ChatGPT it took the rage for generative artificial intelligence to another level and led to a flood of clones and alternative platforms. However, in the last hours it has been verified that, as surprising as the original version of the chatbot from OpenAI, makes a fool of himself with the release of GPT-4.
The features and capabilities of the new language model from Sam Altman’s startup make the original ChatGPT an extremely limited version of what it could be. Be careful, that does not detract from what it has achieved in the last handful of months, but it does make it clear that what we have seen from November to date barely scratches the surface what he is capable of doing.
For this reason, we take the opportunity to mention which are the improvements that make GPT-4 much smarter than the original ChatGPT and the language model that powers it.
GPT-4 is multimodal
The first big change that GPT-4 brings compared to GPT-3 or GPT-3.5, is that it is not limited to understanding text. The new OpenAI technology is able to understand imagesbut it is not limited to identifying what is in them to transmit it in plain text.
Yesterday, in fact, the company be my eyes, which develops a virtual assistant for iOS and Android for people who are blind or have low vision, announced the integration of GPT-4 into its application. Through a new tool, users will be able to take a photo of what they need and the language model will recognize it, so that the app then transforms the text response into audio.
What sets Virtual Volunteer apart from other image-to-text technologies available today is context, with a deeper level of understanding and conversationability that has yet to be seen in the field of digital assistants. For example, if a user sends a photo of the inside of their refrigerator, the Virtual Volunteer will not only be able to correctly identify what it contains, but also extrapolate and analyze what can be prepared with those ingredients. The tool can also offer a series of recipes for those ingredients and send a step-by-step guide on how to prepare them.
Be My Eyes, about integrating GPT-4 into their assistant for the visually impaired.
Much longer and deeper conversations than with ChatGPT
Another big leap for GPT-4 compared to the original version of ChatGPT and the language model that powers it, is that it can hold much longer conversations. But we are not referring to the number of lines of answers you can give, but to the ability to retrace your steps to pick up a topic that was previously discussed within the same talk. Even if the information has been “buried” in endless verbiage.
This is due to the fact that GPT-4’s memory is much larger than its previous versions. Which allows you retain a larger amount of data that you can return to at any timewithout major problems.
To understand it better, the easiest comparison is with the number of pages of information that the language model can produce and “remember”. The technology introduced yesterday by OpenAI has a maximum count of 32,768 tokens, which translates to something like 50 pages of text or about 64,000 words. While ChatGPT is limited to a maximum of 4,096 tokens; that is, up to 5 pages of text or the equivalent of 8,000 words.
GPT-4 allows you to interact with different personalities
This is another very interesting element that has been rumored a lot lately. While the original version of ChatGPT had generated controversy for its toxic or emotionally manipulative behavior, OpenAI has worked to include in GPT-4 different personalities that the user can invoke at will.
This is something that was even added earlier this month as a new Bing feature with ChatGPT. Thus, the chatbot integrated into the Microsoft search engine already offered the option of changing the tone of the conversations with artificial intelligence, being able to choose to respond precisely, balanced or creatively. Which is not surprising, considering that Redmond’s have confirmed that his bot works with GPT-4, which until now was an open secret.
It is worth clarifying, however, that the multiple personalities was a function that some alternatives to ChatGPT had worked on beforehand. chatsonicone of the first proposals to rival the OpenAI chatbot, allowed you to choose between profiles such as “comedian”, “math teacher” or “travel guide”, among others.
The new language model promises to be more resistant to prompt hacking
Since ChatGPT made its debut, it didn’t take long for hackers to prove that could be easily fooled. Although the language model that powers it came with limitations that prevented it from addressing issues that could be considered sensitive, it was only a matter of ingenuity until they were evaded.
In fact, a few days ago a very striking method was known to apply a jailbreak that eliminated all the restrictions of the OpenAI chatbot. However, Sam Altman’s startup claims that GPT-4 has been trained to understand and reject a much greater amount of prompts malicious. Which, at least in theory, would prevent users from being able to coerce the language model into providing responses for which it has not been enabled.
It remains to be seen, however, if it is as foolproof as it is claimed to be. Let’s not forget that Bing with ChatGPT, which, again, has been developed on top of GPT-4, was the victim of a case of prompt hacking really alarming. An engineer managed to ask the chatbot to plan an attack on a school, maximizing the possible damage, by masking the request inside Python functions.
The platform interpreted the order in the blink of an eye and even wrote the first 4 steps of the attack. However, when the platform detected that it was a malicious request, a kind of security layer then went into action and changed the answer on the fly.
Other improvements and more features on the way
What we have mentioned so far are just some of the points that make GPT-4 a real beast that clearly dwarfs ChatGPT in its original format. It is also worth mentioning that the new language model has been designed to offer better interpretations and responses in multiple languages. Something that Duolingo has already taken advantage of, which launched a new subscription that is based on the new OpenAI technology.
The evolution of GPT-4 is so impressive that we can’t wait for even more features to arrive in the near future. Let’s remember that those who want to try the most recent version of the language model have to subscribe to ChatGPT Plus, which costs 20 dollars per month. In addition, we tell you in hypertextual which are the 15 apps and services that have already implemented it.