Insights

What is ChatGPT?

Elon Musk co-founded OpenAI, an artificial intelligence (AI) research organisation, which has since resigned. At the end of 2022, OpenAI unveiled a convincing AI-powered chatbot capable of communicating in plain English, using updated versions of its AI model. As the field of natural language processing advances, Generative Pre-training Transformer (GPT) models such as GPT-3 and GPT-3.5 have the potential to transform a wide range of industries and applications.

GPT-3

GPT-3 was released by OpenAI in May 2020. Trained on a dataset of over 570GB of text data, making it one of the largest language generation models at the time. Literature from Brown et al., (2020) demonstrate how scaling up language models can significantly “improve task-agnostic, few shot performance”. GPT-3 achieved cutting-edge results on language understanding benchmarks such as language translation, text summarization, and question answering (Brown et al., 2020). GPT-3 also demonstrated the ability to perform a wide range of tasks with minimal fine-tuning, including essay writing, computer program creation, and even poetry composition (Brown et al., 2020).

In a recent study by OpenAI, when comparing GPT-3 with other models, the study found that the model was able to answer 70.8% of questions correctly, as opposed to a BERT-based model which answered at 56.4% and a human performance at 83.1% (Raffel, 2020).

GPT-3.5

GPT-3.5 is based on GPT-3, but it operates within guardrails, an early prototype of AI aligning with human values by forcing it to follow policies. GPT-3.5 models use the same pre-training datasets as GPT-3, but with additional fine-tuning. This fine-tuning stage incorporates a concept known as ‘reinforcement learning with human feedback,' or RLHF, into the GPT-3 model (Thompson, 2022).

Microsoft's Azure AI platform was used to train all of the models. Microsoft is a significant investor in OpenAI (OpenAI, 2022) and consequently invested 10 billion euro into Chat GPT this month (Browne, 2023).

GPT-4

With rumours of GPT-4 expected to be released early 2023; built upon its success with GPT-3 and GPT-3.5, GPT-4 will be significantly larger and more powerful, with predictions of 100 trillion parameters in comparison to GPT-3 with 175 billion (Romero, 2021). It promises the breadth of generalist systems like GPT-3 as well as the depth of specialist systems like DALLE (text-images) (general language). Expecting greater accuracy through better optimisation and improved RLHF better human supervised training could help to reduce misinformation (Tech Desk, 2023).

Chat GPT’s Success

For starters, this AI can not only generate paragraphs of well-written English (or French, or Mandarin, or whatever language you prefer), but it can also generate blocks of computer code on command. The bot can create simple webpages and applications in programming languages such as JavaScript, Python, and React. It can also detect code bugs and aid in the creation of new programming languages (Glen, 2022).

In healthcare, GPT can aid in the analysis of large amounts of patient data, which can then be used to improve diagnosis accuracy and treatment efficacy. GPT can also assist in the generation of patient-specific medical reports and documents, potentially leading to a more efficient healthcare system.

GPT can be used in business applications for both internal and external communication between companies and their customers or partners. Customer support, sales support, IT support, and other processes where employees must manage large amounts of information about their customers can benefit from the technology. In the field of content creation via natural language processing (NLP). GPT has already shown the ability to generate human-like text on a variety of topics, making it a valuable tool for businesses and organizations looking to automate the writing of articles, social media posts, and other types of content. This can result in significant cost savings for businesses while also improving the efficiency and scalability of content creation. Customer service, as previously mentioned; is another area where GPT has the potential to be used in the future. GPT models, have shown the ability to understand and respond to natural language input, making them ideal for use in chatbots and virtual assistants. This could potentially lead to customers experiencing a more personalized experience whilst being more time efficient.

Many businesses use GPT for internal communications between departments within their organization or even between different locations around the world. Employees from different locations can collaborate on projects without having to travel back and forth between offices or countries every time they require new information from another person in their department or company. Such as creating e-learning materials or creating clinical notes; to generate personalised treatment plans or summaries with less human error.

Chat GPT’s Drawbacks

Like any other machine learning model, ChatGPT suffers from certain limitation and limitations users should be aware of.

Dependence on Data. The model can occasionally produce plausible-sounding but incorrect and illogical responses. When Meta released its Galactica science model, researchers protested over the same issue (Bastian, 2022). If the model is not trained on a diverse and extensive dataset, it may produce irrelevant or inaccurate results.

Bias. Machine learning models, such as ChatGPT, can exhibit bias in their responses at times. The model is frequently overly verbose and overuses certain phrases, such as repeating that it is an OpenAI-trained language model. These problems stem from biases in the training data (trainers prefer longer answers that appear more comprehensive) and well-known over-optimization problems.

Limited Understanding. While ChatGPT can respond to prompts with high accuracy and fluency, it lacks a deep understanding of the world and the ability to reason like a human (ChatGPT Pro, 2022). When a user submits an ambiguous query, the model should ask clarifying questions. Instead, our current models frequently guess what the user meant (OpenAI, 2022). Generating very logical answer to the end user but is complete nonsense and actual incorrect.

Concluding point

In conclusion, ChatGPT have the potential to revolutionize a wide range of industries and applications in the future. However, it's important to consider the certain drawbacks potentially leading to misinformation, and to develop ways to mitigate these concerns. While GPT is still a work in progress, ongoing research and development in this area has the potential to lead to a wide range of benefits for society. Lets see what GPT-4 has to offer upon its release.