Facebook BlenderBot chat AI no longer has the mental ability of a goldfish


Last April, Facebook’s Intelligence Research Laboratory (FAIR) announced and announced as open your BlenderBot social chat program. While neophyte AI immediately proved itself far less prone to racist outbursts than previous attempts, BlenderBot was not without flaws. First, the system had the ability to remember a goldfish – any subject or data that AI was not initially trained simply did not exist in its online reality, as evidenced by OG BB’s continued insistence that Tom Brady continue to play for the New England Patriots. On the other hand, due to limited knowledge of current events, the system had a strong tendency to hallucinate knowledge, such as the digital Dunning-Kruger effect. But progress BlenderBot 2.0 is displayed, which FAIR unveiled on Friday, should make AI far more sociable, educated and capable.

Although BlenderBot 1.0 could only maintain memory for one discussion, its successor can remember conversation topics over multiple conversations that can last for days, weeks, or even months, thanks to the implementation of a long-term memory module. Moreover, AI can actively update its knowledge base by searching the Internet for the latest news and details on any topic the user wants to discuss.

“BlenderBot 2 asks the Bing API for search results based on the generated search query and conditions its response with the first few results,” Kurt Schuster, a research engineer at Facebook AI, told Engadget. “We rely on Bing to provide high quality search results.” As such, BlenderBot 2.0 is now able to talk coherently about breaking news and new media, not just the data on which it is trained.


“BlenderBot 2 is limited only by what a powerful search engine can provide,” added Jason Weston, a researcher from Facebook AI. For example, if you are more interested in learning about Tom Yewčić (Patriot’s combined QB / Punter from the 1962 season) than about Tom Brady, BB 2.0 has covered that. It’s the same with students, like photosynthesis or redox reactions, Weston continued. As long as the information is available on the web, “there’s no reason BlenderBot 2 can’t discuss this.”

By actively searching for information on the Internet, BlenderBot 2.0 can also reduce the cases in which knowledge hallucinates. “Providing the system with more reasonable thinking will allow BlenderBot to ensure it doesn’t confuse subtle concepts,” Weston explained, “like a film director versus a producer or a throwing coach versus a hitting coach.”



The only wrinkle really happens when discussing non-English based media, such as Demon Slayer: Kimetsu no Yaiba. “It is reasonable to conclude that Bing will release information about it and BlenderBot 2 can use that information accordingly,” he said. “We’re currently focusing on English search results, so non-English references may not be fully covered.” However, the system will recognize this Demon Slayer you are interested and more likely to raise mango-focused topics in future discussions.

FAIR has taken several steps to ensure that BlenderBot does not become the next Tay. “BlenderBot 2 doesn’t learn directly from user input, as Tay did,” Shuster said. “We have taken extensive security steps to ensure that BlenderBot 2 can handle opposing users. In particular, we use sealing techniques and two-phase techniques. BlenderBot 2 can detect for itself whether the incoming context will result in an offensive response, and additional security layers where the security classifier can detect if the user input or bot output is offensive. Each deals with the answer appropriately. “

And while the system is currently focused on chewing through the English language corpus, FAIR sees BlenderBot eventually expanding to other languages. “While it’s not in our immediate plans, our team’s goal is to build a superhuman interlocutor,” Shuster said. “This kind of agent requires multilingual understanding.”

Recent internal benchmarking processes have revealed that BlenderBot 2.0 has surpassed its predecessor by 17 percent in its engagement and 55 percent in the use of previous interview sessions with human evaluators, according to the FAIR blog on Friday. Moreover, BlenderBot’s rate of knowledge hallucinations dropped from 9% (!) In BB 1.0 to only 3% in the current iteration.

Looking ahead, “the people who communicate with artificial intelligence systems through discourse are the future of AI,” Weston argued, “and ensuring that people have an engaging, informative experience is crucial to that future. BlenderBot 2 combines BlenderBot 1.0’s engagement with system knowledge capabilities with full Internet access, so we’re reportedly on the right track. “

All products recommended by Engadget are selected by our editorial team, regardless of our parent company. Some of our stories include associated links. If you purchase something through one of these links, we may earn an associated commission.

Source link


Please enter your comment!
Please enter your name here