GPT-5, the long-awaited upgrade to the ChatGPT AI chatbot, has been released by its maker OpenAI.
It has been one of the most highly anticipated launches in Silicon Valley after OpenAI’s first offering ChatGPT – powered by its GPT-3 model – kick-started the current AI boom in late 2022.
“GPT-3 sort of felt like talking to a high school student,” said Sam Altman, OpenAI’s chief executive.
“GPT-4, maybe it was like talking to a college student. But with GPT-5, now it’s like talking to an expert, a PhD-level expert in anything, any area you need, on demand.”
At the launch event, OpenAI claimed the new chatbot, which will be released to all ChatGPT users on Thursday, was more than a simple upgrade to its previous offerings.
According to OpenAI, the new model exceeds the chatbot competition from the likes of Google, X and Antropic on “benchmarks” – standardised tests used to rank models.
OpenAI claims it has been designed to be easier and more natural to communicate with, better at writing prose and advanced computer code, solving academic questions from mathematics to law, assisting with healthcare-related questions, as well as being safer than its predecessors.
“It’s an incredible superpower on demand,” claimed Mr Altman.
The model is also more intelligent in how it uses its own brain power – and therefore an expensive computing resource – according to OpenAI.
It is a hybrid of previous chatbots and slower, more computing-intensive “reasoning” models like OpenAI’s Deep Research.
Based on a user’s request, the model will decide how much “thinking” is required before answering, rather than requiring the user to switch between different models.
Although AI enthusiasts who had been expecting GPT-5 to represent “artificial general intelligence [AGI]” will be disappointed.
Despite this being OpenAI’s stated goal, Mr Altman billed GPT-5 as a “major upgrade” to GPT-4 and a “significant step along the path to AGI”.
But they’re clearly not there yet.
A real test of GPT-5 will be whether it sells.
OpenAI is projected to spend $8bn (£6bn) this year, on top of $5bn (£3.7bn) last year, and while it is expected to make a profit this year, the business case for increasingly powerful AI models is still not clear to many investors.
Given a single training run for GPT-5 is rumoured to have cost $500m (£373m), there will be an expectation the new model is significantly more useful to business users.
Despite a very slick demonstration of its coding skills at the launch presentation, where it built an online language learning game in seconds, GPT-5 will have to prove its worth for professional coding.
Many in the tech industry prefer Anthropic AI’s Claude model to write code. OpenAI and its investors will be hoping GPT-5 changes that.
AI experts will also be testing GPT-5’s tendency to “hallucinate”, an issue OpenAI claims to have improved with GPT-5.
But erroneous or bizarre answers are a problem that dogs all large generative AI models.
“Shiny things are always fun to play with, and I fully expect GPT-5 to be the shiniest so far,” said Gary Marcus, a cognitive scientist at New York University and AI commentator.
“But that doesn’t mean that it is a critical step on the optimal path to AI that we can trust.”
Source: https://news.sky.com/