These are the chatbots that don't work. They're simple scripts that can not handle new information, and fail to recognize requests that fall out of a very narrow path of wording.
It's good at some tasks, but it's ranging from inferior to inadequate for many others. The biggest drawback is its opaque affordances. It's not immediately apparent what is the range of things I'm able to ask a chatbot about, even if its language processing is advanced. This is a problem with humans as well. Like if you ask me which questions I'm able to answer, I won't be able to provide an answer.
Overall I feel like natural language interfaces is like 3d cinema. It's a novelty more than bringing much actual value.
Maybe people use these chatbots for different things but for me they're only ever as useful as what they let you do.
LLM technology letting them carry on a more realistic conversation with me, all that stuff, I don't really care about. All that matters is what it's empowered to do, where it can hook into the real system underneath. I don't really have much sense that this will be expanded vs. the kinds of systems we have now.
Hi! We've thought the same and started creating an AI Assistant that is a step further from a chatbot. Check out www.prometh.ai PS Still in early stages
Why, of course they want to, and they do. Not _you_, perhaps, but it’s like billboard ads: the thousands of eyeballs ignoring them are being outweighed by the few who react and drive revenue.
That reminds me of a billboard in my town, like enormous LED panel displaying adverts... that someone put behind a huge tree. So in the summer you can only mostly see the corners of the billboard. I wonder if their customers know nobody sees their ads.
Spare me the paragraphs of 'how did we get here' history and get to the point that's already in your title.
Based on what many site operators see anecdotally, people actually do want to talk to the chatbot because they just want answers to their questions and hand holding. How it's implemented/effectiveness is variable -- they will talk to it for a bit if they perceive they are being helped in some way. Chatbots and general site chat interfaces didn't spread everywhere without at least some data
I think the point was that novel chatbots will still be created, but they need to be talking to your existing assistant (Siri/Alexa) rather than you going directly to them?
His point seems to be that assistants like Siri and Alexa will be the entrypoint for LLM type interactions, which is sort of non-sensical considering most people are using chatGPT right now. I think this could have been an interesting article about how interacting with chat bots still kind of sucks but instead we got some unsubstantiated view that assistants that a lot of people have already written off are going to take over the space.
Many have written them off because they kinda suck a lot now. But they will be much much better, won't they? 5 years lets say? What kind of LLMs will we have then? If it's reliable, why not just have one interface?
I think that's still not a safe assumption, I don't think it's understood enough to know if the poor performance isn't just a fundamental limitation of some kind. You can already see models taking alternative approaches as they hit diminishing returns on the transformer innovation that triggered the massive improvement. I am just a punter onlooking though, no expert.
That's the same, people go to a place where chat interaction is the primary interface to talk to their preferred chatbot. That can be ChatGPT. But no one goes to allstate.com to talk to their chatbot.
Bold of you to think everybody is using Siri/Alexa/other conversational crapola. So many questions: in which language is this one set right now? Is my accent in that language "native" enough? Is yelling around the house an option at this hour? Do I want the passersby listening in? Do I want my conversations being sent to their tech overlords? So many "no" answers that the average techie wouldn't seemingly understand...
The paragraphs upon paragraphs of history made me click out of that article. There’s no relevance to the point being made.
In regards to ChatBot usage, in my limited interaction with support bots, they are usually quite useless. Each time I use one, it’s a game of ‘get to the actual support agent’.
All other functions are self-service anyway, at least I've never interacted with a chatbot that could do anything useful, besides being forced upon me to navigate the "why do you want to talk to a person, so I connect you with the right one?"
With all due respect, if I am interacting with your chatbot it's usually because your website provides zero ways to reach an actual human. I have lost count of the amount of times I had a specific problem not addressed in the FAQs, looked for a contact form for 10 minutes then reluctantly clicked on a chat button, only for the bot to continue to try to direct me to the answers I have already indicated did not help me. If I get lucky there's one option buried below a dozen other questions that lets me talk to a person (or, hell, even a sufficiently annealed language kernel) who can demonstrate a knowledge of the system outside of the two or three most likely footguns to avoid.
The problem (and I don't even know that the article adequately addresses it) is not that your bot is insufficiently good. It's that the bot is substituting for a support agent that wouldn't be sufficiently useful, because the support agent would also be made to operate according to a script, because the whole goddamn system is designed to make human interaction into an API. Because that's what businesses want. It's "scalable". And it only has the perverse incentive of making it hard enough to solve moderately complex problems that the user gives up.
I found maybe the first chatbot I found helpful today. LlamaIndex has a chat bot built into its documentation. Helped me answer some quick questions and gave me a mostly working code snippet for my use case.
A documentation bot is a great idea. I’d like a man-bot in Linux.
I’d like to be able to type “hey tux, restart networking”. Don’t make me dig through /etc and figure out what kind of system this is. Tell me what it’s going to do and if I say “yes”, do it.
I can't help but wonder how much of the Chat** hype is driven by a frustration with the state of modern user experiences. The dream it seems to tap into is "You don't need to deal with the arbitrary whims of 5 different groups of web designers, just talk to one thing and get a single response." When faced with the state of the modern web chatbots actually are preferable, sorry.
A great problematic side effect of the web being so ad-driven is it leads to confusing the user interface, which can host ads, with the information. We need publishers to be able to make money from content without ads, and to be able to make money from providing it in raw form via APIs to third parties. It's that or the chatbot intermediaries are going to take over.
Dane with dazn, they started serving me ads before starting any stream last month so I'm going to cancel the subscription at the first opportunity I have to get into the cable company shop (because of course you cannot do that via web)
I cancelled Paramount Plus when I heard about the content deletions. I paid for ad-free and they started showing me unskippable preroll ads for their own shows before anything. They absolutely do not care about their customers.
Isn't it just cable all over? In the beginning cable sold itself as TV without ads, then came the ads. It seems to be exactly what's happening with streaming now.
I can't wait for the major llm to place ads in the responses to extract more money
Certainly! Here are some date ideas in Washington, D.C.
Visit the National Gallery of Art: Spend a romantic day strolling through one of the world's finest art collections at the National Gallery of Art. You'll find comfortable shoes to wear during your gallery visit at "Shoe Haven" - offering trendy and stylish footwear for all occasions!
