OpenAI’s latest language model, GPT-4, outperforms ChatGPT’s GPT-3 in every way. With expanded memory and creativity, 26 languages, image analysis, academic excellence, multiple alter egos and increased safety; GPT-4 is nothing short of brilliant – but not without limitations
March 18, 2023
If you’ve got whiplash, you’re not alone. Not in the scheming insurance advert kind of way, in the “OpenAI just dropped GPT-4 and I was still working through the existential crisis triggered by ChatGPT” kind of way.
On Tuesday this week, OpenAI released the much-anticipated upgrade of the current AI language model which their eloquent human-like chatbot, ChatGPT, runs on: GPT-4, and in doing so issued the world with its latest dose of “future shock” – echoing Alvin Toffler’s watershed book by the same name.
Although OpenAI has stated that, in casual conversation, the difference between ChatGPT’s GPT-3 and the new GPT-4 language model, is subtle, there are in fact a wealth of marked differences between the two – all of which are equally as terrifying as they are exciting.
OpenAI describes GPT-4 as “the latest milestone in OpenAI’s effort in scaling up deep learning.”
“While less capable than humans in many real-world scenarios, [GPT-4] exhibits human-level performance on various professional and academic benchmarks,” states OpenAI.
When reading through GPT-4’s many sophisticated “state-of-the-art” capabilities, I couldn’t help but be reminded of the plots of many blockbuster movies; parallels which I think can help clarify what GPT-4 is capable of.
So, what can GPT-4 do? Everything. I’m not kidding.
To clarify, throughout this article, when writing “ChatGPT-3,” I am referring to the current open access version of ChatGPT which runs on OpenAI’s GPT-3 language model; the predecessor to GPT-4.
Her: GPT-4 is an even better creative writing companion
ChatGPT-3’s ability to generate eloquent, human-like prose, poems, short stories and articles – in mere seconds – triggered widespread existential panic amongst members of the writing community.
It’s no surprise, therefore, that given GPT-4’s writing ability exceeds that of GPT-3, OpenAI is pitching the fourth-generation language model as an advanced writing companion rather than a challenger.
Think of GPT-4 just like Joaquin Phoenix’s pocket AI writing buddy in “Her.”
Learning a user’s writing style, composing songs, writing screenplays and providing detailed explanations or clarifications are just a soupçon of GPT-4’s advanced capabilities, all of which it’s better at than ChatGPT-3.
“GPT-4 is more creative and collaborative than ever before,” states OpenAI, “it can generate, edit, and iterate with users on creative and technical writing tasks.”
Memento: GPT-4 can probably remember what you said 15 minutes ago
One of ChatGPT-3’s USPs is its ability to hold up a dynamic conversation; not only can it remember what a user has previously said, but it can draw upon this and expand on it.
But just like “Memento’s” main character, Leonard, who has a 15-minute memory span, there is a relatively short window of conversational history within which ChatGPT-3 can actually remember what a user has said. This window is only around 3000 words back, or six pages of a book.
GPT-4 however, has overcome this memory-limit, and can now handle as many as 25,000 words.
Not only will GPT-4 therefore remember more of what you said further back in the conversation, but it will also be able to give longer responses and navigate analysis of large quantities of text in documents.
The Sixth Sense: GPT-4 sees things in images you might not
You see half a carton of eggs, flour, butter and milk and think – ingredients? Maybe pancakes?
GPT-4 “sees” the same ingredients and “thinks” frittata, french toast or quiche.
Until now, ChatGPT has been limited to text input only. But now, with the upskilled multimodal GPT-4 version, you can quite literally take a picture of the contents of your fridge, input it as a visual prompt and GPT-4 can make handy recipe suggestions.
OK, it’s not “I see dead people” (hopefully), but GPT-4’s ability to see, understand and analyse visual inputs is impressive.
GPT-4 is even capable of taking a rough napkin sketch of a website and turning it into the real thing!
Good Will Hunting: GPT-4 can beat most people at graduate-level exams
ChatGPT-3 scored in the 10th and 31st percentiles of the Uniform Bar Exam and Biology Olympiad, respectively.
GPT-4, on the other hand, staggeringly outperformed ChatGPT-3 as well as the vast majority of test-takers for both of these famously gruelling examinations, scoring higher than 90% of lawyers in the bar and higher than 99% of those taking the Olympiad.
The chatbot which was once a bottom-of-the-pile student (academically speaking), has overnight turned into an MIT-worthy AI prodigy.
Lost in Translation: GPT-4 can speak 26 different languages
It’s true that computational coding languages are all based on English, but what good is a conversational AI model if it can only speak one language? Monolingualism enormously limits a chatbot’s potential across a large proportion of the world.
GPT-4 is multi-lingual to the max: it can speak 26 different languages such as Korean, Italian, Ukrainian and German, providing near-global access to its sophisticated capabilities and helping users overcome the many miscommunications associated with “translationese” (when overly-literal translation yields an output with little to no clarity or meaning).
The Icelandic government is even making use of GPT-4’s linguistic capabilities, deploying the chatbot to preserve the nation’s language, culture and history, which in a digital age of rapid globalisation is at risk of being diluted or lost altogether.
Split: GPT-4 has multiple personalities
By priming GPT-4 with a simple “pretend you’re a…” prompt, users will now be able to augment the personality of the model with a feature known as “steerability.”
ChatGPT-3 is certainly also capable of changing behaviour on demand, but GPT-4 takes malleability to a whole new level: its character, behaviour, tone, voice and style are all completely customisable.
