Demystifying Generative AI
There’s no doubt that the last year has been a year of innovations & “Generative AI”. This technology has seen a lot of interesting developments since March 2023 & is said to have changed the Future of AI. Gartner predicts that “More Than 80% of Enterprises Will Have Used Generative AI by 2026”. But what exactly is this Generative AI? What is ChatGPT? Is Generative AI the same as ChatGPT? Is Generative AI the same as the AI that we have been reading about all these years? What does it do???? ????
In this article, we will try to answer most of these questions regarding this rapidly evolving technology.
What is Generative AI?
Artificial intelligence (AI) is the field which focuses on completion of human-like tasks with the aid of computers. Machine Learning (ML) is a subset of AI, where models are trained through extensive data to perform dedicated tasks and uses neural networks to mimic the learning process of the human brain. Deep Learning is the specialized form of Machine Learning having multiple layers of neural networks to perform complex operations like image recognition. Generative AI (GenAI) is a revolutionary technology which has the capability of generating new and original content based on input provided.
The Advent of Generative AI
Generative AI relies on neural network techniques such as transformers, GANs, VAEs etc. A neural network is a method in artificial intelligence that teaches computers to process data in a way that is inspired by the human brain. Consider these neural network techniques to be algorithms which eventually helped in building the GenAI tools. Some of these most popular techniques are dated back to 2014 when GAN was developed, followed by Transformers in 2017. Later, in 2018, OpenAI published its article titled “Improving Language Understanding by Generative Pre-Training,” & its foundation model GPT-1 which revolutionized the technology. A foundation model is an AI model built using any of the neural network techniques, trained on large data sets on a scale such that the model can adapt itself to perform a broad range of downstream tasks. OpenAI has been constantly training & improving its foundation models (like GPT 1, GPT 3, GPT 4 etc.) which have been built using the Transformer technique. These foundation models are then used to build Generative AI applications or products/tools (like ChatGPT) that can be directly utilized by the end user. The models are customized & trained on vast data to perform a particular task and eventually build a product/app.
Let’s try to make this easier with the help of an example:
We all are aware of the widely used & popular ChatGPT. Here, ChatGPT is the product/tool being used by you (end user), which is built using the foundation model GPT-3.5 Tubo. (Note: ChatGPT Plus, which is a paid version of ChatGPT has been built using GPT-4). The GPT series of foundation models operates on the neural network technique of Transformers. These foundation models are constantly evolving and improving in terms of accuracy, capabilities and much more. There may be many more models being released as we speak!
This technology has been evolving for years. But the old foundation models delivered inappropriate or inaccurate outputs. With rigorous training, the accuracy & efficiency of these models was improved. And finally, in November 2022, OpenAI launched ChatGPT, a free tool available for public use. This was the inflection point for GenAI. Since then, everybody wants to use GenAI for personal as well as business purposes.
Two of the most recent developments in the market have been OpenAI’s GPT-4 and Google’s Gemini Models.
Generative AI Capabilities
How does Generative AI Generate New Content?
If we ask a GenAI tool to generate the image of a horse on a beach; since the model has been trained on several images of horses and beaches so it is aware of all the details which is required to produce the image of a horse on a beach so it will generate few examples of a Horse on a beach. I can modify the output by giving detailed prompts and ask the tool to generate the image of a Brown Horse wearing shades on a beach. The tool is capable of doing it as well. The more detailed input you provide, the more accurate output is delivered.
What kind of inputs can Generative AI Process?
The input to a GenAI model depends on how the model is designed to process the inputs. There is a wide range of options which can act as an input to a GenAI model. Tt varies from audio, text inputs like comments, documents, research papers, excel sheets, code to visual items like image or video. The output is highly influenced by the input.
There’s more to it.. What are the other capabilities of Generative AI?
Apart from the generation of new and novel content GenAI can perform various tasks which earlier required human intelligence. The image below shows a wide spectrum of capabilities the technology offers. Due to its capability to extract and summarize key points from a large volume of data it finds applications in fields of research and analytics. In contrast to conventional conversational chatbots, GenAI can easily answer unusual queries of customers. To safeguard the integrity of data GenAI can be fruitful for synthetic data generation.
Trad AI vs GenAI
The output generated through GenAI models outperforms existing technological solutions in several ways. Huge computational requirements & large dataset for training is a primary differentiating factor of GenAI, The table below compares how the technology differs from traditional AI models.
Why do I need Generative AI when I can do several tasks with the existing technologies?
The technology offers several advantages in comparison to the existing solutions, since the models are trained on larger datasets, it allows them to generate detailed and superior output. The models offer holistic understanding thus modifying the output based on the context of the inputs provided. Also, the output is easily customized to best suit the user. The models have the capability to handle and process larger volumes of data to generate the output along with providing expandability for the generated output.
Enterprise Adoption of Generative AI
Now that we know what GenAI can do, let’s look at how do we do this & adopt this technology for your organization.
We have divided the adoption plan into 6 major steps:
1. Business Case Identification & Strategy
Identify specific business needs that can be addressed by Generative AI. You may find numerous use cases available like Marketing Content Creation, Meeting Summarization & much more. But to finalize the use cases for your organization, you need to define the business challenges, identify related GenAI capabilities & use cases, and then prioritize them. A systematic framework customizable as per your needs can be built to identify & prioritize use cases. We will soon come up with the next article covering this step, in depth.
2. Foundation Model or Product Selection
Once you finalize the use case to be considered for implementation, you need to pick up a foundation model capable of delivering your chosen use case (eg: GPT 4 , Gemini etc.). In some cases, there may be products or applications readily available in the market. For example, if you want to create images for Marketing, there are multiple products like Akool available where you directly pay the product licensing and usage fees. In this case, you do not have to go through the entire process of selecting a Foundation model & then customizing it.
3. Data Cleansing & Collection
Data preparation and cleansing is an important step in Use Case Implementation. The foundation model needs to be trained on past data so that it provides desired output. This step includes defining the data required for training and fixing or improving the data quality.
4. Model Training & Customization
Since we are taking existing foundation models; these models are already trained on large datasets to perform some defined functions. But now, we need to customize them by training them on more focused data so that they adapt to a narrower subject or can perform more specific tasks. The most common technique of doing this is called finetuning.
What is customization & training?
Training a GenAI model from scratch is a cumbersome process involving huge amount of resources and cost. Customization is the process of enhancing the output by tailoring the foundation model to perform dedicated task. Following are few customization techniques widely used for GenAI models:
- Pre-training: training the model with curated dataset of a particular domain or industry.
- Fine-tuning: training an already trained model with a smaller dataset.
- Prompt engineering: providing the model with detailed input instructions (prompts) to generate the desired output. Like “Generate the image of an Alsatian dog” rather than just asking the model to “Generate the image of a dog”.
- Instruction-tuning: training the model with input-output instructions, thus enabling it to learn through guided instructions.
5. System Integration
Once your product is ready after customization, integrate it into business processes and existing systems. This probably involves deploying the model in a cloud service, creating custom software to interact with the model, or integrating company documents and knowledge databases. You will also have to define if you want to deploy the product on cloud or on premise.
6. Monitoring & Retraining
These models are self sufficient and can operate via self-learning. But they need to be monitored regularly and adjusted or retrained as and when required. Set up processes for data integration, governance (such as a content moderation system), as well as data entry, model output, and error handling. [1]
While all these steps are important, one of the most critical steps is to adopt responsible AI practices. Whenever we talk about AI and any new related technology, there are several recommended practices to be considered while implementation. With AI comes great responsibility & some risks as well!!! We have mentioned some challenges below.
A few challenges associated with GenAI….
There are several challenges associated with GenAI. Since it is trained on a large corpus of data specific care must be taken to ensure that dataset is free from Not safe for work (NSFW) & private content. The potential to use the technology for unfair practices such as deepfakes and cybercrime is very high. The laws governing the protection of sensitive & private data are not well defined and there is a need for formulation of such regulations to guide the adoption of the technology at the global level. There is a direct relationship between the quality of input data and the output, so significant efforts are required to collect data to eliminate hallucinations and biased outputs.
Operate Responsibly !!!!
Hang on with us while we shortly publish the next article on how to adopt Generative AI & identify use cases for your organization!!!!
REFERENCES
