Topic-4-Large Language Models, Claude, Bard & Stable Diffusion

 

 

·       Bard has real time information

·       Bard is free and no costing modal defined yet.

·       Bard has more mature information.

·       Content is more authentic.

·       All responses directly go to your Gmail draft folder

 

Information by Bard is real time, it does not mean that real time modal is trained by our click. It responded at our click. Same thing we are doing at IrfanGpt although this function of bard ai was not known to me.

Another feature is that we can google the desired information. The question here arises, which information would we have on clicking the Google it buttons.

 Would it be before asking the query or after asking the query? You never know, at what time it was updated.

 Difference between bard and ChatGPT was explained with the help of an example. Suppose you ask your servant to serve 5–6 guests in your absence. In case of ChatGPT, it would fail to serve if number of guests are exceeded.

In other case you are in contact with your servant and constantly feeding information to him and updating him about the coke cake etc already store in your fridge and can be served in case the number of guests are increased than expected. Now the servant act like bard and succeed to serve even if the number of guests are increased.

CLAUD

Claude is a next-generation AI assistant based on Anthropic’s research into training helpful, honest, and harmless AI systems. Accessible through a chat interface and API in our developer console, Claude is capable of a wide variety of conversational and text processing tasks while maintaining a high degree of reliability and predictability.

Claude can help with use cases including summarization, search, creative and collaborative writing, Q&A, coding, and more. Early customers report that Claude is much less likely to produce harmful outputs, easier to converse with, and more steerable - so you can get your desired output with less effort. Claude can also take direction on personality, tone, and behaviour

Question: What is difference between Claude – ChatGPT – Bard

Generative AI

Generative AI is a type of artificial intelligence technology that can produce various types of content, including text, imagery, audio and synthetic data.

Generative AI at the most of the time are neural networks. Neural networks we studied in Deep Learning.

Is there any relationship between Machine Learning and Deep Learning?

Deep Learning is sub set of Machine Learning.

There are always two parts in Generative AI. First part is called Encoder and Second Part is called Decoder.

What does Encoder and Decoder Do

Encoder will take input from you and converts it in computer understandable language.

Decoder will take out put from computer and converts it in human understandable language.

The encoder will take the input and make it into a representation, suppose you are creating a concept of a car in the mind of a child. You can represent the car by describing its structure or showing him the picture of the car. These both are types of representation.

Some presentations are easier to understand and some presentations are difficult to understand.

Encoder function is to represent human centric representation and Decoder function is to present machine centric representation.

Generative Models

Examples of Generative Models

• ‌Autoencoders
• Naïve Bayes.
• Bayesian networks.
• Markov random fields.
• ‌Hidden Markov Models (HMMs)
• Latent Dirichlet Allocation (LDA)
• Generative Adversarial Networks (GANs)
• Autoregressive Model.

 

Representations

Earlier it was elaborated that Machine Learning needs numbers, and now I am saying that Machine Learning needs Numeric Representation to function.

What is Regression and Classification?

You predict score in regression an in classification you predict label.

A famous generative model that is being used is called Transformer Neural Networks.

Transformer Neural Networks also has two parts:

Encoder Transformer Neural Networks

Decoder Transformer Neural Networks

When both types have been trained, we can use both separately. We can perform discrimination AI through Encoder separately and can perform Generative AI through Decoder.

It means wherever you will see the Generative AI, it means they would be Decoder part of Transformer.

The Primary Goal of Artificial Intelligence is to teach the computer without programming. Another goal is to reduce the human effort and increase the productivity.

As the human memory and multithreading capacity is limited and conversely, computer has unlimited memory and multithreading capacity. Humans get bored by performing repetitive tasks, but computer can perform these tasks efficiently.

The difference between Regression and Classification is that Regression predicts score and Classification predicts Labels.

GENERATIVE AI

Every Generative modal has two parts:

·       Encoding – Human centric to Machine Centric.

·       Decoding-Machine Centric to Human Centric.

The famous Generative modal that is being used nowadays is Transformal Neural Network.  Transformal Neural Network also has two parts:

It means if we have a Trans-formal Encoder that can convert human information to computer understandable information, is that possible that we can use this Encoder in Classification or Regression. Definitely, the computer would as an output will recognize that incoming information is about dog or cat. It means the Encoder can discriminate through converting human representation to machine representation but cannot generate information.

If we want to create a generative modal, we will first train the modal and then with the help of Decoder generate the information. As we know that Decoder converts Machine representation to human representation so we give the Decoder a random number, it would convert it either into an image or text.

So the Generative Modals would be the Decoder part of the Transformer, as the Decoders are best in converting Machine Language into Human representation.

WHAT IS LARGE LANGUAGE MODAL

So LLM is created by joining so many Decoders. However, by joining many Decoders alone does not create LLM, you also have to train it my provision of large amount of data available on Twitter, Wikipedia, news, etc. It sees the words patterns available in this data to get train itself.

This helps the computer to make sense. For example, what comes after a dog can either “is barking” or “is running”. As the “is barking” frequently used, so computer would generate “is barking”. However, if ask computer to randomly choose from the given ten words. The result would be different. So when we give the selection strategy to the computer, it is like Computer Switching temperature strategy.

EMBEDDING FROM TOKENS

Embedding is coding that computer loves. It is the language the computer likes the most.

STABLE DIFFUSION

What is a Stable Diffusion?

Stable Diffusion is open-source artificial intelligence designed to generate images from natural text. This means that users can make a request using natural language, and the AI will interpret and generate an image that reflects the request.

Imagine you are sitting in a room and room temperature is hot. You turn on the Air Conditioner, and it starts cooling each particle and molecules of the air. After some time, the room temperature changes to chill.

In this example, you replace the AC with the Modal and the air particles with image. The modal acts on the pixels of the image as AC acts on the air particles. With the passage of time, this slow processing of Modal will convert the good image to bad image (Noise).

This technique is being used in Traffic identification or satellite etc.

What if we have to cool an Auditorium with the help of 2 Ton AC. The result will be horrible as its not possible for the 2Ton AC to change the temperature of the room. Similarly, mega pixel image create trouble for the AI modal as a lot of intensive computing is required.

To deal with this problem, we use latent diffuse Modal.

Stable /Latent Diffusion Models (LDMs) 

Latent Diffusion Models (LDMs) enable high-quality image synthesis while avoiding excessive compute demands by training a diffusion model in a compressed lower-dimensional latent space. Here, we apply the LDM paradigm to high-resolution video generation, a particularly resource-intensive task.

Two prompts are used in Stable Diffusion. Positive Prompt and Negative Prompt. Positive Prompt tells what to do and Negative Prompts tell what not to do.