by OpenAI’s new GPT-4 language model was just announced this Tuesday with significant improvements in its ability to analyze complex problems and more communication options with it. The user will no longer be limited to the interaction through short texts such as simple questions, but can upload an image or long texts, papers complete or even short books, for your summary, analysis or search for complex patterns.
Of course, the intellectual capabilities of the service have also been improved, and OpenAI ensures that its ability to reason about complex problems is now more humane and less prone to obvious mistakes like the ones we have been able to detect in its previous version GPT-3.5.
OpenAI offered a demonstration of its capabilities and limitations on its YouTube channel.
What improvements does GPT-4 incorporate?
The new GPT-4 version is capable of handling more than 25,000 words of text, enabling use cases such as long content creation, long conversations, and user-provided document search and analysis. This allows you to translate documents or simplify the work of editors looking for spelling or grammatical mistakes and redundancies. It can even serve to detect general style or possible authorship if these capabilities continue to be expanded. We already know cases where artificial intelligence has been used to rescue and find out the authorship of many manuscripts.
OpenAI has worked on its correct alignment, the main headache for researchers looking for artificial general intelligence, for six months. That is to say, that they behave correctly, do not lie and do not give harmful recommendations to their users as they sometimes ended up doing after “hallucinating”. “GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce objective answers than GPT-3.5 in our internal evaluations.”
The alignment of artificial intelligences is one of the hot topics in Silicon Valley. For some researchers, it is extremely dangerous that these are not correctly aligned with the ethical interests of the human being and may give false or harmful information to society as a whole. However, others believe that a very strict alignment could favor certain political organizations to the detriment of the general interest and pluralism. OpenAI advocates offering freedom to the user within a broad spectrum of what humanity can consider “correct”, but to provide this possibility, they need to improve their alignment, its president Sam Altman assured on numerous occasions.
In this regard, he said that GPT-4 supports a new message via API for developers, which will soon be extended to all ChatGPT users, to customize their behavior. “If you want the AI to always answer you as if it were Shakespeare or only in JSON [un formato de lenguaje de datos usado en programación] you will be able to do it with this new version.” In the demo, a box was shown for the user to customize GPT-4’s personality and capabilities. For example, you can tell him to be a manager who calculates your taxes very carefully. “He’s very good at calculating now,” says Greg Brockman, co-founder of OpenAI.
He also assured last Monday that “we need more regulation in AI”, despite the accelerated pace of updates and innovation of his company with ChatGPT that are causing numerous experts in ethics and disinformation to publicly raise their voices about its dangers.
Despite being less prone to lying, OpenAI says that GPT-4 is more creative than the previous version of ChatGPT because it can generate, edit and help the user in the creation of prose and verse texts with a more refined style and sense of humor such as songwriting, movie scripts or generating articles from one that have done for replicate their style and vocabulary.
Now also understands images
GPT-4 is capable of recommending what to eat if you share a meal with him. Photography inside your fridge. This small improvement and usefulness demonstrates a huge leap in the Cognitive abilities of his language model. Any AI has to be trained with thousands of photographs of each vegetable, fruit or vegetable to recognize them, since its abstraction capacity is not like that of humans. Several models are already being used to catalog forests, detect bird nests or locate where animals are in danger of extinction. But GPT-4 is general in nature and open to the public; It is not only trained to recognize a type of mammal, for example, but any object, product or landscape.
It is now able to interpret a pencil sketch of a web page and transform it into HTML, CSS and Javascript code with functional titles and buttons.
Some rumors indicated that this new version of ChatGPT would be multimodal, that is, that it would allow the creation of images, something that Dalle-2, from the company itself, already does, video or audio, but Altman already warned that people would be “disappointed with such high expectations.”
The ability to upload images for GPT-4 to interpret is not yet open to users, although the model is already trained for it because the company prefers to “go little by little and polish the last details.” During the presentation, they used a Discord channel connected to the API to send the images.
For now, this latest update to the language model is only available to ChatGPT Plus subscriberswhich costs $20 per month.
Microsoft, which has partnered with OpenAI to integrate the capabilities of its language model into products like Bing, will hold an event on artificial intelligence on March 16, where closer collaboration is sure to be featured.
