A new way to discover places with generative AI in Maps
blog.google
blog.google
From a user perspective, it's a much better experience to see a list of places rather than generative AI generating some text.
I think this is just the Maps team getting on the gen AI hype train.
Look at the way it's phrased: not "to improve the user experience" but "to bring generative AI to Google Maps".
This is Google top brass telling everyone explicitly that their products need to incorporate generative AI, and product leads brainstorming ways to do that.
The is not trying to solve a problem that users have, but a problem that Google has.
(FWIW I say this as someone who's optimistic that there is really something there behind the AI hype. But this ain't it).
I don't want GenAI in all of Google's ecosystem. I've already taken broad steps to de-Google many other important things. I guess my next step is to protect myself from Google and go fully back into the alternate Android ecosystem. Thankfully I have a secondary phone that I've been running Graphene on for the last year. Maybe Google's misdirected "everything GenAI" master plan is enough of a catalyst to push folks that have been dabbling on the edge to completely abandon Google.
The piece that really pisses me off, however is to see the head first dive into the shallow end with all of this garbage in Google's AI for the Education space [0]. Kids do not need to be subject to this impending trainwreck.
[0] https://edu.google.com/intl/ALL_us/why-google/ai-for-educati...
> You've finally found a day the whole crew can hang out. The problem? Everyone has different preferences: one friend's vegan, another won't venture uptown, and one has a dog that never leaves their side. With so much to consider, you’re going to need help figuring out the perfect place to go.
I imagine searching "place to hangout downtown that is vegan and dog friendly" is not going to return the results you're hoping for.
It feels like we are optimising ourselves to transactional unhappiness.
Maybe someone in the group has trouble walking.
I tried and it did.
‘Downtown vegan dog-friendly’ works just as well, of course.
And a lot of these queries are valid questions/searches, but don't produce the answer the searcher is looking for. Depending on how this Maps system works, it might work better for those people -- and also provide a new kind of searching for everyone else.
Here's a great example from a different domain: a while back I built a book search powered by LLMs that helps you find books that you remember random details of but can't remember much else ("that one where the woman climbs the mountain and gets stuck in a cave for a week", "that one where two cities on opposite mountains communicate by reflecting light", "that one where the main character finds a dragon egg in his grandpa's garage"). Without indexing the contents, reviews, analyses, etc of each book, an LLM was a great solution for finding the right answer to this kind of query which, anecdotally, almost always fell flat in traditional search engines (google, goodreads, amazon, etc). In Maps, I could see this kind of search working well for locations with queries like "that restaurant in KC with the giant painting of a woman spilling soup out of her mouth", "that museum with the open-air room in the middle", "that coffee shop that only hires women", etc).
[0] Tangential anecdote from the far end of the spectrum: one of my grandma's first google searches was "Hello there, I'm not sure who I'm writing to but I'm hoping you can help me. Years ago, my friend {Alice} told me about a restaurant she went to in Des Moines that had amazing meatballs. I've looked all over and tried {list of places} but I can't find it. She lived near the YMCA near Main..." More than half of the query was ignored (max 32 words on Google) and not a single result was relevant so she just closed my laptop (and rarely used technology ever again); however, I bet an LLM probably would have given a much better result.
[1] There's also an argument to be made that people who didn't grow up around technology are a dying breed, but I think we'll soon find that technology moves fast enough that we don't quite grok all the latest tech -- and in that lens, we might eventually be the ones still using "that legacy way to XYZ" instead of whatever's next.
Also I can zoom in to some strip mall and ask to see all the restaurants and I'll only be presented with a subset that Google decided were worth highlighting. No easy way to see the others.
It seems so simple to fix too. Favor close results over distant ones, and if no results are found within a couple of unzoom levels, pop a suggestion. And in particular if a string of searches is happening for different dishes, there should be a "this user is looking for something to eat right now" flag that engages to prevent serving results on the other side of the planet.
So far I haven't come across a single such opportunity for example in who is hiring threads..
Everything Alphabet has hit with their sugar-stick of AI has turned to a foetid garbage pile in just as much time. search on the google product is less than useless in most cases, its outright harmful to the pursuit of knowledge itself and the security posture of the system accessing the links provided.
Once alphabet tires of the AI hype train im sure whatever adjunct features and functions get added to maps will fester like a withering carbunkle until suddenly maps is replaced with maps(new_name) where the next product is just a reskinned rollback.
That said, it's a neat to see feature. Not a must have, certainly nothing compelling.
1. Type “lunch”. It auto suggests “lunch restaurants”
2. It already has my location but of course it can auto complete a different one.
3. Press the $ button to filter to upscale, aka higher priced, restaurants with $$$ and $$$$.
Very quick. I guess the advantage for AI is finding “favored by locals”.
