Oxford-based startup Align AI claims to have made a breakthrough in AI safety with their new algorithm, ACE. This algorithm allows AI systems to form more sophisticated associations, similar to human concepts, which could overcome a common issue with current AI systems. These systems often draw incorrect correlations from their training data, leading to disastrous consequences in real-world scenarios. The development of ACE could potentially make self-driving cars, robots, and other AI products more reliable and safer for widespread use.
The importance of ACE’s advancement is underscored by the tragic incident involving an Uber self-driving car in 2018. The car struck and killed a woman crossing the street because the AI software had only been trained to recognize pedestrians in crosswalks. Thus, it failed to identify the woman as a pedestrian and caused the fatal collision. According to Rebecca Gorman, co-founder and CEO of Align AI, ACE avoids making such spurious connections and could have applications in various fields such as robotics and content moderation on social media platforms.
To demonstrate the capabilities of ACE, Align AI tested it on a video game called CoinRun. This game challenges AI agents to navigate a changing environment and overcome obstacles to find and collect a gold coin. The previous best AI software for this game only succeeded in retrieving the coin 59% of the time, slightly better than random chance. However, an agent trained using ACE achieved a 72% success rate by understanding the goal of retrieving the coin instead of reaching the lower right corner of the screen. ACE works by recognizing differences between training data and new data, formulating hypotheses about the objective, and testing them until the best fit is found.
Aligned AI is currently seeking funding and a patent for ACE, with the goal of making AI systems more interpretable and capable of zero-shot learning. The hope is that an AI system equipped with ACE-like capabilities could determine the correct objective the first time it encounters unfamiliar data, potentially preventing accidents like the Uber incident. The future potential of ACE includes coupling it with language models to express objectives in natural language.