AI and data: tackling the hype, uncovering the value
Today, so much online chatter is centred on data and the use of AI. But what exactly do we mean when we talk about these terms and what influence are they having on software development. I recently spoke to 101 Ways’ AI and Data lead, Jon Parish to go behind the hype and get some practical answers.
AI: understanding what it is and what it isn’t
Artificial intelligence (AI) is a very broad and very deep subject—taking in everything from data access and analysis through to old school machine learning. Most of the work (and the hype!) right now in dev teams and boardrooms is focused on getting a handle on how best to take advantage of Large Language Models and AI-as-Service offers from the likes of Amazon Q, Google’s Vertex Ai and OpenAI.
Generative AI in its current form, especially with services such as GitHub Copilot, or ChatGPT, certainly offers much and can help with a vast and growing range of tasks – everything from helping individual productivity through to writing code. However, it’s not always clear where the best use cases are or indeed how to turn PoCs into workable realities at scale across the organisation. But it’s new, so there’s always going to be questions. And there’s little doubt GenAI represents a major new category and a significant step change from previous incarnations of machine learning.
The advance of AI vs. the hype
Although some AI models and systems can initially impress, a deeper analysis often reveals gaps and problems and limitations become quite obvious. For example, as useful as GenAI can be, it’s only as good as the data it has been trained on. Bad inputs lead to bad outcomes. Not much surprise there. Similarly, AI systems can inadvertently perpetuate or exacerbate existing biases present in training data. The result is unfair, discriminatory or ‘just plain wrong’ outcomes. And sometimes, particularly in deep learning models, it can be incredibly hard to understand how the AI arrived at its conclusions. The ‘black box’ nature makes transparency and explainability a major challenge. There are lots of other considerations too—hallucinations, ethics, etc.
That’s not to say we should go ‘full luddite’ and write AI off. There’s way too much value to be gained from embracing the technology AI in dev teams and across the business. Rather, as Jon says “validation is critical. And for language models, where validation is more difficult, you need a human in the loop to catch the mistakes.”
One of the most helpful ways to view Gen AI is like a better stack overflow search that generates or retrieves solutions to a variety of programming or technical questions. AI saves time and can generate or adapt simple code, but you wouldn’t put it straight into production without testing and validating first.
Training the data and a question of trust
When considering code generation, it’s worth remembering that signing up for AI-as-a-Service models will often mean sharing some company information with the provider. In theory, it could even become part of the Gen AI’s training data, depending on the terms of the contract signed. Many organisations find that to be an understandably unappealing prospect.
Of course, it is still possible to train your own models or to get a pre-trained model and run that locally. And while AI as a service will give you a quicker and cheaper start, you get more scope and control with training your own. As a result, your outcomes will likely be more contextually aware and accurate, even with a much smaller data set.
That’s not to say that there aren’t challenges when it comes to training a model in-house, however. To train a large language model, for instance, you’d need a huge set of data for it to draw from – literally billions of examples. Also, training your own can be technically complex and require a lot of expensive compute time. And, at that kind of scale, it becomes almost impossible to know whether your model is being fed “good” or “bad” content.
Because of that, you need to be familiar with the data to understand what good data looks like, particularly if you’re going to train your AI on it.
As Andrej Karpathy —one of the cofounders of OpenAI— notes, visualisation can play a critical role in helping you to understand your data. With that understanding, you’ll be able to set a standard, get an idea of the data pipeline, and the degree of normalisation required to remove training set duplicates.
Model training to avoid silent fails
With software, you need to have good testing to be able to know that what you’re producing is solid. It’s this x1000 for anything to do with machine learning. Machine learning models tend to fail silently. They won’t give you an error message. Instead, they will just miscategorise things or start producing garbage. To avoid this, businesses must continually validate their models – and that means having access to ‘clean’ data.
One of the key rules when it comes to training an AI model is that you should never use the data you’ve used for training for validation too. If you collect new data all the time then you’ll always have some ‘fresh’ data to use for validation. If not, then it’s important to hold some back for that specific purpose.
Of course, even a model that delivers great results initially can eventually start to drift. That typically occurs when a model has been trained on a limited dataset – one that relates to a specific point in time, for instance.
“If you trained a model using consumer shopping habits in the middle of a hot summer when everyone was having barbecues, then it probably wouldn’t perform very well come winter,” says Jon. “That’s not a case of the model itself getting any worse, simply that behaviour has drifted away from the original pattern. A model will only ever be as good as the data that it is trained on.”
“Every business changes over time, which means that its underlying data does too,” he continues. “As a result, ongoing validation is important, as it provides a way to ensure that AI models will continue to work even if the data they’re being fed is different to what they were trained on.”
Formulating a strategy and setting expectations
Right now, it’s easy to overstate the benefits and play down the potential pitfalls. Here, Jon sees parallels with cloud computing, where the ease of setup and ability to scale drove the market. Once ‘up there’ many firms found they’d not only massively overprovisioned but also ignored the significant costs of data processing—leading many firms to halt initiatives or bring workloads back on-premise.
As with any new tech, firms and departments need to thoroughly evaluate their needs, understand the particular risks and ask whether AI (of any kind) is the right approach. Is there a business case? Will it deliver value? And significantly, what are the cultural implications? Rather than being driven by hype, understand what it is you are signing up for. In other words, review things sensibly and use control groups and continuous testing to ensure robustness.
When deployed correctly AI can help firms to understand their own business characteristics more effectively, allowing them to scale more efficiently. Similarly in-depth numerical analysis – particularly from own model training – can also unlock new patterns invisible to the human eye and aid in product innovation.
It’s important to remember that AI isn’t a silver bullet for success. But by doing the engineering work and product validation piece, you can avoid the pitfalls and discover just how data and AI can deliver new realms of possibility.
Looking for tips on your data and AI journey? We can help.