OpenAI, the company led by Sam Altman for the development of advanced Artificial Intelligence models, owner of the popular ChatGPT platform powered by its GPT-3.5 model, has just released GPT-4, which can be used from today both through ChatGPT and through the API, there is currently a waiting list.
The company notes that while it still offers sub-human performance in many real-world scenarios, GPT-4 is capable of being at human-like levels in academic and professional settings.
The best machine learning model created to date
For OpenAI, GPT-4 is the most stable model that they have been able to develop to date, becoming the first great model whose training can be predicted with prior precision, and with a view to being able to be reliably scalable, although from the company They will continue working on refining the methodology to predict and launch future capabilities with greater anticipation, something they consider critical at the security level.
Despite the fact that it is a multimodal model, that is, it supports various types of inputs, for now it will only support text input, preparing to support image input in the future with the help of a partner.
OpenAI adds that it is opening OpenAI Evals, its Artificial Intelligence performance evaluation platform, which will allow anyone to report the problems found in some of its models in order to introduce improvements to them.
Perhaps they want to avoid what until now has been called hallucinations, that is, cases in which the models come to invent answers and present them as if they were true answers, this being an aspect that causes a lot of concern.
Outperforming GPT-3.5 and competitive models
How could it be otherwise, OpenAI has come to compare the capabilities of GPT-4 in relation to GPT-3.5, reaching the conclusion that:
GPT-4 is more reliable, creative, and capable of handling much more nuanced instructions than GPT-3.5.
In this sense, GPT-4 has meant six months of refinement taking into account what was previously learned, even with ChatGPT, so a substantial improvement was to be expected, something that they have noticed in their tests through an exam simulation that They were originally aimed at people, offering a technical report that those interested can consult at this link.
And furthermore, they point out that:
We also evaluate GPT-4 on traditional benchmarks designed for machine learning models. GPT-4 considerably outperforms existing large language models, along with most next-generation (SOTA) models that may include additional training protocols or reference-specific elaboration:
With this in mind, it is not surprising that OpenAI’s rivals are working hard to try to improve and get the most out of their respective solutions, something that we will see over the next few weeks as announcements.
Prepared to support multiple languages, even low usage
As an example, OpenAI says that it has trained its new model to work in multiple languages, training with Azure Translate, unlike most machine learning systems, which have used many English-only benchmarks. , noting that:
In all 24 of 26 languages tested, GPT-4 outperforms GPT-3.5 and other LLMs (Chinchilla, PaLM) in English, including for low-resource languages such as Latvian, Welsh and Swahili.
Undoubtedly, it can be said that GPT-4 arrives stomping, where it will show its potential when it is also capable of supporting images as entry points, although in any case, it is preparing to offer responses only in text format. .
More information: OpenAI