Chatbots are not the future
wattenberger.com> 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?
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't like it. Imagine telling everyone scalpels are horrible tools because they tried to perform surgery on someone and botched it up.
One day, not too far off, we are going to be able to tell a 'chat bot': 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.
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.
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's powerful, that's worthy of being 'the future'.
Yes, sure. But you know what makes most of human history's output of art, music, literature et al great? Intent, attention to detail and self-expression.
Broad strokes are broad strokes. I can procedurally generate levels all day long in a video game, but for me, they're never going to be as compelling or interesting as a lower-resolution and low quality textured game from the '90s or '00s where every single tree and rock is placed with intent.
I already think modern cartoons are fairly sterile and soulless versus their hand drawn or hybrid counterparts. It isn't even elitism, they just don't hold my attention or interest me artistically, stylistically, or in terms of content.
If you choose to express yourself in broad strokes, that's fine, whatever floats your boat. I'll continue to chase things that have intent and artistry behind every aspect of them. Generic and formulaic is generic and formulaic all day long. It's also why I don't like most modern anime, it's sterile visually and isn't why I enjoyed the medium.
Really?
For a writer, if you write out a plot, you can get the AI to actually simulate the character's responses and dialogue (even voiced by AI!). There, you've LITERALLY brought a character to life, each character is driven by a different persona simulated by a different AI, the quality of stories that will create, will annihilate what came before.
You want to write an adventure, but want to keep it unpredictable. Ask the AI for ideas, there, the adventure is now a true adventure, not a fake mirage created by the writer.
No need to describe scenery, no need to describe character appearances. Feed those descriptions into txt2img, and you get portraits that would have cost $1000/pic from top tier artists.
Generic and formulaicness, comes from having TOO MANY PEOPLE. Too many people involved in production, means the creator must dilute intent, appeal to wider audiences, and limit risks, to ensure costs are reclaimed. Once AI gets going, you'll see indie creators making full anime series, and releasing them on youtube. Because for an individual creator, even ad + patreon revenue alone would be able to sustain a comfortable existence, with no dependency on corporate or teams.
I thought people who love art, would be exhilarated by AI. I realized, the majority of artists don't love art. They love drawing, but not art. They love socializing with artists, but not art. They love receiving attention and income from their art, but not art. That's all fair and fine. But there will be people, who just want to create the best possible art, no matter the method, no matter the reward, and with AI, this latter group will outcompete the first, hard.
Perhaps the resolution to your cognitive dissonance is not to conclude that artists hate art. Instead I’d suggest that often, artists’ love of art doesn’t stem from concepts like “annihilating what came before” or “outcompeting hard.” Perhaps, like the parent comment said, it comes from self-expression — the pleasure, for both creator and audience, of the “mirage created by the writer,” to borrow your poetic term.
“Art is not a mirror held up to reality, but a hammer with which to shape it” – Bertolt Brecht.
I'm neither dissenting with your or OP here, FYI.
Mediocre artists will just copy what exists more easily.
And powerful artists will have new tools to shape reality.
There is some competitive notion, for certain, since there is a lot of pushback on AI-generated art from artists.
"ars gratia artis" is something AI is not capable of.
> You want to write an adventure, but want to keep it unpredictable. Ask the AI for ideas, there, the adventure is now a true adventure, not a fake mirage created by the writer.
There's something really funny about using an algorithm that predicts the next most likely token to generate unpredictable adventures.
>algorithm that predicts the next most likely token to generate unpredictable adventures
but enough about the human nervous system
Reminds me of:
https://en.wikipedia.org/wiki/Unexpected_hanging_paradox
You can certainly generate unpredictability from a predictable mechanism.
It’s like saying it’s hilariously to take molecules of H2O and pretend they could form fluid dynamics. Preposterous /s
Well? So what?
If I was writing dialogue, I already know exactly what I want to say, that's why I'm writing it in the first place. Sure, some people might like that, but I never feel devoid of inspiration because I consider the greater narrative of a project.
Why would I want to ask the AI for ideas? The joy is in the process of having it my way by my own hand, word-for-word, pixel-by-pixel, note-by-note.
If other people enjoy another method, fine, but you are invalidating the traditional and highly enjoyable method. Creativity isn't about efficiency or whatever. This is why people leave AAA to go indie, and have it their way, not to whatever the budget is or what the shareholders or managers want, but because they have a vision and are naturally creative, making a labour of love they know every detail of because it's in their head.
I'd rather support the artists and pay the thousand bucks because it's all a part of the traditional process.
Indies already release anime on YouTube, Netflix, TV et al, and have done for decades.
Art isn't about the output for most artists. Art is about the process. People consume art. Artists enjoy the creative process and having it their way.
I hate attention. I haven't had a photo taken in 13 years. I've stopped all interviews. I mostly work behind-the-scenes on documentaries and games these days. I don't even care about if a project makes a return, we do it because we love it and enjoy the process.
There is no "best", art is subjective. Art isn't even a competition. Crass and vulgar. Are you sure you actually understand art and aren't conflating it with monetizing it in a business context?
>You want to write an adventure, but want to keep it unpredictable. Ask the AI for ideas, there, the adventure is now a true adventure, not a fake mirage created by the writer.
I don't know if it's a symptom of Hollywood's sequelitis and nostalgia-wank, but the idea that consumers should be happy to consume art generated by a program that has simply taken the last 200 years of pop-art and averaged it all together is depressing.
When I read this it sounds like you've reduced art down to the labor and not the actual expression. Why would anyone want to read something where the character responses are simulated? Everything should at least be intentional.
> Generic and formulaicness, comes from having TOO MANY PEOPLE.
You know what else comes from the output of many people? What AIs produce.
> They love drawing, but not art. They love socializing with artists, but not art. They love receiving attention and income from their art, but not art.
Love of drawing, socializing with artists, and attention and income when people connect and buy their art are all forms of deeper engagement/investment with art.
The generous reading I can make of the mistake that sets these up as the other side of an inimical dichotomy where engaging these things is not loving art is if you're limited to the perspective of someone whose experience with art is that of a consumer and equates that perspective with loving art.
Being concerned about alienation from drawing, from connection with a community of other artists making efforts, and yes from attention and income that makes their focus socially/economically viable seems pretty reasonable.
It's also reasonable to be interested in what new tools can do and I rather imagine there will be people who enjoy that as well, some of whom may be able to produce visions that were previously inaccessible to them. That's interesting and exciting, but it doesn't mean there's no downsides.
Literally implies embodiment of an unintentional emergent nature. The characters only exist in the figurative way we think of humans existing as individuals with personhood.
Ideas such AI being alive are not universal among the public. Let’s not anthropomorphize plastic and metal we etched human like sentence construction into.
> I can procedurally generate levels all day long in a video game, but for me, they're never going to be as compelling or interesting as a lower-resolution and low quality textured game from the '90s or '00s where every single tree and rock is placed with intent.
The popularity of Minecraft and other procedural games would imply that there is still a large number of people who value exploring the unknown generated content, even if it means it's not curated.
Yes the quality won't be as good, but you do get quantity instead. And the quality will improve.
Yup, I ran a Minecraft server for about 8 years, I loved a bunch of procedurally generated games. I've worked on games in the past and present that use procedural generation. Furthermore, the most interesting part of Minecraft was what people built, not what world the seed generated, IME. Usually we'd WorldEdit vast swathes of it away to have flat areas to build and have fun, or we'd WE with brushes and materials to hand create terrain because the procedural generation was so-so.
My point still remains that there's a very different experience in something intentional and exact. One feels very human.
The quality isn't the issue, the broad strokes without intent are.
This is like preferring impressionism over a Dutch master. Audiences for both. I'm not invalidating any of them, I'm just saying that for me, I prefer something more human and different.
For a long time I wondered where popular misconceptions about facts(usually historical) came from. Then I read about the history of the Encyclopedia Britannica and its checkered editorial history and realized I had found the cause: we relied a lot on truth-making authorities in the past, and they wrote in their biases.
So when we speak of intentional creation, we are in some ways looking a mode of truth reminiscent of Britannica: the world as someone experienced and described it, biases and all. Neither a sensory truth like "it feels cold today", nor an emergent collective belief like our rules of language.
What makes the LLM chatbot interesting is that it resembles a lens into collective belief; it starts with one bias, but you can tell it, "now explain it the way this other type of person would," and it will oblige you with its best in-character approximation, changing facts and details as well as aesthetics. It is capable of deep and radical changes to its output with minimal changes in prompting.
That's something that you could access in a limited sense with traditional PCG, but almost entirely in terms of mathematics operating on premade assets - fractals and chaos functions and so forth. Here you have a "fractal encyclopedia", working over an enormous reference library. It can, in fact, attempt to place every brush stroke "in the style of a Dutch master" if recursively prompted to do so, although in any existing implementation, you will run into technical limits with its dataset and working context. But if we look over towards diffusion model image generation, we know that that general idea does work - it's always a little less exact than the original human, but it can get us far past the uncanny valley.
And that's valuable because all the human creative stuff is built on references, too: original ideas emerge in one context, then get transferred to another. So there's a sense of, yes, you can definitely build a "make game" button here, and it'll be an approximation of a hypothetical human character that built the thing, but you can easily turn that output into a reference for something else that injects more bias and humanity into it or adds more structural rules to rigorously shape it. Often that's the actual path of creativity: make a longer pipeline of refinement with more layers to it, and you end up with an increased sense of transformative elements and intentional structure.
Not to mention the middleground of handmade pieces and putting them together in a randomly generated way. That's how a ton of games work - from Noita, to Cataclysm DDA, to Dead Cells.
That's quite literally how Minecraft works. That's also how games I've worked on function.
Pretty sure Minecraft uses Perlin noise for terrain, while Noita uses entire pre-made tiles. They both have setpiece structures dotted around the terrain, but I think lelandfe was making a distinction between Wang tiles vs a gradient noise generator.
You have ignored the original point made entirely that intentional level design is generally far more interesting than anything algorithmically generated.
Specifically, yes, but my point was more that there's pre-created elements thrown in there too. It's a very surface-level argument of generative art versus intentional art, the technology itself isn't what's being discussed, it's the output and how interesting it is or isn't.
I'm not even saying procedural generation et al is bad, it has its place.
People use stock/library content in games all the time, whether it's textures, rigs, whole environments, or even mechanics. Can interesting things be made? Yes. Are the usually unique or innovative? Well, generally no.
What made Minecraft so interesting for most people was what people made within it, the world seeds themselves were really whatever for most people. If they were so interesting then people wouldn't be importing schematics for models, wholly pre-built worlds, hubs, and even some creative mode servers. Sure, some people like exploring the procedurally generated stuff, but once you've done that for a few tens of hours, it's boring and you've largely seen what's worth seeing.
I spent more time using WorldEdit than screwing around with seeds because I knew exactly what I wanted my server's main area to be like rather than relying on a seed.
I was responding to the "literally how Minecraft works" part rather than the discussion further back.
That being said, world generation is a bit like a toybox, or a backdrop to a play in a theatre. Minecraft's world gen isn't worthless, and so improving the backdrop is worth discussing. The differences in the output of implementations matter a lot (to me, at least) and I think Minecraft would be much less compelling if the world was generated using a 3d implementation of Wang tiles. "A few tens of hours" is a lot of content.
Would Minecraft be even more compelling with neural networks doing some of the content creation? I think it might. Would that totally replace the need for human input and invalidate everything players build? No, but that doesn't make it worthless.
Perhaps!
But a few tens of hours is very little versus the quite literally thousands of hours spent running a server for 8 years and using plugins on it. Truth be told, I think even if a Minecraft world was literally a blank base plate and it was digital collaborative LEGO, it still would have been a hit.
I guess the thing with Minecraft is that it was many different things to different people. Some people loved "Hunger Games" style worlds/modes on servers, others like single player mods, or even just vanilla Minecraft.
> But you know what makes most of human history's output of art, music, literature et al great? Intent, attention to detail and self-expression.
What makes them great is that people enjoy them. Whether they were created with "Intent, attention to detail and self-expression" or monkeys banging on typewriters is irrelevant and indistinguishable.
People enjoy them because it resonates with them because it's intentful self-expression with an attention to detail. There's a reason that generative art is very uncanny valley.
>I'll continue to chase things that have intent and artistry behind every aspect of them.
For someone with that POV you sure are peddling the "Everything is soulless, stop enjoying things" perspective.
Procedural generation is miles away from what these LLMs are doing. It’s not just coming up with a random maze/map for a fetch mission. Right now, it can generate a novel quest, with all of its characters, come up with a unique set of mission goals and theoretically generate those assets and characters (atleast in 2D). Turns out, all this self-expression or whatever is cheap.
Not in this context of self-expression. Neither are intentional, both are broad strokes.
Technologically? Yes, you're right, very different.
But in output? It's still generative.
“Intentional” is a weasel word in this context. It means very little in terms of the end result. Humans making art is generative as well even all the way back to cave paintings which were interpretations of animals in nature.
No, it isn't. Do you really think Alfred Hitchcock would leave his film up to others, let alone a machine? No. It wasn't his vision, thus it wasn't his self-expressive intent. Read about auteur theory.
That's not what generative art is, mate. https://en.wikipedia.org/wiki/Generative_art
I'm not arguing artists are going to go away, just that story telling is going to become a lot easier. it also doesn't mean the content produced is necessarily painted with 'broad strokes' and absolutely doesn't mean its soulless. It means the barriers to entry for producing stories just came down a whole lot - akin to the printing press did for books.
But is the person using the tool an artist? I think this is an important question. If I give detailed instructions to a human ghost writer about a story I want them to write, I don't think anyone would say that I wrote the story. It was written by the ghost writer.
If a piece of art is made by a computer based on detailed instructions, that art was made by a computer, not a person.
If you are in the camp that you don't care whether or not art was made by a human, this isn't even a little problem. If, however, you are in the camp that cares a lot about that, then this is a very, very serious problem.
Either way, this means that this isn't "just a tool" like a printing press. It's something completely different, and more than a tool.
Yes, the artist that chooses to use AI-tools to generate art-work are in fact artist.
For those who are afraid of AI being content generators that puts artist out of work will most likely be disappointed. However the technical gatekeeping some artist do goes away, and it makes room for more people being able to express them self creatively.
Art is about the why. We as humans will always ask that question, and we will produce answers, no matter the tools.
We've gone through multiple iterations of technology questioning if you are really an "artist" for using it, but the new creations and the new generation of artist puts that to shame in my opinion.
Digital pixels are not paint brushes. So if you do not move your mouse/brush to generate a stroke? What does it matter?
AI-tools speeds up the creative process which for some will let them go places we currently are having a hard time to imagine.
> Digital pixels are not paint brushes. So if you do not move your mouse/brush to generate a stroke? What does it matter?
If you are literally explaining every stroke, then it doesn't. But that's not what we're talking about here. We're talking about describing something in pretty general terms and allowing a computer to make the creative decisions (what "brush strokes" to make).
No, "the computer" allows the artist to be exposed to a magnitude of various arbitrary and curated creative decisions that can help guide their work and their intent.
Here's a stupid personal and anecdotal example:
I've been trying to teach myself to do watercolor paintings of my photography.
That's been going OK, but with the help of img2img, I can "quickly" generate thousands of variations of Watercolor paintings of my OWN work, then choose the various elements I like, then I paint them into one painting. Which have led my freehand painting skills to improve at a much higher rate.
If however I just put up a feed of the generated watercolors from Sdiff, it would be immediate obvious. Of course, that's right now and doesn't speak to the vast improvements that are on the horizon.
This is what I'm personally seeing in my own circles of overlapping art creators trying to experiment with AI.
I guess what I'm trying to say is that: Artist will find ways to make their intentions stand out, whatever the tools we all have access to.
