Practical applications that can make ChatGPT land and the development direction of ChatGPT

How to implement ChatGPT application, artificial intelligence is just the beginning, the future development direction of ChatGPT

ChatGPTDevelopment direction web3 The reason why web3 is difficult to implement is that it is difficult to find solutions for web3 application implementation. ChatGPT is different. Applications that rely on ChatGPT's powerful language model will be updated and iterated quickly and will eventually be used by users. This may be the reason why ChatGPT chatbots are so popular. ChatGPT is random, unpredictable and creative, which highlights the difference of ChatGPT.

ChatGPT development direction, combining ideas that humans cannot do

If there’s one thing that ChatGPT is amazing about — and clearly superior to humans — it’s how it seamlessly combines disparate ideas into a homogenous whole. I’m talking about things like:

This is arguably ChatGPT's most outstanding skill. Not only is it a great tool, it's the only tool (along with other LMs) that can do this. Even humans - including experts who are well-versed in prompt concepts (e.g., bubble sort algorithms and gangster movies) - can't do this easily.

Here's another example. I asked ChatGPT to write a Shakespearean sonnet, using odd-numbered lines for morning coffee and even-numbered lines for toucans, but in a way that felt cohesive and coherent (just for the sake of curiosity, Shakespeare died in 1616 and toucans were first taxonomically described in 1776, making this poem historically impossible). Here's the result after a few tries (even the rhyme scheme — ABAB CDCD EFEF GG — is nearly perfect):

ChatGPT was trained on a large amount of text data, so it is expected to understand a large number of topics - which are somehow represented in its billions of parameters. In a way, our brains work in a similar way. Memory recall is not that different from using a cue to make the LM output something specific (note that I am not drawing parallels in structure or even functionality, but in principle).

However, seamlessly blending any two or three topics is not something the human brain can do by default. ChatGPT’s ability is “heterogeneous” (in that there is heterogeneity in the ways it gets it wrong). The way it combines people’s writing styles, different topics, or ideas that have never been put together to create unique pieces with semantic cohesion is impressive. To me, this embodies the creative qualities of LM. In contrast, humans—even the most skilled—would find it difficult to consistently create such mixtures. Our brains are not well suited to this task, while ChatGPT is.

This phenomenon is a consequence of how concepts are represented in ChatGPT’s latent space. Prompting allows us to access it in ways that human memory cannot (at least at will — who knows what we could get out of a person’s mind if we could prompt it like we can with ChatGPT).

ChatGPT development direction, modify the content you create

Humans are good at exploring ideas in depth, and this is where ChatGPT struggles. That’s why I never recommend it for writing articles like this, whether it’s an email, a tweet, or a book.

But I am willing to use it to convert between these formats. In routine administrative work, I found the only advantage of this ChatGPT.

Its linguistic form is great, but its coherence is terrible (not to mention its output is neither innovative nor engaging). It's useful for editing things that already exist, but terrible for creating new things that require a minimal connection to reality.

When I say "conversion between formats", the first image that comes to mind is this one, which I think is very practical and safe:

If we extrapolate this idea to everything else, we see that ChatGPT is great for summarizing articles in bullet points (which I recommend doing for articles you write), rewriting paragraphs in grammar style (but not from scratch), or customizing content across platforms (e.g., making a Substack article into a LinkedIn post and then a Twitter thread).

The key here is perfect information. When you have written or created something, you know everything there is to know about it. It is harmless for ChatGPT to modify the content in some way, because you will catch it as soon as an idea or concept does not fit your original intention. This also works well when you are an expert on the subject (although in this case, if you find a mistake, you will doubt whether it is ChatGPT's or the human author's).

If you read an article and want ChatGPT to summarize it, you might succeed, but the further away you are from the content (or knowledge), the harder it will be to spot its errors. Personally, I would only consider doing this with something I wrote myself (I haven't tried it yet).

ChatGPT development direction, AI art model

Have you noticed that my cover images have improved in quality lately? ChatGPT is much better at exploring the latent space of Midjourney than I am. Not because it understands how to craft good prompts, but because it’s good at patterns. All pure deep learning models have this capability (fundamentally, it’s the only thing they do — pattern recognition), and prompt engineering is all about patterns.

To get good Midjourney prompts from ChatGPT (GPT-3 works equally well at another "full load"...) I describe the task, add a few examples (I get image prompts from the Midjourney community feed), and then specify the topic I want. This is the prompt I'm currently using:

“Write a good prompt for an AI system that creates images from text (Midjourney). The image is the cover of an article about [topic of your choice].

