When you pay for ChatGPT, be it Plus, Workforce, or Enterprise, GPT-4 Turbo is now even higher. In a Thursday put up on X (previously Twitter), OpenAI introduced the most recent model of its GPT-4 Turbo mannequin is now dwell for all paid ChatGPT customers, full with some fascinating updates.
So, what’s new? It is all a bit basic, however OpenAI says the most recent iteration of GPT-4 Turbo is now higher at writing, math, logical reasoning, and coding. You may see from a graph within the announcement put up that the April 9 model of GPT-4 Turbo has improved in nearly all classes than GPT-4 Turbo from Jan. 25 of this 12 months. Some metrics are extra dramatic than others: The most important leap seems to be in GPQA, an AI benchmark regarding scientific datasets, the place GPT-Turbo has improved from 39.7% to 49.1%; and the MATH benchmark, which noticed an enchancment from 64.2% to 72.2%.
Apparently, the one space January’s GPT-4 Turbo mannequin excels April’s is within the HumanEval take a look at, which checks an LLMs’ skill to generate code: The previous scored 88.2%, whereas the brand new mannequin scored simply 87.6%. A small distinction, however an fascinating step again that OpenAI would not handle.
What OpenAI particularly focuses on, nevertheless, is the improved basic expertise utilizing ChatGPT with the brand new GPT-Turbo, which is what customers like myself will more than likely discover most anyway. The corporate says whenever you’re writing with ChatGPT, the mannequin’s responses might be “extra direct, much less verbose, and use extra conversational language.” You may see that from the instance OpenAI makes use of in its put up: It asks ChatGPT generate an SMS to remind pals to RSVP to a birthday dinner invite. January’s GPT-4 Turbo mannequin generates a wordy response, that whereas nonetheless probably helpful, does seem AI generated in some areas. The most recent mannequin, nevertheless, generates a shorter, extra informal reminder in regards to the RSVP.
Maybe extra necessary of all, GPT-4 Turbo’s knowledge set goes as much as December 2023. The earlier model was caught at April 2023, which implies it is lacking a full 12 months’s price of context. Whereas the brand new GPT-4 Turbo will not know something about 2024, except you join it to the web, it should have about eight months’ price of present occasions greater than it did earlier than.
