OpenAI this week launched GPT-4, the latest and most powerful and versatile version of its large language model that has captured the imagination of the tech world and beyond.
The new model, introduced Tuesday, is multimodal, meaning it accepts both image and text inputs and releases text outputs.
The large language model was trained on a supercomputer, which the San Francisco-based AI research lab and vendor co-designed with Microsoft to run on Azure.
GPT-4 has passed numerous academic and professional benchmarks, including the bar exam, according to OpenAI, in which Microsoft has pledged to invest $10 billion.
While the vendor launched the upgraded version of the LLM today, it already has allowed several organizations to test the technology for applications in their industries.
Casetext and GPT-4
With GPT-4, CoCounsel enables lawyers to delegate some of the work they might have had to do themselves. This includes reading documents, writing research memos and memorandums and answering questions based on different case statutes.
"We were absolutely blown away," CEO Jake Heller said. "It was obvious that it could do work that GPT-3 and 3.5 could not."
The first capability of GPT-4 that impressed Heller was its ability to read and accurately answer questions based on a given text, he said.
"A lot of legal work is about accurately describing what's in a document and potentially answering questions about it," Heller said. "And when we saw it could do that extremely well, very consistently, that got us really excited."
Heller and his team were also impressed by GPT-4's ability to handle complex and challenging language within a legal document, he said.
Jake HellerCEO, Casetext
Moreover, the team members were struck by the LLM's intelligence.
"Its understanding of what you're saying and asking it to do, and its quality of answer -- all of that is just incredibly impressive," Heller said.
Nevertheless, GPT-4 has room for improvement, Heller said.
For example, he said the technology needs to recognize better when a prompt or query is too vague. This will help keep the answers factual.
"Building that kind of conversation piece before it kicks off a big job and starts spending minutes to research and answer questions, we think is a necessary improvement to giving it the context it needs to answer the question well," Heller said.
And, he added, the model is also neither great at math nor funny.
ChatGPT and GPT-4 restrictions
OpenAI will allow ChatGPT Plus subscribers to gain access to GPT-4 with a usage cap.
While the technology will make ChatGPT more powerful, there will still be some limitations with data, said Forrester analyst William McKeon-White. Apparently, GPT-4 is still bound by the time limitation of being able to access knowledge generated up until only the end of 2021.
"The add-ons are going to be the most transformative elements of it because it still seems to be running off that 2021 database," he said.
However, the inclusion of GPT-4 in ChatGPT should improve the overall accuracy and memory of ChatGPT, McKeon-White added.
Despite the advancements in GPT-4, OpenAI said it still wants users to be cautious. OpenAI said the LLM poses similar risks as previous models in generating harmful advice, inaccurate information and buggy code.
"That this even more powerful LLM is released at the same time that tech companies are shedding AI ethicists is problematic," said Michael Bennett, director of the education curriculum and business lead for responsible AI at Northeastern University, referring to Microsoft's recent elimination of its AI ethics unit.
Meta and Twitter also have disbanded similar internal groups over the past year.
And despite its innovative qualities, the updated model will likely also experience problems such as hallucinations and racial and gender bias.
Anyone wishing to access the GPT-4 API can sign up for OpenAI's waitlist.