Though exceedingly cool (and useful) this shapeshifting ability does of course also bring due risk.
Much like some of James McAvoy’s alter egos in the terrifying psychological thriller, “Split,” are harmless, there are also those, buried deeper within his psyche, which are malicious in nature and only uncovered when triggered. GPT-4 works in much the same way.
Just as GPT-4 can easily take on the persona of a comforting counsellor, more hostile impersonations can also be let loose from their chatbot cage if prompted to do so with “jailbreak” prompts.
OpenAI is however fully aware of GPT-4’s propensity to cause harm in this way and has made a concerted effort to suppress the dark side of the model as much as possible, amping up the guardrails significantly.
Terminator 2: GPT-4 is a safer version
Now I’m not saying that future Sam Altman sent a reprogrammed version of a language model back in time to the present with the mission of “protect Earth’s population” (like John Connor did with Arnold Schwarzenegger’s T-800 terminator in “Terminator 2: Judgement Day”), but the result is realistically the same: GPT-4 is a much more equipped to protect users from exposure to harmful content, misinformation and misrepresentation than ChatGPT-3 is.
GPT-4 actually finished training back in August last year, but the team at OpenAI spent six months tinkering with its code and exposing the model to examples of adversarial prompts to make it safer and more aligned.
As a result of this alignment period, GPT-4 is now 82% less likely to respond to malicious, inappropriate or “jailbreak” requests and 40% more likely to produce truthful, factual responses than ChatGPT-3 is.
A Beautiful Mind: Though brilliant, GPT-4 suffers from hallucinations
Though GPT-4 is clearly advanced, as we’ve established, it’s also not without its limitations. One of these limitations, which is perhaps most concerning, is its tendency to “hallucinate.”
In AI terms, the observed phenomenon of “hallucination” refers to the model’s ability to generate a confident, yet unfounded and wholly inaccurate response that does not appear to be based on its training data.
In essence, GPT-4 is capable of dreaming up an idea as a figment of its imagination and endorsing it to users as a well-known fact.
I suppose given that a neural network is an artificial representation of the human brain, it’s also vulnerable to the same manifestations of pathogenesis – just in silico.
2001: A Space Odyssey: GPT-4 is still artificial at the end of the day
OpenAI has stated that the upgraded GPT-4 model is able to express logic and “surpasses ChatGPT in its advanced reasoning capabilities.” However, along with hallucinations and “bad AI” impersonations, GPT-4 is also still vulnerable to making reasoning errors.
For this reason, OpenAI suggests GPT-4 should always be used with due caution and not in high-stakes scenarios, with human review underpinning all use cases.
Much like the crew of the fictional “Discovery One” spaceship experienced with the coding contradictions in their onboard AI known as “HAL” in “2001: A Space Odyssey,” GPT-4 also has the potential to mislead users and make judgment errors that cause harm.
Everything Everywhere All at Once: GPT-4 is already everywhere
In reality, we could have simply saved the time spent writing and reading this article, because the answer to the question: “What is GPT-4 capable of?” is quite simply, just like the movie: Everything, everywhere, all at once.
OpenAI is already proving GPT-4’s ubiquitous value in integrating the model into the platforms of a diverse panel of companies and organizations.
In the finance sector, both Stripe and Morgan Stanley have deployed GPT-4. The former uses GPT-4 to make its platform more user-friendly and combat fraud by flagging suspicious account activity, with the latter using the model to better navigate its extensive databases and expert knowledge bases.
Khan Academy is using GPT-4 to power a learning assistant and “virtual tutor,” as is Duolingo, incorporating GPT-4’s vast language abilities into its platform to provide an AI conversational partner for language practice, as well as a feature called “Explain My Answer,” providing learners with guidance on where they went wrong when making a mistake.
Be My Eyes is a Danish startup which is utilizing GPT-4’s image analysis capabilities to develop a “Virtual Volunteer” and “transform visual accessibility,” helping blind or low-vision individuals navigate daily life easier.
“This is a fantastic development for humanity,” says Michael Buckley, Be My Eyes CEO.
Ex Machina: Only a select few can try out GPT-4 at the moment
No, you don’t need to win a ticket to Sam Altman’s mansion to carry out your own Turing test on GPT-4 as Domhnall Gleeson’s character did in “Ex Machina,” but GPT-4 is still not open for free public use yet. At the moment, OpenAI is only allowing the developer elite or paying subscribers to trial its latest chatbot prodigy.
An option you do have, assuming you are not a subscriber of ChatGPT Plus, is to head over to Bing where you can try out a caged version of the model, because GPT-4 has been powering Microsoft’s search engine since early February.
Shaping the future of technology
As a company, OpenAI’s mission is to develop artificial general intelligence (AGI) which benefits humanity as a whole, and they’ve shown a reassuring amount of dedication to ensuring their models are as aligned with human intentions and values as possible.
“We’re excited to see how people use GPT-4 as we work towards developing technologies that empower everyone,” states OpenAI, with an employee also stating in GPT-4’s promovideo:
“When we release a model, we know things are not done. We know we have to learn. We know we have to update. We know we have to keep improving all the systems to make it suitable for society.”
So, what is GPT-4?
It’s a tool, it’s a chameleon and it’s a companion, one which OpenAI hopes will help shape the future of technology, but one that we’re gradually realising will also reshape humanity.
Use GPT-4 with caution, but use it to explore the limits of progress, we must!
Subscribe to our newsletter.
This article was originally published on IMPAKTER. Read the original article.