The paid version of ChatGPT, with the GPT-4 Turbo model, goes more to the point when it produces text. A new version of the chatbot has been deployed to make it less chatty. The model has also been improved on other technical aspects.
It’s not GPT-5 yet, but it’s a welcome iteration of the existing one. Friday April 12, OpenAI announced an update to ChatGPT, its famous chatbot, for its paying customers (ChatGPT Plus, Team, Enterprise and via the API). This refresh concerns GPT-4 Turbo mode, which first launched in November 2023.
A new version of GPT-4 Turbo that Sam Altman, the boss of OpenAI, saluted : “ GPT-4 is now much smarter and more pleasant to use. » This reaction, however, does not say much about the changes made. Any iteration of a chatbot is typically described by its proponents as “smarter.”
You have to turn to a Twitter feed of OpenAI to discover slightly more precise indications. The American company explains that it “ improved writing, calculation, logical reasoning and coding skills » of its conversational agent, particularly since its last update, dated January 25, 2024 — still for GPT-4 Turbo.
What’s new in GPT-4 Turbo
The company has unveiled a table showing the evolution of GPT-4 Turbo on five test benches (Drop, GPQA, Math, MGSM, MMLU, HumanEval). MMLU and HumanEval appear very stable, while progress is visible for all others. OpenAI did not specify the unit of the numbers entered on the ordinate.
These benchmarks will not speak to ordinary mortals.
MMLU, for example, is a test that measures the extent of knowledge and problem-solving ability acquired by major language models during prior training, Google recalls. It covers 57 tasks, including elementary math, U.S. history, computer science, law, and more, HuggingFace adds.
Math evaluates a model’s ability to solve complex math problems, requiring reasoning, multi-step problem solving, and understanding concepts. HumanEval is used to measure the capabilities of the code side (success of a functional unit test for programming).
Many benchmarks exist in the generative AI market, in addition to those already mentioned previously. Naturally, companies in the sector may be led to instead highlight tests that praise the progress of their artificial intelligence – in any case, it is not always the same test benches that are used.
An AI that tells less about its life
These tests will not necessarily speak to the general public. On the other hand, what Internet users will be able to note is that this brand new ChatGPT will undoubtedly “tell its life story” less during the generation of a response. The tool is now supposed to be more concise in its written interventions.
“ Responses will be more direct, less wordy and use more conversational language », Promises OpenAI. A screenshot shared by the company compares two results. The first has nine lines of text and two emojis, the second three lines and a single emoji. He is therefore less talkative and goes more to the point.
He who was once lazy, we can say that this is an update that goes in his direction.
If you liked this article, you will like the following: don’t miss them by subscribing to Numerama on Google News.