Featured

Luka

A.I. recommending restaurants in SF in a chat interface

Featured comment

DiscussionYou need to become a Contributor to join the discussion - Find out how.
Eugenia Kuyda
Maker
@ekuyda · Luka
Luka’s conversational technology can be applied to other verticals (restaurants are the first step). Consumer AI will be huge, and chat as a UX is intuitive. Luka is the first startup that is building practical consumer AI based on a chat interface now. We're waiting for an update to be approved by the AppStore this week - Luka will get rid of Facebook authorisation (phew) and will be able to reserve tables.
Stewart Rogers@therealsjr · Journalist/Analyst/Speaker - VentureBeat
@ekuyda Any plans for an Android version?
Gillian Morris@gillianim · CEO, Hitlist
@ekuyda This concept is huge in Asia - WeChat and others already allow consumer AI, and Path's TalkTo functionality also does this somewhat (@tomlimongello could say more on that, I think). And of course Siri was supposed to do all this, though via voice rather than chat. (@libovness's piece linked below has a good analysis of why chat is more promising than voice, at least for now: http://whoo.ps/2015/02/23/future...) But Siri has never been very good, and I'm excited to see another entrant in the space, especially one in my favorite industry, travel. It's worth considering the limits of what AI will be able to parse, though. There's a great discussion of the limits of chat-based AI in Brian Christian's book, The Most Human Human: http://www.amazon.com/The-Most-H.... Essentially, we can probably get to 90% of use cases, but it takes a book to even start explaining why cracking the final 10% will be so hard. On another note, there are distinct similarities to what @xdotai does for email scheduling. I love things like this that don't require another app download or making you learn a new UI.
Jonathan Howard@staringispolite · Growth engineer & founder
@ekuyda I really wanted this to work, but it went the same way Chat AI always seems to: awkward uncanny valley conversation and/or outright failure to get the thing I want. @rrhoover suggested I share this funny exchange I had with Luka on my first try :) Trying to keep an open mind, so let me know I'm "doing it wrong" somehow!
Eugenia Kuyda
Maker
@ekuyda · Luka
@tomlimongello @libovness @xdotai @gillianim Thanks, Gillian! WeChat is definitely the best example of what a messenger can be - not only for consumer AI but also for offering so many services in a chat interface (it's the biggest mobile wallet in the world for now). The Most Human Human is perhaps my favourite book on the subject, gives a very clear perspective on what it really means to be 'human' in a conversation. I believe there are limits at this particular point of history of what AI can do in regards of conducting a human-like conversation, but I strongly believe in assisted AI, where the most difficult cases would be solved by humans (the last 10%), while technology becomes better and better to finally take humans out of the equation. Next week we're launching as a text number - so no need to download the app anymore :)
Eugenia Kuyda
Maker
@ekuyda · Luka
@rrhoover @staringispolite @rrhoover @staringispolite Hi Jonathan and thanks a lot for your comment - these things make Luka better. We're aiming at a 90% quality meaning that for every 100 replies Luka will have adequate and relevant responses in at least 90 cases. However there are still 10% of mistakes (and right now for Luka it's a little more, around 16%, but it's getting better every day), and those could make for a bad experience for the user. We fixed the issue that we found in your conversation (the place you were looking for is called Red Door Coffee, not Red Door Cafe, which is another place in Bush street - we're learning how to handle all those different ways that people spell restaurants and cafes, and we're getting better with all those typos and disambiguations). The more you talk to it, the better it becomes, and hopefully we can make it so that no users get the uncanny valley conversation like you unfortunately had. Thanks a lot for trying the app out, I would be very happy to see it working for you!
Jonathan Howard@staringispolite · Growth engineer & founder
@rrhoover @ekuyda Thanks for the reply! Coffee -> Cafe was actually a voice-to-text error I didn't catch until I took the Yelp screenshot above. Weird that it pretended to know the incorrect Red Door Cafe first, then progressively forget more and more context from the conversation. For what it's worth, this uncanny valley conversation was what made me feel like quitting, not so much the result of finding/not finding any one query. I think I would've kept going if Luka had replied with any of the following instead: 1) Did you mean Red Door Cafe in Pac Heights? 2) Did you mean Red Door Coffee in SoMa? 3) "I didn't find that listing in Soma" 4) "I can't find that - did you mean one of these? [UI list to choose from]") Also note that Yelp handles the incorrect spelling just fine. If the AI were just searching Yelp and guessing the top result that's in SoMa, it would've worked. Keep on fighting the good fight!
ilan kasan @ilankasan · Product and Business Leadership
@ekuyda nice product and great UX. Take a look at www.commersational.com which enables merchants to chat eable their ecommerce catalog and transact via text
Ryan Hoover@rrhoover · Founder, Product Hunt
