Part of my role as a Counter Tech Lead is to keep up to date with the latest trends and technologies and what relevance they may have in the workplace. The advent of Machine Learning, AI in general, and Generative AI (Gen AI) in particular has certainly been hyped as a turning point of the modern era, with many suggesting it to be as seismic a shift as the Industrial Revolution. But is that hype warranted? The last major step forward was the creation of the World Wide Web, something that is now ubiquitous in all of our working lives, but the dot-com bubble is a stark reminder that hype does not always lead to immediate results or success.
Getting Hands-on With Generative AI
I already had a certain understanding of some of the capabilities Gen AI unlocks through personal research and experimentation, but I was keen to know more, and how we can leverage the uses to help our own developers, as well as provide support to our clients and their customers who are finding use cases for it. I learn best by building, and as an AWS Community Builder and a serverless enthusiast, along with carrying out research I wanted to build something of use that incorporated those technologies on the AWS platform. The result was an AWS tutorial on building a CV chatbot, which you can find a link to here, but I am also going to provide some context about the processes used in this article.
Understanding Large Language Models
Large Language Models (LLMs) are the bedrock (remember that word, it’ll become important later) of Generative AI. These are Machine Learning models on a massive scale, trained for thousands of hours on huge amounts of data, allowing them to respond to natural language prompts in a conversational and intelligent manner, or to create pictures, images, or graphs from a simple description. ChatGPT is obviously the best known, but there are many others from various companies – Google, Meta, Amazon all have their own models – as well as AI-specific companies like Hugging Face, Anthropic, and Cohere. There are strengths and weaknesses with each, and for the most part, accessing them in their basic form provides an entertaining distraction, without huge levels of expansion into unique use cases that can be leveraged for business goals. You may have also heard of the term ‘Hallucinations’, where the LLM confidently provides an answer to a question by completely making up facts to support its claim. While that is being improved in subsequent iterations, they aren’t really bugs – the very nature of an LLM is to “people please” and hallucinations are part of that effort to provide a satisfactory response to the user.
RAG: Making AI Reliable and Accessible for Businesses
We can mitigate these side effects, however, by supplying the model with data from which it can find and cite facts in its answer, giving you more confidence in what it is presenting. This process is called Retrieval Augmented Generation or RAG, and it’s where the majority of immediate use cases for Generative AI are becoming apparent – giving businesses the ability to interrogate and find insight in their own data using natural language. There are limitations – the data is searched semantically so context and iterative reasoning by the LLM would be difficult – but it’s a big step forward in providing real-life benefits to business. AWS has made accessing different LLMs and adding RAG data stored to them with Amazon Bedrock (see, told you it was important); a fully managed service which provides you an API endpoint to configure your requests to. This makes iteration incredibly quick – you can try out different models using the built-in sandbox, and with the click of a few buttons, you can stand up a working RAG system in a few minutes. For my tutorial, I made use of AWS’ Infrastructure as Code solution so that the result is repeatable and adaptable, a move away from proof of concept and towards productionisation.
Generative AI for All
If this new world of AI interests you and you’re looking for a way in for your business, feel free to use the tutorial as a jumping off point, either trying it yourself or passing it to engineers in your teams to build. You don’t need to know everything about how LLMs are trained, or the hyperparameters required, or even to fully understand the new discipline of prompt engineering to get stuck in. You can work those things out as you go along, if and when you need to. The most important thing is to see how these tools can be leveraged for your business use cases. And if you need any further guidance, get in touch and see how Counter can help you explore this incredible new frontier.
