OpenAI releases GPT-4 to build upon the success of ChatGPT
OpenAI has recently released a new update to its ChatGPT service, moving it from GPT-3.5 to the new and improved GPT-4 model.
The new model has a few new features whilst also massively building upon the existing ones that made it so popular in the first place. The big new feature of the GPT-4 model is its ability to respond directly to image prompts in a variety of ways. This now makes the GPT-4 a ‘multimodal’ model, meaning it can accept both images and text inputs. For example, the new model, when given an image prompt, could provide a caption, description, respond to questions, or even something like a recipe from an image of a bunch of ingredients.
But it’s not just new features that are making the GPT-4 model of ChatGPT impressive, the team at OpenAI have massively expanded upon the base abilities of ChatGPT. The new model can now process up to 25,000 words at once, a staggering eight times as many as its predecessor, as well as produce a similar level of words for a response.
OpenAI has spent a significant amount of the last six months on improving the safety features for GPT-4, training it on a large amount of human feedback. Despite the progressions made, the team still warned that it could still be prone to making errors or sharing misinformation. Whilst GPT-4 is certainly still a significant step, it is not fully reliable and, like the versions that came before, it may still ‘hallucinate,’ a strange issue that generative AI has where it might invent facts or make critical reasoning errors. The new model is, however, hoped to be more resistant to biases, something that AI has recently been heavily criticised for, addressing biases such as gender or race. GPT-4 is also much better at not falling into the sensitive topic pitfalls that its predecessors often would, building on the data from prior experiences. In sensitive topics such as medical or self-harm advice, the AI responds ‘29%’ more sensitively and refrains from answering disallowed requests ‘82%’ more often.
Initially, the GPT-4 version of the AI will only be available to those who are signed up to ChatGPT Plus, the $20-a-month paid premium subscription version of the service.
OpenAI has stated that GPT-4 has “more advancing reasoning skills,” when compared with the GPT-3.5 model.
Alongside the GPT-4 project, OpenAI has worked closely with commercial partners such as Duolingo to produce additional GPT-4 powered applications. This collaborative effort will be seen in Duolingo’s new ‘Duolingo Max,’ which will offer its English-speaking users an option for AI-powered conversations in French or Spanish, whilst making further use of GPT-4’s power to explain errors learners may have made.
“Artificial intelligence has always been a huge part of our strategy,” says Edwin Bodge, Principle Product Manager at Duolingo. “We had been using it for personalising lessons and running Duolingo English tests. But there were gaps in a learner’s journey that we wanted to fill such as conversation practice, and contextual feedback on mistakes.”
Duolingo has managed to not only now achieve these aims through the use of GPT-4, but also complete them with a significantly higher level of accuracy and efficiency.
It won’t be long until we hear of the next major advancement into GPT-4.5 or even GPT-5, with the service’s reach and popularity showing little signs of slowing down.