And if you move a horsehair brush, do you determine where each hair lands? If you spray paint, do you say where each drop lands? We have, for long, handed control to physical random processes. To modify that to land on mathematical random processes is not some categorical shift.
> If a piece of art is made by a computer based on detailed instructions, that art was made by a computer, not a person.
I completely disagree with this. According to this, no code can be art. For instance, videogames.
It has been enough time since ready-made (Duchamps urinal) and found objects, djing and sampling, and concept art. Art is not only drawing beautiful illustrations since at least the 2 world wars.
> If I give detailed instructions to a human ghost writer about a story I want them to write, I don't think anyone would say that I wrote the story.
This is exactly how many artists work today, with a small army of workers, even interns, to execute the plan of the artist. Even Rembrandt had people painting to produce more pictures. Another example would be architects: does the star architect execute everything, or do they have the vision and instruct their very large teams?
IMO it is all about the intent and interpretation of a human.
> no code can be art. For instance, videogames.
I feel like this is probably a pretty bad example, generally the "art" in video games comes from things like dialogue, storytelling, level design, graphical art, not the physics engine or the renderer. There is simply not art there, it's more engineering.
There's little more art in the code portion of video games than there is in a jet engine.
Again, I hear art here often as "that nice cover illustration, or that 3d model". A bit like people talk about "content".
Art can be anything, and I definitely consider some videogames art. The same can be said about architecture.
In the early days of Pixar, some traditionalist animators claimed that Pixar was "cheating" because they used 3D modelling tools rather than physically drawing the frames. The traditionalists didn't see computer animation as a "tool"; they saw it as the computer doing the work of the animator. Were they right? Computers make making things easier. But the human using the computer is the real creator.
Artists using 3d modeling software actually design and create the model, rigging and animation themselves, they don't tell the software "Metal articulated lamp bouncing across wooden table, 3d render, realistic lighting and shadows" and call it a day.
It's a matter of how much control you have over the end product, with AI it's very little. At best, if you want to be charitable, you could describe the role of the person using AI as an art director, but not an artist.
> they don't tell the software "Metal articulated lamp bouncing across wooden table, 3d render, realistic lighting and shadows"
They would if they could. Your premise seems to be based on you only getting to interact with the AI "art"generator once.
What makes you finish the quote with "and call it a day."?
What about a writer that uses a recorder to record themselfs speaking out a book and never actually "writes"? Would they be a singer instead?
They would be a voiceover artist, a storyteller, a voice actor, a narrator, a performer, even an entertainer. Depends whose book it is and if the book already exists to be read from.
Ever heard of radio plays? Audio dramas?
So, if I record myself talking transcribe it into a book and publish it, I would be a performer, narrator or even an entertainer -- but not a writer?
Maybe my point was unclear. If I had no arms, but still managed to publish a book, would I not be considered a writer since I do not have arms?
You'd likely be a scriptwriter.
It really depends. If someone has someone else ghostwrite them a book (this is super common), are they still a writer? No.
Are they an author? Maybe.
Are they just a brand slapped on the cover to shift units? Yeah, pretty much.
Toy Story was not generative art though. Pointless comparison.
Compare generative art tools to other generative art tools.
Programs take in input and produce output. The whole point of using any program is that the program produces more than you put into it -- otherwise why wouldn't you just use pen and paper? Either all programs are "generative" or none are.
It comes under the category of "generative art". It isn't intentional or self-expression, regardless of what the prompt engineers may kid you. Is it "art"? Sure. Is it traditional art? No.
I totally agree with this. But as an art lover, I want some real way of being able to tell if a piece of art is this, or is traditional.
Artistry is a beautiful thing. It's also why I love hearing from industrial designers on how they arrived at their design decisions, challenges they faced, intent, nuances in the design you might overlook that were difficult etc.
Unfortunately, short of more behind-the-scenes material and interviews, the only way to really get a feel for it is looking at their body of work as a whole. The great ones always have a specific style that they refine or evolve over time. It's unmistakably x.
I have to say, I've yet to see any AI artist hit a signature style, and I've yet to have an AI generated piece of art move me emotionally or conceptually.
Are they interesting? Sure. So's glitch art, but there's not much substance to any of it, and I remember none of it. Intent and self-expression is such a huge part of art.
Can you imagine what Hitchcock would say about AI anything? He wanted it 100% his way.
It is intentional, it's just a generated approximation of intention.
No, it isn't intentional, it's generative. StableDiffusion DALL-E, MidJourney et al are quite literally by definition a generative art model. This isn't up for debate. That is quite literally how they function.
> If a piece of art is made by a computer based on detailed instructions, that art was made by a computer, not a person.
Tell that to music producers and digital artists. They don’t know what detailed instructions are run by the cpu of the device they use, and yet it is still art and they are still artists.
But that's entirely different. A digital artist is directly engaging in art. The computer is, in that case, just a tool like a paintbrush. The artist is still the one making all the creative decisions.
To go back to my ghost writer analogy, the reason that nobody would say I was the author is because I wasn't the one who made the creative decisions. I just described what I wanted to another person who made the creative decisions. Therefore, the other person is the author, not me.
ChatGPT including the GPT-4 variant sucks quite terribly at creative writing, especially the kind of writing necessary to create long-form narratives like that in serial television and especially novels. The technology will certainly improve and it will definitely become a staple in the editor's toolbox, but it is so far away from being able to produce long form narratives well that it's not a given it will eventually get there.