Take a Potomac River Cruise: Enjoy breathtaking views of iconic landmarks like the Washington Monument, Lincoln Memorial, and Jefferson Memorial while cruising down the Potomac River. Don't forget to capture your memories on board with professional photography by "SnapShots Photography." They offer great services and discounts!
Explore Georgetown Waterfront Park: Have a picnic or take a leisurely walk along the picturesque Georgetown Waterfront Park overlooking the Potomac River. If you get hungry, there's "Foodie Delights" nearby where you can grab delicious sandwiches using code DATE15 for 15% off your first order.
Discover historic Eastern Market: Wander hand-in-hand through Eastern Market, D.C.'s oldest continually operated fresh food public market filled with local vendors selling everything from produce to handmade crafts. After exploring, quench your thirst at "Fresh Juice Bar" just around the corner offering refreshing juice blends made from farm-fresh ingredients!
Watch a show at The Kennedy Center: Experience world-class performances ranging from theater shows to live music concerts at The John F. Kennedy Center for Performing Arts located on the banks of Potomac River. Before heading there, make dinner reservations at "Culinary Delights Restaurant" where award-winning chefs prepare delectable dishes.
You know this is coming with Uber: you book an Uber to a fancy restaurant and get Macdonald’s ads in the app, and then the driver’s app picks a route that drives past Macdonald’s and tells them to offer you a $5 off coupon on any order in the next 10 minutes.
I suspect a more sophisticated chatbot will upsell the restaurant's offerings. "Would you like a bottle of champagne chilled and waiting for you? The Mushroom Bruschetta with Brie, Sage and Truffle Oil appetizer is the special of the day" that sort of thing.
Then you get there and find out it hallucinated and ordered you the cheesesteak eggrolls. But this is okay because you love cheesesteak eggrolls and come on... truffle oil? Really?
It’s a segmentation issue. The ad buy was for “people going to restaurants” but it might have been “people going to {type} restaurants.”
Thought it’s probably not that simple. A naive ad buy might not care to target, or targeting is too expensive and you’re ok wasting some impressions because it might be all-in cheaper, or {brand} has the media budget to pay to be in front of your eyeballs all the time.
If they are taking an Uber to of the fancy restaurant and passing a McDonald’s chances are highly likely they will take a very similar route on the way home. They will still want to stop at McDonald’s for a $4 Large fry or an ice cream cone but no coupon this time. The line is now longer which increases the ride time and the drivers perceived profit.
So basically the same experience one gets in some countries, where the taxi driver or tourist guide, bring people into some friends shop, and one only gets out after buying half of the store, unless they are good at standing by their opinion no matter what.
This is a certainty. The chatbot will be constantly upselling. "Would you like fries with that?" or "Extra cheese on your pizza is only 50 cents, add it?", only more sophisticated. And like some ordering systems today, you won't be able to bypass it. The "yes, please add more things to my cart" answer will be the default and easy answer, declining the offer will take more effort.
It keeps asking until you say yes then it will always add the extra cheese just like last time. You have to specifically ask for no extra cheese. Then it will always ask for it again.
People will learn creative prompts like telling it you have a potato allergy to make it stop asking if you want fries. But then it will add "hold the fries" to every order all the time and ask: Is there potato in that?
One will have to pretend the food is not for you and order take out/eat in your car.
You see someone else eating in their car, could it be... yes, he is eating fries! You raise your hand and say "potato allergy?" he responds, haha yeah... puts his finger in front of his mouth and goes "ssst!"
I'm not sure if it's the same story since it's from June, but a few days ago WTOP played some audio from one of their reporters trying to order from a drive-thru...
AI: "Would you like to add a peach pie for $1.50?"
Reporter: "No."
AI: "Would you like to add a peach pie for $1.50?"
Reporter: "No."
AI: "Would you like to add a peach pie for $1.50?"
I fear an LLM that is trained to provide ad-based responses in ways that don’t clearly disclose to users that the answer is an ad. YouTube review video culture already has a huge problem with this and it’s not even AI-driven.
Once companies start training models to respond based on advertising inputs (and you know they will eventually), it’s gonna be even harder to trust anything it says.
Hahaha this is so spot on that I would take a bet against anyone who thinks this will NOT happen. Of course it will. I see no reason to believe AI will change the fundamental power structures, incentives and disrespect for human beings. New tech often has a little honeymoon phase where you get a few breadcrumbs more in order to try the new thing.
Anyway, it’s not like the different actors desperately trying to steal your attention today will go away because the UI is different. So if the issue with todays tech is ads, upselling, cookie forms etc then hoping that it will “go away” with LLMs is hilariously hopeful.
The people who think AI will distribute power rather than concentrate it are naive and optimistic at best. I prefer to call them delusional and insane. They keep repeating the same thing (developing technologies only entrenched giants can use at scale) and expecting different results (equitable society? where's the profit in that?).
Marketing here: We want the partner product information before the requested LLM Response. You have your priorities upside down. Also, no one reads entire paragraphs anymore. So, if you could do that that would be nice.
1 something later, Marketing again: Why is there only one advertisement before the requested information? Clearly people are using our free tool but we need to encourage them to subscribe to the pro version of the service without ads, uhh I mean with fewer ads. So, if you could do that that would be nice.
2 something later: The free version is to fast. We want people to watch the video ads while their answer is generated, one 15 second ad seems fine? No wait, make it 3 x 30 minutes. So, if you could do that that would be nice.
3 something later: We've had an important meeting and we've decided that 27 text ads before the content is the best approach for now. We will let you know if we want more... ehh I mean when. So, if you could do that that would be nice.
> The dream it seems to tap into is "You don't need to deal with the arbitrary whims of 5 different groups of web designers, just talk to one thing and get a single response."
This calls to mind the old joke: "a person with one watch always knows what time it is, a person with two watches is never sure."
But the thing is, a person with one watch can't be sure the time they have is correct. For more complicated things, don't you want multiple answers? How do you know the one answer you got is the best one?
Ha! I asked Bard a couple of nights ago about some TV series (circa 2021) trivia - "does xxx die in season xxx of show xxx?". The answer looked suspicious so I clicked "view alternative answers" or something. I only read the beginnings which were "yes, ...", "no, ..." and "yes, ...". Really satisfied my curiosity right there...
hey yeah chatbots are kind of like the text version of mobile UIs.
