OpenAI, the corporate behind the favored conversational AI mannequin ChatGPT, has launched an replace to enhance the chatbot’s accuracy.
After an prolonged downtime on Tuesday, ChatGPT is up and operating with a brand new mannequin.
That is the primary replace to ChatGPT this 12 months and the second replace to the mannequin since its launch in November.
A pop-up message lists the adjustments in what OpenAI calls the “January 9 launch” replace.
Screenshot from: chat.openai.com/chat, January 2023.
The popup reads:
“Now we have additional improved the ChatGPT mannequin! It must be usually higher and have improved factuality throughout a variety of topics.”
Moreover, now you can cease ChatGPT whereas it is within the strategy of producing a response, which is a extremely requested function.
Particulars within the launch notes are restricted, so I examined a number of prompts in ChatGPT to be taught extra concerning the new replace. This is what I found.
Dates nonetheless restricted till 2021
Whereas ChatGPT could also be extra correct now, its data remains to be restricted to 2021 dates.
When requested what’s included within the new ChatGPT replace, the AI tells me:
“ChatGPT is a big language mannequin developed by OpenAI and educated on a knowledge set of conversational textual content. As it’s an OpenAI mannequin, it’s often up to date with new options and enhancements. Nevertheless, as my data is lower off in 2021, I haven’t got the most recent updates and I can not offer you concrete data on what new options or enhancements have been added to the mannequin within the newest updates.”
Requested differently, I used to be capable of get extra perception into enhancing the accuracy of an AI mannequin:
“Generally, when an AI mannequin like ChatGPT is up to date, it might purpose to enhance its efficiency in a number of elements. For instance, it might have a bigger dataset to coach on, or enhancements in its structure that enable it to generate smoother and extra correct responses. And such updates additionally enhance info, that’s, his capability to know and generate responses based mostly on factual data.”
Accuracy stays questionable
OpenAI claims that ChatGPT can now give extra factual solutions.
I examined this declare by consulting a GitHub repository of ChatGPT errors and operating a number of prompts to see if there can be totally different solutions.
Check one: Failed
Beforehand, ChatGPT couldn’t precisely decide what number of occasions Argentina has gained the FIFA World Cup.
Other than the 2022 World Cup win, since ChatGPT’s data is restricted to 2021, it ought to say that Argentina gained it twice. As soon as in 1978 and once more in 1986.
As proven within the following tweet, ChatGPT didn’t at all times return the right reply:
Requested #ChatGPT who gained the 2022 FIFA World Cup. There may very well be no reply. That’s anticipated. Nevertheless, it appears to be giving mistaken data (concerning the different 2 wins) though the data is there within the system. Any #explanations? pic.twitter.com/fvxe05N12p
— indranil sinharoy (@indranil_leo) December 29, 2022
I ran the immediate via the up to date model of ChatGPT and it returned a distinct one, however nonetheless not appropriate Reply.
Screenshot from: chat.openai.com/chat, January 2023.
Check 2: Failed
Beforehand, ChatGPT could not present an accurate reply when requested who’s the taller basketball participant between Shaq and Yao Ming.
I ran the command immediate via the up to date model of ChatGPT and confidently it returned the identical factor not appropriate Reply.
Screenshot from: chat.openai.com/chat, January 2023.
Going via the ChatGPT errors linked above, I discovered that it continues to wrestle with the identical prompts.
It is tough to pinpoint the areas the place ChatGPT can present extra correct solutions. It will be useful if OpenAI may present specifics within the launch notes of future updates.
Nevertheless, watch out when utilizing ChatGPT as a supply of data. Whereas it solutions many questions accurately, it is at present not dependable sufficient to switch Google.
supply: OpenAI
Featured picture: CHUAN CHUAN/Shutterstock