Why GPT-4?

- [Instructor] Why did the OpenAI researchers create GPT-4? Let's make sure we understand the objectives of GPT-3 and this will make it clear why GPT-3.5 and GPT-4 were developed. GPT-3 was trained on a large volume of text available on the internet and its task was to predict the next word. In one sense, it was like a really expensive text predictor like the one you'd find on your mobile phone. It wasn't supposed to be like an assistant or be good at tasks, but in the research paper on GPT-3 the researchers found that if the model was large enough and you gave it a couple of examples, you could get it to do useful things like summarizing texts it had been given. The downside was that sometimes this didn't always work on the first attempt and so you then have to try a couple of different prompts to get the model to do what you wanted. Unfortunately as it was trained on data from the internet, you could easily get it to generate toxic language and hateful speech and sometimes it would just make up facts. GPT-3.5 was created to address these challenges. GPT-4 then built upon the work done by GPT-3.5 and improved on it. This means it would be able to follow instructions better and it would generate less toxic language and would be less likely to fabricate information. Let's dig a little bit deeper into what we mean by fabricating information, also called hallucination. Let's say you provided some text to the model and asked it to summarize it. As part of its summary, it would include information that wasn't in the original text. So let's say you asked the model to summarize the British National Anthem, "God Save the King." It might summarize it but include France or some other country in the summary when there's no mention of France in the original national anthem. One area in which GPT-4 shines is that it is much better at following instructions. Let me show you an example comparing GPT-3 to GPT-4. So I'm going to head over to the OpenAI Playground, so that's platform.openai.com/playground, and this gives me access to both GPT-3 and GPT-4. So let's stick to the theme of national anthems and let's give the original GPT-3 model the following prompt. Now I have no idea what this will output and if you were to use the same prompt, you'd get different results from me. So write a new national anthem including all of the four countries in the United Kingdom. Now just so you know, that includes England, Scotland, Wales, and Northern Ireland. So let me head over to Submit and let's see what text the GPT-3 model generates. Now you can see that it isn't following the instruction at all. Instead it seems to be generating some sort of a list. So the first bit was the information that I provided in the prompt and then number seven for some reason is get all of the four countries to agree to call their currency the UKP. Number eight is bring back the death penalty. Number nine is changed the name of the Ministry of Defense to Ministry of War and so on. So all of the text that's been generated is in green. Let's go ahead and copy the prompt and let's enter the same prompt so that we can compare the results of GPT-3 versus GPT-4. Now at the time of this recording, one of the ways to access GPT-4 is via the ChatGPT Plus. I'm just going to select a GPT-4 from this menu. And let me go ahead and paste the same prompt. Now again, remember that it needs to have something about the four countries which is England, Scotland, Wales, and Northern Ireland. So let's see what GPT-4 generates. Now these results are pretty impressive. Firstly, it hasn't generated a list like GPT-3. Instead this actually looks like an anthem with a verse and a chorus and so on. And then as I have mentioned, it needs to include information about the four countries but I didn't specify what the four countries are. But you can see that in that first verse, "From the highlands to the valleys and the shore that we adore. United Kingdom, proud and strong, together we stand, forevermore. Four nations bound by history, by heart, and by the sea. In unity, we forge our fate, bound by love and liberty." And you can see that in the chorus, it includes England, Scotland, Wales, and Northern Ireland. And in all of the instances of the chorus, you can see that it's using exactly the same words so that it's maintained the context within the chorus and generating different verses each time including both a bridge and an outro. So this is incredibly impressive because this does look like an anthem as I said before and it's included all of the details that I've asked for. All right, so we've seen that GPT-4 is really good at following instructions and produces excellent output. It still sometimes makes up facts and generates toxic information, but it is much better than its predecessors.