It’s been barely a few months since all the hype for generative AI started: since we first heard of the revolutionary ChatGPT, and the GPT-3 and GPT-3.5 models became a no. 1 topic all over the internet. Shortly after, the leaks about OpenAI’s Foundry followed. And before we even managed to wrap our heads around it all, OpenAI surprised us with another treat, introducing the model’s next generation — the GPT-4.

After relatively quiet releases of previous GPT models, this one comes with a blast, accompanied by various materials showcasing the new model’s capabilities. We aren’t even done with wondering about the exciting use possibilities of GPT-3 and GPT-3.5-turbo; yet, we can take another step forward and start thinking about even more exciting things we can build thanks to the broadened possibilities of GPT-4 integration.

So, what are those possibilities?

GPT-4 — “a new milestone in deep learning development”

As presented by OpenAI’s President and Co-Founder Greg Brockman in a YouTube GPT-4 Developer Livestream on the 14th of March, GPT-4 is a large multimodal model that shows greater accuracy and problem-solving & reasoning capabilities than previous GPT-3 and GPT-3.5. It also seems to understand the context and the meaning of users’ requests, as well as follow-up prompts (either questions, requests, or just comments to previous prompts within the same “conversation”) better than earlier GPT models — and hence, generate more accurate and trustworthy outputs. And it’s more creative!

Example of OpenAI GPT-4's creativity: GPT-4 explaining the plot of Cinderella in a sentence where each word begins with the next letter in the alphabet from A to Z
Source: OpenAI

What else? Even though GPT-4 (like GPT-3.5) was trained on data reaching back only to 2021, it’s actually able to overcome this limitation with a bit of the user’s help. If you provide it with information filling out the gap in its “education,” it’s able to combine it with the knowledge it already possesses and successfully process your request, generating a correct, logical output. 

Another thing that distinguishes GPT-4 from its predecessors is its steerability. It means this model is not limited to one specific tone of voice that would reflect in every output, no matter what you’d ask it to generate. In this case, you can prescribe the model’s “personality” — meaning give it directions (through the so-called “system message”) on the expected tone, style, and even way of reasoning. According to OpenAI, that’s something they’re still improving and working on, but the examples showcased by Greg Brockman in the GPT-4 Developer Livestream already looked pretty impressive.

Another huge advantage of GPT-4 is that the new model’s processing abilities are broadened to over 25,000 words — while previous language models could handle a maximum of only up to 3000 words (which, let’s be honest, was quite a significant downside). This allows it to process and generate much longer forms, such as long content pieces, extended conversations, broad documentation, etc. 

You’d think that’s already impressive enough, but hey, that’s still not all! The last but certainly not least difference in GPT-4, which significantly distinguishes it from its predecessors, is that it accepts images as inputs. And not just in a way it can describe what it “sees.” No, it can also classify and analyze what’s shown on the image and generate various outputs based on that. Give it a meme, and it will tell you why it’s funny. Give it a picture of the ingredients; it will tell you what you can cook with them. Show it your notes; it will draw reliable conclusions, provide analysis, or even write a piece of code based on them. Looks like GPT-4 is pretty capable, isn’t it?

All of the above opens up a really vast range of possibilities for utilizing GPT-4. Let’s then go through some examples of GPT-4-powered artificial intelligence tools!

a screenshot presenting how GPT-4 can analyze a meme and explain to the user what's funny about it
Source: OpenAI

UPDATE 18.06.2024

GPT-4-turbo, the latest version in the GPT-4 family of language models was trained on data up to December 2023 and has an impressive context window of up to 128,000 tokens. It makes it perfect for various tasks requiring processing more data, translating long texts to other languages, analyzing lenghtly articles, etc.

Moreover, on May 13th, OpenAI announced a new model — GPT-4o, with new capabilities reaching beyond its predecessors. GPT-4o is a large multimodal model, meaning it can process (understand and generate) text, image, AND (what’s probably the most exciting here) voice. The voice mode allows you to choose a voice the chat will use to answer questions, making the experience even more entertaining. Funnily enough, one of the options became an object of a little scandal, as it sounds eerily similar to Scarlet Johanson. The problem is, she refused Sam Altman’s request to become ChatGPT’s voice, but somehow, one of the assistants still sounds just like her. OpenAI claims it’s not actually HER voice, but it may be hard to accept and let go when her own family cannot hear the difference.

Same as in the case of previous models, OpenAI doesn’t share information about the new model size, but as it is the first gpt model with “omni” (hence “o” in its name) capabilities, we can only assume it’s probably even bigger than GPT-3.5 and GPT-4 models.

