Late last year, a chatbot called ChatGPT caused a splash because of its ability to tackle complex subjects and also to generate speeches, songs and essays. Now, its creator, OpenAI, has launched a large language model called GPT-4 that it claims is more creative and collaborative than its predecessor, GPT-3.5, which underpins ChatGPT.
The new AI is able to recognise images and provide answers in text. Although image description is arguably its most eye-catching new feature, OpenAI is not releasing that functionality at launch due to fears it could be misused.
Like the original ChatGPT, the new bot is not fully reliable and may “hallucinate” — a phenomenon where AI invents facts or makes reasoning errors.
Here’s what you need to know about GPT-4, including its advancements over GPT-3.5, where you can use it, and how accurate its answers are.
How to use GPT-4?
For now, GPT-4 is only available to those who pay $20 (£16.50) per month for ChatGPT Plus, the subscription service OpenAI launched in February. However, there is a cap of 100 messages every four hours.
The premium plan includes access to its chatbot when demand is high — which it was at the time of writing, probably due to the launch — as well as faster response speeds and priority access to new features.
But, there is a way to access the tool free. In conjunction with the new bot’s launch on Thursday, Microsoft revealed that its Bing AI has been running on GPT-4 all along. The company’s chatbot is currently available in preview, meaning you have to join a waiting list to try it.
Microsoft said there are things you can do to skip ahead in the queue, including making Edge your default web browser, making Bing your default search engine, and making MSN your homepage.
While you wait, you can still use the regular ChatGPT free on OpenAI’s site, and its GPT-3.5 model continues to power chatbots from Snapchat and Slack.
Other companies have confirmed they will use GPT-4 for their products, including language app Duolingo, payments provider Stripe and educational non-profit Khan Academy. More tools built on top of GPT-4 could be in the offing as OpenAI is making its toolset available to developers as an API or application programming interface.
What can GPT-4 do?
GPT-4 boasts several improvements over the free version of ChatGPT. These include the ability to provide more creative answers, deal with longer documents of up to 25,000 words and visual inputs.
Open AI said the system’s evolution is apparent in its performance on a total of 34 academic and professional exams. These include the Uniform Bar Exam, LSAT, SAT Math, and SAT Evidence-Based Reading and Writing exams. In more than half of these tests, it scored higher than the 80th percentile. It also reached the top 90th percentile in the Uniform Bar Exam, versus 10th previously.
GPT-4 also performs better than GPT-3 on a range of machine learning benchmarks, including vision tasks.
How much smarter is GPT-4 than GPT-3.5?
OpenAI said the new system was trained on more data than its predecessor — but it did not divulge how much information was used. These training data components are known as parameters and essentially indicate the skill of the model on a problem, such as generating text.
The largest model in GPT 3.5 has 175 billion parameters. By comparison, Meta recently released its own language model to researchers, which has a maximum of 65 billion parameters.
But, the chatbot still has its limitations. As with GPT-3.5, it lacks knowledge about events “that have occurred after the vast majority of its data cuts off” in September 2021.
Can GPT-4 recognise images?
Notably, GPT-4 can respond to images with answers in natural, conversational text. But, as mentioned above, this feature is still locked away from the public because of safety concerns.
On its website, OpenAI showed the bot replying to an image of eggs and flour with a range of meals and desserts that can be made with the ingredients, including pancakes, crepes, quiche and omelettes.
In a demo, the chatbot was reportedly shown an image from the Hubble Space Telescope and asked to describe the photo “in painstaking detail”. It replied with a four-paragraph description, including an explanation of the ethereal white line that stretched across the photo. A “trail from a satellite or shooting star”, the chatbot wrote.
The expert view on GPT-4
AI expert Dr Peter van der Putten said GPT-4 pushed the envelope for artificial intelligence. He said it laid down the gauntlet for rivals including Google’s upcoming Bard chatbot and Meta’s own AI tech.
“In a nutshell, just the increased performance on the various exams, machine learning test and safety checks of GPT-4 will make it worthwhile moving to this new model,” said Van der Putten, who heads the AI lab at software company Pegasystems, and is also a part-time assistant professor in AI at Leiden University in the Netherlands.
He continued: “And even though not a lot of technical or research innovations were disclosed, OpenAI has set the bar again.”
Is GPT-4 safe?
OpenAI’s decision to hold back GPT-4’s image recognition feature may concern AI sceptics. For its part, the bot’s creator insists that it is safe to use, and has been put through strenuous security checks. These include six months of safety training, with internal tests showing that it was “82 per cent less likely to respond to requests for disallowed content and 40 per cent more likely to produce factual responses than GPT-3.5”.
However, as it was trained on online information, it is still prone to inaccurate responses, social biases and downright odd replies. The last of these largely depends on how you treat the bot; if you decide to use shock tactics in your questions, you can expect the conversation to take a creepy turn.
GPT-4 can also understand follow-up questions, admit its own mistakes and limitations, and reject inappropriate requests.
UK cybersecurity agency warns of chatbot risks
With AI all the rage, the UK’s security apparatus has been forced to address the tech. The National Cyber Security Centre — an arm of the UK’s intelligence, security and cyber agency — has warned that sensitive queries stored by chatbot providers could be hacked or leaked. It also said the tech could be used to create more convincing phishing emails, and allow attackers access to tools they weren’t previously familiar with.
