5 Practical Applications Where ChatGPT Shines
5 Practical Applications Where ChatGPT Shines
Within reach for anyone—no fancy prompt engineering needed
I’m critical of people using ChatGPT for everything. And I’m also critical of people claiming you can use ChatGPT for everything.
One of the most popular articles on Medium last month was “20 Entertaining Uses of ChatGPT You Never Knew Were Possible.” It has 18K likes and 300+ comments. That’s a lot in the platform nowadays. I read it expecting high-quality ideas but got a list that included things like “dating help,” “dealing with loneliness and anxiety,” and “crime fighting.”
If I’m harsh with ChatGPT posts like that one is because they mix drops of truth with rivers of wild exaggerations and plain falsehoods just to amass popularity (or, worse, because they believe their claims).
Today, I won’t criticize ChatGPT but offer solutions. This is my version of what that blog post should have been. I’m going to combat misinformation by pointing to applications for which ChatGPT, flawed as it is, works very well—not more or less like other tools, but arguably better than anything else out there.
There are things not even magic can do
This article is in line with my previous attempts at defining reasonable boundaries for language models (LMs) use. No one has told us how they work or for which tasks they’re well-suited (they don’t know), so our only option is try-and-error. However, people are very bad at making correct inferences from examples.
More so if we’re dealing with magic.
If I found a magic wand, I wouldn’t be able to tell where its limits lie and would presuppose it’s capable of doing a lot of things it may not be able to—influenced by Harry Potter-like fantasy pop culture. That’s ChatGPT in the minds of most people. A magic wand. A piece of advanced alien tech. Something right out of sci-fi Hollywood.
But everything has limits (actually, if we get philosophical, things can only be defined through their limits). ChatGPT is no different.
It can’t do everything but can do some things. And like any other tool, it does some tasks better than others. If we accept these two premises, it follows that there must be a task (or set of tasks) in the applicability space that is the perfect target for ChatGPT (its suitability is optimal). And, in the case of the best—or among the best—tools in a given category (which I believe is true for ChatGPT), it’d be the undisputed best option for doing such tasks.
That’s the set of tasks for which ChatGPT is both perfectly-suited and the best option available. That’s what we’re looking for.
This is obvious for tools designed with a purpose (e.g. a calculator is great for arithmetic and it’s better than an abacus), but it’s tricky for things like web3—a solution in search of a problem. I don’t think ChatGPT belongs to the latter class, but it isn’t straightforward to correctly pinpoint its utility.
Finding those ChatGPT-ideal tasks is a matter of knowing where to look. The chatbot’s unique characteristics reveal hints to discover them. For instance, ChatGPT has a component of stochasticity, which means it’s unpredictable and unreliable. That’s bad if we’re doing a math problem but may be great if we’re writing the lyrics of a new song. Unpredictability goes well with creativity.
Instead of assuming ChatGPT is great for anything related to text, writing, or language, we should approach the problem of its applicability with that top-down perspective: Let’s understand the characteristics that best define the chatbot, and from there draw a line to those tasks where those traits aren’t bugs but features.
From there, it’s our responsibility to avoid the temptation to use ChatGPT for anything else just because other people are doing so.
This article is a compilation of five ChatGPT-ideal tasks that I’ve come up with following this procedure. They’re practical, concrete applications where ChatGPT not only works fine but stands out above other options.
1. Combining ideas like no human can
If there’s something that ChatGPT does amazingly—and significantly better than humans—is combining disparate ideas seamlessly into a homogenous whole. I’m talking about things like this:
This is arguably ChatGPT’s most outstanding skill. It’s not just a good tool for this, but simply the only one that can do it (together with other LMs). Not even humans—including those who are experts in the prompted concepts (e.g. bubble sort algorithms and gangster movies)—could do this with ease.
ChatGPT was trained on tons of text data so it’s expected it’d know about a vast array of topics—they’re represented somehow across its billions of parameters. To some degree, our brain works similarly. Memory recall isn’t that different from using prompts to make an LM output something specific (note that I’m not drawing a parallelism in structure or even function but in principle).
