By Hayden Field
But the technology has drawn criticism from the AI community
Google wants your search queries to look less like a Jeopardy! answer and more like a chat with your friend—filled with the kind of slang and shorthand only a human would understand.
To get there, the tech giant is enlisting a powerful AI tool you all might remember: a large language model, specifically one called MUM (multitask unified model).
- Large language models, which are trained on datasets as large as one trillion words, help computers process and produce human-like language.
But, but, but: The tech has drawn criticism from parts of the AI community. In the past year, Google fired both co-leads of its AI ethics team after a dispute over their research on the dangers of large language models.
- One of experts’ top concerns? Models trained via internet data will naturally learn biases—then can easily replicate those patterns and multiply the resulting harms.
Google hasn’t yet announced a timeline for when it’ll incorporate MUM into live search, but it’s already experimenting with one-off projects.
Then vs. now: In 2020, Google team members spent hundreds of hours compiling the different ways people could refer to Covid in order to accurately route pandemic queries. This year, they wanted to do the same thing for queries about the Covid vaccine—so they used MUM to “generate over 800 names for 17 different vaccines in 50 different languages” within seconds, Pandu Nayak, Google’s VP of search, told Popular Science.
Big picture: Google announced MUM at its developer conference in May alongside other language model initiatives, and now it seems to be doubling down on how important the tech is to its future business model. More sophisticated searches and answers will likely lead to more valuable targeted ads, which could mean big changes for the ad pricing model, reports the FT.—HF
Feature Image Credit: Francis Scialabba