Here are three typical prompts:

1. "The vast sky, all kinds of glittering memories strung together into a huge northern lights, like glowing leaves, white, orange, extremely beautiful, beautiful cosmic distortion, beautiful nightfall, movie-like lighting, beautiful fields, HD, high quality, ultra-detailed."
2. "Kneeling cat knight, portrait, detailed armor, intricate design, silver, silk, cinematic lighting, 4k."
3.  "Super sharp award winning underwater nature photography, woman riding a seahorse with glitter gradient, backlighting, depth of field, sea bottom, lush vegetation, particles, sun rays, corals, goldfish, underwater fashion photography, woman riding a seahorse, super sharp award winning photography."

You can use different words or concepts. Just write one prompt. ”

Now let's look at some images. For illustration purposes, I'm going to use the name of a music album I like as the subject (almost anything can be substituted for the "subject" placeholder):

The Beauty Behind the Madness The Weeknd

“A chaotic swirl of crazy colors and patterns converges into a beautiful and complex butterfly, symbolizing the beauty behind it, dramatic lighting, high resolution, abstract expressionism.” — Section 4

Back in Black The Rolling Stones

“A black and white stylized illustration of a guitarist standing in front of a wall of amplifiers with smoke rising behind them, as if they were playing a concert in an underground club. The player’s attire and guitar should be sleek and stylish, while the amplifier is large and imposing, creating a sense of raw power. Use a combination of line drawing and shading to create a graphic novel style image, with a pop of bright red on the guitar to add emphasis.” — Section 4

Feel Jaymes Young

"This abstract painting captures the feeling of 'feeling something' through a mix of warm and cool colors, strong brushstrokes, and organic and geometric shapes. The focus is in the center, where the feeling is strongest, and the background is blurred and dreamy. High resolution, vibrant, and thought-provoking." — Section 4

The design possibilities are endless. You don’t have to rack your brain for prompts. If you’re in exploration mode, ChatGPT can provide interesting ideas that you can always tweak to fit your intentions.

ChatGPT development direction, assisting creation in the text field

ChatGPT is inherently unpredictable. You can't know what the chatbot will output until it's finished. And, because it's also unreliable (it's not designed to be truthy by default - it's just reinforced to be truthy), that means you don't know when it's going to produce something crazy.

Harvard PhD student Kareem Carr put a positive spin on the debate in a few tweets. I agree with him:

In this context, "idea exploration" means that it does not have to be true (as it may not be subject to the laws that govern our world) and is unlimited (i.e. anything can happen).

Some examples of where I’d like to use ChatGPT (not yet) include brainstorming article topics, titles, and even outlines, as well as recommending things that fall into categories I can assess for authenticity (like a list of gifts for X amount of dollars or a list of books to read next (given my recent history).

Others have also found ChatGPT useful for creative inspiration. Cartoonist Chaz Sutton uses it to create funny comics, which he then draws (another popular post on Medium this month). Guy Parsons, who explores AI art models on Twitter, uses it to write synopses and character descriptions for horror movies, which he then feeds into Midjourney. Jennifer Lepp, who writes an e-book for Kindle about the “paranormal suspense subgenre,” uses 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, then using it is fine, but writing a paragraph is not? It's up to each of us to decide how we want to use AI - and I will do so carefully and in full disclosure (at least until we reach a point - if ever - where ChatGPT is assumed by default to be involved in the process, like Grammarly. Then I'll decide if I want to continue using it).

ChatGPT development direction, further exploration of language models

Using ChatGPT to understand how it works is a key application that will indirectly affect all other applications - although I have to admit it is not that interesting for the average user.

Just like AI art models are not like cameras, ChatGPT is not like a calculator. One reason is the inherent “unexplainability” of large deep learning models. There is no manual that lets users know exactly what to do. With ChatGPT, we can only gather behavioral insights through deductive empiricism (we see examples of how it behaves and draw conclusions).

Using ChatGPT to learn about it is a pretty unique application in itself - I don't know of any other consumer product that falls into this category (out of necessity, not choice). The modus operandi is to run a set of experiments to test the capabilities of the model and come up with a mental model of where it works well and where it breaks (Gary Marcus and others are compiling documentation on ChatGPT's bugs).

My testing wasn’t terribly rigorous, but I found it to be a great source of inspiration to show you the limitations of LM. I wrote a post using GPT-3 and only disclosed it at the end, I had GPT-3 and J1-Jumbo swap outputs to see where the conversation would go without my intervention, and I co-wrote a post with Lex to highlight its inability to make arguments that agree with mine.

As far as I know, someone who has done more public work in this area is Riley Goodside, who has explored GPT-3, ChatGPT, and most recently Claude, perhaps more than anyone outside of OpenAI and Anthropic. He was one of the first outsiders to discover ways in which ChatGPT’s safety filters could fail, and proposed timely injections that, while potentially harmful, could provide companies with models with insights into fixing flaws that might otherwise go unnoticed. These types of tests are among the few we can use to better understand what’s going on inside generative AI models.