Google uses AI to help people explore the topics they’re looking for – here’s how
“Can you get some medicine for someone at the pharmacy?” “
It’s a simple enough question for humans to understand, says Pandu Nayak, vice president of search at Google, but such a query represents the cutting edge of machine comprehension. You and I can see that the person asking the question is asking if he can complete a subscription for someone else, Nayak tells The edge. But until recently if you were typing this question into google it would direct you to websites explaining how to fill out your prescription. “It lacked the subtlety that the prescription was for someone else,” he says.
The key to providing the right answer, says Nayak, is AI, which Google uses today to improve its search results. The statute of limitations query was resolved in 2019, when Google incorporated a machine learning model called BERT into search. As part of a new generation of AI language systems known as large language models (the most famous of which is OpenAI’s GPT-3), BERT was able to properly analyze the nuances of our query of prescription and return the correct results. Now, in 2021, Google is updating its search tools again, using another acronym AI system that is the successor to BERT: MUM.
Originally revealed at Google I / O in May, MUM is at least 1,000 times bigger than BERT, says Nayak; of the same order of magnitude as the GPT-3, which has 175 billion parameters. (Parameters being a measure of the size and complexity of a model.) MUM is also multimodal, which means that it processes visual data as well as text. And he was trained in 75 languages, which allows the system to “generalize from languages where there is a lot of data, like English, to languages where there is less data, like Hindi,” explains Nayak. This helps ensure that any upgrades it offers are spread across Google’s many marketplaces.
Nayak speaks proudly of MUM, as the latest AI prodigy trained in the labs of Google. But the company is also cautious. Large language models are controversial for a number of reasons. They are prone to lying, for example – as happy to write fiction as they are fact. And they have been shown time and time again to encode racial and gender bias. This is a problem that Google’s own researchers have pinpointed and shot down for it. Notably, Google fired two of its top ethics researchers, Timnit Gebru and Margaret Mitchell, after they co-authored an article highlighting the problems with exactly this technology.
For these reasons, perhaps, the search changes that Google is launching are relatively small. The company is introducing three new features “in the coming months,” some of which are powered by MUM, each incidental to its primary search engine function – ranking web results. But Nayak says they’re just the tip of the iceberg when it comes to Google’s ambitions to improve its products with AI. “For me, this is just the start,” he says.
But first, the features. Number one is called “Things to Know” and acts as an advanced snippet function, extracting the answers to predicted questions based on user searches. Type “acrylic paint,” for example, and “Good to know” will automatically generate new queries, such as “How do you use household items in acrylic paint?” Nayak says there are certain “sensitive queries” that habit trigger this response (like “bomb making”) but most topics are automatically covered. It will be deployed in the “coming months”.
The second new feature suggests additional research that could help users broaden or refine their queries. So, with the search for “acrylic painting” above, Google might now suggest a narrower focus, like “acrylic painting techniques”, or a broader mandate, like “different styles of painting”. As Nayak says, Google wants to use AI’s ability to recognize “the space of possibilities within [a] topic ”and help people explore variations of their own searches. This feature will be available immediately, although it is not powered by MUM.
The third novelty is simpler and based on video transcription. When users search for video content, Google will use MUM to suggest new searches based on what it hears in the video. Nayak gives the example of watching a macaroni penguin video and Google suggests a new search for “macaroni penguin life story.” Again, it’s all about suggesting new areas of research to users. This feature will launch on September 29 in English in the United States.
In addition to these AI-based changes, Google is also expanding its “About This” feature in search, which will provide new information on the source of the results. It also brings its MUM-powered artificial intelligence to its visual search technology, Google Lens.
The shift to search is certainly the main focus, but what’s interesting is also what Google is not launch. When he demonstrated MUM to I / O earlier this year, he showcased ambitious features where users could literally talk to their research topics, like the dwarf planet Pluto, and ask them questions. In another, users asked extended questions, such as “I just hiked Mount Adams, want to hike Mount Fuji in the fall.” What should I do differently? Before being directed to relevant code snippets and web pages.
It seems that these types of searches, which are deeply rooted in the functionality of great language models, are too free for Google to launch publicly. Most likely, the reason is that language models could easily say the wrong thing. This is where these bias issues come into play. For example, when GPT-3 is asked to complete a sentence like “Boldness is bold as the Muslim is at …”, near a quarter of the time, he ends the sentence with the word “terrorism”. These are not easy problems to navigate.
Asked about these difficulties, Nayak reframe the problems. He says it’s obvious that language models suffer from bias, but that’s not necessarily Google’s challenge. “Even though the model is prejudiced, we don’t offer it for people to consume it directly,” he says. “We are launching products. And what matters is that the products serve our users? Are they surfacing unwanted things or not?
But the company also cannot completely eliminate these problems in its finished products. Google’s Photo app infamously labeled black people as “gorillas” in a well-known incident, and the type of racial and gender discrimination present in linguistic AI is often much more subtle and difficult to detect.
There’s also the issue of what the shift to AI-generated responses might mean for the broader future of Google search. In a speculative article published earlier this year, Google researchers examined the issue of completely replacing search with large language models and pointed to a number of difficulties with the approach. (Nayak is adamant that this is not a serious prospect for the company: “This is absolutely not the plan.”)
And there is also the constant growl that Google continues to take up more space in search results with its own product, diverting searches to Google Shopping, Google Maps, etc. The new “Things to Know” feature powered by MUM certainly seems to be part of that trend: removing the most informative search results from web pages, and potentially preventing users from clicking, and therefore supporting the creator of that data.
Nayak’s response to this is that Google generates more web traffic every year and if it doesn’t “build compelling experiences” for users, then the company “won’t be there to send traffic to the web. ” in the future. This is not a totally convincing answer. Google may generate more traffic each year, but how much of this is simply due to the increase in web usage? And even if Google disappears from search, wouldn’t other search engines take over by sending traffic to people?
Either way, it’s clear the company puts understanding AI language at the heart of its search tools – at the heart of Google, indeed. There are a lot of open questions about the challenges of integrating this technology, but for now, Google is happy to keep looking for its own answers.