MC Explains | Everything you need to know about AI bot ChatGPT

(Image Courtesy: Open AI)

You may have seen it mentioned around the internet or at least seen extremely funny discourses about pumpkins.

I can't compete with this pic.twitter.com/YdQ87LWIst

Story continues below Advertisement

Remove Ad

— Keith Wynroe (@keithwynroe) December 1, 2022

Or may have run across this hilarious answer to the meaning of life.

what bout this xD pic.twitter.com/VQGodDbt1y

Related Stories

• AI stocks show bubble-like traits as retail buying surges, says McKinsey’s Tim Koller
• 5 ChatGPT settings you should enable for better responses
• Apple rocked by executive departures, with chip chief at risk of leaving next

— Mr Photo (@MrPhoto17) December 2, 2022

The AI bot known as ChatGPT has taken the internet by storm but what exactly is it, and how does it work?

ChatGPT is a conversational dialogue model, trained by AI and machine learning to understand and respond to natural human language.

It gets its name from GPT or Generative Pre-Trained Transformer, which is a deep learning language model that specialises in generating human-like written text.

Deep Learning is a machine learning method with three or more layers of a neural network. This network attempts to simulate the behaviour of a human brain, which allows it to learn, like we do, using data.

The current iteration of GPT (3.5) is what ChatGPT is based on. Unlike GPT 3.5, ChatGPT can generate detailed, human-like written text and can remember earlier conversations for context.

OpenAI, the creators of GPT, say that ChatGPT can answer follow-up questions, admit mistakes, challenge premises and reject requests it deems inappropriate.

OpenAI trained ChatGPT using a training method known as Reinforcement Learning from Human Feedback (RLHF). RLHF uses a reward/punishment system to train AI.

Whenever an AI takes action, it is classified as desirable or punishable. A desired action is rewarded whereas an undesired one is punished. This way, the AI parses a large subset of data and learns like humans, through trial and error.

For ChatGPT, OpenAI used "supervised fine-tuning". Dedicated human AI trainers trained the AI through conversations, in which they played both the user and an AI assistant. The trainers were given suggestions when providing responses.

Like earlier GPT models, ChatGPT can sometimes give plausible-sounding but incorrect answers. OpenAI says that fixing this issue is problematic because there is no source of truth in the data they use to train the model. OpenAI tried telling the bot to be more careful during responses, but then it started declining questions that it could answer correctly.

Supervised training can also mislead the model "because the ideal answer depends on what the model knows, rather than what the human demonstrator knows."

The model is sensitive to phrasing; it can decline to answer a question written in a certain way but will give you the answer if you rephrase it. It can also do this if you prompt it multiple times with the same question.

ChatGPT can be excessively verbose in its responses, and tends to overuse certain phrases. This happens due to the supervised nature of the training. A human prefers longer, comprehensive answers, whereas the model would prefer a succinct one.

While OpenAI does its best to block inappropriate requests, ChatGPT does sometimes respond "to harmful instructions or exhibit biased behaviour."