ChatGPT

What is ChatGPT?

ChatGPT is OpenAI’s latest language model, specifically trained for interactions by conversation. Since the release of GPT-3, many problems have surfaced, such as the AI not being able to respond to some questions very well or making up information. While you can try to change the prompt (i.e. the text given to the AI), it is still you that is trying to cue the AI. But logically, it would be much better if the AI instead cued you. That is, instead of trying to iteratively rephrase a question to make the AI understand what you are looking for and answer correctly, the AI should simply ask you if it doesn’t understand in a conversational manner.

Now, this is the direction ChatGPT moves us towards. It is still more inclined to answer than ask for now, but the conversational format allows for more dynamic interactions. Before you could simulate a conversation with GPT-3, but this isn’t what it was trained for. ChatGPT, on the other hand, has been trained explicitly for this purpose. It uses a technique called reinforcement learning from human feedback. Reinforcement learning is an area within machine learning where agents are trained to complete objectives in an environment driven by rewards. Iteratively, the agent interacts with the environment by performing actions, obtaining rewards and then adjusting its behavior to both explore the environment and obtain better rewards.

Games, in particular, have been a playing ground for these techniques with huge success and impressive results in Chess, the Atari games, Go and Poker. While games have predefined rules and rewards, a conversation does not, thus, human feedback becomes essential. This was done by prompting a model, sampling several responses and then letting a human manually rank the responses. These rankings will then become training data for a reward model. Finally, a fine-tuned language model will be further trained using reinforcement learning to respond to questions so as to optimize the output of the reward model. For more information, check out OpenAI’s blog post:

What Does ChatGPT Mean?

ChatGPT (Chat Generative Pre-trained Transformer) is a complex machine learning model that is able to carry out natural language generation (NLG) tasks with such a high level of accuracy that the model can pass a Turing Test.

ChatGPT was trained on massive amounts of unlabeled data scraped from the internet before 2022. The model is continually being monitored and fine-tuned for specific language-oriented tasks with additional datasets labeled by humans.

Tasks ChatGPT excels at include:

• Providing answers to questions.
• Completing a given text or a phrase.
• Writing fiction and non-fiction content from prompts.
• Producing humanlike chatbot responses.
• Generating computer code.
• Translating text from one language to another.
• Performing calculations.
• Summarizing a given text.
• Classifying text into different categories.
• Analyzing text sentiment.
• Generating text that summarizes data in tables and spreadsheets.
• Responding to user input in a conversational manner.

ChatGPT is built on a large language model (LLM) called GPT-3 (Generative Pre-trained Transformer 3). A large language model is a type of deep neural network that uses billions of parameters and is trained with petabytes of data. The model was developed by OpenAI, a research firm founded by Elon Musk.

Techopedia Explains ChatGPT

Technically, ChatGPT can be categorized as a type of generative AI. Once trained, a generative AI model can produce outputs that are similar -- but not the same -- as the training data.

How was ChatGPT trained?

Data scientists and machine learning engineers used semi-supervised learning to train ChatGPT.

Semi-supervised learning algorithms are trained on datasets that are partially labeled; some of the data points have a label and some do not. The model uses the labeled data to predict the output of the unlabeled data.

According to OpenAI, data scientists gathered the huge amount of unlabeled data required to train the LLM by scraping the internet. They supplemented this with text sources that were either in the public domain or made available for use by researchers and possibly governments.

OpenAI has not revealed exactly how they were able to label the immensely large data sets required to fine-tune the model, but it's known they outsourced some of the labeling and it's likely they also took advantage of crowdsourcing platforms like Amazon's Mechanical Turk.

ChatGPT and Responsible AI

Responsible AI is a governance framework that's intended to help make complex black box AI models like ChatGPT more explainable and therefore more trustworthy. There are several steps OpenAI is taking to help make sure ChatGPT supports the concept of responsible AI.

• Researchers and developers at OpenAI are looking for ways to make this deep learning model more transparent and are trying to be as forthright as possible about the model's capabilities, limitations and potential misuse.
• OpenAI data scientists are continually reviewing user feedback and have put humans in the loop (HITL) to eliminate bias in their training data.
• OpenAI engineers are continuously monitoring ChatGPT user prompts and outputs of the model to help ensure it is being used in a responsible manner.

Business Uses for ChatGPT

ChatGPT is being used as an augmented intelligence tool for a wide variety of business tasks. It is currently being used to:

• Automate the content creation process for social media posts, website articles and product descriptions.
• Build customer-facing conversational chatbots that are capable of going off script.
• Summarize and categorize customer feedback from multiple channels.
• Translate business reports and marketing deliverables from one language to another.
• Write boilerplate computer code and generic scripts.
• Translate the information contained in a table or spreadsheet into text.
• Create chatbots that support just-in-time learning for internal training programs.
• Write meta descriptions and suggest synonyms for SEO keywords.

Is it good?

The question is of course, is it good? The model was made available for free and it quickly went viral to the point where even Elon Musk tweeted about it:

Thus, a ton of people have tested it, most being highly impressed by its capabilities, myself included. Below I’ve aggregated some of the discoveries that have been made so far of its capabilities and limitations: