ChatGPT Maker OpenAI Announces GPT-4, Its Most Advanced System Yet

OpenAI, the artificial intelligence (AI) research and deployment company behind the conversational chatbot ChatGPT, has announced the latest in its line of AI language models, GPT-4.

Like previous GPT models, the GPT-4 base model has been trained using publicly available data (such as Internet data) as well as data with OpenAI. It was trained on an Azure AI supercomputing infrastructure. Its training was wrapped up in August 2022.

Six months of “iteratively aligning GPT-4,” drawing lessons from earlier programmes like ChatGPT, has led to OpenAI’s “best-ever results (though far from perfect),” according to OpenAI.

GPT-4 doesn’t have real-world human capabilities — obviously — but “exhibits human-level performance on various professional and academic benchmarks,” the company says about its latest effort to scale up deep learning.

OpenAI is releasing GPT-4’s text input capability via ChatGPT and the API (with a waitlist), and preparing the image input capability for wider availability. “Image inputs are still a research preview and not publicly available,” the company has said. 

GPT-4 is available to ChatGPT Plus users, who are on a monthly subscription plan with OpenAI.

ChatGPT was announced on 30 November 2022. It was launched as a research preview in order for OpenAI to learn more about the strengths and weaknesses of its system as well as gather user feedback to identify and limit its limitations. Safe to say, it blew up as the buzz grew ever stronger.

Now, GPT-4 promises to do and be more. For example, it can accept and generate up to 25,000 words of text, which is more than eight times that of ChatGPT’s 3,000 words. It can also understand images and express logical ideas about them.

The chatbot is able to accept a prompt of text and image, but responds only in text, such as natural language and code. This gives it the label of a “multimodal model”.

GPT-3.5, which finished training in early 2022, also on the Azure superconducting setup, was an improved version of GPT-3, announced in 2020. The latter was the first demonstration of an AI model conversing like an error-prone human and later GPT-3.5 powered ChatGPT.

GPT-4 is a further improvement on GPT-3.5. According to OpenAI, “GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5.”

The company has contrasted GPT-4’s abilities against that of its predecessor with an example: the new chatbot is said to have passed a simulated bar exam with a score that puts it within the top 10 per cent of test takers; GPT-3.5’s score, on the other hand, would relegate it to the bottom 10 per cent.

GPT-4 was also found to considerably outperform existing large language models when evaluated on traditional benchmarks designed for machine learning models.

Yet, it does carry forward some of the limitations of the earlier GPT models. As OpenAI chief executive Sam Altman wrote in a tweet, “it is still flawed, still limited,” and still more impressive than before.

“Most importantly, it still is not fully reliable (it “hallucinates” facts and makes reasoning errors),” OpenAI says, adding that “great care should be taken when using language model outputs, particularly in high-stakes contexts...”

The model continues to have various biases in its output, “generally lacks knowledge of events that have occurred after the vast majority of its data cuts off (September 2021), and does not learn from its experience,” OpenAI acknowledges.

Further, GPT-4 users can expect simple reasoning errors and a ready acceptance of obvious false statements at times. The chatbot can also be confidently wrong in its predictions.

Notably, now that GPT-4 has picked up more advanced capabilities, the risk of its misuse, too, is enhanced. OpenAI realised this would be the case and thus engaged experts in various domains, such as cybersecurity and international security, to adversarially test model behaviour in high-risk areas. 

For example, says the company, “we’ve collected additional data to improve GPT-4’s ability to refuse requests on how to synthesize dangerous chemicals.”

The company says they have also decreased the model’s tendency to respond to requests for disallowed content by 82 per cent compared to GPT-3.5 and made the model more aligned with their policies for sensitive requests, such as involving medical advice or self-harm.

Despite its limitations, naturally, the hopes are high for GPT-4. “We think that GPT-4 will be the world’s first experience with a highly capable and advanced AI system. So, we really care about this model being useful to everyone, not just the early adopters or people very close to technology,” an OpenAI executive says in a video introducing GPT-4.

Education has been billed as an area that would feel the impact of GPT-4 greatly, as the latest model can teach a huge range of subjects. It would be very useful to school kids who may be learning or revising at home.

OpenAI says it has already partnered with companies like Duolingo (language-learning app) and Khan Academy (education company) to integrate GPT-4 into their products.

Ever since ChatGPT was launched, the competition in the AI-driven chatbot space has heated up. Google in early February unveiled its own offering to rival Microsoft-backed ChatGPT. “Bard” is said to leverage Google's own Language Model for Dialogue Applications (or LaMDA).

The Google AI chatbot will draw "information from the web to provide fresh, high-quality responses", Sundar Pichai has said, which may suggest that it could pull up information from recent events — not a ChatGPT strong suit.

Rest assured, innovating in AI modelling is set to continue.