“Can you get 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 questioner is asking if they can fill out a subscription for another person, Nayak tells The BlueHillco. But until recently, if you typed this question into Google, it would direct you to websites explaining how to fill out your prescription. “It missed the subtlety that the prescription was for someone else,” he says.
The key to delivering the right answer, says Nayak, is AI, which Google is using today to improve its search results. The prescription query was solved in 2019, when Google integrated 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 parse the nuances of our prescription query correctly and return the right results. Now, in 2021, Google is updating its search tools yet again, using another acronymized AI system that’s BERT’s successor: MUM.
Originally revealed at Google I/O in May, MUM is at least 1,000 times bigger than BERT, says Nayak; on the same order of magnitude as GPT-3, which has 175 billion parameters. (Parameters being a measure of a model’s size and complexity.) MUM is also multimodal, meaning it processes visual data as well as text. And it’s been trained on 75 languages, which allows the system to “generalize from languages where there’s a lot of data, like English, to languages where there’s less data, like Hindi,” says Nayak. That helps in ensuring that any upgrades it provides are spread across Google’s many markets.
Nayak speaks of MUM with pride, as the latest AI wunderkind trained in Google’s labs. But the company is also cautious. Large language models are controversial for a number of reasons. They’re prone to lying, for example — as happy writing fiction as fact. And they’ve been shown time and time again to encode racial and gender biases. This is a problem that Google’s own researchers have highlighted and been shot down for doing so. Notably, Google fired two of its top ethics researchers, Timnit Gebru and Margaret Mitchell, after they co-authored a paper highlighting problems with exactly this technology.
For these reasons, perhaps, the changes to search that Google is launching are relatively restrained. The company is introducing three new features “in the coming months,” some powered by MUM, each of which is ancillary to its search engine’s primary 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. “To me, this is just the start,” he says.
First, though, the features. Number one is called “Things to know” and acts as an advanced snippet function, pulling out answers to predicted questions based on user’s searches. Type in “acrylic painting,” for example, and “Things to know” will automatically generate new queries, like “How do you use household items in acrylic painting.” Nayak says there are certain “sensitive queries” that won’t trigger this response (like “bomb making”) but that most topics are automatically covered. It will be rolling out in the “coming months.”
The second new feature suggests further searches that might help users broaden or refine their queries. So, with the “acrylic painting” search above, Google might now suggest a narrower focus, like “acrylic painting techniques,” or a broader remit, like “different styles of painting.” As Nayak puts it, Google wants to use AI’s ability to recognize “the space of possibilities within [a] topic” and help people explore variants of their own searches. This feature will be available immediately, though it is not powered by MUM.
The third new feature is more straightforward and based on video transcription. When users are searching for video content, Google will use MUM to suggest new searches based on what it hears within the video. Nayak gives the example of watching a video about Macaroni penguins and Google suggesting a new search of “Macaroni penguin life story.” Again, it’s about suggesting new areas of search for users. This feature will launch on September 29th in English in the US.
In addition to these AI-based changes, Google is also expanding its “About This” feature in search, which will give new information about the source of results. It’s also bringing its MUM-powered AI smarts to its visual search tech, Google Lens.
The change to search is definitely the main focus, but what’s interesting is also what Google isn’t launching. When it demoed MUM and another model LaMDA at I/O earlier this year, it showed off ambitious features where users could literally talk to the subjects of their searches, like the dwarf planet Pluto, and ask them questions. In another, users asked expansive questions, like “I just hiked Mt. Adams, I want to hike Mt. Fuji in the fall. What should I do differently?” before being directed to relevant snippets and web pages.
It seems these sorts of searches, which are rooted deeply in the functionality of large language models, are too free-form for Google to launch publicly. Most likely, the reason for this is that the language models could easily say the wrong thing. That’s when those bias problems come into play. For example, when GPT-3 is asked to complete a sentence like “Audacious is to boldness as Muslim is to …,” nearly a quarter of the time, it finishes the sentence with the word “terrorism.” These aren’t problems that are easy to navigate.
When questioned about these difficulties, Nayak reframes the problems. He says it’s obvious that language models suffer from biases but that this isn’t necessarily the challenge for Google. “Even if the model has biases, we’re not putting it out for people to consume directly,” he says. “We’re launching products. And what matters is, are the products serving our users? Are they surfacing undesirable things or not?”
But the company can’t completely stamp out these problems in its finished products either. Google’s Photo app infamously tagged Black people as “gorillas” in one well-known incident, and the sort of racial and gender-based discrimination present in language AI is often much more subtle and difficult to detect.
There’s also the problem of what the shift to AI-generated answers might mean for the wider future of Google search. In a speculative paper published earlier this year, Google’s researchers considered the question of replacing search altogether with large language models and highlighted a number of difficulties with the approach. (Nayak is definitive that this is not a serious prospect for the company: “That is absolutely not the plan.”)
And there’s also the consistent grumbling that Google continues to take up more space in search results with its own product, shunting searches to Google Shopping, Google Maps, and so on. The new MUM-powered “Things to know” feature certainly seems to be part of this trend: filleting out the most informative search results from web pages, and potentially stopping users from clicking through, and therefore sustaining the creator of that data.
Nayak’s response to this is that Google delivers more traffic to the web each year and that if it doesn’t “build compelling experiences” for users, then the company “will not be around to send traffic to the web” in the future. It’s not a wholly convincing answer. Google may deliver more traffic each year, but how much of that is just a function of increasing web use? And even if Google does disappear from search, wouldn’t other search engines pick up the slack in sending people traffic?
Whatever the case, it’s clear that the company is putting AI language understanding at the heart of its search tools — at the heart of Google, indeed. There are many open questions about the challenges of integrating this tech, but for now, Google is happy to continue the search for answers of its own.