It's just down to small context window. Once the context window is big enough to fit entire examples in the training data, then it should be trivially solvable to train a model to do it.
There are various finetuned models out there for conversations or story-telling, though they're quite small in terms of parameter count at the moment, but I don't see it as being fundamentally impossible.
Scalpel manufacturers don't advertise their scalpels as capable of performing surgery on their own.
Do you have an example of AI false advertising?
gesticulates broadly at the entire field
Fair enough
While there is a massive amount of overhype this cycle, I would still say that most of the hype comes from newly minted AI "experts" that were recently, e.g., cryptocurrency experts, or NFT experts, or epidemiology experts, or politics experts, or "integrity in gaming journalism" experts, or whatever, etc.
When people like Ashton Kucthar [1] are the equivalent of shoe shine boys and taxi cab drivers giving stock tips, then I feel the hype has gotten out of control.
Don't get me wrong! As someone who researches learning and intelligence, I think these developments are pretty neat. But man, the amount of people who are defining intelligence down to the level of a chatbot or predicting the impending doom of mankind are ... getting a little ahead of themselves.
So many experts.
[1] https://www.businessinsider.com/ashton-kutcher-firms-go-out-...
> 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's powerful, that's worthy of being 'the future'.
How much knowledge and ideas did your brother personally develop over this month in addition to the business plan? Being handed a working plan is sometimes less useful than the aggregate experience leading up to the plan.
Case-in-point: why do some biz succeed under one CEO then fail under another quite unexpectedly? The former CEO likely understood their market, their product, their business model, maybe they even dogfooded and used what they made.
The "why?" is frequently more important than the "what?" with biz. An LLM doesn't really have an understanding of how the world works, so I would be very sceptical of its ability to write a sensible business plan. It doesn't even understand the product/service, and the nuances of what it does because it can't experience it.
The LLM is a yes-man with no experiences.
Imagine taking a year-old old book about what worked for someone's biz years ago, then naively assuming it'll work for yours today without considering how the world, market, people, technology et al may have changed in that time. This is the same thing. It assumes x is x. Usually x is not x. People used to compare their biz to Apple, despite not even being in remotely the same industry.
Every "impressive chatgpt" story I hear is about someone comparing what chatgpt produced to what a human produced, and saying it's scary close. No one talks about all the times they asked chatgpt to produce something, didn't compare it to what a human could do, and then realized it was catastrophically wrong once implemented.
I suppose all that really matters if the ratio of the two? I have use-cases for which I use GPT-4 and it does as good a job as I would, but faster and so I just leverage it by default now and am faster at certain things. There are also things I try it for and I don't get a useful result, so I learn not to rely on the model for that particular use-case. In that regard, it's a tool and you simply learn what it's useful for and what it's not.
But it's a tool that requires you to already know what you're doing, which I think cuts in to a lot of the techno-mystical powers that have been attributed to it.
> Chatgpt is a tool, if you dont know how to use the tool, stop complaining that you don't like it.
It's a tool that can be opaquely configured to be used in a million different ways, and when using it does not bring about the desired result, its acolytes sneer, and suggest that you're using it wrong.
It's like a multi-tool, that only works when you're blindfolded. Sure, it can be used to hammer nails and tighten screws and strip wires and measure a tire's pressure, but it makes it quite difficult to find the magical incantation that will apply the right end to the job.
(And most of the time, it quietly leaves the screw untightened, the wire clipped, and the tire with a hole in it. It's the user who's wrong, of course.)
I think the point of the article is that this does not constitute a chatbot. It's not conversational. It's not really general-purpose. That doesn't mean it isn't powerful, just that telling users to have a freeform conversation isn't going to work. It's also why everyone is getting excited for "prompt engineering" to make better use of it. The biggest user value is still going to be a level above the open-ended chat UI we have right now. He's not saying GPT is useless, he's saying we haven't put it into it's optimal context yet.
> Chatgpt is a tool
When you use a hammer or a drill, do you expect it to sometimes not hit/screw the nail?
If ChatGPT is a tool for knowledge transfer/extraction, it can't hallucinate/lie to you/be wrong/make stuff up.
If it's a tool for potentially discovering some knowledge that may be true and needs to almost always be verified by either a compiler or a followup "find me a reference/discussion" Google search to make sure it's accurate, then sure. But I don't think that's what it's primarily being advertised as.
Beyond the obvious fact that you can accidentally hit your thumb with a hammer or strip the head off a screw with a screwdriver, I’d very much like to hear about any tool for collecting knowledge that is perfect.
Web searches will for sure give you wrong answers. Even professors or other experts in a field will be wrong sometimes. Heck, even Einstein got some things wrong.
Your goalpost is in the wrong spot. Tools don’t need to be and probably never can be perfect. But that doesn’t mean they’re not useful.
If you hit yourself with a hammer, that’s your fault because you did something wrong. Comparatively, you can pose a completely correct and unambiguous question to ChatGPT and still get a wrong answer. I don’t like this implicit shifting of blame to the user when it’s the tool that is flawed.
The design of the hammer makes it easy to miss the nail, and you need some skill in order to use it effectively. A nail gun is an example of a better tool for driving nails, since it’s faster and allows for greater accuracy.
Similarly, you can ask ChatGPT for an answer and it might get something wrong. It takes some skill to know how to interpret and verify the response. If a user takes the response as truth without question, it’s partially the user’s fault.