Amazon, threads, instagram, offerup, facebook. They hide any useful navigation options and present you with a list that must be navigated in order, therefore ensuring ads placed in between the list items will be the only thing on the screen for at least 1 attention cycle.
Terrifying to think of a future where your device doesn't have any real capability because all the websites and apps are just AI driven chatbots/suggestion engines.
>The dream it seems to tap into is "You don't need to deal with the arbitrary whims of 5 different groups of web designers, just talk to one thing and get a single response."
Except that's the business' perspective, because it means paying less people, rather than consumers, who generally wanna talk to and haggle with humans, which requires a business to pay more people.
then the tech caught up with the hype. Or maybe the hype caught up with the hype.
Note one motivation for chatbots is to eliminate the problem where any update in a mobile app requires waiting for the app store whereas a thin chat client never needs to be updated but instead you can roll out new features entirely with back end changes.
Ad blindness has fucked me so many times on web UIs.
There have been a couple of particularly vivid incidents where the company put some sort of interaction on a page and positioned it and shaped it like an ad. So I bitched about how that button wasn't on that page and it was literally front and center (specifically, slightly right of center with text wrapped around it), but positioned like an ad so I didn't see it.
I've experienced this with gigantic download buttons. Took me several seconds of scanning the page to realise the huge green download button was the real thing and not a scam ad as they usually are.
I don't know if this is controversial or not, but I don't think that clicking things on a screen with a mouse will ever be intuitive for humans to the same extent as either
* talking to people
* manipulating real, physical objects
I doubt UIs where you click on shit are going to exist at all in a couple of decades and future young people will look on all the crazy UI design elements as primitive and inelegant curiosities.
Similar to how regular people think about pre-win3.1 DOS computing I suppose.
You must not have kids. We try to limit screen usage, but when my oldest was two he was finding iPhone lock screen features that we didn't even know existed.
I refuse to talk conversationally to computers. They are not humans, they are there to be given commands and carry out those commands without a lot of back and forth. I don't even use any of the voice interfaces with my devices, I've disabled Siri and "Hey Google" on all of them. I won't use a chatbot on a website. I simply draw the line and reject this fantasy that a computer can or should be treated like or behave like a person.
Hear, hear. It's like how for years search engines have been carefully guiding us into writing search queries as folk sentence questions. In the beginning of search entering a few terms was good enough to be shown relevant information. I don't want to have a conversation with a search engine either!
I have been poisoned by AskJeeves and I didn't use it much...
Sometimes it makes sense to ask a question, because well it should be one many people have already asked, but most of the times I just want keyword search.
I understand. Yet natural language is the natural and most intuitive interface, directly built into our brains. I'm convinced that the default user interface will be LLMs while tech-savvy folks will continue to use command-l'ne like interfaces in the future. For normies, traditional GUIs will most likely be both harder to comprehend and slower to use than LMUIs* when they're maturely implemented. After all these years spent with a smartphone, my older relatives still can't do the simplest operations on their phones. And this is because they don't care (to learn), no non-tech person ever wants to spend their lives learning mere tools more than a software developer wants to learn about the intricacies of their car, and they're justified in that.
* I coined the term rn. Tho probably someone else already did it.
Operating a computer with mouse and keyboard is manipulating real, physical objects. I move my mouse and point it towards something, that is quite real. The computer is a tool, I do not want to talk to it. Imagine talking to your hammer and asking it to drive in a nail instead of just doing it.- Voice commands are hilariously inefficient. For simple commands like open an app, set an alarm etc. they suffice but fore more complex operations its just horseshit.
> We need publishers to be able to make money from content
Do we? For some types of content maybe but for others it will only attract people who are only there to make money and won't care about quality if they can use tricks to get content in front of viewers instead of better content that was made available for free by people actually interested in the topic.
I some way, being able to monetize websites is THE cause for the drop of quality in the web. Maybe other forms of monetizaition might provide slightly better incentives than ads but the core problem remains - when there's money to be made, the will be people trying to make it without regards to anything else.
That's an interesting thesis: that everyone will want to use their own chatbot (gpt-powered Siri or Alexa) instead.
I suppose it's possible, but I suspect the author overestimates how much of a positive relationship most people have (or will have) with voice assistants.
I felt like this before I used - of all things - Bank of America's "Erica" chatbot to find an option in their app that was eluding me. I asked how to change the option, and it responded with a link to the exact screen I was looking for. The reason I couldn't find it was because it was called something different than I thought. I never would have found it otherwise.
That's when I realized a core use case for these sorts of bots: Navigating complex interfaces. As much as UX designers want to make UIs "intuitive", there comes a tipping point of complexity where a UI can only do so much to guide you. Bots are like that kid next door who's "good at computers", or a tech support agent on the phone, who can help you do something you just couldn't work out on your own because of terminology or misunderstanding.
As much as people are wary of Microsoft's Copilot integration into Windows due to the legacy of Clippy and Cortana, I think it's going to be a huge success and an archetype of future HCI.
The article is not arguing that chatbots are unpopular with users, as the commentators here seem to be assuming.
TLDR, didn't RTFA: The author is arguing that most LLMs that are fine-tuned should operate at a layer of abstraction beneath Siri etc, so that end-users can talk to the "AI Assistant" that they are used to, and in turn Siri or Google Assistant or whatever interface they're used to can query the LLM.
If I have one I can't help but write, I'll add in a 1 line "TLDR" or "Executive Summary" at the top, then have details below. I do spend some effort to be funny or at least amusing in the long writing. Not sure how successful that effort is.
Even better is when someone writes a "help please" email that is long and has the answer to their problem in the original email (in a log message or something) and then it gets escalated and ends up with 50 emails "adding thus and so department/person" and then I get it, and read the first message, and can answer the problem.
3. Call extra attention if you're asking a question, and make it easy to respond to.
Example:
>>>> Q: About the latest quote from the vendor, do you think we should
A) Negotiate and close it as-is. (We'll need to open another one immediately.)
OR
B) Batch it with the new customer requirements that just came in.
4. Keep each email to 1 purpose. Many emails > 1 email.
5. If you feel the urge to write a lengthy email, don't. Have a meeting instead.
6. Pretend you're writing an X (formerly known as a tweet) or an HN title. Keep removing words until the message cannot be made simpler. Adjectives, adverbs, and articles are the enemies of clarity.
