ChatGPT: we tried GPT-4, and there is still room for improvement
The launch of ChatGPT in November 2022 had the effect of a bomb, the whole planet having rushed to the OpenAI site to ask questions of its “artificial intelligence”. Remember, however, that ChatGPT and natural language processing algorithms are not intelligent, as they rely primarily on statistics. In short, ChatGPT does not understand what it writes, but the texts generated are statistically close to the expected answers.
Improvements and still limits
By trying ChatGPT, we have all experienced some risky answers, which clearly show the limits of these algorithms. Thus and with a certain aplomb, ChatGPT powered by version 3.5 of GPT could affirm that the number “six” comprised four letters (when you are told that it does not understand what it writes…). However, the conversational agent quickly recognized his mistake.
The answer offered by ChatGPT version GPT-4 provides more details, while remaining somewhat ambiguous. Indeed, the conversational agent indicates that“to date, there has never been a female Prime Minister in France”while continuing on the fact that“Édith Cresson was the first woman to hold the post of Prime Minister in France”. Confusing. We can imagine that the algorithm is deceived by the duality “women” And “Prime Minister”. The same question with “Prime Minister” might have given a more specific answer.
Overall, it seems that GPT-4 provides somewhat more subtle answers than the previous version, although it is difficult to give a definitive opinion on this point. We also tested the multimodal aspect of the robot. Thus, you can ask ChatGPT to describe a submitted image. Our various tests, carried out with images from vehicle tests, again leave us doubtful.
By submitting the image above to the artificial intelligence of OpenAI, here is what it proposed to us.
Difficult to be convinced by this return. If the model and color (and again, the vehicle being orange) are well identified, it should be remembered that this information is present in the test.
New test with a photo taken from the test of the Sony WH-CH520 headphones published this morning on the site.
Here again, the result is ambivalent. On the one hand, ChatGPT correctly identifies the headset, the brand and the model (again, the information is present in the text surrounding the image), but it gives us a rather fanciful description mentioning an audio cable which does not exist just not. The rest of the information simply cannot come from an analysis of the image (“cushions […] upholstered in faux leather for maximum comfort”), but probably comes from other articles or tests of the product.
ChatGPT “in the sauce” GPT-4 does not therefore seem to revolutionize the game, but allows OpenAI to regain control of the communication component while global competition between Web giants is in full swing. The fact remains that the algorithms are refined version after version, the answers appearing more and more relevant; and in this field, competition allows dazzling progress.