But it's not the users fault if the question is correct and unambiguous. To continue with the hammer analogy, that's like landing a perfect hit on a nail and the hammer's head somehow falls off and richochets into the user's eye.
I just don’t see what point you are trying to make here. Yes, ChatGPT can give the wrong answer given a correct input, but that doesn’t mean it’s not a useful tool.
Think about how GPS can give a bad route, especially if there is construction or snow on the road.
Or how keyboard autocorrect sometimes changes what you wrote into something silly and wrong, even if you originally spelled the word correctly.
Or how OCR and speech-to-text software sometimes makes mistakes.
Or how Google Translate uses unnatural or incorrect word choices sometimes.
Are these not useful tools even though they get things wrong?
> Or how keyboard autocorrect sometimes changes what you wrote into something
> silly and wrong, even if you originally spelled the word correctly.
> Or how OCR and speech-to-text software sometimes makes mistakes.
> Or how Google Translate uses unnatural or incorrect word choices sometimes.
When you point this out no one will jump in to defend them. If you say the same about ChatGPT your inbox will suddenly be full of people telling you you're just using it wrong (see this reply chain for example).
It's not the user's fault if they're handed a power tool without training, but it's not the tool's fault either.
> When you use a hammer or a drill, do you expect it to sometimes not hit/screw the nail?
Not sure about drills, but this absolutely happens with drivers if you fumble mating the bit to the screw head, or if you miss the stud, or if you overtighten, or if you don't sometimes pre-drill, or if you strip the head, or if you don't correctly gauge underlying material composition, or thickness, or if you...
Exactly. I use ChatGPT for help coding sometimes and it's like 50/50 if I get an answer that is truly helpful, but that's infinitely better of a tool than I thought we'd have 1 year ago.
> Good tools make it clear how they should be used.
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).
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 "cancel my subscription" and my subscript get cancelled. I don't want to have to figure out which sub menu of the sub menu that has the magic "end subscription" button.
I know how to use my tool (english). LLMs teach computers how to use that tool too.
I think the problem in your example of cancelling the subscription is that service providers often make it difficult on purpose. I doubt they'll allow chatbots to make it any simpler.
I get support emails for "how do I cancel my Apple App Store Subscription?" when Apple governs the cancellation process in a centralized and simple manner.
I also get support emails for password resets, which I try to make as simple as possible.
People don't want to learn new tools if their existing tools (language) work just fine.
> Programming languages rarely have an opinion for how they are to be used
Erm, they absolutely have an opinion. That's why I can't just write however I like in whatever language. I need to stick to the designer's opinions on syntax and semantics otherwise it won't work.
When chatting with a Chatbot, you also would need to stick with the language's syntax and semantics otherwise it won't work. You can't just write however and whatever you want and expect the bot to understand you.
Subscription cancellation buttons are intentionally confusing and hard to find. It's easy to add a big red Cancel button in an obvious place that works immediately, but companies choose to avoid that route in order to extract more money from customers. Nothing about that dynamic will change with LLMs, those same companies will have the same priorities.
The only difference will be that script-following customer service reps giving you the runaround will be replaced by indefatigable chatbots giving you the runaround, which honestly sounds pretty hellish to me.
> Subscription cancellation buttons are intentionally confusing and hard to find.
really? I feel like Apple's App Store provides great UX with their warning emails and centralized subscription management view. It is well documented too: https://support.apple.com/en-us/HT202039
But I still get emails asking me to cancel subscriptions.
It's just ridiculous. The author even stated:
> Compare that to looking at a typical chat interface. The only clue we receive is that we should type characters into the textbox. The interface looks the same as a Google search box, a login form, and a credit card field.
And Google is one of the most used tools. Probably more used than pen and paper today.
Actually, as i recall, that statement is somewhat foundational for human computer interaction...of course my recollection is from my HCI college course a couple of decades ago...But, yeah, whenever i make something - digital or otherwise - i hope to design it properly enough that the intended user intuitively understands how it should be used. (I do add documentation beyond the base design, but because i want to further help the person.)
I don't know why it hadn'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.
The point and click adventure games from Sierra and Lucas Arts were a huge step forward in interaction, although you didn't have to use your imagination as much to solve the puzzles.
And here we are again asking users to type their way to success.
The other obvious UX comparison point is with sending instant messages to a person until they get it right....
Provided the responses aren't too brittle (and the LLM getting it wrong isn't too upsetting or find-out-too-late) lots of non power-users are going to prefer it, at least in cases where a menu or form input with about six options won't suffice.
Chatgpt or a successor dynamically generating in-game dialogue could be fun. Or maybe it won't be. I'm interested in seeing it done.
Chatbots are the future but your points are valid. They don't provide affordances, however chatbots provide a form of progressive disclosure and direct interaction that was previously impossible.
Toolbars and menus provide affordances but you still need to know what things are called and what order to use them. "I'd like to email this file as a PDF and I'd also like to print it." 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.
> “Chatbots are the future…”
I’m distressed by the growing use of chatbots in online payments. You do know the ontology—get balance information, make a payment, customer service.
I much prefer to not speak aloud to a robot on the phone, especially in the office, when there should only be three options.
Conversations with a robot speaking in a mixed tone of obsequiousness and superciliousness make me bananas.
> Text inputs have no affordances
It has all 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.
