Two weeks into its release, ChatGPT-4 is showcasing significant improvements over GPT3.5

ChatGPT is good at many complex NLP tasks because it has been trained on a lot of knowledge and has been fine-tuned based on human feedback. For example, it is good at translating natural language to code, completing highly masked text, and making up stories based on user-defined elements and styles.

It is also good at typical NLP tasks like text classification, entity extraction, translation, etc. Also, the researchers have trained ChatGPT in a conversational manner. ChatGPT can respond to follow-up inquiries, acknowledge mistakes, challenge incorrect assumptions, and reject improper requests.

All told, when OpenAI launched ChatGPT it was not a sudden leap in machine learning but a steady evolutionary process over months and years that led to enough confidence among OpenAI honchos to go public with their revolutionary product some months ago.

Generative pre-trained transformers (GPT) are a family of OpenAI language models typically trained on a huge corpus of text data to generate text that resembles human language.

2018- GPT-1 (It learns a large corpus of continuous text.)

2019-  GPT-2 (It translates, answers queries, summarizes, and generates text.)

2020- GPT-3 (It produces human-like text and prompt-continuing text.

2023- GPT-4 (It  is a multimodal large language model and it is 10 times more advanced than its predecessor, GPT-3.5.)

Enter ChatGPT-4

It is worth noting that the evolutionary process that is now giving us GPT-4 has come within five months of the sensational November launch of its predecessor, the GPT3.5 version. GPT-4 is better than GPT-3.5 for a primary reason: Its ability to produce more creative answers to questions. GPT-3.5 is, of course, very creative. But GPT-4 makes things even more sophisticated.

Try asking both models to do something creative, like write a poem with lines in both English and French, and ChatGPT powered by the latest GPT-4 model will do a better job. For example, GPT-4's answer would use both languages on every line, but GPT-3.5 would switch between them instead, using one language on one line and the other on the next.

In the same way that GPT-3.5 can only respond to text prompts, GPT-4 is multi-modal and can respond to both text and visual prompts. To be clear, when we say "visual inputs," the image doesn't have to be a picture of a typed prompt. It can be a picture of anything. So, GPT-4 can understand and describe almost any image, from a handwritten math problem to a Reddit meme.

GPT-3.5’s Limitations

One of GPT-3.5's weaknesses is its inclination to create nonsensical and false information confidently. It is referred to as "AI hallucination" in AI lingo and might lead to mistrust of AI-generated knowledge.

In GPT-4, hallucinations continue to be an issue. Yet, the GPT-4 technical study indicates that the new model is 19 to 29 percent less likely to experience hallucinations than the GPT-3.5 model.

Even apart from what the technical study says, the GPT-4 model's responses on ChatGPT have been observed to be substantially more factual. Likewise, the context window and context size between GPT-4 and GPT-3.5 have also seen an improvement.

A context window is how much information a model can "remember" and for how long during a chat session. The context size and window of GPT-4 is much better than those of its predecessor. In real life, this means that GPT-4 can remember the context of a conversation and the instructions given during the conversation better and for longer.

Glimpse Of The Future

Researchers have recently asserted that GPT-4 passed the bar exam in the United States. In their paper, the researchers compare the zero-shot performance of a preliminary version of GPT-4 to that of previous generations of GPT on the entire Uniform Bar Examination (UBE), including not only the multiple-choice Multistate Bar Examination (MBE) but also the open-ended Multistate Essay Exam (MEE) and Multistate Performance Test (MPT) components.

GPT-4 greatly surpasses human test-takers and earlier models on the MBE, displaying a 26 percent improvement over ChatGPT and outperforming humans in five of seven topic areas. These results demonstrated not just the rapid and spectacular progress in the performance of large language models in general but also the potential for such models to help deliver legal services in society.

Without a doubt, GPT-4 is a significant improvement over the models that came before it. Even though it still has some of GPT-3.5's problems, the model is a big step forward in the quest for brilliant AI language models because it is much better than GPT-3.5 in many ways and has new features.

Nivash Jeevanandam writes stories about the AI landscape in India and around the world, with a focus on the long-term impact on individuals and society. Views are personal and do not represent the stand of this publication.