And winnowing the list. Yelp's results in the Bay Area are strongly skewed by people who go to a fine-dining establishment, have a terrific meal, and then leave one star because it's priced as fine dining.
What I'm really looking forward to is describing what you want midjourney prompt style then being teleported to a hyper realistic world you can explore with quests and journey.
I see ads for Google phones where they brag about how rather than taking actual pictures, they have their AI hallucinate better pictures, and I wonder who exactly these ads are supposed to appeal to?
Already Google Maps is something of a nightmare to use -- searching for something will often refocus the current view in wild and useless ways. This is doubly true when you're trying to do a multi-part search (e.g. "I'm going to be at this store, I wonder if there's a good Thai restaurant nearby" will zoom out and show you Thai restaurants thirty miles away, and you've lost the pin from the first search). And the "explore" option just seems kind of useless and intrusive -- it is literally never what I am using Google Maps for. And the near-extortive labelling -- "in order to get your restaurant labelled as 'LGBTQ friendly' you have to create a Google account and claim the restaurant" is mind-boggling.
Adding AI to Google Maps could conceivably help with keyword labelling (like when a search for "pharmacy" fails to show up a given pharmacy because it's marked as a "drug store" or something). Maybe if I search for a good "salad place" it'll find restaurants that serve good salads but don't have them in the restaurant name. But I'm terrified of the kind of generalized hallucinations that it will provide. Might be interesting to do a "vibe"-based search from time to time ("find me a clean, well-lit, quiet bar" or a "bar with a pool table") but I suspect this will get old after a couple of false positives.
Everyone that takes pictures where: someone closed their eyes, wasn’t looking, something annoying in the background, unflattering pose, etc
Most peoples pictures are not a statement of record - we just want to remember the good times with loved ones
But on that note, having a flawed high fidelity record is no worse for remembering the good times. The hallucinations are mostly harmless, until the AI changes your eye color or adds extra teeth or all of a sudden there's an elephant in the background, even though you told the AI that you specifically wanted a picture without an elephant.
Or, more whimsically, your friend who memorably never smiles in pictures is smiling in all of the pictures with your new phone.
This doesn't describe most avid photo takers, who are presumably the people who will choose a smartphone based on its camera features.
Maybe so, but personally if I'm taking a picture, I want that picture to be of real things. I don't want it to be altered or to have artificial things added. That just adds an additional layer of separation between my and reality.
If I want a picture to be artificially enhanced, I'd be using photoshop or similar.
But I understand that's just me. One of the things I worry about with how generative AI seems to be going is that it's being used for things that further blur what is real from what is not, and we already have far too much of that as it is.
You end up in this situation where the app is adding features useful to visitors, but the vast majority of users just want a drive time estimate.
Finding the optimal route given the current and historical traffic conditions along with information of what is happening that day (e.g. A festival or sporting event) is an AI problem.
Can AI be trained to solve it? Sure, but it's a weird and perhaps un-needed use case when there are several better understood and more efficient algorithms that approach lower bounds.
I work in the AI space, love the challenges it presents and the use cases it enables. We have to be careful to not hype into AI -- applying it to places where it isn't necessary or could never work -- because it will absolutely continue to technically improve but may be overly scrutinized from a regulatory perspective.
So is the rest of AI because we are using computers which operate on numbers. OR algorithms are one tool to emulate intelligence.
Searching for a place can be done with the help of AI classifiers, but in the end, you just get a point on a map, nothing generative here. Other things that Maps can do, like pathfinding is most likely better served using traditional algorithms.
Edit: but is it possible that it makes up things that do not exist and presents it as local trivia?
A year ago a normie wedding planner was scouting locations for a couple that had a list of specific needs, nothing weird, just rooms for the night, sea view, allowed catering and had a garden space, but even if the requirements were pretty standard it's not something you can get at glance from maps or Google searches.
So it was taking significant time to open each location web site and validate them because one has to be sure and often locations doesn't explicitly have a list of features, you'd have to see photos or find the right snippet of text in whichever subsection it could have been, and the hit rate was petty low
So I tried asking gpt and quickly had 20 matches for her ready. Some were hallucinated, some we couldn't confirm the features, but more than two thirds where legit candidates, which saved a lot of busy work.
So in the end the planner opened an account and never used gpt ever again because it was weird and alien and couldn't figure out prompting.
On the positive side, Google will probably kill it in a year anyway.
Language processing gets me an immediate answer to, "hey, I'm looking for a cute date spot with pizza, high quality pizza but where you still eat it as slices".
That's a massive benefit.
But if they're training it on the reviews people leave as well as on the user-submitted photos, I see no reason why it wouldn't.
After all, when I do an analagous query about historical events or technological products in Google Bard it works just fine.
We don’t have the ability to extract the ideas you’re discussing from photos imo.