However, mixing any two or three topics seamlessly isn’t something the human brain can do by default. There’s an “alienness” to ChatGPT’s ability (as there is alienness in the way it makes errors). The way it combines people's writing styles, disparate themes, or ideas that have never been put together to create a unique piece with semantic cohesiveness is impressive. This embodies, for me, the creative idiosyncrasy of LMs. In contrast, humans—even the most skillful—would find it extremely hard to create such concoctions consistently. Our brain isn’t well-suited for this task whereas ChatGPT is.
This phenomenon is the result of how concepts are represented in ChatGPT’s latent space. Prompts allow us to access it in a way human memory is inaccessible (at least at will—who knows what we could get from a person’s mind if we could prompt it as we do with ChatGPT).
2. Modifying content you’ve created
Humans are great at exploring ideas in depth which is something ChatGPT struggles with. That’s why I’d never recommend it to write essays like this one, and either emails, tweets, or a book.
Yet I’d be willing to use it to translate between those formats. Within routine administrative-ish work, I find this ChatGPT’s sole strength.
It’s great with the form of language but horrible with coherence (not to mention its outputs aren’t innovative or engaging). It’s useful for editing things that already exist but quite bad at creating new ones that require a minimum degree of association with reality.
When I say “translate between formats” the first image that comes to mind is something like this, which I think it’s high in usefulness and safety:
If we extrapolate this idea to everything else, we find that ChatGPT is well-equipped to summarize articles in bullet points (I recommend doing this with those you’ve written), rewrite paragraphs Grammarly-style (but not write them from scratch), or customize content across platforms (e.g. take a Substack article and make it a LinkedIn post and then a Twitter thread).
The key here is perfect information. When you’ve written or created something, you know everything there’s to know about it. ChatGPT modifying the content in some way is inoffensive because as soon as an idea or concept doesn’t match your original intention, you’ll catch it. This also works well when you’re an expert on the topic (although in this case if you catch an error you’ll wonder if it was ChatGPT’s or the human author’s).
If you’ve read an article and want ChatGPT to summarize it, you may succeed, but the more you distance yourself from the content (or the knowledge), the harder it’ll be to catch its mistakes. Personally, I’d only consider doing it with things I’ve written myself (I’ve yet to try it).
3. Prompting AI art models
Have you noticed that my cover images have improved in quality lately? ChatGPT is much better than me at exploring the latent space of Midjourney. Not because it understands how to craft good prompts but because it’s good with patterns. This feature is shared by all pure deep learning models (at a fundamental level, it’s the only thing they do—pattern recognition), and prompt engineering is all about patterns.
To get a good Midjourney prompt from ChatGPT (GPT-3 works just as well when the other is “at capacity”…) I describe the task, add a couple of examples (I take image prompts from the Midjourney community feed), and then specify my desired topic. Here’s the prompt I’m currently using:
“Write a good prompt for an artificial intelligence system that creates images from text (Midjourney). The image is the cover for an article about [topic of your choice].
Here are three typical prompts:
1. "A vast sky with various glowing strings of memories shaping an enormous aurora borealis like glowing foliage, white, orange, extremely beautiful, beautiful universe twist, beautiful nightfall, cinematic lighting, beautiful field, high definition, high quality, hyper detail."
2. "Kneeling cat knight, portrait, finely detailed armor, intricate design, silver, silk, cinematic lighting, 4k."
3. "Ultra sharp award winning underwater nature photography of a woman riding a glistening gradient sea horse, backlit, depth of field, ocean floor, lush vegetation, particles, solar rays, coral, golden fishes, under water fashion photography, woman riding a seahorse, ultra sharp award winning photography."
You can use different words or concepts. Write just one prompt.”
Now let’s see some images. For illustrating purposes, I’ll use as topics the names of music albums I like (pretty much anything can substitute the “topic” placeholder):
Beauty Behind the Madness
“A chaotic swirl of colors and patterns, representing the madness, converging into a beautiful and intricate butterfly symbolizing the beauty behind it, dramatic lighting, high resolution, abstract expressionism.” --v 4
Back in Black
“A black and white stylized illustration of a guitar player in front of a wall of amplifiers, with smoke rising up behind them, as if they are playing a concert at an underground club. The player's outfit and guitar should be sleek and stylish, while the amplifiers are large and imposing, creating a sense of raw power. Use a combination of line drawings and shading to create a graphic novel-style image, with a pop of bright red on the guitar to add emphasis.” --v 4
Feel Something
"An abstract painting that captures the feeling of 'feeling something' with a mix of warm and cool colors, intense brushstrokes, and a mix of organic and geometric shapes. The focus is on the center, where the feeling is most intense, and the background is blurred and dreamy. High resolution, vibrant, and thought-provoking." --v 4
The possibilities for design are endless. You don’t have to rack your brain searching for prompts. If you’re in exploratory mode, ChatGPT can provide interesting ideas that you can always tweak afterward to align with your intent.
