We’re going to make something of a prediction – that this year’s word of the year will have something to do with AI. It might be generative AI. It could be ChatGPT. Or maybe it’ll be deep learning. Whatever it ends up being, it’ll capture the thing that many of us will be thinking about, working with and trying to make sense of in 2023 – large language models.
We can all thank OpenAI for this, which has, largely thanks to ChatGPT, put AI on the mainstream map. Hot on the heels of that launch, we now have GPT-4. And it’s a big deal. Now that the frenzy phase of that announcement has settled, we thought we’d answer all the pressing questions you might still have about this game-changing technology. Check them out.
What is GPT-4?
GPT-4 stands for generative pre-trained transformer, which, if you’re not that familiar with the technology – don’t worry, you’re in good company – doesn’t really tell you a lot. In short, it’s artificial intelligence (AI), a large language model that can deliver a wide-range of text-based solutions and responses, from writing creatively to generating code (and a whole lot more besides).
It can do all of this because it has been fed – or trained on – a massive amount of data. You can describe what it does as learning. It identifies patterns, starts to join dots and begins to“understand” how things work.
And, like an incredible polymath, it can, when promoted, apply this knowledge across numerous disciplines in relatively new and innovative ways. As for the number four, in case you were wondering, it simply represents what version of the model it is.
What happens to GPT-3 now, if anything?
Nothing. You can still use it. It’s just now a slightly older version of a particular type of model. GPT-4 is just a more up to date version of this particular family of models. Its predecessor, GPT-3, will still deliver similar results, but as is the case with any sort of technological development, the latest iteration is, of course, much improved.
Here’s what ChatGPT – which is “fine-tuned from GPT-3.5” – has to say about the matter, which we’ve had to edit ever so slightly because it doesn’t have a clue about about the world post September 2021). Its response is: “If your application is already working well with GPT-3 and you don’t need the additional capabilities of GPT-4, then there may not be a compelling reason to switch.”
Is GPT-4 better than ChatGPT?
Yes but not necessarily in the way you might think. You see, the highly popular, hyped-up and very misunderstood ChatGPT has been built using GPT-3.5. It’s an application that uses the large language model to deliver a very compelling chatbot experience.
You can think of GPT-4 and all previous versions as the source material, the essential building blocks behind solutions like ChatGPT. As it itself explains when asked: “ChatGPT is powered by a variant of the GPT architecture that includes the GPT-3 model.”
How big of an improvement is it compared to GPT-3?
It’s a big improvement,but one that you can think of more as a small step forward rather than a huge leap. OpenAI claims that it performs better in exams, that it’s more creative and that it can now deal with significantly more words (over 25,000).
One of the most notable updates is its ability to now handle visual inputs*, meaning you can upload an image and, for instance, ask it to explain what it can “see” or use it to generate code, as demoed by OpenAI, for a website based on a rough hand drawn image (more on this below).
*Do note that at the time of writing you can’t yet make visual inputs.
What’s the reception been so far?
Predictable. We’ve had the usual response from the optimists who talk about GPT-4 as a new dawn, a disruptor, a revolutionary leap forward in technology and so on. And then we have the doomsayers who wince at the destructive powers of large language models and can’t help but worry about the inevitable rise of machines.
And then there’s everyone else in-between, who range from “it’s early days, relax, I’ve got work to do” to “okay, I see the potential, let’s see what we can do”. Everyone has an opinion.
Is anyone using GPT-4 already?
They are indeed. As OpenAI has been busy working behind the scenes to get the model into shape, it has also been collaborating with partners to demonstrate how it can be used, which also serves to illustrate how powerful and innovative it can be.
Some examples include Stripe, which is using it to improve the user experience; Morgan Stanley, which is using it to make information more discoverable; and Be My Eyes, which is using it to power what it calls a Virtual Volunteer.
How can I start using GPT-4?
For now you’ll need to either be subscribed to ChatGPT Plus or signed up to the API waiting list. The latter is mainly aimed at developers, agencies or organisations with the capability of building applications and solutions using GPT-4.
And that’s an important point. Unlike ChatGPT, which anyone can use – to varying degrees of success – unlocking GPT-4’s potential belongs to experts. It requires technical know-how. It requires experience of working with models. And it requires a lot of resource to bring ideas to life.
Can I create images and videos with GPT-4?
Nope. It’s all about words. GPT-4 can only respond to prompts with text. That might sound limited, but it’s anything but. Getting to this point and sophistication has been a long journey, with the “development of the transistor, the computer, the internet and the semiconductor industry and all the programming languages” all paving the way for AI.
