In its current usage, artificial intelligence (AI) refers to computer algorithms that can accurately mimic human cognitive functions including learning, decision-making, problem-solving, and even creativity.
Generative AI enters the scene in this final and maybe most human quality. Models for generative AI are trained on data, just like other modern AI. Following the instructions and directions that they have learnt, they use that data to produce new data.
It will learn, for instance, that a cat has four legs, two ears, and a tail if you train it with photographs of cats. Then, you can instruct it to create its own image of a cat, and it will create as many versions as you require while still adhering to those fundamental principles.
Understanding the contrast between generating and discriminative (or predictive) AI is important. Discriminative AI primarily focuses on classification, identifying differences between “things” like cats and dogs. This is what Netflix and Amazon’s recommendation algorithms use to differentiate between content you might want to watch or purchase and content you are unlikely to be interested in. Or to identify between safe paths from point A to point B and those you should probably avoid in navigation software.
Instead, generative AI focuses on deciphering data’s structure and patterns in order to generate new data that resembles the original.
So What Can Generative AI Do?
Creating text and images was often the first use case for generative AI, but as the technology has advanced, a wide range of new applications have become possible. Here are a few examples:
Images: A natural language (i.e., human language) prompt can be used to generate a picture using a variety of generative AI methods, such as Midjourney or Stable Diffusion. Tell it you want to see a vision of a two-headed dog piloting a spaceship into a black hole while wearing an Elvis costume, and watch as it—or something similar—appears before your eyes.
Text: While ChatGPT undoubtedly started the current intense enthusiasm around generative AI, there are additional generative text tools like Google’s Bard and Meta’s Llama. Anything from essays and articles to dramas, poems, and novels can be written using them.
Coding: In addition to ChatGPT, resources like Microsoft’s GitHub Copilot and Amazon’s CodeWhisperer make it simple for anyone with little to no technical expertise to generate computer code.
Audio: Voice synthesis, a feature of generative AI tools, enables computers to say words that have never before been said by a human, as well as produce music and sound effects.
Video: Although still not as sophisticated as text or image production, tools are starting to appear that enable us to create and edit video by just expressing what we want to see.
Data augmentation: Generative AI makes it simple to produce completely synthetic data sets for use in training other AI models that adhere to real-world norms without imposing obligations on people who store and utilise it with regard to privacy and data security.
Virtual environments: Think of virtual reality (VR) environments or video game worlds that can be explored and engaged with, or the rather overhyped concept of the metaverse. Designing these is a highly difficult undertaking that can be substantially accelerated with the help of generative AI.
How Does It Work?
Generative AI, like all of the AI we see today, arose from the discipline of AI study and practice known as machine learning (ML).
While traditional computer algorithms are written by humans to instruct a machine on how to perform a specific task, machine learning algorithms improve as they are fed more data.
Put a bunch of these algorithms together in a way that allows them to generate new data depending on what they’ve learnt, and you’ve got a model – effectively an engine designed to generate a specific sort of data.
Examples of models applied in generative AI applications are as follows:
Large Language Models (LLMs) – By taking in a lot of text, these models learn the semantic connections between words and use that data to produce new language. OpenAI’s GPT-4, which powers the ChatGPT tool, is an illustration of an LLM.
Generative Adversarial Networks (GANs)- These pit two competing algorithms against each other to determine which can produce data that is similar to its training data. The other algorithm must determine whether the output is produced or real. Images, sounds, or even videos can be produced using this kind of generative model.
Variational Autoencoders – This sort of model learns how data is created by encoding it in a straightforward manner that captures its fundamental properties, then working out how to reconstruct it. It is frequently employed to produce artificial data.
Diffusion models- function by adding random data (referred to as “noise”) to the data they are learning about, then working out how to eliminate it while keeping the original data. This enables them to learn what information is crucial and what can be ignored. In the creation of images, diffusion models are most frequently used.
Transformer Models- This name serves as a sort of catch-all for a category of models, which includes LLMs but also refers to any model that learns the context and relationships among various parts in its training data.
Generative AI in Practice
The use of generative AI to produce fantastic (and occasionally horrible) things has already been demonstrated in many amazing ways.
Consider the Coca-Cola “Masterpiece” commercial, which was developed in collaboration with human and artificial intelligence and recreates some of the finest pieces of art ever created on television.
The use of generative AI to produce fantastic (and occasionally horrible) things has already been demonstrated in many amazing ways.
Consider the Coca-Cola “Masterpiece” commercial, which was developed in collaboration with human and artificial intelligence and recreates some of the finest pieces of art ever created on television.
By reconstructing partially recorded lyrics by John Lennon and adding fresh material by Paul McCartney, it has also been utilized to write a new Beatles song.
In an emerging subject known as “generative design,” artificial intelligence is utilized to generate blueprints and production methods for new items. For instance, General Motors designed a new seatbelt bracket using generative tools developed by Autodesk that is 40% lighter and 20% stronger than its present components.
A UK business recently announced that it had developed the first AI-generated immunotherapy cancer treatment, demonstrating how it is also being used to speed up drug discovery.
Deepfakes, a recent phenomenon that blurs the line between fact and fiction by making it seem as though real people have said or done fake things, are also made possible by generative AI.
Deepfake One of the first and most well-known examples was Tom Cruise. More perniciously, prospective contenders for the upcoming 2024 US presidential elections on both sides of the aisle have appeared in deepfakes meant to disparage them for political reasons.
The Ethical Issues with Generative AI
While generative AI may accomplish some fantastic things, it is also obvious that it is forcing us to consider some challenging problems and questions.
When it becomes impossible to distinguish between what is real and what is created by AI, it may be one of the major concerns.
It’s probably going to happen sooner rather than later given how quickly innovation is occurring in the industry.
So, the question becomes: What should we do about it, if anything? Should the rest of the world adopt legislation similar to those established by countries like China, making it unlawful to deepfake someone without their consent?
There is also the issue of how this will impact human jobs; would the livelihoods of creators be in jeopardy if the businesses that employ them can produce as many images, sounds, and films as they require just by instructing a computer to do so?
Copyright is an additional problem that needs to be resolved. Who owns a piece of art produced by an AI? Who created the artwork using AI? who designed the AI in the first place? Or all the (likely) hundreds of artists whose creations were (in actuality, frequently without permission) utilized to train the AI?
Given the quickening rate at which this technology is being developed, all of these problems require quick answers. The future of generative AI in our world and in our daily lives may very well depend on how we respond to these questions.
Source link
