ChatGPT took a giant step forward in its marathon run to become one of the most powerful AI chat rooms in the world. On March 14, the company OpenAI, creator of this prototype, announced that the new version of the chatbot is ready: it is ChatGPT-4.
To understand the differences of this latest update with its predecessor (ChatGPT-3) it is important to take into account that prior to reaching this level there was an intermediate version called ChatGPT-3.5which was the model that became popular and was available from November 30, 2022 to March 13 of this year.
In recent months, as people have been interacting with ChatGPT, the model has grown: it has trained and learned a lot more. However, its knowledge database is still limited to the information that was on the internet as of September 2021according to the GPT-4 Technical Report published by OpenAI on March 16.
Regarding the information with which the most recent version of the model was built, what happened with the number of parameters (data used for training) with version 4 is important.
To size: in 3.5 about 175,000 million parameters were needed, while 4 is estimated to have a size five times larger (about 100,000,000,000,000 parameters). Here the size does matter because basically are the parameters where the word information is stored and this translates into greater precision when delivering the queries that are made to it.
Joseph Vergel, a machine learning engineer at Guane Enterprises, highlighted something key in the evolution of ChatGPT. He said that in 3.5 there was evidence of a change in the feedback with the human: The possibility of taking advantage of the language model was opened because it already had a very good number of words learned and the answers were delivered with context (it related the previous answers). He gained greater fluency and coherence in the conversation.
Going to 4 shows a considerable improvement in the number of backward messages that the model begins to take into account. This means that context words were increased to answer. Imagine that you make a query and another and another and if one of those is among the 32,000 above, the chat has the ability to understand it and relate it to those that have already responded. Like when the human mind remembers something from the past.
In practice: There is a scenario where an interview is simulated, the chat is asked in the first message to be the President of the United States, Joe Biden. As the conversation evolves, the topic of Ukraine was touched on and there he gives a context of what has happened, but then in the eighth message he was asked again, for example, what would the president’s opinion be if the next world war blow up with Korea, so at that point ChatGPT-4 has the ability to rewind the entire dialogue and give a response with the context of what has been delivered during the interaction.n, take into account your previous answers.
“What we are seeing is that the amount of backwards text that he could review and summarize increased, to relate it to the new one. It is like when you have short, medium and long-term memory,” said Vergel. In short, what they did was they increased the length of that memory, as if 3.5 had short-term memory and 4 is already in the medium term.
“One of the notable differences is that when faced with the same question, the new version is capable of responding with fewer words,” highlighted Sergio Gutiérrez, professor at the Faculty of Engineering at the University of Antioquia.
To this is added that within the same fourth version there are a couple of differences: there is the GPT-4 (no vision) unimodal that only works with text and the multimodal GPT4 that processes texts, images and videos, and in turn answer with images and videos beyond the text.
Precisely, this last type of language was the one that Microsoft incorporated into its Excel, Word and Outlook programs so that it is capable of not only understanding and generating text, but also now doing the same with images.
There are more news
Continuing with the comparison between both versions of the models, it is found that 3.5 was very repetitive in certain answers: it repeated phrases and added some context. This improved in the most recent update.
“The 4 has a very important characteristic and that is that the tool begins to have a shield in terms of issuing outputs on sensitive issuesSo that may generate controversy, are those related to medicine or geopolitics, for example. The new version (ChatGPT-4) has a probability of less than 80% to respond to topics that are not allowed, but a probability of 20% is still maintained and that is tricky” said Gutiérrez.
On the other hand, according to engineer Vergel, in addition improved performance to do predictive tasks how to identify which are the entities (person, object, organization or location) within a sentence.
That was evident in GPT-4’s performance on several of the world’s most recognized academic and professional aptitude exams. According to the Technical Report, on the GRE Quantitative test, 3.5 scored approximately 23%, 4 (unimodal) scored approximately 62%, and 4 (multimodal) scored approximately 80%.
“In this type of exam (quantitative reasoning) generally the tests are not only textual, questions are also asked where the student analyzes images, that is why multimodal GPT-4 achieves better performance.” He had the ability to interpret the test graphs.
If these percentages are translated into an educational system grading scheme, such as a 0-10 scale, the 4 multimodal obtained 8.0 points, which is an outstanding performance. Bottom line: GPT-4 won the test.
Definitely, subsequent versions of the generative model will continue to improvebecause as you learn, at the same time its creators are identifying shortcomings, especially those that are related to the accuracy of the answers.
And this leads us to continue reflecting on the ethical implications around the use of the tool: opens the possibilities of plagiarism in the construction of texts. We will see more and more texts built with ChatGPT instead of being written by an author, which will lead to permanent discussion.
The new ChatGPt for now is only available paying. It is 20USD for the plus package. The free version, meanwhile, remains GPT3-3.5. So unless you pay, you won’t see a change.