That said, GPT-4 can now respond to visual inputs by “reading” and interpreting images much in the same way as a human. That’s a major breakthrough. In one of the examples given by OpenAI, GPT-4 was asked to describe three images on a grid and explain why it was funny. And it did just that, accurately and authoritatively.
How can GPT-4 help me with digital marketing?
Good question. In a nutshell, you can use it across all aspects of digital marketing. Work in SEO? Ask it to help you categorise keyword research. Copywriter? Get it to brainstorm some ideas based on specific goals. PR specialist? Tell it to help improve your outreach emails. The list goes on.
All of this can help improve your day-to-day, but if you really want to make the most of GPT-4, then, as noted earlier, you’ll want access to the API. This is where the magic happens, where you can create new, bespoke and powerful solutions (think custom-made models, new apps, new services and so on).
This requires a significant level of expertise and knowledge. If you have that, then brilliant. If not, well, it’s as simple as collaborating with specialists. And yes, that’s a mini plug for Melt AI.
Is GPT-4’s dataset up to date?
Somewhat surprisingly, no. The training data for GPT-4 only goes as far as September 2021. And because it can’t, for example, access the internet or other, more recent dataset, it remains stuck in the recent past. It has no idea what is actually going on in the world (the proverbial living under a rock).
The reason for this is that OpenAI has been working on the latest version of its autoregressive language model for some time now, meaning the data was never going to be as recent as, for example, 2023. Why it’s not any later than 2021 is anyone’s guess.
What limitations are there?
Plenty. As noted above, GPT-4 is very behind the times, which, given how quickly technology develops and how much information is now generated, does mean that it has a slight handicap.
It’s also, like its predecessors, still unreliable and prone to “hallucinations’ as OpenAI aptly puts it. It will make up facts and it will not notice mistakes, failing, for example, to take care “to double-check [its] work”.
It isn’t, therefore, fully autonomous. And that’s fine. The real value of this kind of technology comes from working with it. From start to finish, it requires human input, guidance and oversight. At Melt Digital, we call this human-powered AI.
What about risks?
While GPT-4 is markedly better than earlier versions, the same risks are still there. There is no easy fix – getting machines to approach ethical and philosophical questions and operate within certain boundaries is an extremely complex and ongoing task.
OpenAI hasn’t shied away from this fact, with its CEO Sam Altman tweeting that GPT-4 is “still flawed [and] still limited”. It does exhibit biases. It will churn out misinformation. And, it will generate harmful advice. Altman also recently said “people should be happy that we are a little bit scared”.
How safe is all this technology then?
This is a very important question, one that always seems to result in us picturing the likes of Hal 9000 in 2001: A Space Odyssey and Skynet in Terminator. While that tendency to lean towards a dystopian future used to seem predictable and a little silly, today it’s being treated more seriously, not least by OpenAI.
So, for now, it’s relatively safe. It’s where the technology is possibly heading that poses real challenges. OpenAI is fully aware of this, which is why it is vocal in its commitment to make sure that the currently theoretical idea of artificial general intelligence (AGI) – which, in its simplest form, is AI demonstrating “true” human-like behaviour and capabilities – is kept contained. Or as it describes it, “aligned with human values”.
Should I quit my job and live off grid?
Far from it. This is, from a certain perspective, just the beginning. And specific to large language models like GPT-4, the technology is far from perfect, a lot still needs to be figured out and there are still some huge limitations (such as its data being dated). Many so-called cognitive and creative professions are still safe.
All that said, the technology will change things – more so in the coming years as it gets better and better. That’s just the way it is: new technologies reshape the world as we know it. On the plus side, that means new ways of working, new jobs and new ideas. On the negative side that does mean some professions will become obsolete.
But alongside all of that, many professions will endure and be strengthened. Think same job, new job description and a more productive and fulfilling experience. It’s all about perspective.
What is ChatGPT-4?
Don’t even get us started. There seems to be some confusion online about what OpenAI has just announced, what preceded it and how everything fits together. We’re more than happy to clear that up right now: there is no ChatGPT-4.
So if you see it mentioned somewhere, be cautious. A lot of people haven’t done their homework, don’t understand what they’re talking about and have rushed to get their content out into the world without it being sense checked. ChatGPT-4 does not exist..
On a slight side note, why does nobody talk about InstructGPT?
Interesting question! In all honesty, we don’t know. But perhaps it’s because it’s the wrong question. Maybe what you’re really asking is what is InstructGPT? If so, that’s much easier to respond to.
InstructGPT is just another model, one that OpenAI says is “better than GPT-3 at following instructions”. Described in plural by OpenAI, they’re also the default language models on OpenAI’s API (try saying that fast). We hope that helps.
We are on a mission to help brands transform how they perform organically through AI-powered content – tailored to match your exact needs. To find out more about how we can work together, drop us a message here.