Read more: GPT-4 vs. GPT-3. OpenAI Models’ Comparison

GPT-4 can produce factual responses better than previous GPT models (GPT-3 and GPT-3.5)

Real-life GPT-4 examples: the model’s early adopters

Even though GPT-4 was announced less than two weeks ago, some were lucky enough to get early access to its exciting capabilities — so we can already observe it in action in some of the early adopter projects:

GPT-4 in Duolingo

This well-known language learning app uses the model in its brand new subscription variant (announced the same day as the release of GPT-4), Duolingo Max. The plan introduces two major features (Explain My Answer and Roleplay) that bring the in-app learning experience to a whole new level. 

The first one, Explain My Answer, puts an end to the frustration of not understanding why one’s answer was marked as incorrect. The feature allows you to chat with Duolingo’s mascot, Duo — which, thanks to GPT-4, now becomes an interactive virtual tutor — and get clarification on why our answer was wrong, along with additional examples on the topic so that we can understand it better.

The Roleplay feature, in turn, allows users to practice their language skills in a real conversation. Well, it is as real as chatting with an artificial intelligence model can get — but we already know it can get pretty real. The talks never repeat, allowing for a more realistic and effective learning experience that mirrors real-life communication scenarios.

GPT models now help people learn other languages; creators of Duolingo utilize OpenAI's new model in their app.

GPT-4 in Be My Eyes

It’s a Danish mobile app that strives to assist blind and visually impaired people in recognizing objects and managing everyday situations. The app allows users to connect with volunteers via live chat and share photos or videos to get help in situations they find difficult to handle due to their disability.

Until now, in-app assistance has been provided by an online community of human volunteers. Now, GPT-4 has joined the ranks of those volunteers! 

a screenshot of a tweet by Be My Eyes about how they utilize GPT-4 in their app
Source: Twitter

GPT-4 in Khan Academy

This non-profit educational organization’s mission is “to provide a free, world-class education to anyone, anywhere.” Khan Academy provides a wide range of tools and materials covering a variety of subjects, from math and finance to history and art, to bring equal access to education for people all over the world. Now, they’re making it even more accessible by adapting GPT-4 (in a small AI pilot, for now) to their learning environment! 

“It [artificial intelligence] can guide students as they progress through courses and ask them questions like a tutor would. AI can assist teachers with administrative tasks, which saves them valuable time so they can focus on what’s most important — their students.” — writes Sal Khan, founder and CEO of Khan Academy, in his article about the GPT-4-based innovations.

GPT-4 in Morgan Stanley

It’s an American multinational financial services and investment management company supporting people, institutions, and governments in raising, managing, and distributing the capital they need to achieve their goals. Morgan Stanley joined the ranks of GPT-4 early adopters and utilized the model to manage its vast knowledge base. As we can read in Finextra’s article:

“Morgan Stanley Wealth Management will use the technology to access, process and synthesize content to assimilate its own range of intellectual capital in the form of insights into companies, sectors, asset classes, capital markets, and regions around the world.”

GPT-4 in Stripe

Another GPT-4’s early adopter is Stripe, a financial services, and SaaS company that created a payment processing platform supporting building websites and apps that accept payments and send payouts globally. Stripe uses the model to make documentation within their Stripe Docs tool more accessible to developers. With GPT-4 integration, developers can ask questions within the tool using natural language and instantly get summaries of relevant parts of the documentation or extracts of specific pieces of information. This way, they can focus on building the projects they work on instead of wasting energy reading through lengthy documentation.

GPT-4 and the Government of Iceland

This one is slightly different from the above examples, as it’s not about an app or a tool utilizing GPT-4. In this case, we’re talking about the cooperation — between OpenAI, the Government of Iceland, the Citizens Foundation, and an Icelandic language technology company, Miðeind ehf — to preserve and popularize the native Icelandic language by training GPT-4 on proper Icelandic grammar and cultural knowledge. 

Despite the new model’s broadened capabilities, initially, it showed significant shortcomings in understanding and generating materials in Icelandic. Which, as you can guess, made it unreliable to use in this language. To change that, Miðeind ehf assembled a team of 40 volunteers on a mission to train GPT-4 on proper Icelandic grammar and cultural knowledge.

a screenshot of a tweet by Citizens Foundation (Iceland) about how they utilize GPT-4 to preserve native Icelandic language
Source: Twitter

Wondering if GPT-4 is safe for your use case? You can read all about its safety implications here: A quick guide to GPT & ChatGPT security

A step further: potential use cases of GPT-4

It’s no longer a matter of a distinct future to say that new technologies can entirely change the ways we do things. With GPT-4, it can happen any minute — well, it actually IS happening as we speak. And it’s not only about changing how we work. This transformation can, and most likely will, affect many various aspects of our lives. 

