Why Google's AI Overviews being so horribly wrong is O.K.
Ok, so, here's what happened.
During their annual developer conference called I/O, Google released a new feature called "AI overviews." When you search on Google, apart from the regular search results, you also get an AI-generated summary at the top. This summary provides the gist of what you're looking for based on the search results.
Some people even claimed that this feature will kill Perplexity, the AI startup trying to compete with Google. (the video announcement, which was cool, is in the X post)
However, as millions early access users started stress-testing "AI overviews", all kinds of crazy answers began to surface. Here are some of the legendary ones for your eyes only.
Some of the most ridiculous answers were obviously faked, but no one fact-checks before sharing these days, so everyone just spread the fakes, causing an outrage by millions of people who thought these overviews were real.
Even without the fakes, there were many silly and wrong answers. Reactions ranged from "this is funny" to "Google is being irresponsible and this is very harmful." The latter type of comments were fueled by answers like the one below, where the AI purportedly suggested that women should smoke during pregnancy.
The thing is, AI is a new technology, and no one fully understands or knows how to deal with it. Google's security teams tested the feature extensively (source), but it's impossible for any group of engineers, no matter how smart or numerous, to anticipate all the searches that millions of people will execute once they get their hands on a new feature.
It's easy to underestimate how hard it is to detect every corner case that can generate wrong or harmful answers. Google processes billions of queries daily. Even a 0.0001% failure rate means thousands of bad answers every day.
Releasing features early to the public is the only way to get massive feedback on what can go wrong while AI is still in its early stages.
Given this, I think it's OK, even great, that Google releases these features early. It helps the public understand that today's AI systems are far from perfect and should not be blindly trusted. It also reminds everyone that answers from Gemini, ChatGPT, Claude, and others should be taken with a grain of salt. This public testing is also good for Google and other AI companies. It provides invaluable information to help us understand AI, and make it safer and more effective for everyone.
We need to figure out how to control and align AI with human values, and the most effective way to do this is to let everyone try it, observe when it behaves badly, and then fix it.
One of the worst outcomes would be if companies like Google stopped sharing their work out of fear of criticism. We would then lose the chance to see how AI performs in unexpected situations, and increase the risk of creating something actually dangerous.
I mean, if you can't see that an AI-generated sentence advising you to smoke 2-3 cigarettes per day is clearly a malfunction, you might be better off avoiding the internet.
So, please Google, keep releasing things early and fix them as you go. It's for your own good, and ours as well.
My personal favorite I/O announcement
Despite AI overviews getting all the attention, my favorite part of Google's I/O announcements was Project Astra. It's an AI model that can process audio, video, and text all at once (making it multi-modal), paving the way for the era where everyone might have their own personal JARVIS. I liked this so much that I made a naive recreation.
It's incredibly exciting that we live in a time where someone can create a demo like this with just one hour and 100 lines of Python code.
โ
๐ From the Internet
โ
|
|
|
โMaxime, one of my favorite humans, sent me this word game: https://neal.fun/infinite-craft/. Neal, the game's creator, has a lot of tiny apps that are funny and smart in one way or another. This is one of my favorites.
Can you create a spaceship to time travel through the Multiverse? It's possible. I did 179 words ๐
|
|
|
A company called Sightful released a new computer they call "Spacetop". It's a combination of a keyboard and AR glasses, which allows you to work on a 100" virtual monitor. I don't know how to feel about the form factor, but IT'S SO MUCH FUN that after decades of tasting the same hardware designs people are finally trying out new things. |
