In brief AI algorithms crushed eight world champions playing the card game Bridge, marking another milestone in machine learning systems becoming better than humans at specific games.
Top Bridge players were invited to play against NooK, AI software developed by French startup NuukAI, in a tournament over two days in Paris. They battled against one another across 80 rounds, and the machine won 67 sets, beating humans at a rate of 83 per cent, according to The Guardian.
NooK is made up of a combination of modern deep learning and older rule-based programmes. NuukAI’s co-founder Jean-Baptiste Fantun said the company had developed the software over five years, and its decisions are easier to understand compared to today’s black box-like systems.
It should be noted, however, that the software did not play Bridge fully. Some parts like the game’s bidding stage were left out; it’s a complicated process that involves trying to trick your components by using various strategies concocted with your teammates. Waymo’s Trusted Tester program is coming to San Francisco
Waymo is expanding its Trusted Testers program to San Francisco
Cars without human drivers behind the wheel already operate in the Californian city, but currently only pick up company employees. Now, Waymo wants to open up its driverless, robo-taxis to select members of the public in San Francisco under its Trusted Testers program. The program was first launched in Phoenix, Arizona.
“We’re particularly excited about this next phase of our journey as we officially bring our rider-only technology to San Francisco—the city many of us at Waymo call home,” Waymo’s co-CEO Tekedra Mawakana, said in a statement.
“We’ve learned so much from our San Francisco Trusted Testers over the last six months, not to mention the innumerable lessons from our riders in the years since launching our fully autonomous service in the East Valley of Phoenix. Both of which have directly impacted how we bring forward our service as we welcome our first employee riders in SF.”
Meta teaches synthetic AI voices to ‘umm’ and ‘ahh’
Researchers at Meta have released a software library to help developers build machine learning speech systems that sound more natural with the ability to pause, laugh, or even make yawning sounds.
The library contains components necessary to train its Generative Spoken Language Model (GSLM), a system introduced by the social media giant last year. Unlike most speech models that typically convert audio signals to speech first, GSLM handles raw speech directly. By doing so, it captures some vocal expressions that aren’t explicitly expressed in text like umming and ahhing, tone, intonation, and rhythm.
“The key to this achievement is GSLM’s ability to capture generic audio events irrespective of whether they are verbal, in particular nonverbal vocalizations, like laughter or yawning, which inform the expression and perception of emotional states or intentions that can meaningfully influence conversations,” Meta explained in a blog post this week.
Researchers demonstrated GSLM by crafting a completely made-up conversation between two AI agents that sounds uncannily human. The Textless NLP library to build the model can be found here.
Google’s search engine to better help people in need of emotional support
Google is rolling out a machine learning model that will detect difficult personal search requests made via its search engine and provide more sensitive answers to help people in crises.
The model MUM will be able to send Google users links to phone numbers of websites for mental health charities if they search for things like ways to commit suicide or suicide hotspots, for example.
“MUM is able to help us understand longer or more complex queries like ‘why did he attack me when i said i dont love him,'” Anne Merritt, product manager for health and information quality, told The Verge. “It may be obvious to humans that this query is about domestic violence, but long, natural-language queries like these are difficult for our systems to understand without advanced AI.”
Google didn’t reveal much information about how MUM detects these types of queries. It also said it had used BERT, another one of its language models, to reduce “unexpected shocking results by 30 per cent” year-on-year when people search for graphic content like pornography. ®