The OpenAI startup formalized this Tuesday GPT-4, the new version of ChatGPT (GPT-3.5). And the gap is huge.
GPT-4 is here. The co-founder of the star up OpenAI, Greg Brockman, announced the news on Tuesday on Twitter, two days before the conference of Microsoft – his partner – which was precisely to be the opportunity to find out more.
OpenAI has finally decided to take the lead and unveil, before the release to the general public, what the successor to ChatGPT – technically called GPT-3.5 turbo – will offer.
A subtle improvement
An example of the new mount’s abilities? GPT-4 passed the bar exam with a score close to the top 10% when its predecessor was around the bottom 10%.
But OpenAI prefers to announce it from the start: there will be no “wow” effect with this new tool.
“In casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtle,” the company explains on its site.
The difference appears when the complexity of the task increases.
“GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5,” OpenAI promises.
Here again, its creators had fun making it pass a battery of exams, in this case tests in biology, law, economics or even literature.
The results show that GPT-4 fares much better than its predecessor in law or chemistry. They also show that they cannot, on the other hand, overcome the mediocre results of his eldest on certain tests such as a university mathematics test.
Risks “similar” to those of ChatGPT
The other novelty is also the arrival of visual inputs. Concretely, you can therefore add images to the text that GPT-4 will be able to recognize. We can thus imagine that he will be able to identify an object, to explain its usefulness or its use, or even to understand a written language.
“Despite its capabilities, GPT-4 has similar limitations to previous GPT models,” OpenAI would like to point out. “Most importantly, he’s still not completely reliable – he’s imagining facts and making errors in reasoning.”
But again, OpenAI promises that its inventions are far fewer than for ChatGPT. Like its predecessor, GPT-4 has to be content with a database dating from before September 2021, even if its creator suggests that he may know of certain later events.
“GPT-4 presents risks similar to those of previous models, such as the generation of harmful advice, buggy code or inaccurate information”, insists OpenAI which nevertheless announces that it has tested it with around fifty experts to track dead ends.
GTP-4 should also stand a little wiser than its elder and more difficult to “jailbreak”, that is to say to make him circumvent the rules.
The new version is not currently available to everyone, you must request access on the waiting list.