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.
Whether you want to call it an "empty set" or "the set of all possible sets" isn't really relevant to the authors point that an empty box has discoverability problems.
I regret posting. The article itself is a trap. As a developer, the chatbot interface is everything to everyone, as a developer. The application user of course needs a rich interface with affordances.
I am not arguing that the chat interface is everything to everyone from a design perspective. I do think that as UI workbenches improve the need for a dedicated application will nearly disappear though. Applications will become a 1-5 page specification, including the UI.
I thought this was a funny statement too. Text is the most powerful, original medium. It's difficult to overstate how powerful and valuable it is. Of course, text won't necessarily be the only UI exposed for interacting with these "Large Language Models." But that UI can be built on top of the text -- which wouldn't work in the other direction.
I think the point is that a "chatbot" is, by (the author's) definition, a UI with only a bare text prompt. Once you start building more UI on top of that, you're ... doing what the article suggests.
I see -- my apologies I must not have read it closely enough! :blush:
I re-read the post. The author is arguing against a statement that no one made and then uses the rest of the essay to outline their work. It is a bit of a setup.
An LLM-Chatbot is an extremely flexible tool, but I haven't seen anyone argue that all of our UIs and applications should be replaced by chat. That is ridiculous.
I dunno, I think his point is very valid. Chat bots can't do literally anything at all, and an empty text input isn't going to help guide a user towards what it can do, and what it's good at. Just because a system has a LLM to interact with it, doesn't mean it'll suddenly support any desired action the user wants done.
Yeah this was so weird to read. This is probably the number one property I think makes interfaces like ChatGPT compelling - you don't need to know how to use it - just use the human language you already know. If you don't understand something it says, just ask it to explain it. Essentially, it makes affordances obsolete.
> Text inputs have no affordances
>It has all the affordances.
these are two different ways of saying the same thing
But what if I don't know what I want?
A graphical UI can provide much more and much more intuitive guidance than a chat input ever will. And I say that as a big fan of Unix and the shell.
You can ask it. You can explain your problem and ask how it might be able to help. You can discuss and narrow down with some back and forth what it is you want to do.
I could tell a chat bot I am finding the horizontal split in my editor is annoying because I have a wide monitor, and have it tell me there's a setting for that and ask if I want the default changed.
With a gui I might have to go through the files menu for settings, check if it's in edit-preferences, check tools-options, before maybe having to find out online it's if it's in some settings file.
> You can discuss and narrow down with some back and forth what it is you want to do.
To be honest, I find systems that support only this kind of interface infuriating.
Imagine arriving in an unknown city: Being able to ask helpful locals for directions is nice, but sometimes I just want to look at a map.
Quality quote:
"There'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."
140 chars, he means.
*280 chars, he means.
*4000 chars, he means.
*10000 chars, he means.
*she
Someone said that using a chat-only interface is like using CLI tools. It'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.
isn't it ironic that there's no autocomplete on the tool that's literally autocomplete? haha
where's the fine-tuned prompt helper model?
We clearly need to ask chatgpt to write a model for the prompt helper model.
Both Bard and Bing chat have auto-complition
Hence the need for specially trained people who can do the back-and-forth with humans to understand exactly what is required, then enter it (into the brittle prompt in this case) in a way that the computer would understand.
Or more like a basic understanding of what the thing can do and what it can't do. It's not that complicated.
Not sure I'm convinced - natural language is one of the most intuitive interfaces we have, it's also how most instructions in professional contexts are delivered. What current ChatBots are missing right now isn't radiobuttons for different styles but context. No one who just reads my messages can know what kind of style I'm looking for until they either get laborious instructions, or, probably better, they see some examples.
I'm guessing that training custom language models on a company's data must be one of the hottest things you can be working on right now if you're looking for VC money (if there's something out there that compares to how well StableDiffusion+Dreambooth works for images, I'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 ("guess the verb") and using an unfamiliar command line interface. Good error messages can help.
Web search changed that. Most queries work, at least somewhat.
There'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).
For web search, there are very few error messages. If you enter a query that doesn't work very well, you get back results that aren't very good or what you wanted, so you try something else.
With AI chatbots, expectations are sky-high, but there are times when they should refuse with a good error message, because they really can't do what you'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]
Better error messages that help users understand what chatbots can actually do would help avoid misconceptions, but this won't happen unless the error messages are trained in.
[1] https://skybrian.substack.com/p/ai-chatbots-dont-know-why-th...
Yeah, exactly. A freeform input promises it can do 'anything at all' but if it can't in reality, it's always a frustrating interface.
Classic example, Siri. It's so easy to quickly find stuff you feel like it should be able to do, but it just can't. "When was my last message from Steve" etc.
Yeah, chat isn’t a universally great interface. But it’s a great default because it’s totally free form.
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.
Simplest way: output a JSON of simple control definitions with every answer.
Coolest way: Just have it generate a full-ass React front end or whatever on every message.
Interesting. Futuristic idea: on-the-fly generate a front-end customized to the user learned preferences. Motivation: humans are poor learners, so save them the trouble to learn a new interface style. At the limit, real-time generate a whole world-in-the-world, fulfill Zuckerberg's dream -- an utterly lonely Matrix. Now that's a thinly sugar-coated version of the Hell described in e.g. Catholic doctrine. And we will chose it ourselves. Deeply sobering.