4. Assisting in creative work
ChatGPT is intrinsically unpredictable. You can’t know what the chatbot will output before it’s done. And, because it’s also unreliable (it isn’t designed to be truthful by default—only reinforced to be so), it means you can’t know when it’ll generate something crazy.
Harvard Ph.D. student Kareem Carr tweeted something that turns this argument on its head to see the positive side. I agree with him:
“Idea exploration” in this context means that it doesn’t have to be true (because it may not be tied to the laws that govern our world) and is boundless (i.e. anything goes).
Some examples that I’d be willing to use ChatGPT for (haven’t yet) include brainstorming article topics, headlines, and even outlines, and recommending things that belong to categories which truthfulness I can assess (e.g. a list of gifts for X amount of dollars or a list of books to read next given my history of recent ones).
Others are also finding ChatGPT useful for creative inspiration. Cartoonist Chaz Sutton used it to create funny comics that he then drew (another hit post on Medium this month). Guy Parsons, who explores AI art models on Twitter, used it to write the synopsis of a horror film and the descriptions of characters he then fed into Midjourney. And Jennifer Lepp, who writes e-books for Kindle on the “paranormal mystery subgenre,” is using it “for titles and plots.”
Another question is where to draw the line between right and wrong when it comes to letting ChatGPT do creative work. If it comes up with a catchy headline it’s okay to use it but writing a paragraph isn’t? It’s up to each of us to decide how we want to use AI—I do so conscientiously and on full disclosure (at least until we reach a point—if ever—where it’s assumed by default that ChatGPT takes part in the process, like Grammarly. Then I’ll decide whether I want to keep using it or not).
5. Understanding itself
In the same way that AI art models aren’t like cameras, ChatGPT isn’t like a calculator. One reason for this is the inherent “uninterpretability” of large deep learning models. There’s no manual for users to know precisely how to do what. With ChatGPT we can only gather behavioral insights through deductive empiricism.
Using ChatGPT for understanding itself is quite a unique application—I don’t know of any other consumer product that falls under this category (by necessity, not choice). The modus operandi is to run a set of experiments to test the model’s abilities and come up with a mental model of where it works fine and where it breaks (Gary Marcus and others are compiling a document with ChatGPT’s errors).
I haven’t been super rigorous in my testing, but I’ve found it to be a great source of inspiration to show you LMs’ limitations. I wrote a piece using GPT-3 only disclosing it at the end, I made GPT-3 and J1-Jumbo exchange outputs to see where the conversation would go without my intervention, and I co-wrote a piece with Lex to underscore its inability to come up with arguments aligned with mine.
The person who, as far as I know, has done more public work on this area is Riley Goodside who has explored GPT-3, ChatGPT, and more recently Claude, possibly more than anyone outside of OpenAI and Anthropic. He was one of the first people to find ways to bypass ChatGPT’s safety filters and come up with prompt injection methods.
Regarding future improvement on ChatGPT and other LMs, using them to understand how they work is their valuable application.
To wrap up, here’s a nice summary of this article in the form of a visual graph:
Right off the bat I have to comment and say that the art piece "Feeling something" exuded a very strong Kandinsky vibe. I loved that one for sure.
Other than that this is an excellent and helpful article.
I have not spent much time with either Chat Gpt 4 or Bard, but I have used both and what I enjoyed most spending time doing with them was acting like a guide to get it to "personalize" its answer to my prompt questions.
With the start of the new year I will be signed up with the paid Chat Gpt service and making prolific use of it as an intellectual stimulant among other things.
https://twitter.com/YirenLu/status/1620171019788120066