I just gave it a test run (made with Tailor): Very impressive. I see you have integration with Yelp (and probably other 3rd party services) but how are you deciding which restaurants to recommend and keeping your database up-to-date?
Eugenia Kuyda
Maker
@ekuyda · Luka
@rrhoover Thanks! We collect all data automatically from more than 15 sources - Luka parses data from user reviews (Yelp, Foursquare, TripAdvisor etc) and professional reviews and blogs (from SF Chronicle to Eater) and guidebooks (Michelin, Zagat). We plug all that data into our data model, which is pretty elaborate (we have more than 300 data points for each restaurant, plus menus and photos) and also use some relevant quotes to show our users. The recommender system is based on a knowledge-based recommender system. Luka's knowledge graph stores all the connections users may have with any possible entities (whether you only eat gluten-free food, or like only the places that are walking distance, or hate chinese) and matches that with all the data we collect about the restaurants (that are part of the graph too) - that makes recommendations hyper personalized and relevant.
Shaun Trennery@shauntrennery · Founder X&Go
Love everything about this product. AI is not ready to take the controls alone thou. Here is an example of it going wrong. https://pbs.twimg.com/media/B9JU...
Eugenia Kuyda
Maker
@ekuyda · Luka
@shauntrennery Thanks for your comment! That's an issue we've just fixed - by tomorrow yo won't see negative reviews when Luka recommends a place. Thanks for your feedback!
Irving Torres@irvingtorresyc · Freelance Designer
@shauntrennery I completely agree. My experience thus far has been fun but every time I ask a clarifying question it recommends a new place instead. The grammar is also a little off. Very interesting to see where this goes though!
Eugenia Kuyda
Maker
@ekuyda · Luka
@shauntrennery @irvingtorresyc Thanks, Irving! It gets better the more people talk to it - it's still very raw, but improving every day. The more you talk to Luka, the better the experience becomes for you... Thanks for your feedback - send me whatever you think works wrong at kuyda@luka.ai
Frederick Tubiermont@ftubiermont · ceo adsy.me
@shauntrennery what about employing real people instead of just relying on AI? There are people out there who would be ready to become chat assistants I think ;-)
Shaun Trennery@shauntrennery · Founder X&Go
@ftubiermont Agree. Prompt agents with one or more options.
Eugenia Kuyda
Maker
@ekuyda · Luka
@ftubiermont @shauntrennery it's great when it's assisting AI in rare and hard cases, but we figured that users don't want to wait for real people's answers as it takes significantly longer (at least 10-15 seconds for an answer, up to a few minutes if you need to give a recommendation, quote a review and send the menu). We tried a Wizard of Oz experiment at the very beginning in Russia to prove the concept, but we figured it would be hard to replicate chat bot experience with humans, as strange as it sounds.
Ben Tossell@bentossell · Community Lead, Product Hunt
Like I said on Twitter...I think this product is awesome. Such a natural way to discover recommendations. Would love to read about your story and how Luka came about! Is there a way (or going to be a way) to review after you have been to a restaurant, so that Luka learns your likes/dislikes - i.e X needed better service
Eugenia Kuyda
Maker
@ekuyda · Luka
@bentossell Thanks, Ben, it's really nice to hear that! There is a way to teach Luka - it usually would ask you how it went and what you liked at a place you went to, but that's something we're working a lot on now. You will soon be able to send Luka photos and chat about the restaurant while you're there too. The main thing about Luka is that people share a lot in a conversation - so the recommendations eventually become hyper personalised. Yelp will never know stuff that people tell Luka - like whether they eat only gluten-free food or like places with hipsters.
Ben Tossell@bentossell · Community Lead, Product Hunt
@ekuyda Awesome, got ya! Really interested and keen to see how this develops :)
Eugenia Kuyda
Maker
@ekuyda · Luka
@bentossell Thank you! We'll keep you posted :)
Jonathan Libov@libovness · I can lift 100 lbs. right over my head
Looks very cool. Added it to a blog post that I've been drafting over the past few days on all the innovation that's happening in text as a medium: http://whoo.ps/2015/02/23/future...
Eugenia Kuyda
Maker
@ekuyda · Luka
@libovness Thank you Jonathan! The blog post is super useful! QuickType experience is awesome. We are actually planning on adding some new cool keyboards to Luka - so that sometimes you get a GUI in a keyboard box (works for picking a time slot for a reservation, or a price range for instance). Your idea about integrating services that could respond directly to your input is awesome. I believe that's how AI might work - as part of your messaging thread (I think Emu tried a similar approach). I really believe in texting more than in making Luka voice-operated. Texting as you mentioned allows for much more different types of interactions - for us that means we can add group chatting soon, for example, and many other features that would be impossible if that was all by voice. In the end of the day, voice is only necessary when you're driving, which is a pretty limited use case.
Jonathan Libov@libovness · I can lift 100 lbs. right over my head
@ekuyda Always bet on text! http://graydon2.dreamwidth.org/1...