Chatbots are not the future

&gt; When I go up the mountain to ask the ChatGPT oracle a question, I am met with a blank face. What does this oracle know?<p>I think if your attitude is that its an oracle, then you already have the wrong attitude for using the tool. Chatgpt is a tool, if you dont know how to use the tool, stop complaining that you don&#x27;t like it. Imagine telling everyone scalpels are horrible tools because they tried to perform surgery on someone and botched it up.<p>One day, not too far off, we are going to be able to tell a &#x27;chat bot&#x27;: make me a cartoon, it takes place in a steam punk fantasy. make it 2 seasons with 22 episodes each season. great, add a cliff hanger at the end of episode 11. add a love story component to it. reduce it back down to 1 season but 44 minute episodes.<p>content creation is no longer going to be tied back to knowing how to draw cartoons, or have armies of writers. yes, we can get garbage out of the system. but its a tool, plenty of tools produce garbage results if you dont know how to use them.<p>As an experiment, I asked chatgpt to write a business plan (one my brother started). The business plan was very close to what my brother produced, after working on it for a month. That&#x27;s powerful, that&#x27;s worthy of being &#x27;the future&#x27;.

&gt; Good tools make it clear how they should be used.<p>This is such a weird statement from someone in the tech space. Programming languages rarely have an opinion for how they are to be used (for example JS MUST only run the browser or which code style to use).<p>When I chat with customer support, I wish they could meet me where I am instead of me needing to learn their tools. For example, I want to say &quot;cancel my subscription&quot; and my subscript get cancelled. I don&#x27;t want to have to figure out which sub menu of the sub menu that has the magic &quot;end subscription&quot; button.<p>I know how to use my tool (english). LLMs teach computers how to use that tool too.

I don&#x27;t know why it hadn&#x27;t occurred to me before now, that using ChatGPT is quite to similar to playing Zork and Infocom games from the 1980s, with less trial and error needed to get something out of it.<p>The point and click adventure games from Sierra and Lucas Arts were a huge step forward in interaction, although you didn&#x27;t have to use your imagination as much to solve the puzzles.<p>And here we are again asking users to type their way to success.

Chatbots are the future but your points are valid. They don&#x27;t provide affordances, however chatbots provide a form of progressive disclosure and direct interaction that was previously impossible.<p>Toolbars and menus provide affordances but you still need to know what things are called and what order to use them. &quot;I&#x27;d like to email this file as a PDF and I&#x27;d also like to print it.&quot; may be much easier in a chat UX than in a menu based UX. Often these things can co-exist but chatUX has access to much more nuanced UI that would otherwise be too complex to build or expose.

&gt; Text inputs have no affordances<p>It has <i>all</i> the affordances. You can turn it into anything you want. Want it respond with json, check, it can do that. Turn a wall of text into a Python data structures, check, it can do that too.<p>You can take your LLM text interface and put what ever api you want on top. You start with clay and mold it into anything you need. You construct a parser so that the way data is already constrained and validated. Same goes for the output.

Quality quote:<p>&quot;There&#x27;s an ongoing trend pushing towards continuous consumption of shorter, mind-melting content. Have a few minutes? Stare at people putting on makeup on TikTok. Winding down for sleep? A perfect time to doomscroll 180-character hot takes on Twitter.&quot;

Someone said that using a chat-only interface is like using CLI tools. It&#x27;s even worse, I would add, because there is no autocomplete or man command to help you out. Most of us here probably get it, but my dad had a hard time getting a good answer from GPT when he first tried it, even with GPT-4 model.

Not sure I&#x27;m convinced - natural language is one of the most intuitive interfaces we have, it&#x27;s also how most instructions in professional contexts are delivered. What current ChatBots are missing right now isn&#x27;t radiobuttons for different styles but context. No one who just reads my messages can know what kind of style I&#x27;m looking for until they either get laborious instructions, or, probably better, they see some examples.<p>I&#x27;m guessing that training custom language models on a company&#x27;s data must be one of the hottest things you can be working on right now if you&#x27;re looking for VC money (if there&#x27;s something out there that compares to how well StableDiffusion+Dreambooth works for images, I&#x27;d be thankful for any pointers).

Learning how to use freeform text input can be pretty annoying when only a few things work. Some examples are playing a text adventure (&quot;guess the verb&quot;) and using an unfamiliar command line interface. Good error messages can help.<p>Web search changed that. Most queries work, at least somewhat.<p>There&#x27;s a point where freeform text input becomes better than structured input. A simple search box is what people mostly use instead of an advanced search form, let alone a web directory (like Yahoo! back in the day).<p>For web search, there are very few error messages. If you enter a query that doesn&#x27;t work very well, you get back results that aren&#x27;t very good or what you wanted, so you try something else.<p>With AI chatbots, expectations are sky-high, but there are times when they <i>should</i> refuse with a good error message, because they really can&#x27;t do what you&#x27;re hoping to do. An example is when you ask it to explain its reasoning. An LLM never knows why it wrote what it did, but it will try to invent a plausible explanation anyway. [1]<p>Better error messages that help users understand what chatbots can actually do would help avoid misconceptions, but this won&#x27;t happen unless the error messages are trained in.<p>[1] <a href="https:&#x2F;&#x2F;skybrian.substack.com&#x2F;p&#x2F;ai-chatbots-dont-know-why-they-did" rel="nofollow">https:&#x2F;&#x2F;skybrian.substack.com&#x2F;p&#x2F;ai-chatbots-dont-know-why-th...</a>

Yeah, chat isn’t a universally great interface. But it’s a great default because it’s totally free form.<p>It should be pretty easy (even with today’s APIs and technology) to have an LLM design a user interface for you for your current task.<p>Simplest way: output a JSON of simple control definitions with every answer.<p>Coolest way: Just have it generate a full-ass React front end or whatever on every message.

Contextually relevant suggested options (that can be acted upon with a single click), alongside the free form input box will emerge as the norm.<p>Down the line people will expect applications to be chat ready. They will see an input box and expect the application to understand natural language and respond in the most helpful way. Which might be showing an error message or suggesting relevant next steps.

Amelia presented a bit more on her demo at our meetup last week:<p><a href="https:&#x2F;&#x2F;www.latent.space&#x2F;p&#x2F;build-ai-ux" rel="nofollow">https:&#x2F;&#x2F;www.latent.space&#x2F;p&#x2F;build-ai-ux</a><p>full recording in the video at the bottom!

A lot of people are projecting for ChatGPT to come for the search market, but I wonder how that will play out for lowly cognitive queries.<p>Quite a few people use search by awkwardly typing on mobile one or two words, probably misspelled and&#x2F;or auto-completed as they type it. The query isn&#x27;t sophisticated, lacks a lot of context and parameters, which the search engine then tries to guess.<p>When you use ChatGPT in that way, you&#x27;ll get useless generic answers. It seems to shine specifically when being more specific, detailed, which also suggests users are willing and able (education level) to give such rich input.<p>The idea that it&#x27;s better than search for this specific normy behavior, I openly question. And let&#x27;s not forget about the economics. More expensive to run, vastly less ad space, and content owners (the whole web) are going to be pissed and will put up ever higher walls.

Why even write a book when every insight in them can be shaken out of an LLM? You know, shorter content doesn&#x27;t have to mean mind-melting right? Strip away the self-marketing poison found in every social media post and lets see.<p>Books were very simple machines to dress up our thoughts and the community reoriented itself well to the demands of the machine. But progress marches on the graves of obsolete machines as it did quills, book presses, typewriters, telegrams, libraries, word processors. Joe Reader has the same access to mind blowing dynamic text that creatives only wish to see as a finishing tool. He won&#x27;t settle for the old glass window over static text just to please the artisan book writer.

