Researchers from the French A.I. research firm NukkAI held a competition last week in Paris in which the startup’s A.I. defeated some of the world’s top Bridge players in a tournament. The victory follows a string of recent conquests in which A.I. has defeated humans in various games, such as the ancient Chinese board game Go and the video game Grand Turismo Sport.

The win was noteworthy because Bridge requires players to make strategies without knowing many of their opponents’ cards, making it a more difficult challenge than games like Chess, in which players can see their opponents’ pieces.

But about that defeating humans thing. It’s only sort of what happened.

The A.I. did not defeat human Bridge players in direct one-on-one games. Instead, the A.I. outperformed human players when competing against other computer opponents during a series of matches, or deals, in which the conditions were similar. 

Additionally, the tournament did not include the “bidding” process that takes place during the beginning of a Bridge match. During bidding, players must tell their opponents how many “tricks” it will take for them to win based on the cards that they were dealt; the Telegraph, which reported on the tournament, referred to the bidding process as the “most quixotic part of bridge.” 

Noam Brown, an A.I. research scientist at Facebook parent Meta, downplayed the significance of the A.I.’s victory on Twitter, noting that the A.I. did not beat humans in one-on-one victories and was not involved in “bidding.”

“Bridge remains an unsolved challenge for A.I.,” Brown wrote. “I’m not working on it, but I think the team that actually beats world champions in the game deserves fair recognition for it.”

The neurosymbolic approach

Still, the A.I. Bridge tournament was noteworthy because it demonstrated a different approach to developing sophisticated A.I. software.

Previous A.I. victories primarily involved the A.I. technique of deep learning, in which researchers use neural network software to analyze and act on patterns found in enormous quantities of data. NukkAI, however, used a so-called neurosymbolic approach to creating it’s A.I.

To create their software, researchers specializing in neurosymbolic A.I. use a blend of deep learning techniques and other A.I. concepts that have generally fallen out of favor over the past few years. 

In previous academic papers, the NukkAI researchers described how they used a statistical approach known as Probabilistic Logic Programming to help their A.I. derive strategies at winning bridge even when the A.I. is unaware of its opponents’ cards.

The NukkAI researchers also used a variant of the popular Monte Carlo “search” technique to help the software better evaluate possible moves, as they descriube in an another research paper.

Meta A.I. researchers previously used a combination of deep learning and Monte Carlo techniques to create software that mastered the card game Hanabi.