Simply Tech - ChatGPT – GPT 4

In this series of “Simply Tech” we cover about ChatGPT which is having an “iPhone moment”.

What is ChatGPT?

ChatGPT is an AI based language model developed by OpenAI. ChatGPT uses a Generative Pre-trained Transformer (GPT) large language model that is based on a neural network architecture for Natural Language Processing.?

‘Generative’ – is the ability to generate new text.

‘Transformer’ – type of a neural network architecture to process sequential data like languages.

‘Pre-trained’ – the model learns structure of a language and statistical patterns from large data sets and in turn generates new texts.?

A neural network for Natural Language Processing can be thought of as a brain of the computer system. It processes words, sentences, and tries to understand the meaning, just as our brain does when we read or hear language. This large language model is based on a neural network architecture trained on vast amounts of data to learn patterns and relationships between words and phrases.

Background

GPT-3 was the language model released in 2020 and was trained on 570GB of text data and had 175B parameters, making it one of the largest and most powerful language processing models to date. The evolution from GPT-1 in 2018 to today has been a marvel that equates to the ‘iPhone moment’. ChatGPT reached a million users in just five days compared to TikTok which took 14 days to achieve the same.

Capabilities

• Ability to have a casual conversation and continue conversations without repeating the history of the topic.
• Allows to make follow-up corrections to previous discussions.
• Ability to recommend code snippets.
• Translate text.
• Answer questions.
• Solve difficult problems with certain level of accuracy.
• Responsible AI established with certain guardrails to ensure proper adversarial prompts, unwanted content and privacy concerns are appropriately addressed.

Limitations

• Not completely dependable. OpenAI states, “it “hallucinates” facts and makes reasoning errors”.?
• ChatGPT may be prone to sharing disinformation.
• Currently, no knowledge of events that have happened after the data cut-off date (September 2021).
• Is susceptible to simple reasoning errors and fail at hard problems, such as introducing security vulnerabilities in code snippets.?
• Might generate harmful advice.
• Might produce codes with bugs.
• The established guardrails, adversarial prompts, unwanted content, and privacy concerns could potentially be exploited.
• Absence of citation of the information source.

GPT-4

GPT-4 is the successor of GPT-3, released on March 14, 2023. OpenAI has further enhanced the capabilities of ChatGPT to enable more complex tasks. OpenAI developed supercomputers built for purpose and used deep learning stack to train GPT-3 followed by GPT-4. Deep learning is a technique that uses artificial neural networks to solve complex problems. These networks learn and improve based on many examples that are fed in.

What is new?

• GPT-4 is a large multimodal model that now accepts images and texts (previously only texts).?
• GPT-4 can process 25,000 words which is about eight times ChatGPT.
• GPT-4 can collaborate, generate, edit, and iterate on creative and technical writing or learn a user’s writing style (screenplays, poems, songs).
• Better stable performance compared to GPT-3.
• GPT-4 is more creative and can handle further nuanced instructions than GPT-3. Thus, resulting in better performance, such as on tests (Olympiads, Uniform Bar Exam, LSAT, SAT Math and many others) designed for humans.?
• GPT-4 significantly reduces ‘hallucinations’ relative to the previous version.
• Significantly better safety features within GPT-4 as compared with GPT-3. In addition, GPT-4 is more sensitive to respond to sensitive queries.?

Application and Use Cases (certain features are limited in availability)

• Recommending dishes based on a picture uploaded of available ingredients.
• Turning a picture of a sketch to a functional website or a web app.
• Code snippets to make games, websites, browser extensions and more.
• Social Media analytics.
• Research and analysis.
• Drug discovery.
• Customer service, virtual assistants and content creation.?
• Review code for vulnerabilities or edit code for improved code scripts.
• Completes text passages or generate text, stories in a specific tone and style.
• Writing a synopsis about a book, movie, audio, video, article or an event.
• Duolingo has partnered with ChatGPT to enable deeper conversations with two features, ‘Role Play’ an AI conversation partner, and ‘Explain my Answer’ by breaking down the rules when a mistake occurs.
• Khan Academy has partnered to explain difficult to understand concepts adapting learning to personalized skill levels.
• Advanced reasoning capabilities for example to find a common time slot to schedule an appointment between several participants.
• Be My Eyes is developing an app to generate same level of context and understanding as a human to help people with low or no vision.?
• Stripe has various applications of GPT-4, such as fraud detection and scanning and writing summaries about websites.?
• GPT-4 used by the Government of Iceland to preserve its language.
• Morgan Stanley is experimenting the usage of GPT-4 to help its wealth managers navigate through vast amount of information to locate relevant information. Thus, ensuring better knowledge management.

The capabilities of ChatGPT will keep evolving with time, while people start using and getting adapted to new developments and territories previously unexplored. Experts state that the actual usage of an application like ChatGPT is only limited to the reaches of the human mind. Getting comfortable in unchartered territories will take us some getting used to. Is the time of mass adoption of AI models to capitalize the value of data already here? What is next that can be achieved, is something that becomes clear as we move ahead.

To read the first article in the series of “Simply Tech” you can click on this link:?Differentiation and Integration with Terminology.