The case against conversational interfaces « julian.digital

Comments

• aesthelete@lemmy.world ⁨3⁩ ⁨weeks⁩ ago

  The obsession with conversational interfaces likely stems from two places: sci-fi and CEOs who are used to ordering people around.

  source

  • paraphrand@lemmy.world ⁨3⁩ ⁨weeks⁩ ago

    “Do this, do that, and read between the lines!”

    source
  • taladar@sh.itjust.works ⁨3⁩ ⁨weeks⁩ ago

    Maybe the tendency for LLMs to shower the user with praise for their prompts also makes them attractive to egocentric CEO type of personalities?

    source

    • aesthelete@lemmy.world ⁨3⁩ ⁨weeks⁩ ago

      When I used those things, I absolutely understood why a CEO would want those to be the future. It’s everything they’re looking for: a strident, confident yes man machine who will produce without consternation any kind of spin (unethical or not) to any kind of content requested.

      source
• taladar@sh.itjust.works ⁨3⁩ ⁨weeks⁩ ago

  Every couple of years a shiny new AI development emerges and people in tech go “This is it! The next computing paradigm is here! We’ll only use natural language going forward!”. But then nothing actually changes and we continue using computers the way we always have, until the debate resurfaces a few years later.

  Reminds me a bit of graphical programming. Every couple of years someone comes up with the idea of replacing textual programming with some kind of graphical interface with arrows between nested boxed of various shapes and it inevitably fails.

  source

  • jorm1s@sopuli.xyz ⁨3⁩ ⁨weeks⁩ ago

    Except there’s Simulink, which has been around since the 80’s, and is anything but a failure. For a few specific usecases, like modeling complex physical systems and developing control algorithms for them, it’s far better than any traditional text based language. Especially when it comes to maintainability of that code.

    Though I have to admit that if you try to use it as a general programming language, you’ll learn that while that’s possible, it’s also very painfull. And even while implementing said control algorithms you’ll occasionally run on to some bits of logic that prove to be annoyingly difficult to implement with it compared to any text based language.

    source

    • taladar@sh.itjust.works ⁨3⁩ ⁨weeks⁩ ago

      I think the problem is that you can’t create new abstractions very well in graphical languages. It works for something like fixed domains (e.g. Blender node editor or your example) but for a general purpose language you need the ability to define abstractions that never existed before.

      The other problem is that you can’t really apply any of the tooling to it that works with other languages, e.g. version control, formatters, linters,…

      source

      • -> View More Comments
• sugar_in_your_tea@sh.itjust.works ⁨3⁩ ⁨weeks⁩ ago

  It’s a great article, actually click through and read it if you haven’t already.

  My favorite example of truly effortless communication is a memory I have of my grandparents. At the breakfast table, my grandmother never had to ask for the butter – my grandfather always seemed to pass it to her automatically, because after 50+ years of marriage he just sensed that she was about to ask for it. It was like they were communicating telepathically.

  That is the type of relationship I want to have with my computer!

  The author’s point is that natural language is a slow way to communicate, and it’s not even our preferred way, so why are we pushing so hard for it?

  One of the best productivity tools for me is my CLI shell, which predicts what I’m about to type based on what I’ve done in the past. There’s no AI here, just simple history search. It turns out i do the same thing a lot.

  None of this is to say that LLMs aren’t great. I love LLMs. I use them all the time. In fact, I wrote this very essay with the help of an LLM.

  The author argues that LLMs are an augmentation to existing tools, not a replacement. Just like the mouse didn’t replace the keyboard (my example), LLMs won’t replace existing workflows, it’ll merely help in the knowledge retrieval stage.

  For this future to become an actual reality, AI needs to work at the OS level. It’s not meant to be an interface for a single tool, but an interface across tools.

  This is where I partially disagree.

  Yes, I think some level of AI makes sense at the OS layer, but its function should be to find the right tool, not to be a tool. For example, “open my budget” would know from context which file that is (family, company, client, etc), which program (GNUCash, Excel, or a URL in a browser), and then pass on context to the app-specific AI, which would know which part to open and be ready for context-relevant questions (is it payday? Was I just looking at concert tickets? Is someone’s birthday coming up?).

  But even then, the usefulness of a system-wide AI is pretty limited. Most people can efficiently navigate to what they want. Indexes work well to find files (and full text search is feasible), file extensions work well to open the right application, and applications remembering what they were last doing is usually sufficient.

  So I see it as more of an accessibility feature at the system level instead of an actual, useful system in itself. However, I really like the idea of different models passing context in some standard way to each other so I can seamlessly move between apps.