Thanks to its abilities — and let’s keep in mind this is just the beginning, and we have already jumped from “generating human-like text” to way more impressive skills — GPT-4 can unimaginably boost the effectiveness and even capabilities of both individuals and organizations. It can help us get things done faster, plan better, be more efficient (while possibly less tired), come up with more (and better) ideas, find more accurate solutions (and do it faster), get a better understanding of complex matters, and, as the example of the Be My Eyes app clearly shows, even make the world a better place for people with disabilities.

We used to think that the internet and search engines like Google were the biggest revolution in the accessibility of information. And sure, it was — at one point. But that’s not the pinnacle of our possibilities anymore. Now, we have moved another step forward, where researching things can be even easier, faster, and more effortless because we not only have access to the immensity of information, but we also have the technology that can analyze all this information for us and bring it to our screens on a silver plate, in an easily digestible, natural language.

Inspiring, isn’t it? But I feel like the above use case examples, although already impressive, still don’t draw the whole picture of what you can achieve with GPT-4. And it makes sense. They’re early adopters projects, so it’s all new and probably not yet as developed as it could be. Let’s then broaden this perspective by discussing a few more — this time potential, yet realistic — use cases of the new GPT-4.

Read also: Open-source vs. OpenAI. 8 best open-source alternatives to GPT

GPT-4 can be used to create a highly advanced virtual fitness assistant that could help people plan their week including training, suggest a diet and analyze photos of their posture

GPT-4-powered personal training assistant

I assume we’re all familiar with recommendation engines — popular in various industries, including fitness apps. Now imagine taking this to a whole new level and having an interactive virtual trainer or training assistant, whatever we call it, whose recommendations could go way beyond what we knew before. “Someone” who would give you their undivided attention, carefully analyze all known aspects of your health, physical condition, daily routine, training data, and eating habits, and provide you with high-level assistance in incorporating healthy habits and physical activity in your everyday life. In a way that really suits you.

Imagine such an assistant could:

  • Analyze photos of your body figure and suggest what exercises you need to achieve your goals. Then, adjust those suggestions over time as you update them with your progress, changes in your body, or any other new information.
  • Base those training recommendations on your physical capabilities and limitations (including historical training data), health condition, and even medical recommendations (that you could just upload to the system).
  • Perfectly understand your dietary requirements, including your allergies and recommendations from your dietician or a doctor (which, again, you could just scan and upload), as well as what you like and dislike — and, based on that, provide you with meals ideas that are perfectly adjusted to you and your needs. Never the same (unless you want it), so you can have a healthy and varied diet even if you lack creativity in this matter and/or have a long list of products you don’t eat.
  • Analyze your weekly schedule — meaning all your plans and duties — and help weave your training sessions into it so you don’t have to struggle to find time for everything.

How does it sound? It could be a perfect solution for both those who love being active and eating healthy but could use some assistance in organizing their routine, as well as those who struggle with their diet or to motivate themselves to exercise regularly.

natural language processing is the fundamental ability of all generative AI models

GPT-4 as a high-level analysis assistant

Considering GPT -4’s advanced analytical skills, a pretty natural conclusion is that it could provide invaluable support in data analysis. Especially that, thanks to its ability to accept images as inputs, it can analyze all sorts of queries, from text to tables and graphs and everything in between.

With these abilities, GPT-4 could serve as a highly efficient assistant for human analysts, helping them work through complex information, drawing even more valuable conclusions (in quality and quantity), and generally optimizing their work. On the other hand, it could support teams that lack dedicated analysts, where domain experts may not have sufficient analytics experience but still need to rely on data and make data-driven decisions.

In this form, GPT-4 could also be a game-changer for education, especially for aspiring data analysts. Imagine a tool allowing students to check their reasoning and conclusions and even discuss any uncertainties they may have with the model. This way, they would be able to quickly identify errors in their approach, avoid mistakes that could interfere with their learning process, and, hence, learn faster.

Building a generative AI platform in practice

Learn from the hands-on experience of Caju AI!

Read the interview

GPT-4 in marketing — engagement prediction tool

When it comes to GPT -4’s possibilities in the marketing area, the easiest thing to say is it can do everything previous models could — AND more. But besides bringing significant improvements to the applications I described in my previous article about GPT-3 use case ideas, thanks to its broadened capabilities, GPT-4 can be utilized for many more purposes. And again, it’s mostly thanks to its new multimodal capabilities.