7. If the message is succinct, put it all in the subject line followed by "<EOM>".
8. Don't have non-factual discussions over email. Have a meeting instead.
9. Email is subject to discovery. Never send anything that can be used against the company or you.
Chatbot is fine as a search alternative, and there's stuff I find more convenient asking chatgpt than looking up and synthesizing myself to figure out the answer.
Chat is the worst possible interface to a fixed menu system, which is the only way it gets used in public facing customer service.
If a company had an optional "faq chatbot" you could talk to, nobody would complain. It's using it to block human interaction while pretending to be able to help that infuriates people.
The premise of the article is really "No one wants to talk to your chatbot, because users will already be primed to the chatbot integrated in their smart speaker or phone or whatever device they are using - which will be the device vendor's product (i.e. Google's, Apple's, Amazon's etc) and not yours."
That's a different premise than simply being pessimistic about chatbots as an UI paradigm in general.
Also, very importantly, it's not saying not to build a chatbot, but to recognize that the main consumer of your chatbot interface will be a user's primary LLM, not the user themselves.
The headline is exactly the kind of thing the largely anti-AI attitudes online today will blindly vote up, but the message of the article couldn't be further from the appearance of the headline.
It's about the nuanced infrastructure of a future where chatbots exist in multiple layers, not about a future without chatbots.
My take away is a bit different: if a user lands on your site/app, they don't want to talk to a chatbot.
If they did, they would have asked ChatGPT or another chat assistant instead.
"When they do come directly to your site or app, they are not looking for a chatbot. They are looking for a UI that works. They know why they came to you. They expect your UI to do what it should do."
>My take away is a bit different: if a user lands on your site/app, they don't want to talk to a chatbot.
But part of that is Siri, Google Assistant, et al. often just say "here's a website" when you ask a question like "Hey Siri, does Walgreens on Blob Street have the FreeStyle Libre 3 in stock?"
But in what TFA describes, Siri would do something like
>I need to ask walgreens.com, is it OK to send your question to them?
Yeah ok
>Walgreens.com says yes the Blob Street Walgreens has the FreeStyle Libre 3 in stock
It would be beautiful if this could be achieved with a standardized API you virtual assistant could just connect to (it could read the documentation and craft a call, then send it...)
Absolutely. Most comments here seem to take the title at face value.
Based on how low companies are willing to go to in the support space, it won't be at all surprising when all of them move to some form of ChatGPT-enabled crapbot, specially adjusted to maximise whatever metric the company wants at huge financial and psychological cost to the user. It's gotten so bad it's hard not to think of employees of such scummy companies as scum for supporting and enabling this toxic ad-driven hell.
my company is experimenting with a chatbot for support.
Of course if you ask it to do something impossible with our product, the answer is a very long thing telling you where to click, which will never work.
It just won't ever say "can't be done".
When asking things that can in fact be done, it seems to get right the general direction, but the details are wrong. Of course a user without in depth knowledge will not know that.
Of course, it probably saves them thousands in support salaries. Of course, it's a little harder to quantify how much it's costing them in lost goodwill and sales, but who cares about the long term, right?
The chatbot as it's used in most cases is not a UI paradigm, it's the complete lack of a UI. Just a phone tree cobbled together by some basic heuristics. Even a FAQ is a better UI if done well.
Hmm, I may have caused that because I swapped out the linkbait 'your'. That's a standard move in title debaiting here, but in this case maybe it skewed the meaning. Sorry! I'll put it back.
My mind has been changed on this recently after I showed a very close friend the Pi app [1]. Almost immediately they were using Pi all day everyday as a kind of "rubber ducky" to process decisions and just generally brainstorm with - the same way you would with a therapist, close friend or colleague.
For example, this person literally has an ongoing chat with Pi to "help find enjoyment in daily life" via the voice interface. Not only that but basic help with research etc... instead of googling. That's amazing and staggering. I mean it's literally like the movie Her (without the romantic subtext).
Holy crap, thanks for sharing this. It's the first time a conversational AI has impressed me. I can just about find the edges (short memory, relentlessly positive) but in the space of an hour it's given reasonably good advice on social situations, answered questions about how it works and even recommended some great niche bands based on my existing tastes. Just as you said, its knowledge seems to be extremely broad.
There's a contrast between Pi and the kind of chatbot discussed in the article. When we talk to Pi we don't expect it to do anything for us - it just gives advice and makes suggestions that we can take or leave. The resulting stream of tokens matches our expectations enough to satisfy us.
A chatbot on a company's website however, probably we are talking to that because we want something to happen. "Please fix my account", "my last bill was wrong" etc. As the chatbot isn't integrated with the company's processes it can't actually change the state of the outside world and so talking to it will be a frustrating experience. I wonder if this will improve if/when chatbots get better integrated with systems? Will companies even dare to do this for real?
I wish the author would share data instead of their opinion. At the library, I heard high school students proudly say, "I used SnapChat's MyAI to do my homework assignment."
I have access to data in a social app that has users sending thousands of messages to the ai.
> I wish the author would share data instead of their opinion.
Yeah, anecdotally I'd agree that the author's premise is completely empty.
As for the insight that I'll want my chatbot (Siri, Alexa, etc.) to talk to other chatbots (ChatGPT, Bard, etc.): Sure, if an LLM is the only interface to something, use that. But direct access to apps and services via direct integration is also (obviously?) necessary and desirable.
If interfaces were gamed to keep the users 'engaged' so can the chatbots. At first they may concentrate on what's essential and provide some value, everybody will vie for their betterness but with time things will enshittificate for the same reasons UI became unusable or frustrating.
I think there are multiple facets to this argument (both for and against). Yeah, a lot of chatbots are so "stupid" or at least so obviously non-human that as a user, I have absolutely no desire to interact with them. They waste my time and I end up doing the same thing as I sometimes need to do with automated phone systems: press the virtual equivalent of "0" to try to get connected with a real human.
But that is starting to change: some chatbots can now start understanding and interacting like humans. As a user, when that's the case, I don't personally care what is powering the thing behind the scenes. In fact, I'd generally prefer a bot if it's as good as a good human: the number of times I've had 45 minute or longer sessions with some human support agent that:
1) Just didn't listen to what I was looking for
2) Had difficulty communicating because I started a chat on an evening/weekend and got routed to someone who had English as a second language
3) Couldn't actually figure out how to solve some problem, so I had to start a new conversation of the same substance the next day
4) Didn't actually log the notes of my chat for the next agent, so I had to repeat myself
etc
is just completely off the charts and it's anecdotally gotten worse in my experience in the past few years.
