The Defense Advanced Research Projects Agency (DARPA) recently announced that an AI algorithm piloting an F-16 Fighting Falcon in a simulated dogfight against a seasoned US AIR Force pilot achieved a flawless 5-0 win, with the human pilot never scoring a single hit.
Regardless of how realistic “Computer vs. Human simulations” can be in such a complex environment, the most significant aspect is the fact that the specific AI system was developed less than one year ago using so-called “deep reinforcement learning”: starting with a complete lack of understanding about basic flight, the AI software autonomously learned fast, gaining the equivalent of 12 years of experience over the course of 4 billion simulations.
This AI revolution has been decades in the making, but only in the last decade have advances in computing power enabled a new era of AI training. The most recent technological disruption is enabled and accelerated by “Deep Learning”, the most advanced AI state in which machines can learn autonomously by analyzing vast amounts of unstructured data.
As the cost of memory and compute came down dramatically, Deep Learning breakthroughs enabled computers to process a wide range of information across different and complicated data-driven applications, as in image recognition tools, speech recognition (NLP), image recognition (GPU), data discovery, and extraction.
The era of Deep Learning, source Nvidia and JP Morgan
One of the most impressive examples of advances in machine learning is Alpha Zero, developed within the DeepMind division of Alphabet, which obliterated the highest-ranked chess program in the world (Stockfish). Given only the rules of the game, Alpha Zero learned how to play chess within four hours learning from playing against itself. The crucial advancement of this technology is the capacity for life-long learning: this AI system can acquire new information and keep in mind those already experienced to solve progressively more tasks, without losing information previously learned, a hurdle known as “catastrophic forgetting”.
The exponential advancements in AI models are supported by ever larger AI chips which are embedded with significant amounts of fast memory to handle the demands of AI algorithms. The race for manufacturing the largest AI chips includes both public and private companies: while Xilinx has announced the chip with the highest logic density on a single device ever bult, featuring 35bn transistors, private startup Cerebras has recently showcased the largest chip ever built, with 1.2 trillion transistors and 3000x more in-chip memory. Finally, Alphabet announced a major breakthrough in quantum computing, whereas their processor “Sycamore” took about 200 seconds to complete a task that would take a state-of-the art supercomputer approximately 10’000 years.
Regardless of potential winners and losers, technologic advancements are attracting large amount of investments. According to MarketsandMarkets, the overall Deep Learning market is estimated to reach $18.16B by 2023 from $3.18B in 2018, at a CAGR of 41.7% over this period.
The investment implications are staggering and long-term in nature since companies are increasing technology investments to survive the rapidly changing landscape of the Fully Connected Economy. As a result, AI spending is projected to grow 28% annually from $19bn in 2018 to almost $100bn in 2023.
The information in this article should not be regarded as a description of services provided by Delian Partners SA. The opinions expressed in this article are for general informational purposes only and are not intended to provide specific advice or recommendations for any individual or on any specific security or investment product. It is only intended to provide education about the financial industry. The views reflected in this article are subject to change at any time without notice.
