ChatGPT (Photo: Reuters)
Let's start with a clarification: although ChatGPT was originally described as GPT-3.5, i.e. a late iteration of GPT-3, in practice it is all about a user interface, which is not limited to one language model or another. In other words, GPT-4 is not a replacement for ChatGPT but a language model that it relies on to perform operations.
Now, the system that exploded online in months, is being upgraded with a new set of intriguing capabilities and an ever-wider context window, capable of "remembering" about 50 pages of content, 5 times what the previous version could hold in memory. Why is this significant
? Because large language models are indeed trained on millions of web pages, books and other texts, but when they have a conversation with a user, there is a limit to the amount of information they can 'remember'.
Models with small context windows tend to "forget" the content of even very recent conversations, leading them to deviate from the topic.
After a few thousand words, they also forget their initial instructions, and base their behavior on the last command received in their context window instead of the original request.
In fact, the limited connection window is the main reason for the inaccuracies we've seen in ChatGPT to date, among other things leading to ChatGPT being boycotted by the academic community.
This limit in previous versions of GPT was in the region of 8,000 words, which is about five pages of a book, with each new conversational interaction relying on the previous messages.
Exceeding this amount would cause the bot to 'forget' what it was talking about and return to the starting point, making it difficult for users to train it as they wished.
The GPT-4 on the other hand has memory equivalent to up to 64,000 words or 50 pages of text, which is enough to write an entire play or short story.
This means that in a conversation or text production, he will be able to retain much more content, and more importantly, remember what you talked about even 20 pages later.
In this way, GPT-4 can actually gain knowledge and insights as the creative process progresses and achieve better results.
For example, suppose we are writing a story that opens with a long monologue recounting the events of the main character.
35 pages later, when you want to add dialogue between the same character and another character, GPT-4 will be able to remember who it is and write the dialogue accordingly, even 35 pages later.
Alan Pike, a former software engineer at Apple, explained it to TechCrunch simply: "The old model will forget anything you try to teach it. It will forget that you live in Canada. It will forget that you have children... if neither of you mentions its name for a while , he'll forget that too. If you talk to a GPT-enabled character for a while, you'll start to feel some kind of connection with them... But eventually, the fact that he has no medium-term memory becomes clear, and the illusion is shattered."
There is no doubt that the enlarged context window is the significant upgrade of this version, but that's really not all.
So what else is new in GPT-4?
Better maintenance of the sector boundaries
Pike notes that one of the reasons chatbots like Bing Chat can degenerate into bad behavior is that their initial guidelines—to be a helpful chatbot, respond respectfully, and so on—are quickly pushed out of their context windows by additional prompts and responses.
For all that today's chatbots do right, with a little coaxing they can be tricked and 'convinced' into abusing their power.
The abundance of malicious instructions CHATGPT received from users during the runtime was enough to train the new language model so that it knew how to refuse inappropriate commands such as writing malicious code or instructions to break into a bank.
According to the company, the new model is much better than its predecessors in "practicality, steering ability and refusal to leave the safety barriers".
Ability to identify and analyze images
The most notable change in the fourth version of GPT is that it is "multimodal", meaning that it is able to understand and analyze information of different types.
If ChatGPT and GPT-3 were limited to text only, GPT-4 is also able to recognize and process visual content, such as images and videos.
In the launch video, OpenAI demonstrates the multimodality feature using an image of a bouquet of balloons tied with string.
If we tell GPT-4 to cut the wires, it already knows by itself that the balloons will fly up.
One of the interesting applications of this feature is the company's new partnership with the accessibility application Be My Eyes, thanks to which the blind and visually impaired will be able to receive a verbal description of everything around them, ask for directions and more.
The demo proved beyond any doubt that Gpt-4 really understands what it sees, and all the user has to do is ask the right questions.
More in Walla!
The treatment that prolongs the lives of lung cancer patients
In collaboration with the Israeli lung cancer association
GPT-4 knows more languages
While the world of artificial intelligence is dominated by English speakers, the capabilities of large language models are applicable to any written language.
At the launch of GPT-4, the company demonstrated how the bot is able to accurately answer thousands of multiple choice questions in 26 languages.
This initial test of language capabilities may seem promising, but it is far from the answer to multilingual AI.
First, because multiple choice questions don't really represent normal speech and second, because the test criteria were translated from English in the first place.
However, he did a great job at something he wasn't really trained for, which is a good sign going forward.
Improved steering options
'Steering' is a concept in AI, which refers to the bot's ability to change its behavior on demand.
This can be useful when you want to dictate the tone of voice of the writer, for example an enthusiastic fan, a frustrated customer, a bouncy bunny or Donald Trump.
Everything goes.
Now, users can customize their bot personality and create a permanent personal style for it, just like a human.
This is especially useful for brands that want to take advantage of GPT-4 without losing their unique voice.
Another example can be when you write a script and want to insert the unique character of each character into the dialogues.
This can be done with prompts like "Answer like a 19th century Turkish general" or "Answer like you're being interviewed for the news."
If until today these features were only available to us in a basic way, today we can adjust the nature of the 'writer', his perspective, his style of conversation and his tone of voice in each interaction and even set defaults.
So where can you try GPT-4?
Currently, GPT-4 is only available to users of ChatGPT Plus, OpenAI's premium package, but the company promises that soon it will also be available to developers via API, and it is very likely that we will see a free version soon.
Although we haven't tried it yet, it's not hard to imagine how conversations with it could be much more compelling than with the previous models.
With a larger "memory", GPT-4 should be able to hold coherent conversations for hours or even days.
And perhaps more importantly, his chances of derailment should be at least significant.
Marketing and digital
in the headlines
Tags
artificial intelligence