No, they&#x27;re not, but my reasons for reaching that conclusion are somewhat different:<p>1) I don&#x27;t think chatting with anything, human or machine, is a learning experience, particularly since the machine veracity is poor, untrustworthy, and Hinton&#x27;s resignation today tells you everything you need to know about the narrative inside big research orgs right now.<p>2) Recognition vs. recall. Given that it&#x27;s the equivalent of an informal language CLI, which I prefer by the way; but there is no recognition (as in symbols) only recall.<p>Long story short, I think the emergent need is for written communication, with a tip of the hat to Daniele Procida:<p><a href="https:&#x2F;&#x2F;ubuntu.com&#x2F;blog&#x2F;engineering-transformation-through-documentation" rel="nofollow">https:&#x2F;&#x2F;ubuntu.com&#x2F;blog&#x2F;engineering-transformation-through-d...</a><p>Except that what&#x27;s missing is a human-computer collaboration, i.e. sensemaking with another tip of the hat to Peter Pirolli:<p><a href="https:&#x2F;&#x2F;www.efsa.europa.eu&#x2F;sites&#x2F;default&#x2F;files&#x2F;event&#x2F;180918-conference&#x2F;presentations&#x2F;20-3_07_Pirolli.pdf" rel="nofollow">https:&#x2F;&#x2F;www.efsa.europa.eu&#x2F;sites&#x2F;default&#x2F;files&#x2F;event&#x2F;180918-...</a>

I think the future interface is a smart assistant for your life that gives you suggestions on what you should be doing (both for work and personal life). Sure, there may be a prompting text box, but the assistants will be so good at suggesting that you won&#x27;t need it very often (besides searching for the occasional thing or giving feedback).<p>Driving these suggestions is all of your data as well as your goals and values that you can give to the assistant in natural language.<p>At work the goal might be: &quot;I want to sell $100,000 worth of widgets this quarter&quot; and it will break down step by step how that might be possible.<p>For personal life it might be &quot;I want to get involved in the kayaking community&quot; and it will recommend activities, clubs, etc.<p>Once these assistants are good enough, it will be reckless to not use one (especially at work). We will then live in a world where AI and human live together and make decisions together hand in hand. Buckle up.

When it&#x27;s about the future, limitations of current implementations aren&#x27;t a strong argument.<p>ChatGPT can be confidently stupid, but what if it gets better?<p>You need explanations&#x2F;affordances of what it can and can&#x27;t do only when its capabilities are limited. If it really could do whatever you asked, you wouldn&#x27;t need it. Just say what you want.

A contextualized chatbot is still a chatbot. I think they’re going to stick around for a while… we’ve effectively been trying this out on the web since the AskJeevs days, and that dream is mostly realized now.

Not a very convincing argument, your gizmo sliders and checkboxes or whatever aren&#x27;t going to replace chatbots but only slightly extend them and really aren&#x27;t needed if AI gets better over time.<p>Some neural connection to the brain that will interpret your thoughts is the only logical thing that will supersede chat ai, but that won&#x27;t happen for a while. Maybe a connection to all your data will happen first so the AI will better understand what type of person you are and what you want, that&#x27;s already probably happening based on past responses.

&gt; I&#x27;ve convinced you that chatbots are a terrible interface for LLMs.<p>I was already convinced of this. What I&#x27;m not convinced of, and the article has little to say about, is<p>&gt; Chatbots Are Not the Future... chatbots are not the future of interfaces.<p>Chatbots are a terrible interface to LLMs, and yet they are absolutely going to be the future of every third godawful website I must visit.

I agree that prompts are mostly just context. Although you can get quite detail with that, to a degree that it doesn&#x27;t feel that way. That&#x27;s why I&#x27;m building InventAI, to help with that process: <a href="https:&#x2F;&#x2F;inventai.xyz" rel="nofollow">https:&#x2F;&#x2F;inventai.xyz</a>

Google Search got to be a pretty successful business and probably still the single most popular information retrieval tool on earth - it was done using with a single input box.

If you think the far future (100s of years) involves being able to talk to a synthetic humanoid using spoken language, then Chatbots are almost certainly a point on the curve.

&gt; The interface looks the same as a Google search box<p>Indeed, and like any other text box, like whatsapp, like word — all tools that no one uses because they lack affordances.

Very very offtopic:<p>He called ChatGPT oracle, nice but not enough.<p>I want someone to name chatbot their `oracle of delphi` plz. thank you.

TL;DR, more targeted tools could be more helpful for specific tasks than an unstructured text interface for everything.

one thing I can say for certain: scroll handlers definitely aren&#x27;t the future