Researchers from OpenAI have achieved a significant breakthrough in machine learning by using self-play to improve the performance of their system. They applied this approach to Dota 2, a popular multiplayer game, and found that it can catapult the performance of machine learning systems from below human level to superhuman. The system went from barely matching a high-ranked player to beating top professionals in just a month. This improvement is attributed to self-play's ability to generate new data as the agent gets better, unlike supervised deep learning systems which are limited by their training datasets. --- Why it matters: This matters because it shows that self-play can be an effective way to improve machine learning performance, potentially leading to breakthroughs in other areas of AI research. The impact on engineers and researchers is significant, as this approach could enable the development of more advanced and capable AI systems. Source: https://openai.com/index/more-on-dota-2