Most virtual assistant makers like Google and Amazon have spent their past few years in building more conversational A.I. bots that could go beyond the usual jokes and commands and essentially talk to you as a real human. Google today shared what it has been up to, and in a published paper, details a “human-like” chatbot called Meena that can “engage in conversation on any topic.”
Meena, unlike its peers such as Google’s own Assistant, is an open-domain chatbot. What that means is that Meena is not built around a limited set of data that is hand-picked to accomplish specific tasks. Instead, Meena is designed to contextually and constantly converse with you — no matter the topic.
This is made possible thanks to Google’s vast collection of data. The company claims Meena has been trained “on 40 billion words mined and filtered from public domain social media conversations.”
More importantly, with Meena, Google is tackling the perplexity shortcomings common voice assistants suffer from. The machine learning models Siri and Google Assistant use cannot handle multiturn dialogues like humans. When in doubt, they simply tell you they can’t understand the query and pull up a web result.
To ensure Meena doesn’t fall victim to the same stumbling blocks, Google is adding a second parameter to their algorithms, which in addition to being clever enough to not respond with gibberish, can also come up with a specific answer.
“For example, if A says, ‘I love tennis,’ and B responds, ‘That’s nice,’ then the utterance should be marked, ‘not specific’. That reply could be used in dozens of different contexts. But if B responds, ‘Me too, I can’t get enough of Roger Federer!’ then it is marked as ‘specific’ since it relates closely to what is being discussed,” added Google in the paper.
Based on these metrics, an average human scores 86%. Google says it has so far managed to bring Meena up to a staggering 79%. In comparison, Pandora Bots’ A.I. agent that has been winning the Loebner Prize, an annual competition that rewards the programs that are the most human-like, for four years got 56% in Google’s tests.
Google isn’t the only one racing to figure out conversational bots. Through its Semantics Machines acquisition, Microsoft has been working towards engineering multiturn dialogue in chatbots for over two years. Samsung, at the Consumer Electronics Show earlier this year, showcased a virtual human.
Google’s Meena likely won’t make it to your devices anytime soon. But it’s that clear in the next few years, voice assistants will undergo dramatic, fundamental upgrades. We might hear more about this at Google’s upcoming annual developer conference in May.
San Francisco startup Dimensional advertised for people to walk robot dogs through city streets. The New York Times reported that the job could include carrying a roughly 33-pound robot back to the office if it overheated. A lift home is an awkward requirement for a labor-saving machine.
Keeping the hardware running is only one challenge. Anthropic's robotics tests found that language models mostly failed when controlling joints directly, but did better when directing trained controllers. Giving an AI agent a body doesn't teach it to handle ordinary household objects.
Soon, a mattress protector may be more than a bed cover. Australian medtech company Sleeptite and researchers at RMIT University have received a US patent for FlexiSense, a smart bedding technology that tracks body position and posture to help caregivers identify when older residents may be at risk of falling.
The technology is being developed for Sleeptite’s in-bed monitoring system, REMi, and was recently on display at The MedTech Conference in Boston. A peer-reviewed study in Flexible and Printed Electronics says the sensors can also be made at scale without sacrificing performance or durability.
AI companies spend enormous amounts of time convincing us their technology is safe and secure. But what happens when those same companies become targets of increasingly sophisticated AI-powered attacks? Anthropic, the company behind Claude, is now facing that question. According to a Washington Post report, cybersecurity startup Cathedral discovered vulnerabilities that let its researchers access sensitive information in Anthropic's internal systems.
The researchers used AI agents to identify potential entry points in less than 12 hours. Interestingly, they didn't need to crack Claude itself. Instead, they exploited weaknesses in third-party business applications that Anthropic uses for everyday operations. Anthropic's security testing program disclosed the findings, which allows outside researchers to identify and report vulnerabilities.