  But I absolutely agree with the main point here: AI should be seen as an add-on, not a replacement.

  source
• Mbourgon@lemmy.world ⁨3⁩ ⁨weeks⁩ ago

  An interesting/useful article, if only because I was unaware of the Mac equivalence of Launchy (Linear, Superhuman,etc). The biggest problem is that a Conversational interface could excel at certain tasks, but just using it Willy-Nilly is asking for disappointment, plus the voice recognition still needs to get better, especially with context clues, which would require more integration (watching whatever you’re watching, hearing whatever you’re hearing, what page are you looking at, etc. )

  But if you’re in front of a keyboard, then directing the computer to do something has got to require less context switching then just bringing up on the keyboard. Even if it’s only 60 words a minute, if all you’re doing is typing in a handful of strokes, then it’s probably faster than coming up with all the keywords necessary to tell the computer to do it.

  source

  • taladar@sh.itjust.works ⁨3⁩ ⁨weeks⁩ ago

    As a primarily CLI user on Linux I wouldn’t even think of most commands as “words per minute” unless I am composing a complex pipeline or run a command with dozens of parameters at which point typing speed is not my bottleneck limiting the speed of input anyway and a free conversational interface would be totally fucked trying to figure out what I want it to do.

    source
  • paraphrand@lemmy.world ⁨3⁩ ⁨weeks⁩ ago

    Quicksilver is the original one of these apps. And it appears it predates all of those (launchy, etc)

    I prefer Alfred these days myself.

    You can even do similar with just Spotlight on macOS and the Start Menu in Windows 10+.

    source

    • darkkite@lemmy.ml ⁨3⁩ ⁨weeks⁩ ago

      i replaced my windows start menu with the offical powertools run. which is pretty much the alfred/spotlight solution now

      source
• TheUniverseandNetworks@lemmy.world ⁨3⁩ ⁨weeks⁩ ago

  Interesting article, I agree with his analysis, not sure (yet) that I agree with his conclusions. My brain needs to think about it in the background for a bit (just the way mine works).

  TLDR: we should expect conversational interfaces to be an addition to the workflows we currently use.

  source

  • drspod@lemmy.ml ⁨3⁩ ⁨weeks⁩ ago

    Computer upvote this post. I mean comment. No, I meant the comment. Computer remove the upvote from the post. Computer upvote the comment.

    Computer compose reply.

    Dear Aunt, let’s set so double the killer delete select all

    source

    • Imgonnatrythis@sh.itjust.works ⁨3⁩ ⁨weeks⁩ ago

      In my own real world usage I estimate a comprehension rate of about 92% with voice agents. I’m no linguist, but I’d guess that you’d need to achieve at least 98% comprehension to not feel like a conversation is frustrating. I’m also trying instantly irritated if my computer is delayed and nothing happens when I click on something, or if I go to use someone else’s computer and they have double-clicking enabled for some reason (why?!) so my tolerance is probably on the low end.

      Anyway, I thought this was an insightful read and the key to me is that the bar is pretty high now for Man-machine interfaces, so any implementation of newer tech needs to be both thoughtful and bug-free as possible in this realm.

      source

      • -> View More Comments
• DemonVisual@lemm.ee ⁨3⁩ ⁨weeks⁩ ago

  I always refer back to this video, he has a lot of excellent points. Of course all of this varies depending on the person, personally I don’t know if it’s actually be able to write 36 words per minute on a traditional phone keyboard but a Swype enabled one I might actually get close… I really like the idea of a pie launcher for the desktop, I use one for the phone, and while it’s limited to a single “layer” it’s still faster when doing routine tasks on the phone (music,maps, notes, internet).

  youtu.be/AItTqnTsVjA

  Again, phones seem to be the forerunner for these concepts to be implemented, since AI is creeping in quickly - I haven’t really found a great use case, but maybe actually listening to the sales pitches; “create a calendar note”, “compose short message of ETA” could be something that can accelerate the day to day, maybe even run locally on the phone or computer?

  source
• tynansdtm@lemmy.ml ⁨3⁩ ⁨weeks⁩ ago

  A fascinating read. It inspired me to look further into the StarCraft voice integration. Other games have tried it, using voice commands to direct computer companions as an additional layer of realism. But I’m not aware of any game that’s done it well. Might be nice for applications like picking from a long list, sometimes “build unit X” is way faster than paging through buttons, but again we have keyboard shortcuts for that. Keyboard shortcuts wouldn’t work for dynamic menus though, and voice commands do.

  Sorry for the stream of consciousness.

  source