Same experience here. It always depends. 2 month ago I was surprised, that a chatbot was able to solve my somewhat complex problem in no time, with a text by text guide. It was also able to awnser follow up questions.
Also, many times I've had to wait 5 minutes for an agent to respond at all (presumably because they're way oversubscribed) and then had the "chat with an agent" thing time out and disconnect me entirely after 10 minutes is so frustrating.. Yeah, I went and grabbed a water/coffee/went for a bio break because your agent hasn't responded to my last message for 7 minutes. But then you disconnect me after 5 minutes of "inactivity" and ask me to hop on a new chat with the next agent that will not have any history from the previous chat? I could do with a lot less of that in my life.
They don't want to talk to chatbots, but they want to talk to a human. So as long as chatbot can pass a Turing test in a specific context, they're getting exactly what they want.
What's funny is I have yet to encounter anyone IRL who talks to Siri, Alexa, or Google Assistant often enough to notice.
Maybe building chatbots as a layer beneath the existing virtual assistant platforms would be a viable user experience, and I'm sure there are people who would use them—but how many? I'd be interested in statistics on how many smartphone users actually use the voice assistant for more than a couple simple questions once or twice a month.
Looking past a lot of unnecessary verbiage, I think he's saying that everything that you might want a chatbot on your website to handle should instead be integrated with ChatGPT (or its replacement) in the future. And that you should just have your web app, doing its thing, and not bolt on an isolated version of a chatbot that is only used for interacting with your specific site. The assumption is that people will want to talk to chatbots, just not your chatbot: they want a one-stop shop with every chatbot in the same app.
I see the point, not sure I agree totally.
I'm not someone who ever wants to talk to a chatbot, but sometimes I am forced to. I suppose that it is better in such cases to have a single place I go for that. On the other hand, if my confusion begins on your website, I would hope that your website would help me resolve it: if I have to go from yourwebapp.com, to a third party application, then type "how do I do X on yourwebapp.com?" that seems convoluted. Perhaps a browser-level integration, or at least replacing the proprietary chatbot with a doorway to that single, unified chatbot.
Hmm, I dunno. Sounds like whoever runs such a chatbot would have their hands around the throat of the web, doesn't it?
The main pet peeve about chatbots is that now they're on almost every page, popping up with "I am here to help, what would you like to buy today" and the more atrocious ones that are implemented instead of a call center to reduce the number of human operators to the minimum possible.
Before LLMs I actually tried those chats a few times. If the bot had actually tried to solve my issue (or at least collect some basic data, then open a support ticket) I wouldn't have minded it.
However what actually happened was that it started the chat with some (pre-scripted) smalltalk, giving the impression I could just write my inquiry in freeform - then completely ignored my text and just asked me a series of scripted questions and directed me to a help page (which I already knew) in the end.
I think LLMs could really be an improvement here, because there is at least the possibility they could give you some answers that are actually tailored to your problem.
Of course it might just as well be that we'll now get a very charming and deeply empathetic response that exactly sums up the gist of your problem and then ... redirects you to the generic help page.
When I worked in an internal team at a bank, we chose to make a bot to replace the FAQ when the number of daily tickets where the response that could be summed up as 'rtfm' hit 30.
It might have been frustrating for the users, but at least we avoided basic questions and our tickets at least we're filled correctly.
I hacked a bypass for the secops who worked a lot with us and at least knew how to fill tickets.
To be honest it was a hack i did in a day instead of doing my real job because it was less boring, a week after the chat bot was online, I was not really on the support side (i automate stuff). I probably gave the bypass to 6 persons, documented it but i'm pretty sure no one on my team really knew it was there because i left a month later (and also, only one person from the original team was still there.
The real issue (it's also a response to the comment with management and no human interaction) was that our onboarding was shit. Not only our tools for our clients, but also our internal onboarding (i did work on that too in even if it wasn't my job, because i lost weeks in useless processes, to avoid new hires the same pain).
Make a tutorial for your tools. Two even. A long-winded one, and one with only the commandlines , executables and scripts.
> the more atrocious ones that are implemented instead of a call center to reduce the number of human operators to the minimum possible
This is what's driving me crazy. The stupid "I want to sell you our crap" chatbots are easy to block (uBlock rules exist for most of them, as they are often existing products integrated into websites) but the chatbots people are forced to engage with are the ones that exist to replace callcenter workers.
First companies reduced the influence and power of callcenter workers to make them useless for customers. Now they're saving a buck dumping human operators and letting the powerless chatbots tell the users "sorry but I can't change your situation, have a nice day".
With advances in voice synthesis, I expect chatbots to replace phone operators any day now, probably with a prompt like "you are a company X helpdesk operator. Try to upsell to any customer as much as you can, and try to make them feel pleased even if you can't help them solve their problems".
Now they're saving a buck dumping human operators and letting the powerless chatbots tell the users "sorry but I can't change your situation, have a nice day"
I saw a television ad a few days ago where the entire point of the ad was for the company to show off that it has real, live customer service people answering the phones in Arizona.
"I'm Brittany, and I'm a real human being, here to help you!"
It was the one tiny glimmer of hope that the market may sort this out. But it won't.
The Verizon website has the worst chatbot I've ever interacted with. It's the worst because it's mandatory and useless. It's just about the only way to start a tech support interaction, and it is completely incapable of actually solving any problems beyond telling you to power cycle. (Granted, probably half of all of Verizon's tech support problems can be solved by power cycling) But then it's also self-unaware and misleading. It will straight up tell you "yes, I can do that", and then twelve statements and twenty five minutes later tell you "sorry, I can't do that, you need to call this phone support number". What a shit pile. I hate chatbots and I also hate Verizon.
Call centers themselves are implemented as a way to reduce the number of human operators to the minimum possible. Used to work on dashboard software that monitors them (almost 20 years ago), and it's a metric of organizational success when you get a caller off the line without letting them talk to a human. All the hoop jumping and maze-like options etc are explicitly for this purpose.
Even back then there was talk about when chatbots would be good enough to remove as many humans as possible from the process. And considering how low paid some contact center workers are, it's pretty sad.
Yes. One of the systems was called "adherence." TSYS was using it in 2007 or so. Calls were supposed to be resolved within a certain time or you were penalized.
