An Introduction into Generative AI
Some people are surprised to learn that Artificial Intelligence (AI) is really nothing new. It has evolved since the 1940s, and has made continual, steady progress since then. But what is different about AI now is that it is easily accessible to anybody that wants to use it, something that never occurred before. The primary catalyst for this happening is the dawn of Generative AI. Here is a technical definition of Generative AI:
“Generative AI enables users to quickly generate new content based on a variety of inputs. Inputs and outputs to these models can include text, images, sounds, animation, 3D models, or other types of data.”1
So, there are a number of key differences between Generative AI and the traditional AI. Here are some of them:
- Traditional AI is merely viewed as “Garbage in and Garbage out.” This simply means that the produced output is 100% dependent upon the inputs that are fed, which are the datasets. In contrast, Generative AI can look beyond the datasets that it has been trained upon. For example, it can find on its own other resources on the Internet through the superior Unsupervised Learning that it has achieved.
- Generative AI can produce many other kinds of outputs, as explained in the definition above. For example, it can provide the outputs as images, video, or even as audio. Or if the end user prefers, they can even get the traditional text-based format as well.
- Unlike traditional AI, Generative AI can create fresher levels of content for its output. This simply means that the output will be far more robust, given its ability to search for different resources in response to the query.
- With traditional AI, you are rather limited to how your query can be structured. As a result, this will greatly limit the quality of the output that you will receive. But with Generative AI, there is much more freedom and flexibility to create the query that will get the output you are looking for. This is also technically known as “Prompt Engineering.” But the drawback here is that you must have some higher level of expertise in order to create the “perfect query.” This can take a great deal of time to accomplish, as it can only be learned via trial and error.
- Traditional AI is best suited for heavy data analysis and finding hidden trends in large data sets (also known as “Big Data”), and is pretty much limited to just that, although this functionality can serve a wide variety of industries, such as that in healthcare and insurance. On the contrary, Generative AI can be used in a wider range of applications, thus it can serve more markets and industries in this regard.
Generative AI And Large Language Models
There is often a lot of confusion between Large Language Models (also known as “LLMs”) and Generative AI. While the latter is used to power the former (especially when it comes to the creation of the “Digital Person”) there are differences to them as well. A technical definition of an LLM is as follows:
“A large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content.”2
In a more focused sense, the goal of an LLM model is to try to understand human language, both written and spoken, and try to emulate those processes in having a real-life conversation with an end user.
In this regard, there are four kinds of LLM models, which are as follows:
- The Zero Shot Model: This is meant for just general use cases, and only general datasets are used for them. It is typically the GPT3 algorithms that are used to power this kind of model.
- The Domain Specific Model: This is a type of LLM model which serves only one purpose, and is discarded after its use is over.
- The Large Representation Model: This kind of model makes use of the “Bidirectional Encoder Representations from Transformers” (also known as “BERT”) that is used to drive Natural Language Processing (also known as “NLP”) models.
- The Multimodal Model: This kind of model can produce sophisticated text and image outputs. It is the GPT4 algorithms that power these.
To view an LLM model in simpler terms, it can take petabytes of data (and possibly even more) to try to emulate human language, whether the inputs are text, audio, image, chat, audio, video, etc. But what makes an LLM model unique is that it attempts to look for the meaning and the correlation of the individual words and phrases that are both spoken and written. Because of the sheer volume of datasets that they can ingest, LLM models can understand to some degree how humans speak with another, whether in a remote or face to face setting.
The matrix below describes the four major LLM algorithms:
Algorithm Features/Characteristics
|
|
This was developed exclusively by OpenAI and is far more powerful than its predecessor, the GPT3 algorithms, which laid the foundation for ChatGPT. For any use case, it can generate up to 25,000 words. It has been estimated that there are some 1.76 trillion parameters that make up the GPT4 algorithm. |
|
The Generalist Language Model (also” known as “GLaM”), |
This is a set of LLM algorithms that was developed by Google. It has some 1.2 trillion parameters that are attached to it. This kind of algorithm is used to create human like responses to any query that is posed to it. |
|
The Bidirectional Encoder Representations from Transformers (also known as “BERT”). |
Also developed by Google, it is not nearly as powerful as GlaM is. It is only associated with 340 million parameters, and is used primarily for answering simple queries, whether in text or in speech. |
|
The Large Language Model Meta AI (also known as “LlaMA”). |
This algorithm was developed by Meta. It is rumored to have “billions of parameters” attached to it. It is trained in 20 different foreign languages, and thus, is heavily used for foreign language translation works. |
Up Next: Generative AI and Natural Language Processing
The next article in this series will address what is known as Natural Language Processing, a key component to Generative AI.
Sources/References:
- NVIDIA glossary
- TechTarget.com
Ravi Das is an Intermediate Technical Writer for a large IT Services Provider based in South Dakota. He also has his own freelance business through Technical Writing Consulting, Inc.
He holds the Certified In Cybersecurity certificate from the ISC(2).
