The AI behind popular chatbot ChatGPT has been updated to a new version known as GPT-4 – and many people have already been unknowingly exposed to the newest AI’s supposedly improved capabilities for weeks prior to the announcement.
OpenAI, the company that developed , says it “spent 6 months making GPT-4 safer and more aligned” so that the AI is less likely to produce “disallowed content” in response to human users’ queries. GPT-4 delivers “human-level performance” and outperforms its predecessor, GPT-3.5, on many simulated exams for university admissions and professional fields such as law and medicine, according to an and technical report. For example, GPT-4 passed a simulated bar exam for lawyers with a score among the top 10 per cent of test takers, whereas GPT-3.5 previously scored in the bottom 10 per cent.
The technical report about GPT-4 shared minimal details about the AI’s architecture and hardware or the computing power requirements for training and the data used to train it. OpenAI described this lack of disclosure as being due to “the competitive landscape and safety implications of large-scale models like GPT-4”.
Advertisement
But the missing information also makes it difficult to independently check claims about GPT-4’s performance accuracy and safety. “If anything, OpenAI has provided even less detail with which to evaluate its new system,” says at the AI Now Institute, a research centre in New York. “But what we’ve seen so far in the roll-out of generative AI is that these systems are prone to error… so we should be wary of any claims that these issues are resolved.”
OpenAI itself reiterated the increasingly familiar warning about even the most advanced language AIs by cautioning that GPT-4 can still “hallucinate” facts and make reasoning errors – meaning that “great care should be taken when using language model outputs” and especially in “high-stakes contexts”.
Microsoft, a major backer of OpenAI with a , also revealed that its Bing search engine chatbot had already been running on GPT-4 for five weeks prior to the public 14 March announcement. OpenAI’s announcement also highlighted cases of GPT-4 now being used by the language-learning app Duolingo, the payment service Stripe, the online learning service Khan Academy and the financial services firm Morgan Stanley.
The rollout of the updated AI service comes as Silicon Valley tech giants such as Microsoft and Google – along with Chinese companies such as Baidu and Tencent – are competing in an AI arms race to develop and deploy AI chatbots capable of understanding language and performing many different tasks based on written prompts.
