What is generative AI? post #2
This is the second of two posts on Generative AI.
Late last year (November), OpenAI released ChatGPT to the public. However, it took a few months for the internet at large to catch on and realise the potential of this new tool. The trend started to take off in December, and by February everyone was talking about it.
ChatGPT is an example of a "Generative AI". To recapitulate, let's see what this means. For convenience, let's call AIs that are not generative, "regular" AIs.
- Regular or "Narrow" AI (Artificial Intelligence): Regular AI, often referred to as "narrow" or "specific" AI, focuses on building systems that can perform specific tasks or solve specific problems. These AI systems are designed to operate within predefined boundaries and excel at specialised tasks. For example, a regular AI could be created to classify images, play chess, or process natural language. Another example is the AI that chooses your words in predictive text. Narrow or regular AI is typically trained using supervised or unsupervised learning techniques and relies on large datasets for training. It aims to optimise performance on a specific task and often follows a rule-based or algorithmic approach. Supervised learning means that a human monitors and/or controls the learning process. Unsupervised means that the computer is left alone to do its own thing.
- Generative AI: Generative AI, on the other hand, refers to a class of AI models that are capable of generating new content, such as images, text, or even music. Generative AI models, like ChatGPT, are trained to learn patterns and relationships in data and generate new samples that are similar to the training data. These models use techniques like deep learning and neural networks to generate content by capturing statistical dependencies and structures in the data they have been trained on. They can create new and original outputs based on learned patterns, and their ability to generate realistic content has improved significantly in recent years.
- A GAN, or Generative Adversarial Network. This is a system which generates new content by setting two systems to compete (adversaries). Imagine there are two artists in a competition: a forger and a detective. The forger's job is to create fake artwork, while the detective's job is to identify which artworks are fake and which are real. They both learn and improve through the competition. In the case of a GAN, the forger is called the "generator," and the detective is called the "discriminator." The generator's job is to create fake data, such as images or text, that look as realistic as possible. The discriminator's job is to examine this data and determine if it's real or fake. The generator and discriminator go through this process repeatedly, each learning from the other's successes and failures. Over time, the generator becomes more skilled at creating realistic fake data, while the discriminator becomes better at detecting fakes. In the end, this competition between the generator and discriminator results in the generator being able to create fake data that is very close to the real thing. GANs have been successfully used to generate realistic images, produce synthetic voices, and even generate text that resembles human writing.
Here is a list of useful sites which make use of GAN (generative adversarial network) technology:
URL | What it does | Type of tool |
Stable Diffusion - an image generation AI to create new images of anything - just describe what you want, also available at https://stablediffusion.in/ | Art | |
DALL-E - an image generation AI to create new images of anything - just describe what you want | Art | |
Currently a research paper, but to make 3d models from flat images | Art | |
Midjourney - an image generation AI to create new images of anything - just describe what you want. Click on the eyeball icon on the home page for some examples | Art | |
Makes websites for you | Coding | |
To help you with coding | Coding | |
To help you with spreadsheets | Coding | |
To help you with coding | Coding | |
To help you with coding | Coding | |
An algebra solver | Coding | |
To help you with coding | Coding | |
To make videos with voice of anyone (deepfake) | Multimedia | |
To get the voices out of a music track | Multimedia | |
Create deepfake videos for e.g. training videos | Multimedia | |
https://chrome.google.com/webstore/detail/eightify-youtube-summary/cdcpabkolgalpgeingbdcebojebfelgb | Eightify, summarises youtube videos into text | Multimedia |
Repairs photos | Photo repair | |
Repairs photos by enlarging them | Photo repair | |
To colourise old images | Photo repair | |
Repairs photos by enlarging them | Photo repair | |
To change facial expressions in photos, currently in research phase | Photo repair | |
To turn your photo selfies into cartoons or art of yourself in a new style | Photo repair | |
To cut out shapes in photos | Photo repair | |
To try different "looks" on your selfies | Photo repair | |
To remove watermarks from photos | Photo repair | |
To get summaries and ask questions about a document that is too long to read | Research | |
To get an understanding of the scientific consensus on anything | Research | |
To find research papers on any topic | Research | |
To help you write academic papers | Research | |
To get an understanding of the scientific source citations | Research | |
Plugins for ChatGPT 4.0 | Research | |
To get summaries and ask questions about a document that is too long to read | Research | |
Plugins for ChatGPT 4.0 - a howto guide | Research | |
An alternative to ChatGPT | Research | |
Many academic tools e.g. plagiarism checking | Research | |
ChatGPT - an interactive chatbot that answers any questions (Except private confidential questions) | Research |