News

GPT-4: OpenAI’s Generative AI Great Leap Forward

Generative Artificial Intelligence systems have only exploded in popularity a few weeks ago, but they are already grabbing all the headlines. Also the language models used by those that work through text, either as a means for the requests that the systems receive, or to return the responses and results to these requests. Of them, the most popular, perhaps, is GPT, a large language model capable of imitating human writingas well as his reasoning.

This model is able to do all this because it has been trained with a large amount of communication and texts written by humans. From works of classical literature to messages and texts collected on the Internet. GPT is a type of AI model that bases its operation on predicting which letter, number or character follows the one that has already appeared.

Currently, its creator, the company OpenAI, has already released its version 4although on certain systems, such as the free version of the generative AI system ChatGPT, continues to maintain 3.5. GPT-4 represents a giant leap compared to previous versions of the model, as we will see below.

GPT-4: what it is and access prices

GPT-4 is the fourth generation of the foundational GPT model, and it is already a multimodal model capable of creating not only text, but also works of art, video or sound, just like humans do. In addition, it is capable of solving problems that are supplied to it in writing in a more elaborate way than GPT-3 and GPT 3.5. It is also capable of solving much more complex problems than previous versions of GPT.

Its launch became official on March 14, 2023, and its availability through the paid version of ChatGPT, for a price of 20 dollars per month for individual users. Companies wishing to use ChatGPT Plus have to sign up to request access through a waiting list. Still, its access is limited, and for example, OpenAI has still hardly given enterprises access to the visual terrain-related model functions. To use your API to integrate it into your products, you have to sign up for this list.

Pricing for the CPT-4 text-only API starts at $0.03 per 1,000 tokens, each about four English characters, prompt. The price of the exit tokens starts at $0.06 per thousand. Approximately, a token is equivalent to about 75 words, on average. OpenAI provides more details on how it accounts for tokens, and what they are, on this page.

In addition, Open AI already has plans to launch a new subscription plan for those who use GPT-4 quite frequently, as well as a free access portal to GPT-4. Of course, with limitations, since it will only allow a limited number of questions and requests to be made to its users. But we will still have to wait until it is available, since OpenAI has not yet provided deadlines or dates for its availability. Meanwhile, users of the free version of ChatGPT continue to use a system that incorporates GPT-3.5.

What this model allows to do

The main use of GPT-4, like its predecessors, is answer questions asked in natural language through text, as well as other similar requests. According to its creators, it is capable of following complex instructions in natural language, as well as solving complicated problems with a high level of precision. A precision that is also much higher than what the model can achieve with previous versions.

Specifically, GPT-4 is capable of solving mathematical problems, answering questions, telling stories or drawing conclusions from various premises that are provided. It is also capable of generating summaries of texts and contents of a certain length, which can be useful, for example, to summarize presentations or make a synthesis of the results observed during a doctor’s visit to a patient. These are just two of the possible use cases that can be identified for GPT-4.

Furthermore, GPT-4 has been able to achieve higher scores than GPT-3.5 and other versions in various tests. Thanks to his ability to repeat information in a coherent order, he has achieved, for example, scores that are around 90% correct on the United States Bar Practice Exam, and 93% on the SAT Reading and Writing test. based. in evidence.

Of course, it must be taken into account that these tests do not measure general knowledge of the model. What is achieved by running GPT-4 to answer test questions that are standardized is the model’s ability to generate responses that seem correctgenerated from all the data and information with which you have been trained.

The limitations that GPT-4 still has

Although GPT-4 allows you to get more answers, and with more variety and precision, than previous versions of the model, there are quite a few things that it is still not capable of. For example, it doesn’t check that everything it says is true, which leads to failures. In addition, due to your training, your answers may not make sense, and even what you say may generate controversy due to controversial statements.

To avoid the latter, OpenAI relies on digital controls, as well as human trainers, to try to make the model’s responses as useful, and as appropriate for business use as possible.

In addition, GPT-4 tends to generate what are known as hallucinations in many cases. With this term, in the case of language models, we refer to imprecise or erroneous responses. In these cases, the words and phrases that it returns in response to a prompt or question make sense within the phrase.

This is because they are based on probabilities established by the way the system has been trained. But they are not connected to real events, nor has their accuracy been proven. In fact, the hallucinations of GPT-4, in many cases, are correctly composed phrases, but in many cases without meaning, or with errors of a certain weight. OpenAI is currently working to reduce the number of false assertions that the model generates in response.

Another of the main limitations of the model is the lack of clarity from OpenAI regarding whether the sensitive corporate information supplied to GPT-4 will be used on other occasions to train the model, and therefore will be exposed to third parties. Microsoft, for example, which has a reseller agreement with OpenAI, plans to offer private instances of ChatGPT with GPT-4 to enterprises in the short term. Therefore, by then it will be important to have information in this regard, and for many companies, with the certainty that the information they contribute to the model is not accessible to third parties.

On the other hand, as with GPT-3.5, GPT-4 does not have information after the last months of 2021 in its knowledge base. It is an important limitation of the model with respect to other competitors. This is the case of Bard from Google, which has recent information because data and information have been used to train it practically up to the present.

GPT-4 vs. ChatGPT

GPT-4 offers a amount of knowledge greater than that of GPT-3.5, the model that ChatGPT incorporates. Of course, it is not known to what extent, since OpenAI has not revealed the size of the model used to train GPT-4. The only thing they have claimed is that it has more data and computation than the billions of parameters used to train ChatGPT:

The GPT-4 model has shown greater ability than ChatGPT in writing a greater variety of content. Among them, fiction texts. This latest version of GPT has also achieved higher scores than ChatGPT, in standardized tests, as we have mentioned. In addition, the responses to the requests supplied to the model tend to be more concise, and easier to analyze.

This version of the model is also better than the previous one, and therefore better than Chap GPT, in making important decisions for companies, such as planning or summaries. GPT-4 is, according to OpenAI, responds 82% less often to requests for disallowed contentand is 40% more likely to generate factual responses.

But without a doubt, one of its main advances is that GPT-4 can work with images and visual content. Thanks to this, it can serve as a visual support, describing objects in the real world, among other applications. Therefore, GPT-4 is undoubtedly a leap forward in the field of large language models. For now, it is only possible to access it by paying an amount or through applications from companies associated with OpenAI. But it won’t take long, probably, before we can see its full potential.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *