Over two consecutive days in late July 2021, amid the usual squall of press notices, two announcements, one from DeepMind and one from Google, claimed unprecedented accomplishments.
Google said it was currently keeping track of time crystals its researchers had created in one of their quantum computers, and DeepMind wanted the world to consider a gaming environment they just built called XLand (above) that will bring deep learning even closer to human levels.
The title of the blog post from DeepMind quietly asserted that a new approach to training AI agents had successfully created a more human-like capable agent: “Generally capable agents emerge from open-ended play,” released July 27, 2021.
The London-based AI research group has had some of its most important developments realized through its AlphaZero general-reinforcement learning algorithm, which “beat world-champion programs in chess, shogi, and Go after starting out with knowing no more than the basic rules of how to play.” The company explains, “Through reinforcement learning (RL), this single system learnt by playing round after round of games through a repetitive process of trial and error.”
But despite tournament wins over human grandmasters, AlphaZero is limited in the same way that all machine learning is limited. It’s very good at learning about one thing at a time, unlike humans who can bring to a learning task knowledge from all sorts of directions.
To open up their algorithm’s tunnel vision, the researchers began work on AI agents with more general and adaptive behavior, which “instead of learning one game at a time, these agents would be able to react to completely new conditions and play a whole universe of games and tasks, including ones never seen before.” The DeepMind paper describes such an agent and the vast game environment called XLand, “which includes many multiplayer games within consistent, human-relatable 3D worlds.”
The agent’s learning process proceeds without needing human interaction data. The paper, still in preprint (not yet peer edited), explains, “The agent’s capabilities improve iteratively as a response to the challenges that arise in training, with the learning process continually refining the training tasks so the agent never stops learning. The result is an agent with the ability to succeed at a wide spectrum of tasks—from simple object-finding problems to complex games like hide and seek and capture the flag, which were not encountered during training.” The goal is a more general AI agent with the ability to adapt rapidly within constantly changing environments.
The DeepMind agents operate in 3D first-person avatars in a multiplayer environment, and “this enables [the researchers] to include billions of tasks in XLand, across varied games, worlds, and players.” The results have been reassuring.
On its blog, DeepMind reports, “After training our agents for five generations, we saw consistent improvements in learning and performance…. Playing roughly 700,000 unique games in 4,000 unique worlds within XLand, each agent in the final generation experienced 200 billion training steps as a result of 3.4 million unique tasks. At this time, our agents have been able to participate in every procedurally generated evaluation task except for a handful that were impossible even for a human.”
How far a step this might be toward artificial general intelligence (AGI) is unclear, but it’s definitely a positive move in a direction away from the wall that some researchers believe AI might hit due to the current limitations on machine learning.
The research paper “Observation of Time-Crystalline Eigenstate Order on a Quantum Processor,” released July 28, 2021, was posted by Google the day after the XLand news. The report was the product of a collaborative effort of physicists from Stanford, Princeton, and several other universities. Simplified, the title claims that researchers have used Google’s quantum computer to create and then observe the responses of an actual time crystal.
The idea of a time crystal is almost a decade old, and what it is is obfuscated by its name. A “time crystal” sounds like something that’s brilliant and dynamic—something you can hold, or at least watch, something that has enclosed time. Actually, it isn’t so much a thing or particle, but rather a state created and maintained.
The Google experiment was originally reported in a scientific paper published last month on Cornell University’s arXiv.org. One hundred researchers set up an array of 20 qubits (quantum particles that act as bits in quantum computers) to serve as a time crystal. In Google’s paper, the time crystal was created on the company’s Sycamore quantum computer. But it wasn’t the first; others have also done this recently with scientists at Delft University in Netherlands, creating a time crystal out of carbon atoms within a diamond.
Time crystals don’t occur in nature. They began as a theoretical possibility in 2012, first conjured by a Nobel laureate professor at MIT, Frank Wilczek. Most scientists at the time dismissed the idea out of hand because their existence would seem to contradict the second law of thermodynamics governing the normal way matter absorbs, uses, and expels energy.
Described by some researchers as “a new phase of matter,” a time crystal has very peculiar properties especially regarding energy and movement. Atoms in normal crystals are arranged in repeating patterns in space. Time crystals answer the question, can you create a crystal whose atoms are arranged in repeating patterns in time?
Well, it appears you can, if, like Google, you fire a finely tuned laser at the atoms, which might then flip into another state, then flip back, and then repeat that motion without absorbing any energy from the laser. And theorists say you can do this because quantum computers, like Sycamore, can simulate quantum particles and then study them in their unique quantum state.
The system can’t lose energy to the environment and come to rest. Because of this, the motion is “motion without energy.” In a commentary in Quanta Magazine, Natalie Wolchover, editor for Quanta Magazine, adds, “Like a perpetual motion machine, a time crystal forever cycles between states without consuming energy.”
Gabriel Perdue, a quantum researcher at the Fermilab in Illinois in the United States explains, “The thing that is most exciting here, for me, it’s a demonstration of using a quantum computer to really simulate a quantum physics system and study it in a way that is really novel and exciting.”
That’s fine for the theoretical physicists, but computers are useful as much more than just laboratories. And as it turns out, there’s a possible practical application for time crystals in quantum computers that could, according to science journalist Rahul Rao, be as important as silicon is to classical computers. “They might help make quantum computers become more robust. Engineers have struggled for years to create something that could serve as memory in quantum computers; some equivalent to the silicon that underpins traditional computers. Time crystals, physicists think, could serve that purpose.”
So, unlikely as it seems, there were two possibly pivotal developments in AI and quantum computing, one following the other over two days in July. And once again, the lessons from the accelerating winds of change in tech have reminded us all—Don’t blink!