Did DeepMind just make a big step toward more human-like A.I.?

The company's researchers created software that learns to transfer skills and strategies across games.

This is the web version of Eye on A.I., Fortune’s weekly newsletter covering artificial intelligence and business. To get it delivered weekly to your in-box, sign up here.

In January 2020, in a Fortune magazine cover story, I chronicled the corporate race for artificial general intelligence, a kind of human-like or even superhuman A.I. that is the staple of science fiction. The pursuit of AGI, as it’s more commonly called, has led to many of the machine learning innovations that underpin the current A.I. boom.

Subscribe to unlock this article and get full access to Fortune.com

This is the web version of Eye on A.I., Fortune’s weekly newsletter covering artificial intelligence and business. To get it delivered weekly to your in-box, sign up here.

In January 2020, in a Fortune magazine cover story, I chronicled the corporate race for artificial general intelligence, a kind of human-like or even superhuman A.I. that is the staple of science fiction. The pursuit of AGI, as it’s more commonly called, has led to many of the machine learning innovations that underpin the current A.I. boom.

But that boom is centered around narrow A.I—software that can perform one, specific task well. AGI, if ever realized, would involve a single piece of software that can handle multiple tasks.

The difficultly of building AGI hasn’t stopped big tech companies from investing big money in the quest to create it. Google-parent Alphabet has poured billions into its subsidiary DeepMind, which has AGI as its stated goal, and has spent billions more on Google Brain, its own in-house A.I. research lab. Microsoft has invested $1 billion in OpenAI, a company also dedicated to birthing AGI. Elon Musk, who helped found OpenAI, frequently muses about AGI and, Tesla’s recently announced plans to create robots may indicate that he’s still interested in creating it. Meanwhile, Facebook says it doesn’t really believe in AGI, but the tantalizing prospect remains a twinkle in the eye of some of the top machine-learning scientists it employs.

As I argued in that Fortune cover story, to these large tech companies, AGI is a moonshot in which the moon itself is somewhat beside the point. The quest generates lots of ancillary benefits: marketing panache that helps attract top research and engineering talent, as well as technology advancements that make commercially-useful narrow A.I. applications possible. AGI is a race these companies cannot afford to lose, even if none of them ever actually manages to win.

That said, it is worth highlighting when someone actually makes a concrete step towards AGI, even if only a baby step. And in late July, in research that has so far received less attention than it deserves (although my colleague Jonathan recently mentioned it briefly in this newsletter), DeepMind took such a step.

The company’s researchers created software agents that navigate a three-dimensional game world, called XLand, observing that world from a first-person perspective. These agents learn from experience, by trial and error, how to play a vast number of simple games in XLand—moving objects of a certain color and shape to specific locations or hiding an object from another player’s view. Many of the more complex games involved two or three players. Some of the games were competitive and some were collaborative.

Then, based on that experience, these agents were able to employ skills and strategies they’d learned in new games, without specialized training. In other words, these software bots had acquired general skills useful for playing any game in that simulated world, even ones they had not encountered before. That’s very different from software DeepMind has created previously in which an algorithm could be trained to perform at superhuman levels at a variety of board games and Atari video games. In those cases, the algorithm had to be retrained for each game—it didn’t learn strategies that worked across games.

Max Jaderberg, the DeepMind researcher who led the new research project, which uses what DeepMind calls “open-ended learning,” says he sees the research as a “one of the essential steps” to creating general agents that can perform many tasks. (You can read DeepMind’s research paper and accompanying blog post for more details.)

It’s not just DeepMind that thinks this is a breakthrough. “This is the G in AGI,” says Chris Nicholson, CEO at Pathmind. Nicholson and I have chatted about AGI in the past, and he called me as soon as he saw DeepMind’s research. He sees potential near-term commercial applications: Pathmind sells A.I. systems that help manufacturers, oil refineries, and the like improve their operations. It often trains that software using the same reinforcement learning methods – that is teaching based on experience, usually acquired in a simulator – that DeepMind is known for. Nicholson says DeepMind’s new methods could be used to create A.I. software, and even robots, that would be more adaptable and easier to train than those that exist today. They wouldn’t be human-like, but they may be more capable and cost effective than existing systems.

In fact, Jaderberg says that the focus on superintelligence is the wrong one. “We have to do a little mindset shift to stop chasing absolute performance in one narrow domain and focus instead on wide performance across as many tasks as possible at the sacrifice of expert performance in any particular one,” he says. In other words, the software is kind of a jack of all trades, master of none. A key aspect of human intelligence is our adaptability. Maybe we should aim for this same flexibility in our A.I. software.

But, in the real world, where many problems are framed narrowly, how useful would this be?  Wojciech Czarnecki, another DeepMind scientist who worked on the research, says general agents may be key when there isn’t enough data or a good simulator for a highly-specific task. The generalist software would have some skills that would enable it to perform okay, if not expertly, at the task, and then, with just a little bit of training and experience, become an expert.

AGI may still be years or decades away. It may never arrive. But the possibility suddenly seems closer than it did before.

**
How can data make your business more resilient? Hear stories from the frontline in a special Fortune Brainstorm AI webinar (in partnership with Accenture) on September 27, from 12-1 pm ET. Among those sharing their experiences will be:
•Susan Doniz, chief information officer and senior vice president for information technology and data analytics at Boeing
•Sandra Nudelman, head of consumer data and engagement platforms at Wells Fargo
•Bonnie Titone, chief information officer at Duke Energy
•Joe Depa, global managing director and data-led transformation lead at Accenture
Apply to attend here.
**

Jeremy Kahn
@jeremyakahn
jeremy.kahn@fortune.com

A.I. IN THE NEWS

Facebook apologizes after A.I. confuses videos of Black men with those of “primates.” The company’s video recommendation system suggested to users who had watched a video of Black men involved in confrontations with white police officers that they may like to watch further videos of primates, The New York Times reported. Facebook apologized for what one of its product managers called “an unacceptable error” and said it was investigating the cause to prevent it from happening again. The social media giant’s racist stumble follows a similar issue that plagued Google’s photo labeling A.I. in 2015.

