META researchers claim they have observed that the company’s artificial intelligence (AI) systems improve without human input.
“Over the past few months, we have begun to give a glimpse into how AI systems improve themselves,” Meta CEO Mark Zuckerberg said on the company’s website in a policy paper published on July 30th. “The improvements are slow so far, but I can’t deny it.”
Zuckerberg added that self-improvement AI is the first step to achieving artificial tension (ASI). This is an artificial intelligence that can surpass human cognitive abilities.
You might like it
Broadly speaking, AI researchers are examining AI capabilities in three levels. Current AI models can achieve extraordinary superhuman feats, such as predicting protein structure, but these functions are very narrow. In such cases, the system can only overtake humans in one particular area, but there is no general intelligence that this ability applies to different areas.
The next layer is called artificial general information (AGI), and refers to models that can be understood, learned and adapted in almost the same way as the human brain.
The ASI is the final state that Zuckerberg refers to in his blog post, and is related to a model that has evolved far beyond human potential, improving itself at an exponential rate, leading to a “intelligence explosion.”
Scientists refer to hypothetical moments where AGI is achieved as technical specificity.
Related: AI is in a “unprecedented system.” We should stop it – and we should stop it before it destroys us –
This phenomenon is not the first time that Zuckerberg has mentioned visually seeing AI self-improvement. In October 2024, University of California researcher Santa Barbara published a paper in the Preprint ARXIV database, focusing on the concept of self-improvement.
In it, they defined, created and tested an AI framework based on the ideas of the Gödel machine. This is a theoretical device that allows you to improve yourself by rewriting your own code and instructions. Importantly, the machine only implements the change if it can produce formal evidence that it is beneficial.
Related: AI can quickly think in ways we don’t understand – avoid efforts to align it – top AI scientists warn
The team experimented with AI frameworks based on that concept, demonstrating that Gödel agents can improve their own performance in a variety of tasks, such as coding, science, mathematics, and inference.
Almost all models are usually not authorized to rewrite their own code. The Gödel agent has access to not only the entire codebase, but also the code used to develop improvements, indicating that those improvements are beneficial.
AI consistently performed better in areas of importance than the human-designed agents it compared to the comparisons. Zuckerberg said that ASI represents a transformative step in human technological advancement, leading to “discovering new things that are unimaginable today.”
That said, he added that he would be much more careful about the models that Meta chose to make public under an open source framework.
“I am very optimistic that humanity will help us accelerate the pace of our progress. But perhaps even more importantly, superintelligence could launch a new era of personal empowerment with larger institutions to improve the world in the direction people choose,” writes Zuckerberg.
“It may be as deep as the richness created by AI, but even more meaningful impact on our lives will come from personal closeness that will help you achieve your goals, create what you want to see in the world, experience adventures, become better friends with those you care about, and become the person you want.”
Source link
