ChatGPT has dominated global conversations for weeks on end and there’s no shortage of uses, funny gags, and - depending on your disposition - a little bit of dystopian horror. The fun conversational language bot wears many hats, from assistant writer or coding tutor to basic party planner, and with the new GPT-4 version we’re bound to see a plethora of new uses pop up as people get to know OpenAI's new darling.
The model's debut showcases not just its new capabilities but all the new software and projects it’s already powering, like being a virtual assistant to the visually impaired and boosting your conversational skills in Duolingo. Just from those two examples - including the incredibly cool way Iceland is getting involved - we already get the sense that GPT-4 is not here to play around and is definitely going to step things up compared to its predecessor.
As a quick refresher, ChatGPT is a language-learning bot that draws from a vast and deep pool of data to answer queries and come up with responses to prompts. With the release of the GPT-4-powered version (available for Plus subscribers first), we can’t wait to see what cool and quirky things we can use the new-and-improved bot for - but for now, let’s have a look at the key differences between the two versions, and what the new beast brings to the table.
The Gift of Sight
The main limitation of ChatGPT 3.5 was the fact that you could only input and receive plain text. It could read and write, but that’s about it. The chatbot could only deal with text and code, and you couldn’t get it to generate images.
GPT-4, however, can be given images or sounds and process them to pick out relevant information, which is why it’s super useful as a virtual assistant when paired with Be My Eyes, an app that allows visually impaired or blind people to interact with volunteers and describe what they’re seeing. The GPT-4 demo video shows how people can now have the app describe what's in their fridge, help differentiate between clothing items, and even read maps.
The possibilities are endless with this new facility, you could have GPT-4 dissect a piece of artwork, help writers describe and set different scenes, or identify things like plant types, animal breeds, or landmarks. This alone elevates GPT-4 above its predecessor.
GPT-4 is a polyglot
For those who haven’t come across this term before, a polyglot is a person - we’re using that word loosely in this context - that knows and uses multiple languages, and GPT-4 qualifies for this category. If you’ve been following the ChatGPT hype for a while you’ll notice almost every screenshot, tutorial and breakdown is in English. Much of the data, testing, and research is conducted in English, which isolates non-English speakers as the capabilities of the AI bot transcend language.
Now, GPT-4 has response capabilities across 26 languages, including Korean and Italian. It’s best around the Romance and Germanic languages but still performs well in others. The bot is able to answer multiple-choice questions with pretty high accuracy across these languages, and while most people don’t talk in multiple-choice questions (maybe you do), it does open the door to being a lot more accessible across language barriers, which again is a step up from GPT 3.5.
Remember how we used to reminisce
Language models like this are trained on millions of books, webpages, and a whole load of text data which it’ll draw upon when conversing with users. When you ask ChatGPT to summarize a novel or a film, it ‘goes back’ to its database and pulls relevant information to that query, and presents it to you in a conversational way.
As you talk to it, there’s a limit to how much ChatGPT can remember and ‘keep in mind’ from previous conversations (something I find very relatable). With ChatGPT 3.5 the limit was around 8,000 words, which means once your conversation exceeds that limit it may ‘forget’ bits from older chats and need reminding in some way. It’ll lose track of things when it goes too far back.
But with GPT-4, that maximum memory bank word count is now 64,000 words, eight times more than before. This equates to about 128 pages of single-spaced text, which is a lot! This basically means that GPT-4 will be able to keep just under 130 pages ‘in mind’ during interactions. So if you talked about something 30 pages back and it may be relevant to the current conversation, it’ll remember that.
This is a very simplified explanation, but it should give you a general idea of how the attention mechanism has been massively boosted between the versions. GPT-4 will be capable of adeptly carrying much longer conversations than its predecessor.
It has a great personality
The capacity to change or alter an AI’s behavior on command is called ‘steerability’, and that concept had clearly been implemented on a small scale in GPT-3.5. We’ve previously talked about the ways you could improve your interactions with it and noted that giving the bot a set personality or ‘role’ in conversations could help improve results. For example, if you’re a student looking to improve your essays and want ChatGPT to mark it - definitely the only thing students are using the bot for, surely - you’ll get better results by giving it the role of a teacher or examiner. Having a set perception with which to look at your text allows ChatGPT to be more thorough.
ChatGPT definitely doesn’t lack personality - in fact, in some cases one could argue it may have a little too much at times. According to OpenAI, GPT-4 will be able to have its tone and style changed “within bounds”.
So, the possibility of creating your perfect digital best friend with their own unique personality isn’t so far away. You can customize your ChatGPT bot to fit your vibe, be that a pirate, a feeble Victorian child, or an angry caveman.
Fool me once
There’s no shortage of jailbreaks and hissy fits and just general goofiness that has come out of ChatGPT recently. There’s been a handful of complaints that the bot safeguarding is a little too ‘woke’, mostly by an angry and rather loud billionaire but for the most part a lot of the oddities have sprouted from people trying to break the bot. A little mean, but valid I suppose.
GPT-4 has been trained on the malicious prompts helpfully provided by users over the years and is now better able to refuse to go outside of set guidelines in an effort to create a safe experience for all users. We experienced this a little a few days before GPT-4 launched when messing around with the bot, asking it to summarise horror novels. When the bot received certain novel titles to break down, it refused to go into detail about the plot but instead discussed thematic reasoning behind specific plot points, or did not provide a response altogether. So, hopefully, this means fewer inaccuracies and tantrums.
New technologies on a scale like this, with such a strong thread of personality and intelligence, can be scary, as new things often are. While we are obligated as contemporary, online communities to be responsible and reasonable with things like ChatGPT and its new souped-up capabilities, there’s no reason we cannot enjoy this breakthrough in human intention.
Seeing how it can be used to preserve a language, assist disabled users, and improve the way we learn and grow is definitely impressive and something to be commended. However, we do have to look at these advancements with some kind of scrutiny: people will undoubtedly use these updated features for the wrong reasons, and there’s only so much safeguarding OpenAI can do. If ChatGPT can pump out AI-generated pictures and videos for example, we have to consider how people are already making deepfakes and revenge porn of the people in their lives (and celebrities, of course). Cheaters, misogynists, racists, and scammers have already made a home with AI-generated text, so with these advancements, we have to question ourselves and if there will ever truly be ‘ethical AI’.
All in all, we are cautiously optimistic and excited to get a hold of GPT-4, grab a coffee, and have a nice conversation.
Sign up to receive daily breaking news, reviews, opinion, analysis, deals and more from the world of tech.
Muskaan is TechRadar’s UK-based Computing writer. She has always been a passionate writer and has had her creative work published in several literary journals and magazines. Her debut into the writing world was a poem published in The Times of Zambia, on the subject of sunflowers and the insignificance of human existence in comparison.
Growing up in Zambia, Muskaan was fascinated with technology, especially computers, and she's joined TechRadar to write about the latest GPUs, laptops and recently anything AI related. If you've got questions, moral concerns or just an interest in anything ChatGPT or general AI, you're in the right place.
Muskaan also somehow managed to install a game on her work MacBook's Touch Bar, without the IT department finding out (yet).