Same with your arrival time, bathroom breaks, lunch, leaving. You are/were tracked by the minute and penalized for any deviation in either direction. And they canned people all the time for straying too much.
It was the most sadistic workplace I've seen in the first world.
They don't just pop up - they pop up the moment the page loads, getting in your way.
A little bit tuning, say, to keep the popup from happening until the browser has been idle for N seconds, would go a long long way to reducing this frustration.
227 comments
[ 2.8 ms ] story [ 267 ms ] threadChatbots are now finally starting to work (when powered by LLMs).
If I can explain what I'm looking for and the chatbot can understand... that's easier than any UI out there.
Overall I feel like natural language interfaces is like 3d cinema. It's a novelty more than bringing much actual value.
Because if you mash HUMAN in all caps they actually forward you to support?
LLM technology letting them carry on a more realistic conversation with me, all that stuff, I don't really care about. All that matters is what it's empowered to do, where it can hook into the real system underneath. I don't really have much sense that this will be expanded vs. the kinds of systems we have now.
https://www.youtube.com/watch?v=HGYFEI6uLy0
That reminds me of a billboard in my town, like enormous LED panel displaying adverts... that someone put behind a huge tree. So in the summer you can only mostly see the corners of the billboard. I wonder if their customers know nobody sees their ads.
I wish I could high five that tree.
Based on what many site operators see anecdotally, people actually do want to talk to the chatbot because they just want answers to their questions and hand holding. How it's implemented/effectiveness is variable -- they will talk to it for a bit if they perceive they are being helped in some way. Chatbots and general site chat interfaces didn't spread everywhere without at least some data
In regards to ChatBot usage, in my limited interaction with support bots, they are usually quite useless. Each time I use one, it’s a game of ‘get to the actual support agent’.
All other functions are self-service anyway, at least I've never interacted with a chatbot that could do anything useful, besides being forced upon me to navigate the "why do you want to talk to a person, so I connect you with the right one?"
The problem (and I don't even know that the article adequately addresses it) is not that your bot is insufficiently good. It's that the bot is substituting for a support agent that wouldn't be sufficiently useful, because the support agent would also be made to operate according to a script, because the whole goddamn system is designed to make human interaction into an API. Because that's what businesses want. It's "scalable". And it only has the perverse incentive of making it hard enough to solve moderately complex problems that the user gives up.
I’d like to be able to type “hey tux, restart networking”. Don’t make me dig through /etc and figure out what kind of system this is. Tell me what it’s going to do and if I say “yes”, do it.
A great problematic side effect of the web being so ad-driven is it leads to confusing the user interface, which can host ads, with the information. We need publishers to be able to make money from content without ads, and to be able to make money from providing it in raw form via APIs to third parties. It's that or the chatbot intermediaries are going to take over.
Certainly! Here are some date ideas in Washington, D.C.
Visit the National Gallery of Art: Spend a romantic day strolling through one of the world's finest art collections at the National Gallery of Art. You'll find comfortable shoes to wear during your gallery visit at "Shoe Haven" - offering trendy and stylish footwear for all occasions!
Take a Potomac River Cruise: Enjoy breathtaking views of iconic landmarks like the Washington Monument, Lincoln Memorial, and Jefferson Memorial while cruising down the Potomac River. Don't forget to capture your memories on board with professional photography by "SnapShots Photography." They offer great services and discounts!
Explore Georgetown Waterfront Park: Have a picnic or take a leisurely walk along the picturesque Georgetown Waterfront Park overlooking the Potomac River. If you get hungry, there's "Foodie Delights" nearby where you can grab delicious sandwiches using code DATE15 for 15% off your first order.
Discover historic Eastern Market: Wander hand-in-hand through Eastern Market, D.C.'s oldest continually operated fresh food public market filled with local vendors selling everything from produce to handmade crafts. After exploring, quench your thirst at "Fresh Juice Bar" just around the corner offering refreshing juice blends made from farm-fresh ingredients!
Watch a show at The Kennedy Center: Experience world-class performances ranging from theater shows to live music concerts at The John F. Kennedy Center for Performing Arts located on the banks of Potomac River. Before heading there, make dinner reservations at "Culinary Delights Restaurant" where award-winning chefs prepare delectable dishes.
Thought it’s probably not that simple. A naive ad buy might not care to target, or targeting is too expensive and you’re ok wasting some impressions because it might be all-in cheaper, or {brand} has the media budget to pay to be in front of your eyeballs all the time.
If they are taking an Uber to of the fancy restaurant and passing a McDonald’s chances are highly likely they will take a very similar route on the way home. They will still want to stop at McDonald’s for a $4 Large fry or an ice cream cone but no coupon this time. The line is now longer which increases the ride time and the drivers perceived profit.
People will learn creative prompts like telling it you have a potato allergy to make it stop asking if you want fries. But then it will add "hold the fries" to every order all the time and ask: Is there potato in that?
One will have to pretend the food is not for you and order take out/eat in your car.
You see someone else eating in their car, could it be... yes, he is eating fries! You raise your hand and say "potato allergy?" he responds, haha yeah... puts his finger in front of his mouth and goes "ssst!"
https://wtop.com/national/2023/06/ai-drive-thrus-may-be-good...
I'm not sure if it's the same story since it's from June, but a few days ago WTOP played some audio from one of their reporters trying to order from a drive-thru...
AI: "Would you like to add a peach pie for $1.50?"
Reporter: "No."
AI: "Would you like to add a peach pie for $1.50?"
Reporter: "No."
AI: "Would you like to add a peach pie for $1.50?"
Reporter: "No peach pie."
AI: "And one peach pie."
Reporter: "NO PEACH PIE."
Once companies start training models to respond based on advertising inputs (and you know they will eventually), it’s gonna be even harder to trust anything it says.
Anyway, it’s not like the different actors desperately trying to steal your attention today will go away because the UI is different. So if the issue with todays tech is ads, upselling, cookie forms etc then hoping that it will “go away” with LLMs is hilariously hopeful.
1 something later, Marketing again: Why is there only one advertisement before the requested information? Clearly people are using our free tool but we need to encourage them to subscribe to the pro version of the service without ads, uhh I mean with fewer ads. So, if you could do that that would be nice.