Considering what GPT-4 is capable of, together with the AI Team, we came up with an idea for a GPT-4-powered tool that could analyze photos, pictures, etc., and predict their potential for generating high engagement in social media.

Imagine a fashion brand aiming to attract more people from its target audience and generate more buzz around the brand. They could use GPT-4 to analyze trending content on social media and the materials they prepared for a new campaign to assess which photos or images have the highest potential to attract the most attention — or to get suggestions on what should be improved to achieve that goal. Thanks to GPT -4’s steerability, users of such a tool could precisely determine the perspective in which the model should analyze the images and hence receive highly accurate recommendations.

It could be an excellent tool for helping businesses and individuals broaden their ability to reach desired target audiences and boost engagement — powering up their marketing efforts.

Read also: Llama 2 vs. GPT-3.5 vs. GPT-4 model comparison

today, large scale models such as GPT-4, can process visual inputs

GPT-4 as a career advisor

What about an app that could analyze the user’s whole education and career history, interests, and even their successes and failures, and suggest what steps they should take to advance or — in case that’s what the user is looking for — advise on new career paths they could embrace. 

Imagine someone is completely burned out and feels like they don’t want to continue in their current profession, but they struggle to determine what else they could do. There are so many people facing this problem. In most cases, they are well aware of their skills and interests; they also know what tasks come smoothly for them and which are just a source of frustration — but they don’t know what all those skills, interests, and preferences could be suitable for. So they need some guidance.

Such an app could provide this much-needed guidance, suggest what professions might be aligned with one’s skills and interests, and even brainstorm those options with the user. And once there’s some conclusion on what might be the best direction, the app could advise the user on what courses they should take, what they should learn, and what skills they should polish to succeed on their new career path. 

GPT-4 on the guard of mental health

GPT-4 in mental health support

Considering that GPT-4 already has incredible abilities that will most likely get broader and broader with time, why not go a step further and consider it as a technology that could help us spot red flags in people’s online activity — behaviors that can suggest that someone is struggling with mental illness or have suicidal thoughts?

Of course, the form of such a monitoring tool is a complex matter that would require analyzing all the ethical aspects and creating a whole, well-thought-through system around it. But let’s leave this aspect aside for a moment, and let’s assume there is general social consent to introduce a system that would track our online activity for alarming, mental-illness-related behaviors and which — in case certain conditions are met — would be passing such information to a dedicated mental health organization, that would be reaching out and offering help to “flagged” users. Such a system could help us start noticing signs that used to pass unnoticeably before. Signs that, in many tragic cases, became “visible” to friends and family only when it was already too late.

Or, to make this idea more realistic, it could be an app that one can install on their phone when they kind of feel that something is not right but are not ready to ask for help just yet. Such an app could help them track their mood, plus it would monitor their online activity and many other things — even the music the user listens to. Then, the app would analyze collected data and alert the users themselves if the conclusions imply there are reasons to believe this person requires at least professional assessment.

The app could be interactive or include a chat feature, so the users could always talk to the virtual assistant and, for example, ask questions about therapy or psychiatric treatment. Or any other questions they might be ashamed of asking anywhere else in fear of “revealing” their mental issues. The superior goal of such a GPT-4 powered assistant would be to familiarize the users with the concept of therapy and psychiatric treatment and help them start feeling more comfortable with the idea of using them.

Imagine that all these people who are afraid to speak up about their struggles or reach out for help could take this “half-step” in the right direction — a step that could eventually lead them to start proper treatment and maybe even save their lives. Isn’t it a beautiful vision?

Use cases of GPT-4

Use cases of GPT-4 — conclusions

It’s incredible to witness generative AI developing so quickly. Within a few months, we went from being impressed that large language models can generate human-like text to GPT-4 standing on par with human volunteers supporting visually impaired people.

At this point, nobody doubts that this technology can revolutionize the world — probably in a similar way that the introduction of the Internet did years ago. Or even faster, as the competitive landscape of the AI industry results in exciting advancements being announced nearly every month.

Of course, we should keep in mind that, even though more and more impressive, generative AI models are still imperfect and prone to hallucinations — and even their creators carefully underline this fact. But considering all the capabilities OpenAI’s models already have, imagine how beautifully all the GPT-powered solutions will grow as the models are further developed!

Got an idea for a GPT-4-based app?

We’re here to help you bring it to life. Complete the form below, and let’s talk about GPT-4 development!