Beijing unveils sweeping algorithm law. China’s Internet watchdog published landmark draft legislation that would, among other things, ban the use of algorithmic recommendation systems that encourage “addiction or excessive consumption.” It also says such algorithms should actively encourage “the spread of positive energy.” The proposed rules could further hamper the runaway growth of several big Chinese tech companies as well as set an interesting precedent globally, but as my colleague Eamon Barrett writes, they may not work as intended in practice. 

Only humans can hold patents in the U.S., judge rules. A Virginia federal district court judge’s ruling that only a human, not a piece of A.I. software, can hold patent rights is the first such decision in a U.S. court. The decision was reported by Bloomberg News. The case was brought against a patent registered as part of the Artificial Inventor Project, run by Ryan Abbott, a law professor at the University of Surrey, in England, that has sought to get a computer listed as an inventor in various jurisdictions. Abbott has secured favorable rulings in South Africa and Australia, although the Australian patent office is appealing that decision. Abbott told Bloomberg he would appeal the U.S. judge’s decision.

Tech companies competing to help ICE build a huge data dragnet. Business Insider reported on the scramble between Amazon, Google, Microsoft, and others, to win contracts to help the U.S. Immigration and Customs Enforcement agency build a tool called RAVEn that would help it draw connections between vast troves of information gathered from everything from social media to surveillance cameras to biometric databases. In the past, workers at these tech companies have protested their employers’ relationships with ICE. By the end of September, the agency is expected to award three contracts, worth a total of $300 million, to create and maintain the system through 2025.

EYE ON A.I. TALENT

Consulting giant Accenture has hired Portia Crowe as chief data strategy director for defense at Accenture Federal Services, its U.S. government consulting division, Washington Technology reported. Crowe was previously the chief data officer for the network cross functional team in the U.S. Army Futures Command and she also had several other civilian technology roles within the U.S. Army.

Swiss-based Zurich Insurance Group has hired Peter Kasahara as group chief data officer, according to Insurance Business MagazineKasahara was previously managing partner for PwC Digital Intelligence at the global consulting and accountancy firm PwC.

Colby College in Waterville, Maine, has named Amanda Stent as the new director of its Davis Institute for Artificial Intelligence, The Portland Press Herald reported. Stent had led the People and Language A.I. team at Bloomberg L.P.

EYE ON A.I. RESEARCH

Can machine learning be used to manage Internet traffic better than existing static algorithms? This question was posed by a joint team of researchers from Huawei’s Network Technology Lab in Beijing and researchers from the Barcelona Neural Networking Center at the Polytechnic University of Catalunya in Spain. Traffic engineering is a key part of building any network, including those that carry Internet traffic. The key is to figure out how to ease the flow over the network and eliminate congestion, even if the configuration of the network changes due to problems such as equipment failures. The researchers decided to try two A.I. methods in combination, multiagent reinforcement learning and graph neural networks, to see if they could beat the existing best system for traffic engineering, which is called DEFO (short for Declarative and Expressive Forwarding Optimizer). Their conclusion—presented in a paper that has been accepted to November’s International Institute of Electrical and Electronic Engineers (IEEE)’s International Conference on Network Protocols—was the machine-learning methods could achieve comparable or better performance than DEFO, even over network configurations it had not encountered in training. What’s more, the neural network-based approach could come up with solutions in seconds, compared to minutes with DEFO. The results could have implications not just for telecom companies trying to find the best way to route data across their networks, but also for those who have to find the best ways to route other kinds of traffic, such as vehicles through a crowded city. 

FORTUNE ON A.I.

Commentary: With A.I., business leaders must prioritize safety over speed—by Francois Candelon and Theodoros Evgeniou

Audi gives glimpse of its first self-driving car, a private jet on wheels—by Christiaan Hetzner

Beijing’s new algorithm laws are a global first, but they might not work—by Eamon Barrett

Here’s what Beijing’s sweeping new data rules will mean for companies—by Yvonne Lau

BRAIN FOOD

The demise of radiologists has been greatly exaggerated. In 2016, Geoff Hinton, considered one of the godfathers of today’s neural network-based A.I. revolution, famously quipped that “people should stop training radiologists now,” because it was “just completely obvious that within 5 years deep learning will do better than radiologists.” Well, those five years are up. And guess what? There are still plenty of radiologists. Hinton has since said he regrets his 2016 statement. He now says what will actually happen is that “radiologists will spend less of their time looking at CT scans and trying to interpret them, and more of their time interacting with patients.” But Hinton may turn out to be wrong again. Evidence keeps mounting that not only are radiologists not going away, our A.I. systems are actually a lot worse at reading complex medical imagery than Hinton and his colleagues thought. The latest evidence came this past week from a major study of A.I.-based mammography screening software published in the British Medical Journal. The study found that 34 of the 36 A.I. systems it looked at were less accurate than a single human radiologist and that all of them were less accurate than the consensus of two or more radiologists. (In the U.K. and many other countries, all mammograms must be read by two radiologists to avoid missing possible cancer diagnoses.) “AI systems are not sufficiently specific to replace radiologist double reading in screening,” the study’s authors wrote. Which goes to show that as good as many machine learning scientists like Hinton are at building prediction machines, they are lousy at making predictions themselves.