Natural selection paved the road for unicellular organisms to transform into the sentient human race we know today. Researchers at OpenAI now believe artificial intelligence may be capable of learning, growing, and expanding intellectually in a similar capacity.
The San Francisco based for-profit AI research lab believes that artificial intelligence can adapt to its environment similar to the way humans and animals do. Although many believe AI evolvement is already a reality, Open AI has sought to scale the technology dramatically through its recent research.
This form of experimentation encompasses two ideas currently permeating throughout the AI industry. First, multi-agent learning uses algorithms to pit agents against one another in a competitive fashion to determine winners. Think Darwin’s theory of natural selection in a simulated, digital realm. Second, reinforcement learning uses trial and error mechanisms to help reach desired outcomes.
The results of OpenAI’s experiment showed significant findings related to AI learning capabilities. Although many people believe AI is unable to exhibit intellectual growth, the new results say otherwise.
Setup of the Experiment
OpenAI set up a game of hide-and-seek amongst two separate teams and put them through millions of trials using sophisticated strategies. Their focus was on scaling AI technology while reviewing outcomes periodically to measure growth and development.
The virtual environment consisted of an enclosed area filled with items such as blocks, ramps, and barricades that were either stationary or mobile. Reinforcement-learning algorithms controlled two groups with the respective goals of finding or hiding from members of the opposing team.
Millions of simulated rounds occurred during the experiment. Researchers discovered that strategies became more sophisticated amongst AI agents as time went on.
After approximately 25 million rounds, hiders began to move objects and build forts so they couldn’t be found. The individual players also start to work together as a cohesive unit to improve time and efficiency.
Once the games were 75 million rounds in, the seekers developed advanced methods of finding the hiders. Ramps were used to climb walls and reach members of the opposing team. Hiders discovered a way to counteract this strategy by locking ramps to prevent encroachment of the hiding spaces.
Researchers believed the growth capabilities of the AI plateaued past the 100 million round mark, but continued gameplay showed progressive advancement at the 380 million round mark. The seekers developed a method of breaking into the other team’s forts using locked ramps to climb over unlocked boxes. The final phase ended with hiders learning to lock ramps and boxes before encapsulating hiding areas.
Future Implications for AI Technology
The results of the experiment indicate that AI is not a rudimentary technology. AI can learn, grow, and evolve in a capacity similar to humans. The AI participants in the games were given no directions or tasks beforehand. As the games went on, ongoing competition caused AI adaptations to occur much like natural selection.
The study leads researchers to believe there is value in testing the limits of existing technologies. AI may be capable of learning complexities far beyond what researchers initially intended. Although the simulated environments may be simpler than on earth, there is significant evidence to show that intelligent behavior can respond and adapt through learning mechanisms.
There is reason to believe advanced AI may be able to pick up new traits and abilities through continued learning without the need for human assistance.
The continued expansion of AI technology has profound implications for future technologies. Advanced AI techniques continue to help accomplish challenging tasks in today’s complex world.
Have questions about Bitvore's AI technology? Checkout our FAQ on Unstructured Alternative Data in Predictive Modeling.