Popular AI Applications, and the ML Type, Algorithms, and Models used in/by them.
ChatGPT
What is ChatGPT
ChatGPT is a very popular GenAI Application, which is now being used by millions of people in some way or the other.
It can be used to write and fix code, ask financial and technical questions, cook, and do a variety of other things.
Model versions
The following table lists the main model versions of ChatGPT, describing the significant changes included with each version:
| Version | Release date | Description | Status | |
|---|---|---|---|---|
| GPT-3.5 | November 2022 | The first ChatGPT version used the GPT-3.5 model. | Discontinued | |
| GPT-3.5 Turbo | 2023 | An improvement over the legacy version of GPT-3.5, GPT-3.5 Turbo in ChatGPT offered better accuracy in responses while using a similar model. | Discontinued | |
| GPT-4 | March 2023 | Introduced with the ChatGPT Plus subscription, the March 2023 version is based on the more advanced GPT-4 model. | Active | |
| GPT-4o | May 2024 | Capable of processing text, image, audio, and video, GPT-4o is faster and more capable than GPT-4, and free within a usage limit that is higher for paid subscriptions.[107] | Active | |
| GPT-4o mini | July 2024 | A smaller and cheaper version of GPT-4o. GPT-4o mini replaced GPT-3.5 in the July 2024 version of ChatGPT.[108] | Active | |
| o1-preview | September 2024 | A pre-release version of OpenAI o1, an updated version that could “think” before responding to requests.[109] | Discontinued | |
| o1-mini | September 2024 | A smaller and faster version of OpenAI o1.[109] | Discontinued | |
| o1 | December 2024 | The full release of OpenAI o1, which had previously been available as a preview.[103] | Active | |
| o1 pro mode | December 2024 | An upgraded version of OpenAI o1 which uses more compute, available to ChatGPT Pro subscribers.[103] | Active | |
| o3-mini | January 2025 | Successor of o1-mini.[110] | Active | |
| o3-mini-high | January 2025 | Variant of o3-mini using more reasoning effort.[110] | Active |
source: wikipedia
How does ChatGPT gather data?
ChatGPT was trained on a large training dataset that consisted of books, articles, and web pages that are available publicly on the internet.
This is one of the largest training dataset available.
The data collection process was also called Common Crawl, where all the publicly available information was gathered and then fed to ChatGPT.
Training Used:
Unsupervised Learning + Supervised Learning during fine tuning process.
The outcome of the training was a model called Generative pre-trained Transformer model or GPT.
Hence, GPT is the model that powers all the versions of ChatGPT.
ChatGPT is a large language model (LLM), because it has been trained on a large training dataset, which contains billions of instructions.
DALL·E
What is DALL·E
DALL·E is an AI model that can generate realistic images, and art from a description in natural language.
It is developed by OpenAI.
How does DALL·E gather data?
The training data for DALL·E consists of a vast collection of text-image pairs sourced from the internet.
These pairs include captions and corresponding images, allowing the model to learn the relationship between the textual description and visual representations.
Training Used:
Unsupervised Learning + Supervised Learning during fine tuning process.
DALL·E is also a large language model (LLM) because it has been fed a large training dataset.
The output model that is used for DALL·E is GPT-3 model, and it is specifically designed for image generation.
GitHub Copilot
What is GitHub Copilot?
According to GitHub Documentation, "GitHub Copilot is an AI coding assistant that helps you write code faster and with less effort, allowing you to focus more energy on problem solving and collaboration".
GitHub Copilot is an AI tool developed by OpenAI in collaboration with GitHub.
Copilot suggests code as you type, just like having a coding assistant right in your development environment.
How was it trained, and How does it generate code?
GitHub Copilot has been trained on a large dataset on the publicly available code from repositories, coding websites, forums and documentation available on the internet.
Training Used:
Unsupervised Learning + Supervised Learning during fine tuning process.
The outcome of the training was codex model.
GitHub Copilot uses codex model, a descendent of GPT-3, based on transformer architecture.
Thanks for Reading 😊
Please drop a 👍 if you liked the post!
Also, feel free to reach out if you need any other info around this, or any other topic. Will be happy to share.
Top comments (0)