observing could be the foothills of a much larger mountain, the ascent of which will have far-reaching consequences for both the tech industry and society as a whole.