Google I/O has been a bit low-key in the past few years—no surprise with the ongoing global pandemic. This year's Google I/O still wasn't the in-person soirée to which we've become accustomed, but there were a surprising number of hardware announcements like the Pixel 6a and Pixel Watch that attracted the most attention. However, one announcement might end up being the most important in the future: Its auto-summarization announcement, which debuts in Google Docs, leverages years of Google's AI advances to potentially save us all a lot of time. It's also fascinating from a technical perspective.
Attention to detail
Google has spent years leaning heavily on machine learning, but unfortunately, this stuff is deviously complex and opaque to users. Google and other firms regularly report huge breakthroughs in the way AI understands the world, but it's hard to grasp the importance of these advances until they filter down into features we use every day. When they do, though, hoo boy.
The AI magic of Google Photos comes to mind—It's incredible at classifying images. On numerous occasions, I've gone from vaguely recalling a photo to actually finding the exact image in Google Photos in just a few seconds. This is a demonstrable example of AI making technology more useful, and automated summarization in Google products could be similarly innovative as it moves from tech demo and esoteric AI theory to a real product.
The lineage of Google AI summaries reaches back to the "transformer" model devised by its researchers in 2017, which led to language processing and generation systems like GPT-2 that are capable of self-supervised pre-training. That means much of the early AI training happens without human oversight, so it's faster and more adaptable than older methods like recurrent neural networks. More recently, Google refined these techniques for the Pegasus text summarization model in 2020.
Transformers have been a big deal for natural language understanding (NLU) and natural language generation (NLG) because they can replicate something you and I take for granted: attention. Our brains have the ability to focus on distinct tasks, filtering out irrelevant information... or at least what your brain thinks is irrelevant. That's what lets you zero in on a difficult problem, but it can come back to bite you, for example, if you're tinkering with your radio or (god forbid) texting while you're supposed to be operating a motor vehicle. Getting a machine to do this as well (or hopefully better) is no simple feat, but Google has created a model with Pegasus that allegedly focuses on the right data.
To train Pegasus, Google provided the AI with news articles and other long web documents, but it removed entire sentences and challenged the machine to recover them. Yes, that sounds like something even we humans would have trouble doing, but a computer can churn through data 24/7 until it begins to understand language. While Pegasus is capable of self-supervised learning, Google says that the model still needed a little fine-tuning. However, it only took about 1,000 supervised examples to tweak the model, whereas older transformers might need 10,000 or more to do a worse job. With some refinement to make Pegasus smaller and more efficient, we end up with a hybrid AI that can digest and summarize long documents.
Making sense of your digital life
Google CEO Sundar Pichai talked about the upcoming summarization feature near the start of the keynote, which suggests to me that Google sees this as an important initiative. Automated summarization will come first to Docs, but that's just the start. Google also showed an early example of automated summaries for Google Chat, with short synopses for missed conversations. I would kill for something like this in Slack, which makes it almost impossible to catch up on multiple channels in the morning or after a long meeting. Google even plans to generate summaries of meetings in Google Meet, presumably with voice transcription.
I'm not a developer or AI expert, but I do spend long stretches of my day editing and analyzing long documents. Creating a coherent summary is hard and takes time, and the fact a machine could do that instantly is amazing. If, indeed, it can do that. We don't know how good these summaries will be or if there will be ways to fool the AI. I could imagine people plugging in text with the intention of getting humorous or offensive summaries. This still feels like the start of something important. Might it even be smart to tweak the way we write long documents to make them easier for an AI to summarize for us? Maybe!
Even if the summaries aren't perfect, Google will give you the chance to address that. AI summaries will be optional, and you'll be able to edit them before anyone else sees them. Perhaps they'll get more accurate over time. The Chat example shows up and downvote buttons to rate the summaries, similar to the way human volunteers evaluated Pegasus in testing. Allegedly, testers didn't rate human-generated summaries higher than the AI-generated ones, so we might be surprised at how accurate the AI already is. We're all bombarded with information all day, and an AI that can understand your personal data and summarize everything could be life-changing.