After the long walk to the third restaurant, I looked closely through the results to see why Google was giving me useless results. It was reviews! It had matched variations on "Great food but no vegan options."
I'm sure Google Maps used to have a "Cuisine" filter, which isn't really correct in this case but is a lot closer than what they're doing now.
But a lot of the time you're looking for things like "cute date spot", "trendy", "chill atmopshere", "cozy", "good for a large birthday group", and so forth.
There aren't filters for these. That's my whole point.
There's a ton of information that can be gleaned from people's reviews and photos, but that simply aren't exposed in anything filterable. And keyword search often isn't available but also doesn't often work -- e.g. searching for "budget" gives you lots of reviews that say "definitely not budget". But AI is perfect for this.
Extremely unreliable information. Isn't that made worse if what you're looking for is subjective descriptors like "cozy", "chill", "cute", etc? Those are the kinds of things that there won't be a lot of agreement on.
There's enough common agreement on what cozy or chill mean. There's plenty of signal in the noise of reviews.
> There's enough common agreement on what cozy or chill mean.
Maybe? I don't know. I do know that I often see people using the same subjective descriptors in ways that are contradictory to each other. What one person thinks is "cozy", another person thinks it "claustrophobic", for instance. Unless you know the mindset of the person using those sorts of words, it's hard to know how to interpret them.
you can't glean good information from that, not about the people but about the restaurant?
that's still so many words, though, and i'm a lazy person. i want to just type "pizza" and, if it's a really fancy app, hit a single button to filter by "casual" "takeout" "formal" etc
That said, I expect the worst from this update. My expectation is that it will serve recommendations based on AdWords or other advertising tech. And this will then transform Google Maps into even worse of a product that will be even more covered in ads than it already is.
I hope this beats my expectations, but I’m not optimistic.
Here's a talk that google gave on improvements happening in Maps 10 years ago: https://youtu.be/HrLyZ24UcRE?si=t0J0FToLO5fBsdA7
It feels like when people say "they used to make cars better in the 1950s; now those were REAL cars"
I honestly don't think it's that at all. It's that google maps has gotten worse at addressing my particular use case.
>I think every major update to Google maps in the last decade or so has made it worse. I expect no different with this one.
I just hold our community to higher standards.
You know offline maps came out in 2015? Because I suppose that wasn’t in your rubric for what’s good and not.
I'd agree in the last 4-5 years the changes have been worse. Post-COVID they seem to have focused on making a lot of negative changes by stuffing it with too many ads, worse color schemes and a lot of "features" that are half baked to try and capture users and send them to paying businesses on the platform
Interestingly, that list is mostly what I had in mind when I said that maps was getting worse.
I suppose that it entirely depends on what you want out of a map application. For me, all that stuff just gets in my way. That you think those things are great tells me that maps is no longer aiming to solve the problems that I want solved, and is addressing other problems instead.
That's not saying Google Maps is bad or wrong, just that it's a different product now, and one that isn't going to provide what I want.
So I should just pipe down. I'm not the target demographic anymore. I just need to move on to a different application and be done with it.
The problems most people have are finding places and ways to get to those places. All the changes I outlined, including continuing improvements to their directions, served that purpose for quite a while. Reviews, the commercial areas highlighting and local guides helps find shopping and food spots and I used this extensively during multiple trips with great success. Same with the transit directions and landmarks highlighting and I fail to see any reason how these made maps worse. I cannot possibly think of why adding bus and rail timings made maps worse. They are incredible features. Things I didn't originally include are also traffic conditions and suggesting routes based on that by factoring in timing which is somewhat hit or miss but still leaning towards hit in my experience
I can point specifically to things like their new Business Messaging, the ramped up difficulty to contributions and reducing them primarily to photos and reviews, and the new experimental color schemes as decidedly worse changes
My problem is very basic: I want to see an uncluttered map, to be able to search for a particular address or street and be shown it on that map, and to get driving directions.
The driving directions are OK, but the uncluttered map isn't a thing anymore. Now, the map is obscured with with quick-search buttons, that pull-up thing showing "places of interest" at the bottom, and is populated with irrelevant (to me) pins showing random businesses.
It all just makes actually using the thing as a map more difficult.
> The problems most people have are finding places
Fair enough. That's why I say that Google Maps is not aiming at people like me. I want basic navigation, in a way that is as friction-free as possible. I also have zero interest in things like reviews, local guides, and that kind of thing.
Again, I'm not saying that what you want isn't important, or even that what you want isn't what most people want. It's just not what I want, and the presence of those things makes using the app for my use case more difficult.
So I just need a different product that is aimed more at my needs.
And, of course, Android Auto.
