Artificial intelligence company OpenAI has announced a new version of its chatbot ChatGPT, called GPT-4.
The company says it has “broader general knowledge and problem solving abilities”.
Here’s what we know.
The recent wave of ‘chatbots’, including ChatGPT and Google’s Bard, are digital tools that can give original answers to user questions.
They are ‘trained’ on a large database of online text, allowing them to give realistic-sounding answers. ChatGPT’s last edition could write essays, recipes, emails and jokes to a reasonable standard.
However, by its own admission, it did not ‘understand’ what it wrote and often produced nonsense answers.
What’s in GPT-4?
According to OpenAI, the new GPT-4 can do a wider range of tasks and is capable of more advanced ‘reasoning’.
It can respond to visual prompts (e.g. writing a recipe from a picture of ingredients).
Unlike previous versions, it can also learn from input you give it. This means for example that it could learn and copy your writing style if you gave it examples of your writing, or could summarise a paper or article you ask it to read.
OpenAI also claims it is 40% more likely to produce factual responses and 82% less likely to respond to unsafe requests.
It performs much better on a range of cognitive tests than the previous version. Whereas ChatGPT would rank in the bottom 10% of humans taking the Uniform Bar Exam for U.S. lawyers, GPT-4 would be in the top 10%.
Can I use it?
GPT-4 can be tested on the OpenAI website for paid customers only.
GPT-4 is also being used to power the chatbot on Microsoft’s Bing search, via its browser Edge.
Some companies have also started using GPT-4 in their own products, including Duolingo, which is testing its use to have more interactive conversations with language learners.