Small hypothesis: I bet people like AI chat interfaces in some part because they are “clean” – simple text, easy to process, consistent visuals, no ads, no pop-ups, etc.
To use a cliche example: Even if it wasn’t in any way “smarter,” it’d still be nicer to ask ChatGPT for a recipe than go to a webpage to read that recipe. Its interface is a natural “reader mode.”
But… that’s not going to last.
By the way, wanted to applaud everyone for having an actual conversation, rather than overreacting!
But maybe I’m applauding myself to some extent for blocking everyone who couldn’t handle it last time I posted something about AI. 😀
Do you mean the search results? Yeah, those used to be so good. So good. Now if I really REALLY need something, Gemeni mysteriously has all the results that the normal google search does not. Wonder how long that will last.
EDIT: Oh, right. Yeah, the web pages were also a chore to read. I completely agree.
Check this out https://lite.cnn.com/
@miblo I’m not sure I fully see the connection, unless you mean specifically the seductive confidence and the ease with which you can get to an answer, even if it’s wrong or oversimplified. This ↓ resonated with me the other day.
(But I was more thinking about the UI mechanics itself.)
@mwichary Gosh, yeah, that is an apt post here!
So am I right that the UI mechanics are: user types natural language into prompt box, presses Enter, interface displays (what appears to be) a full answer also in natural language?
I think that counts as "intuitive" by the standards of those posts that crop up from time to time in favour of making things intuitive. They'd sit more on the "super simple" side of the quoted post, right, requiring no, let's say, cognitive development of the user?
@mwichary Years ago and before AI; there was a lot of talk about conversational ui and how one app can be used for everything with just this one interface.
A command line interface is easy to understand, it doesn't have to be unix cli or AI.
100%: I keep asking some free AI (duckduckgo) for things about a programming language I don't know well. Not for full code, but for the pointers to the functions and functionalities that I expect exists. Easy to check whether its hallucination or not, just try a line of code and/or cross check in the manual.
Interestingly: as I learned more, the answers degraded rather quickly to barely usable.😜
@mwichary that is a great point actually.
I find it annoying how verbose and intricate it is and that is why I don't use it though.
A good way to get the clutter out of the way on the web is to install ublock.
@delthia @mwichary I too found that very annoying. I didn't like its jolly, cheerful tone. I am a cranky, sullen New Yorker. I am also an introvert.
I found that I was able to write a default prompt for Claude that made it much less annoying. For example, I have told it:
“I am a computer technology expert with decades of experience in many areas, so you don't have to explain every little thing. Answers should be concise and to the point. If I want elaboration, I will ask for it. If a question is a brief technical question, try to keep the response to one or two sentences. If a question doesn't specifically request a detailed example, do not supply one.”
“Don't describe my ideas as clever, brilliant, or inspired. It embarrasses me. Instead raise specific points about their technical merits and flaws.”
“Default Claude style is to end every exchange with a question inviting the user to continue the conversation. This is not necessary. I'm confident in my ability to lead us where we need to go.”
Somewhat to my surprise, this has actually worked well.
I have still not internalized how to deal with the LLM. I keep looking for the configuration knob that makes it less sycophantic. There isn't one. Instead, the way to make it less sycophantic is to tell it to be less sycophantic. The way to make it more concise is to tell it to be more concise.
@mwichary @delthia For programming work, I have put it in a file called CLAUDE.md, which Claude is supposed to read automatically every time it starts up.
On the website, I have defined a “default project” with default instructions that are read automatically at the start of each conversation. It's less programming-specific, but the style notes serve the same purpose.
@delthia @mwichary I wonder if there is a statistical connection between people who reject AI and people who keep the regular web somewhat usable by blocking the heck out of popups, ads, cookie-banners and so on.
And on the flip-side, between people who embrace AI and those who usually have to click through 5 layers of ads and other BS to even get through to the content of any website they are trying to use.
@StreetDogg @mwichary yeah, it would be very interesting indeed.
Another thing that I think would be related is the device that they are using, at least for me, and when looking for specific (for example technical) information, it is much quicker to do a quick web search and then ctrl+f my way through the page.
@Julianoe @mwichary @joachim This is a great insight. I interact with a few people who I would describe as having a fair amount of phobia for new technology. Any new app or software update is a source of great deal of stress and resistance. The process of exploring, discovering and learning a new UI is not pleasurable or fun for them.
I was surprised to see when they adopted AI chat apps for regular use before I did.
@mwichary I think it very well might, because of the difference in their nature.
Recipe webpages are strictly transactional. Their main purpose is to squeeze as much juice out of you *on this visit* as they can, and they compete in an extremely brutal, low-margin economy against other pages which have the same content and are attempting the same thing. There's 0 brand loyalty, there are no switching costs, if the page falls down in SEO rankings, they'll have 0 revenue after a week. This means that if they already have a customer "hooked", they need to shove as many ads in their face as they can. Next week, that customer won't remember whether the awful domain was online-recipes[.]com or recipesonline[.]us anyway.
Chatbots are different, as most people only use one. That means there's value in keeping the customer hooked, and giving them access to useful info when nobody else will is a great way to do this.
In an AI chatbot, the goals of squeezing as much money out of that customer and keeping that customer for themselves are at odds with eachother, and both of these goals are important.
I don't think they'll stay as pleasant to use as they are, but I also don't think they'll ever get anywhere near as bad as recipe websites.
@mwichary For a search engine, a text input box is a default option, which maybe is why people most often associate AI with search and use it accordingly.
However, for almost everything else, it’s a terrible interface because it doesn’t present any affordances that help users discover its functionality. Maybe that’s why people have difficulty using it in other ways than a search tool.
Scariest would be maybe if the inevitable advertising and manipulation is done quietly without anyone realising they are reading advertising.
There are so many people and groups willing to pay money to manipulate people's view of the world, and LLMs make it extremely easy to manipulate them.
@mwichary A so-called chat interface is the holy grail that users have always asked for. i.e. “Why can’t I just tell the dang computer what I want and it gives it to me?”
But users are discovering what designers knew a long time ago:
1. Users don’t know what they want as much as they think they do
2. Users don’t want to deal with details from a lower level of abstraction
@mwichary
A direction being explored—badly—by Facebook.
When reading an article, email, message, or thread in any source, I’d like to be able to open a side channel conversation about *that* source—and its context—using my chatbot of choice. Essentially, ask a follow up question.
I currently use Claude based on: its performance; abllity to use instructions to tune its reply style; and its background refinement of contextual Memory that’s easily to read, correct, or refine manually.
Marcin Wichary
Rosetta Type/Research
Madagascar_Sky
Gregatron5
Glenn Fleishman
Matt Mascarenhas
Jack Wellborn
Tamir Bahar
Chris Armstrong
Michael Clayton
Prem Kumar Aparanji 👶🤖🐘
Em & future cats 🇺🇦🐈🏳️🌈
Abdulla Almuhairi
Harald
Iago
Mark Dominus
StreetDogg
grybelfix
Jeremy Kanter
Nolwenn Maudet
Julianoë
jpl99
miki
Kit Paul
Jeffrey P. Bigham 🔥🔥
Arlo Godfrey
moopet
Wenzel Massag
Artem R 🇺🇦
R.C.
FediThing 
Moos-a-dee 🫎
<smitty />
Severák
Daniel Lyons
Greg Lloyd