With large language models such as OpenAI and Microsoft's ChatGPT catching off around the world, Google, once a leader in artificial intelligence, has had to rush into action. In order to catch up with the times, they once again sought the help of DeepMind's artificial intelligence research lab, which is another collaboration after DeepMind was acquired by Google.
DeepMind, a London-based artificial intelligence lab, has made a global name for itself with the development of AlphaGo. Alpha Dog successfully defeated the top human chess player, showing the world the great potential of artificial intelligence. Since then, DeepMind has launched an upgraded version of Alpha Dog, AlphaZero, further expanding its influence in the field of artificial intelligence. However, due to Google's lag in AI research, they turned to DeepMind again in the hope of a breakthrough.
The latest breakthrough comes from DeepMind's new artificial intelligence system, AlphaDev. The results, published in the scientific journal Nature, focus on the discovery of faster computer algorithms. Computer algorithms play a vital role in software development and are used trillions of times a day by companies. AlphaDev is based on AlphaZero, an improved version of AlphaGo, and utilizes reinforcement learning methods that enable computers to learn and develop strategies autonomously to discover faster algorithms, including computer science functions such as sorting and hashing. Sorting algorithms are used to sort data, such as ranking web search results and financial institutions' back-end systems. Hashing algorithms convert the data into a unique string, allowing users to find what they need in places like databases. Since these algorithms are widely used by companies, making them faster will significantly reduce the resources required for computation.
Colin Murdoch, chief business officer at DeepMind, said: "This means we can do the same amount of computation with fewer resources. When applying AlphaDev to the C++ sorting library, DeepMind said that in smaller-scale sorting tasks, AlphaDev is 70 percent faster than traditional algorithms, and in large-scale sorting tasks, the speed is improved by 1.7 percent. For hash functions, AlphaDev also found an algorithm that is 30% faster than traditional algorithms in the range of 9 to 16 bytes. Both algorithms are available for developers to use in open-source libraries.
As part of DeepMind's ongoing efforts to improve the efficiency of computer systems, they have partnered with various units at Google and Alphabet to apply artificial intelligence systems similar to AlphaDev to optimize network resources, keep data centers cool, and share computing resources among servers. According to DeepMind, in the trial, the AI system reduced underutilized hardware (i.e., servers that failed to fully utilize server capacity) in Google's data centers by up to 19%.
For businesses, unused computing resources mean wasted energy and money. "If you can allocate resources more efficiently, it will increase the speed of the business because you will be able to use digital resources that would otherwise be being taken up," Murdoch said.
DeepMind's first application project was about four years ago to optimize YouTube's video compression process, allowing users to watch videos with less data without sacrificing video quality. Daniel Mankowitz, DeepMind's AlphaDev principal investigator and DeepMind employee research scientist, said that through the success of the project, they turned the team's attention to code optimization. While optimization techniques themselves are not new, and mathematical methods of finding optimal solutions for resources with certain constraints have been around for decades, DeepMind's "envision the results of potentially efficient algorithms in the future" approach has yielded faster results than sorting algorithms previously developed by engineers.
In April, Google combined its Brain and DeepMind research into one unit, led by DeepMind CEO Demis Hassabis. After fierce competition with Microsoft-backed OpenAI, which developed projects such as ChatGPT, Google aims to accelerate its AI and generative AI efforts. Murdoch said that in addition to its research in the life sciences, DeepMind is also focused on generative artificial intelligence, including developing large-scale language models and helping enterprises apply these models.
"More research is needed to make large language models run more efficiently in the cloud, perhaps even on mobile phones and devices." "This is going to be a very important research question in the coming months," Murdoch said. ”
With the launch of AlphaDev, a new artificial intelligence product launched by DeepMind based on the original architecture of AlphaGo, Google is expected to surpass OpenAI and disrupt the existing situation in the competitive landscape of artificial intelligence. This breakthrough will have a huge impact on the development of the computer field, accelerate the application and innovation of artificial intelligence technology, and also demonstrate the determination and strength of DeepMind and Google to push the frontier of science and technology.