According to OpenAI, the latest version of the groundbreaking AI system that powers ChatGPT, should be more creative, less likely to make up facts, and less biased than its predecessor.
GPT-4 is “multimodal”, meaning images as well as text can be entered to generate content, prompting concerns that AI tools could be used by criminals to generate deep fake photos, videos or audio files.
GPT-4 upgrade
For now the text-input feature will be available to ChatGPT Plus subscribers and to software developers, with a waitlist, while the image-input ability remains a preview of OpenAI’s research.
Concerns have also been raised about the ethics of using AI tools for creating academic papers or published works, since the material generated may be plagiarised from other sources.
Teachers have reportedly begun spotting AI generated homework assignments as well.
Into this OpenAI has said that GPT-4 is “less capable than humans in many real-world scenarios,” but it “exhibits human-level performance on various professional and academic benchmarks.”
“We’ve spent 6 months iteratively aligning GPT-4 using lessons from our adversarial testing program as well as ChatGPT, resulting in our best-ever results (though far from perfect) on factuality, steerability, and refusing to go outside of guardrails,” it said.
“Over the past two years, we rebuilt our entire deep learning stack and, together with Azure, co-designed a supercomputer from the ground up for our workload,” OpenAI stated. “A year ago, we trained GPT-3.5 as a first “test run” of the system. We found and fixed some bugs and improved our theoretical foundations. As a result, our GPT-4 training run was (for us at least!) unprecedentedly stable, becoming our first large model whose training performance we were able to accurately predict ahead of time.”
“We are releasing GPT-4’s text input capability via ChatGPT and the API (with a waitlist),” it said. “To prepare the image input capability for wider availability, we’re collaborating closely with a single partner to start. We’re also open-sourcing OpenAI Evals, our framework for automated evaluation of AI model performance, to allow anyone to report shortcomings in our models to help guide further improvements.”
So what improvements does GPT-4 deliver over the previous version?
Well the new version can process up to 25,000 words, about eight times as many as ChatGPT, and the new model can also respond to images.
For example it can provide recipe suggestions from just a photo of ingredients.
GPT-4 will initially be available to ChatGPT Plus subscribers, who pay $20 per month for premium access to the service.
Chatbot concerns
The widespread usage of chatbots such as ChatGPT has already generated plenty of feedback from cybersecurity experts, but the addition of imagery has prompted fresh concerns.
“Allowing AI to create multimedia outputs from simple text inputs comes with a worrying new level of cyber threats,” said Jake Moore, global security advisor at ESET.
Jake Moore, ESET
“Deep fakes are already making headway with targeting companies, but enabling new multimodal models at the fingertips of threat actors potentially opens up a new door to more security fears,” Moore cautioned.
“Not only could companies be targeted with fake video and audio but there is the potential of this technology being used among nation state armoury to target national security,” said Moore. “Although the technology may still be young and not be completely fool proof in this version, it will undoubtedly knock trust levels across nations and regions with inevitable illicit use.”
“Seeing can no longer be relied upon as believing but helping people to question what they see will be a large scale task,” Moore concluded.
Another expert touched upon AI chatbots within the academic, legal and software development environments.
“Strikingly, the launch paper highlights GPT-4’s capabilities against various human academic and professional exams, with the AI exceeding the scores of the vast majority of human test takers, with a big increase in performance in sciences, maths and economics,” noted Dr Andrew Rogoyski, from the Institute for People-Centred AI at the University of Surrey.
“Notably, it passes a simulated version of the Uniform Bar Examination with a score ranking in the top 10 percent of test takers,” said Dr Rogoyski. “Buried in the paper is a comment, ‘we expect GPT-4 to impact even jobs that have historically required years of experience and education, such as legal services’, something that will send a chill up the spines of legal firms worldwide.”
“It’s also better at computer coding, which will open up another avenue for commercial exploitation,” said Dr Rogoyski. “Imagine just having to ask your computer to create a programme that, for example, predicts how your latest product will perform in its market. No programming required. This could be transformational for many businesses.”