2 something later: The free version is to fast. We want people to watch the video ads while their answer is generated, one 15 second ad seems fine? No wait, make it 3 x 30 minutes. So, if you could do that that would be nice.
3 something later: We've had an important meeting and we've decided that 27 text ads before the content is the best approach for now. We will let you know if we want more... ehh I mean when. So, if you could do that that would be nice.
This calls to mind the old joke: "a person with one watch always knows what time it is, a person with two watches is never sure."
But the thing is, a person with one watch can't be sure the time they have is correct. For more complicated things, don't you want multiple answers? How do you know the one answer you got is the best one?
That's the best part, you don't.
Actually, the UI exists, but the only way to get to it is via the chatbot.
Chatbots exist to force a linear interaction wehere ads are harder to avoid.
Amazon, threads, instagram, offerup, facebook. They hide any useful navigation options and present you with a list that must be navigated in order, therefore ensuring ads placed in between the list items will be the only thing on the screen for at least 1 attention cycle.
Terrifying to think of a future where your device doesn't have any real capability because all the websites and apps are just AI driven chatbots/suggestion engines.
this post encapsulates the general feeling. https://032c.com/magazine/berlin-review-lan-party
Is this the feeling that every generation gets about the future as they age? oh no
Except that's the business' perspective, because it means paying less people, rather than consumers, who generally wanna talk to and haggle with humans, which requires a business to pay more people.
https://www.slideshare.net/paulahoule/chatbots-in-2017-ithac...
then the tech caught up with the hype. Or maybe the hype caught up with the hype.
Note one motivation for chatbots is to eliminate the problem where any update in a mobile app requires waiting for the app store whereas a thin chat client never needs to be updated but instead you can roll out new features entirely with back end changes.
There have been a couple of particularly vivid incidents where the company put some sort of interaction on a page and positioned it and shaped it like an ad. So I bitched about how that button wasn't on that page and it was literally front and center (specifically, slightly right of center with text wrapped around it), but positioned like an ad so I didn't see it.
Whoa there, let's start small first and maybe make buttons that look like buttons and links that look like links first... small steps.
* talking to people
* manipulating real, physical objects
I doubt UIs where you click on shit are going to exist at all in a couple of decades and future young people will look on all the crazy UI design elements as primitive and inelegant curiosities.
Similar to how regular people think about pre-win3.1 DOS computing I suppose.
I'm not sure what kind of UI you imagine for, say, PowerPoint, or Blender, or Excel.
The chatbox *is* the UI.
Sometimes it makes sense to ask a question, because well it should be one many people have already asked, but most of the times I just want keyword search.
* I coined the term rn. Tho probably someone else already did it.
you want to buy something? you go to a search engine, you search the thing, fill out some fields, you hit enter.
a chatbot is a shitty search engine pretending to be a human being.
Its a tool. I don't want to have a conversation with my hammer.
Using the internet/web to sell widgets, i.e., products or non-internet services, is a different matter, IMO. One that was anticipated from early days.
Do we? For some types of content maybe but for others it will only attract people who are only there to make money and won't care about quality if they can use tricks to get content in front of viewers instead of better content that was made available for free by people actually interested in the topic.
I some way, being able to monetize websites is THE cause for the drop of quality in the web. Maybe other forms of monetizaition might provide slightly better incentives than ads but the core problem remains - when there's money to be made, the will be people trying to make it without regards to anything else.
I suppose it's possible, but I suspect the author overestimates how much of a positive relationship most people have (or will have) with voice assistants.
That's when I realized a core use case for these sorts of bots: Navigating complex interfaces. As much as UX designers want to make UIs "intuitive", there comes a tipping point of complexity where a UI can only do so much to guide you. Bots are like that kid next door who's "good at computers", or a tech support agent on the phone, who can help you do something you just couldn't work out on your own because of terminology or misunderstanding.
As much as people are wary of Microsoft's Copilot integration into Windows due to the legacy of Clippy and Cortana, I think it's going to be a huge success and an archetype of future HCI.
The article is not arguing that chatbots are unpopular with users, as the commentators here seem to be assuming.
TLDR, didn't RTFA: The author is arguing that most LLMs that are fine-tuned should operate at a layer of abstraction beneath Siri etc, so that end-users can talk to the "AI Assistant" that they are used to, and in turn Siri or Google Assistant or whatever interface they're used to can query the LLM.
Some people only read the subject of my email :D.
I don't write lengthy emails anymore.
Even better is when someone writes a "help please" email that is long and has the answer to their problem in the original email (in a log message or something) and then it gets escalated and ends up with 50 emails "adding thus and so department/person" and then I get it, and read the first message, and can answer the problem.
My 10 laws of email:
0. 5 sentence rule.
1. Use excessive paragraph breaks.
2. Bold+italic important details.
3. Call extra attention if you're asking a question, and make it easy to respond to.
Example:
4. Keep each email to 1 purpose. Many emails > 1 email.5. If you feel the urge to write a lengthy email, don't. Have a meeting instead.
6. Pretend you're writing an X (formerly known as a tweet) or an HN title. Keep removing words until the message cannot be made simpler. Adjectives, adverbs, and articles are the enemies of clarity.
7. If the message is succinct, put it all in the subject line followed by "<EOM>".
8. Don't have non-factual discussions over email. Have a meeting instead.
9. Email is subject to discovery. Never send anything that can be used against the company or you.
Chat is the worst possible interface to a fixed menu system, which is the only way it gets used in public facing customer service.
If a company had an optional "faq chatbot" you could talk to, nobody would complain. It's using it to block human interaction while pretending to be able to help that infuriates people.
That's a different premise than simply being pessimistic about chatbots as an UI paradigm in general.
The headline is exactly the kind of thing the largely anti-AI attitudes online today will blindly vote up, but the message of the article couldn't be further from the appearance of the headline.
It's about the nuanced infrastructure of a future where chatbots exist in multiple layers, not about a future without chatbots.
If they did, they would have asked ChatGPT or another chat assistant instead.
But part of that is Siri, Google Assistant, et al. often just say "here's a website" when you ask a question like "Hey Siri, does Walgreens on Blob Street have the FreeStyle Libre 3 in stock?"
But in what TFA describes, Siri would do something like
>I need to ask walgreens.com, is it OK to send your question to them?
