ChatGPT is a conversational AI interface created by OpenAI, which combines the capabilities of genera-purpose large language models, like GPT-3, and it has been fine-tuned with a model called InstructGPT, which added a supervised learning method, which is human-in-the-loop, to increase truthfulness, and reliability of the conversational interface.How does ChatGPT work?
To grasp large language models, it’s critical to understand a few key concepts.
Current large language models, like GPT-3, have been trained on a massive amount of data and billions of parameters, which were used to pre-train those models.
In the pre-training phase, through unsupervised learning, it’s possible to give the model a straightforward goal, like predicting the next token in the sequence.
In short, the main goal of a large language model like GPT-3 in the pre-training phase is the text-to-text prediction of what comes next in the text sequence.
Once the model has been pre-trained, it’s pretty good at general tasks. Though it can be misleading and untruthful, thus hallucinating often.
To “put some breaks” and enhance the guardrails of a general-purpose model, like GPT-3, that can be fine-tuned.
In short, it can be trained, with a supervised learning approach, on a much smaller dataset, which is human-labeled, and where humans show the general-purpose engine how to get better at specific tasks and be more truthful.
In the specific case of ChatGPT, it has been used a specific human-in-the-loop model, called InstructGPT, to smooth out some of the negative aspects related to GPT-3 to make ChatGPT viable as a conversational interface.
ChatGPT can often still be misleading, yet it can improve over time as it learns how to deal with more and more edge cases.How does ChatGPT Make money?
ChatGPT launched as a free tool at the end of November 2022; it’s now getting monetized via a premium subscription model.
In short, ChatGPT follows a freemium model.
Stability AI Ecosystem
* This article was originally published here