The 10-minute guide to GenAI for Software Professionals

In an era where technology evolves at lightning speed, staying ahead as a software professional means embracing the latest advancements. Among these, Generative AI (GenAI) stands out as a game-changer. However, with the vast ocean of information surrounding AI, it’s easy to feel overwhelmed or unsure where to begin. This guide cuts through the noise, offering you a concise yet comprehensive overview of GenAI tailored specifically for software professionals.

Before we dive into the specifics of Generative AI (GenAI), let’s first ensure a foundational understanding of what Artificial Intelligence (AI) even means, and it’s relation to other buzzwords that you might have come across.

What is Artificial Intelligence

In it’s most basic form, Artificial Intelligence (AI)is a field of computer science that enables computer systems to perform tasks that typically require human intelligence. Some examples of these tasks could be reasoning, learning, problem-solving or decision-making. It is an umbrella term that encompasses several other technologies such as Machine Learning, Deep Learning, and most recently GenAI.

Early days of AI were limited to hard-coded rules and algorithms. Think of playing a chess game against a computer where AI would brute force, or look ahead for all possible moves that you could make next. One would throw more computing power at the problem to get same results faster.

Interacting with a chatbot on the other hand would mostly lead you to canned responses, that the software developer considered for programming the system. As you can imagine, the extent of intelligence was limited to what these systems have been trained around, or limited by the knowledge of engineers themselves. Or more importantly, these systems did not have a capability to think on their own.

Over the past years there have been many advancements which now allow us to make the systems more intelligent, enabling them to go above ad beyond. But how do you go about training machines to develop this capability — enter Machine Learning!

Using data and algorithms to train machines

Machine Learning (ML) refers to the process of making a machine learn facts, associations and relations from existing data. The output of this process is a model — an artefact that is later used to make predictions or decisions when responding to queries from the end-user. Needless to say, the quality of the input data pretty much defines the quality and accuracy of the output — what you give is what you get. Therefore, frequent testing and iterations are logical next steps that would follow after a model is created for the first time.

The bigger challenge with data, however, is the vast quantity and varying levels of quality that we deal with. Oftentimes data is structured, such as CSVs, databases etc., but sometimes the processes also need to deal with text, audio and video content, which are largely unstructured. In some cases, the data might be labelled (for example, images of cars with embedded metadata like the different components, engine type, color etc.), whereas in other scenarios it might be un-labelled — requiring the processor of that data to comprehend it in different ways and derive relationships automatically. All of these factors can play a big role on the efforts needed to cleanse or prepare the data before a ML training process can even begin. A model trained on 1000 Orange images will identify the object better than one whose input data consists of random fruits without any additional context or metadata.

Learning mechanisms that use labelled input data are known as supervised learning, while the ones using unlabelled data are known as unsupervised learning. Wait — can we not adjust the learning process as we go? Yes, that’s possible as well — with reinforcement learning! In this case, we use a feedback mechanism that is provided to the machine in form of rewards or penalties for the output it gave, and over time, this leads to better outcomes.

To put it simply: Machine learning is like training a very eager student by showing them many examples. The more data/examples this student sees, the better they can grasp the patterns and relationships, and eventually make smart predictions about new situations.

As discussed at the beginning of this post, AI is all about making computer systems function like a human brain. Another concept which takes all of this to the next level is Deep Learning!

Thinking like the human brain

Just like how our brain is formed of many many neurons which process information and establish connections to each other, deep learning works with nodes (neurons) and multiple layers — each formed of multiple nodes. Think of each node connected to multiple others that represent related information. They have one input layer, many hidden layers and one output layer.

During the training process, deep learning works towards establishing relationships around existing information, inferring new connections, and patterns that can be leveraged to deduce new outcomes that might not have been seen already. This is key, because in traditional ML training processes, you were limited by what the data represented. Now, the machines can infer more than that.

It should be clear by now that the real value with everything AI is not working with known facts, but rather exploring and automatically deriving new information, new patterns, and brand new outcomes — something that might not have been thought of, or experienced before. Generative AI is exactly this! Writing stories, composing music, generating code and creating art are some of the fundamental use-cases that are heavily influenced by recent innovations in the GenAI space.

Getting creative with AI

It’s kind of interesting that investments in Machine Learning space have been happening since many years now but the GenAI rocketship launched just recently. This can be largely attributed to the vast amounts of data now being available, cheaper compute and infrastructure resources, and finally, the willingness to explore uncharted territories.

As an individual user, the most exciting part of all this is the general availability of models that have been trained on internet-scale data. You no longer need to collect and work on classification of input data, or bother yourself with pre-training tasks. All this has been done by companies that had access to high quality data — both labeled and unlabelled, resulting in creation of several groundbreaking models — also known as Foundation Models or FMs. These foundational models can be used for myriad of tasks such as text generation, chatbots, text summarisation and audio generation, to name a few.

For example, models like DALL-E can create original images just from a text prompt like “a surreal painting of a cat drinking tea on Mars.” On the text side, GPT-3 can write creative stories, poems, scripts and more based on a given starting point.

Note: All the images you see in this post have been created by using some of the famous FMs such as Amazon’s Titan Image Generator and Stability AI. If you are an AWS user, it’s very easy to use these models (and others) through the Amazon Bedrock service. Tip: Use regions like us-east-1 for more options.

But, there must be a lot of steps behind making a Foundational Model actually available for all these diverse use-cases, right? Let’s learn about them.

Typical lifecycle of a foundational model (FM)

A foundational model is trained on internet-scale data which is generally unlabelled but sourced from variety of trusted input sources. The ML algorithm then leverages the structure within these datatypes to autogenerate some labels by itself. This helps the algorithm to learn the meaning, context and relationship between different words of the dataset.

A logical next step is to evaluate the performance of such a model by using appropriate benchmarks and metrics. The main goal here is to understand if the model meets the needs of a business. At this stage, there are additional pre-training mechanisms that can be used to enhance the model further.

And finally, as you can imagine, the model is deployed for use in a production system. It can be embedded within the target application itself, consumed as an API, or abstracted behind a GUI. The most important takeaway here should be that training and optimising the model is an iterative process. This is achieved by sharing feedbacks at every stage back to the model, so that subsequent iterations can improve even further.

There are different categories of Foundational Models that can be used for variety of tasks such as Large Language Models (LLMs), Stability Diffusion Models, and Multi-modal Models, but lets discuss them in another post!

While generative AI unlocks new frontiers in creativity, it also raises concerns around potential misuse like spreading misinformation, perpetuating biases, or displacing human creators/artists. As these models become more powerful, developing guardrails through responsible AI practices will be crucial.

As of Aug 2024, there are a bunch of FMs offered by companies like Anthropic, Meta, Mistral AI and Amazon, to name a few. You can use these foundational models as-is or even fine tune them for your specific use-cases. There are many methods that can be used to shape them in a specific way, such as Prompt Engineering or RAG etc.

Thank you for reading!

I hope this guide provided you with valuable insights into the world of Generative AI for software professionals. If you found this post helpful, feel free to share it with your network or leave a comment below — I’d love to hear your thoughts and experiences.

For more content on cloud, AI, and software development, you can also connect with me on LinkedIn for more updates and discussions on the latest trends in tech.

Stay curious, and keep learning!