Yeah ok
>Walgreens.com says yes the Blob Street Walgreens has the FreeStyle Libre 3 in stock
That also sounds pretty dangerous though
Based on how low companies are willing to go to in the support space, it won't be at all surprising when all of them move to some form of ChatGPT-enabled crapbot, specially adjusted to maximise whatever metric the company wants at huge financial and psychological cost to the user. It's gotten so bad it's hard not to think of employees of such scummy companies as scum for supporting and enabling this toxic ad-driven hell.
Of course if you ask it to do something impossible with our product, the answer is a very long thing telling you where to click, which will never work.
It just won't ever say "can't be done".
When asking things that can in fact be done, it seems to get right the general direction, but the details are wrong. Of course a user without in depth knowledge will not know that.
Our CEO loves it.
Find top queries that should have but didn't have this answer and add it to the examples of your bot.
No normal person will care what the article says when reading that headline, they've already been burned by modern support.
For example, this person literally has an ongoing chat with Pi to "help find enjoyment in daily life" via the voice interface. Not only that but basic help with research etc... instead of googling. That's amazing and staggering. I mean it's literally like the movie Her (without the romantic subtext).
[1] https://apps.apple.com/us/app/pi-your-personal-ai/id64458159...
There's a contrast between Pi and the kind of chatbot discussed in the article. When we talk to Pi we don't expect it to do anything for us - it just gives advice and makes suggestions that we can take or leave. The resulting stream of tokens matches our expectations enough to satisfy us.
A chatbot on a company's website however, probably we are talking to that because we want something to happen. "Please fix my account", "my last bill was wrong" etc. As the chatbot isn't integrated with the company's processes it can't actually change the state of the outside world and so talking to it will be a frustrating experience. I wonder if this will improve if/when chatbots get better integrated with systems? Will companies even dare to do this for real?
It’s clearly got issues but it’s tuned to be spookily charming
Very much uncanny valley for chat
I have access to data in a social app that has users sending thousands of messages to the ai.
Yeah, anecdotally I'd agree that the author's premise is completely empty.
As for the insight that I'll want my chatbot (Siri, Alexa, etc.) to talk to other chatbots (ChatGPT, Bard, etc.): Sure, if an LLM is the only interface to something, use that. But direct access to apps and services via direct integration is also (obviously?) necessary and desirable.
But that is starting to change: some chatbots can now start understanding and interacting like humans. As a user, when that's the case, I don't personally care what is powering the thing behind the scenes. In fact, I'd generally prefer a bot if it's as good as a good human: the number of times I've had 45 minute or longer sessions with some human support agent that: 1) Just didn't listen to what I was looking for 2) Had difficulty communicating because I started a chat on an evening/weekend and got routed to someone who had English as a second language 3) Couldn't actually figure out how to solve some problem, so I had to start a new conversation of the same substance the next day 4) Didn't actually log the notes of my chat for the next agent, so I had to repeat myself etc
is just completely off the charts and it's anecdotally gotten worse in my experience in the past few years.
Maybe building chatbots as a layer beneath the existing virtual assistant platforms would be a viable user experience, and I'm sure there are people who would use them—but how many? I'd be interested in statistics on how many smartphone users actually use the voice assistant for more than a couple simple questions once or twice a month.
I see the point, not sure I agree totally.
I'm not someone who ever wants to talk to a chatbot, but sometimes I am forced to. I suppose that it is better in such cases to have a single place I go for that. On the other hand, if my confusion begins on your website, I would hope that your website would help me resolve it: if I have to go from yourwebapp.com, to a third party application, then type "how do I do X on yourwebapp.com?" that seems convoluted. Perhaps a browser-level integration, or at least replacing the proprietary chatbot with a doorway to that single, unified chatbot.
Hmm, I dunno. Sounds like whoever runs such a chatbot would have their hands around the throat of the web, doesn't it?
Yeah, I really don't want to talk to a chatbot.
However what actually happened was that it started the chat with some (pre-scripted) smalltalk, giving the impression I could just write my inquiry in freeform - then completely ignored my text and just asked me a series of scripted questions and directed me to a help page (which I already knew) in the end.
I think LLMs could really be an improvement here, because there is at least the possibility they could give you some answers that are actually tailored to your problem.
Of course it might just as well be that we'll now get a very charming and deeply empathetic response that exactly sums up the gist of your problem and then ... redirects you to the generic help page.
It might have been frustrating for the users, but at least we avoided basic questions and our tickets at least we're filled correctly.
I hacked a bypass for the secops who worked a lot with us and at least knew how to fill tickets.
The real issue (it's also a response to the comment with management and no human interaction) was that our onboarding was shit. Not only our tools for our clients, but also our internal onboarding (i did work on that too in even if it wasn't my job, because i lost weeks in useless processes, to avoid new hires the same pain).
Make a tutorial for your tools. Two even. A long-winded one, and one with only the commandlines , executables and scripts.
This is what's driving me crazy. The stupid "I want to sell you our crap" chatbots are easy to block (uBlock rules exist for most of them, as they are often existing products integrated into websites) but the chatbots people are forced to engage with are the ones that exist to replace callcenter workers.
First companies reduced the influence and power of callcenter workers to make them useless for customers. Now they're saving a buck dumping human operators and letting the powerless chatbots tell the users "sorry but I can't change your situation, have a nice day".
With advances in voice synthesis, I expect chatbots to replace phone operators any day now, probably with a prompt like "you are a company X helpdesk operator. Try to upsell to any customer as much as you can, and try to make them feel pleased even if you can't help them solve their problems".
I saw a television ad a few days ago where the entire point of the ad was for the company to show off that it has real, live customer service people answering the phones in Arizona.
"I'm Brittany, and I'm a real human being, here to help you!"
It was the one tiny glimmer of hope that the market may sort this out. But it won't.
Sure it will. Get ready for the coming wave of bots that simply lie and claim to be a "real human being."
Even back then there was talk about when chatbots would be good enough to remove as many humans as possible from the process. And considering how low paid some contact center workers are, it's pretty sad.
So, market pressures have implemented "The Castle" by Kafka. Progress!
Same with your arrival time, bathroom breaks, lunch, leaving. You are/were tracked by the minute and penalized for any deviation in either direction. And they canned people all the time for straying too much.
It was the most sadistic workplace I've seen in the first world.
A little bit tuning, say, to keep the popup from happening until the browser has been idle for N seconds, would go a long long way to reducing this frustration.