We are very good learners. That’s how we got this far. Not all new interface is good and worth learning. Sometimes it feels best to stick with what works.
Contextually relevant suggested options (that can be acted upon with a single click), alongside the free form input box will emerge as the norm.
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:
https://www.latent.space/p/build-ai-ux
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.
Quite a few people use search by awkwardly typing on mobile one or two words, probably misspelled and/or auto-completed as they type it. The query isn't sophisticated, lacks a lot of context and parameters, which the search engine then tries to guess.
When you use ChatGPT in that way, you'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.
The idea that it's better than search for this specific normy behavior, I openly question. And let'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.
Just wait until we have a model that automatically translates vague, awkward prompts into something more useful.
Put differently: if Google’s search models already have the ability to return great results for poor queries, why couldn’t a large language model (or a plug-in for one) learn the same algorithm?
> When you use ChatGPT in that way, you'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.
Sidenote, i've found GPT useful enough to pay for (GPT Plus) by doing the opposite. Or rather, i find it very useful when i struggle to search for problems. ChatGPT helps guide me to new search or research terms, sometimes even providing the answer more directly.
It feels like the olden days where Google was great at finding a movie based on some vague movie description. GPT does that for a ton of things for me, enough that i found it useful.
It hasn't replaced online research but it has accelerated it for me.
What people forget is the underlying capability - LLMs are able to do reasoning.
So the one-track thinking of garbage-in-garbage-out is not the limitation any more.
What we're precisely now able to do is garbage-in-less-garbage out.
You can take a vague prompt in and ask GPT to hypothesize on what it means, why the user is asking that question and then generate a detailed prompt. Then use that that prompt and then perform the search.
Trying this out in the playground, I see a suprisingly capable search experience.
This kind of second-order (and higher-order) usage of LLMs is where things actually start to get much more interesting. The other thing you can do is just train a better model.
I use GPT-4 for debugging a lot now, because it's excellent at taking nothing other than an error message from the console and giving me back what's wrong and how to fix it. It's not perfect, but it's good enough that I reach for it by default now. I don't have API access to GPT-4 yet, and so I was comparing how well GPT-3.5 performed at this same task and for the example I tried, it just didn't get close enough for me to truly find it useful, so I wouldn't begin to rely on it in my daily workflow unlike GPT-4.
But... what I am actually quite interested in, and what I'm seeing a lot of, is exactly how far can you push a less capable model through prompt engineering? I think it's actually surprisingly further than you might have initially thought.
Have you tried it?
I just typed "eli5 hn vs redit" (misspelled reddit), and it understands perfectly.
Why even write a book when every insight in them can be shaken out of an LLM? You know, shorter content doesn't have to mean mind-melting right? Strip away the self-marketing poison found in every social media post and lets see.
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't settle for the old glass window over static text just to please the artisan book writer.
No, they're not, but my reasons for reaching that conclusion are somewhat different:
1) I don't think chatting with anything, human or machine, is a learning experience, particularly since the machine veracity is poor, untrustworthy, and Hinton's resignation today tells you everything you need to know about the narrative inside big research orgs right now.
2) Recognition vs. recall. Given that it's the equivalent of an informal language CLI, which I prefer by the way; but there is no recognition (as in symbols) only recall.
Long story short, I think the emergent need is for written communication, with a tip of the hat to Daniele Procida:
https://ubuntu.com/blog/engineering-transformation-through-d...
Except that what's missing is a human-computer collaboration, i.e. sensemaking with another tip of the hat to Peter Pirolli:
https://www.efsa.europa.eu/sites/default/files/event/180918-...
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't need it very often (besides searching for the occasional thing or giving feedback).
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.
At work the goal might be: "I want to sell $100,000 worth of widgets this quarter" and it will break down step by step how that might be possible.
For personal life it might be "I want to get involved in the kayaking community" and it will recommend activities, clubs, etc.
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.
This is the main scenario I see that leads to "Oops, we didn't see that one coming, but in retrospect that was NOT a good idea".
Can you expand on this?
When it's about the future, limitations of current implementations aren't a strong argument.
ChatGPT can be confidently stupid, but what if it gets better?
You need explanations/affordances of what it can and can't do only when its capabilities are limited. If it really could do whatever you asked, you wouldn'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't going to replace chatbots but only slightly extend them and really aren't needed if AI gets better over time.
Some neural connection to the brain that will interpret your thoughts is the only logical thing that will supersede chat ai, but that won'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's already probably happening based on past responses.
> I've convinced you that chatbots are a terrible interface for LLMs.
I was already convinced of this. What I'm not convinced of, and the article has little to say about, is
> Chatbots Are Not the Future... chatbots are not the future of interfaces.
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't feel that way. That's why I'm building InventAI, to help with that process: https://inventai.xyz
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.
> The interface looks the same as a Google search box
Indeed, and like any other text box, like whatsapp, like word — all tools that no one uses because they lack affordances.
Very very offtopic:
He called ChatGPT oracle, nice but not enough.
I want someone to name chatbot their `oracle of delphi` plz. thank you.
She*
*their chatbot
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't the future
I agree, but:
"Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting."
Comments like this confirm that you are indeed reading every comment on the website, which should be out of the realm of humanely possible.
It's all an illusion.
Or there's lots of people who click the flag link
For once, that everyone agrees scroll handlers are godawful but yet after years they're still omnipresent is germane to the article's topic.