But for discovery? When I look around my neighborhood, I see Circle K, Walgreens, and Wendy's, all with their colorful logos. I don't know for sure, but now it feels like a yellow pages of advertised businesses where it used to be a map.
When I use OSM or Google Earth, I see the parks and street names. I can search for Wendy's if I want that, but it's not promoted info.
Imagine finding guides on HN about how to make your shop appear first when searching for any of the top 500 words.
This is an incredibly hard problem since businesses have such strong incentives to game the system, so it’s hardly new but I think AI is going to make it even harder to spot fake accounts.
"We fed it into this black box, and this is what we got, we don't know how it works". Of course the training data will probably be a weighted output of who pays google, rather than what will make the user happiest.
It's pretext to violate reviewing systems without looking obviously corrupt.
I just want to see the top 10 most rated things per category, then I want to see the top 10 things over a given threshold of reviews. I don't want googles mangling of the data, I just want to be able to query the review data.
If whatever google displays isn't better than the top 10 results in a given area sorted by number of reviews, it's clear there is a reason why. Whether it's devs creating unnecessary complexity to justify their employment, or a system designed to weight results in favor of who pays google, or the need to inject 'randomness' in order to be able to inject paid advertisements without them being 100% ignored, or to provide a reason to keep scrolling allowing more ads to be displayed, it's all anti-user behavior. This will just be another tool to manipulate a user rather than help them.
That is the problem. You do not want a web search engine that gives you a WEB PAGE for your query. You want an answer AI.
Google can not find a page nobody has written, and if several people write about a topic, there is just no way Google can figure out which one should be the first result correctly.
When Google talks about finding relevant results, they just mean if you type "hacker news" on Google the first result will probably be this website instead of news about hackers who hacked a bank or something. That is what "relevancy" means. When there is a single, canonical, unambiguous web page that matches a query exactly.
Everything else is just expecting more from Google than it can feasibly provide and then getting angry when it tacks AI on everything because it can't do magic.
As a separate service distinct from web search, sure. But I also want to be able to search the web and get lists of pages.
> That is what "relevancy" means. When there is a single, canonical, unambiguous web page that matches a query exactly.
This is what I don't want in a search engine. Unless that search engine figures out a way to read my mind, anyhow, which would bring up a whole host of other issues. Google trying to guess at what I want rather than just giving me a list of pages that match my query is the main problem that I have with Google search.
It's a SEARCH engine, not a BROWSE CATEGORY engine.
And I get it, I get frustrated at this too. Sometimes I try to search for a category of something, like free image editors, and Google just can't give me an answer I'd like.. And it's really not a hard query to answer. It just lists me a bunch of names with zero info about what platform they are for, or what features they have. It's an useless list. It's simply beyond the scope of what it is made to do.
I mean, seriously, I know Google is a blackbox, but Google is really just a single textbox. We're expecting too much from a single textbox.
Not at all, as evidenced by other search engines that are successful at this and the fact that Google used to be successful at this.
> It's a SEARCH engine
Precisely. I want it to search the web and return a list of sites that match my search criteria. I don't want it to to answer a specific question for me, nor to try to guess what I "really" want. That should be done by a different tool.
> We're expecting too much from a single textbox.
I'm not sure what you think that I'm expecting, but what I'm expecting is much simpler than what Google seems to be trying to do.
I suspect that I've failed to adequately explain my use case and we're talking about different things, but I'm not sure where I went wrong so don't know how to clarify.
"You don't want a tool, you want to remodel your bathroom" is what the other person is arguing. That's kind of true.
We're arguing that we don't want someone tell us how to remodel our bathroom, we just want a tool we can use to remodel our bathroom ourselves.
Google said that providing a tool doesn't make enough money, so they transitioned into being a contractor.
The crux of the issue is the "-" operator. As a tool, it removed any search result with the word after the "-" operator on the page in a very specific way. If you weren't getting the results you wanted, it's because you didn't think of how to use the tool in the right way. As a contractor, they try to interpret what you want to subtract and then give you what they think you want.
The person you are responding to might be too young to remember that google was actually a very good tool at one point.
Like all hype cycles, the syrup ends up everywhere. You need an actual user need to have the long-term product justification to add features.
There probably is one with maps. It just may not be this early.
edit: maybe I was thinking of Google (https://www.2gb.com/mans-family-sues-google-maps-after-it-di...)
Some Apple maps events: https://www.theregister.com/2012/12/12/another_apple_maps_me...
So, yeah, I am not sure Google “supercharged” AI id going to work for a product that requires a precise but fast user feedback loop.
More discussion earlier: https://news.ycombinator.com/item?id=39218367
If the answer is "regular search results are crap because of irrelevant sponsored placements being put before the thing I want", that's not a function of the search engine; that's revenue-maximising choice the search engine provider makes... so why do you think that won't happen to LLM results by the company that provides it?