Let’s Talk About Generative AI and AIoT

Well, something big happened last week: we announced that SpinDance is merging with Mutually Human. Check out the press release to read all the details. The tl;dr version is this: they’re a phenomenal digital engineering firm specializing in artificial intelligence, data and software development. Together we’ll be able to provide a comprehensive approach to collecting, organizing and leveraging data in this increasingly connected world.?

As we continue to share notes on the market, the same observation continues to come up time and time again: people are really curious about generative AI, and how it can help maximize their investments in things like IoT. Given that Gen AI is white hot, this makes all the sense in the world. As the next “big thing,” people want to be ahead of the curve.

It’s also clear that most people are still forming an understanding of what Generative AI can–and can’t–do. For example, the other day a customer asked me about using Generative AI to perform anomaly detection for IoT devices. He was disappointed to hear it’s not a great fit.?

We’re sure he isn’t alone, and thought it would be a perfect topic for this month’s issue of The Intelligent Device.?

Generative AI vs. Discriminative AI?

When we talk about artificial intelligence, we’re really talking about two broad types: generative and discriminative AI. Discriminative AI, for example, is fantastic for classifying or labeling data. It’s what powers a lot of common applications, like spam filters or – in the IoT world – anomaly detection models. Imagine monitoring an HVAC system for unexpected patterns in temperature or airflow; a discriminative AI model can spot irregularities and flag them, helping prevent system failures before they happen.

On the other hand, generative AI doesn’t just classify or label; it creates. Think of it as a type of AI designed to generate content, whether it’s text, images, or even code. For instance, when you ask ChatGPT a question, it’s generating a response based on the information it has learned, rather than just categorizing your question.

Generative models can only produce content based on the patterns they’ve been trained on, which means they’re drawing from a mix of generalized data. This is where a challenge arises for IoT applications: since anomalies are often unique to a specific type of IoT device or setup, a generalized generative AI model might struggle to grasp the intricacies of an unusual situation. In cases like this, the model might even produce information that doesn’t match reality, something we sometimes call “hallucinations.” These hallucinations can lead to misinterpretations or even false insights, making generative AI less reliable for scenarios requiring precise, device-specific anomaly detection.

Generative AI and Discriminative AI?

An important concept to grasp is that generative and discriminative AI aren’t an either/or choice—they’re two sides of the same coin. Each is powerful alone, but together they offer even greater value. For example, imagine a discriminative model detects an anomaly in an industrial machine’s performance. A generative model can then interpret that data, providing a detailed response with potential causes and recommended actions. This way, users receive not just an alert, but actionable insights.

Think of it like a team with specialized roles: discriminative models excel at identifying what’s wrong, while generative models explain or contextualize those findings. Understanding each model’s strengths lets you design AI systems that are both accurate and user-friendly.

On a factory floor, for instance, a discriminative model might flag a temperature spike in a motor, indicating a potential issue. But not all temperature spikes mean the same thing—some might result from increased load, others from maintenance needs. Here, fine-tuning a generative model trained on the specific machine’s data becomes essential. This model can then offer a more tailored response, such as, ‘Overheating likely due to prolonged high load; inspect cooling system or consider reducing runtime.’

Fine-tuning allows the model to adapt to the unique context of each machine, making AI insights more accurate and relevant. Over time, as the generative model learns from real data, it produces increasingly reliable insights. This approach not only helps with immediate troubleshooting but also contributes to a deeper, long-term understanding of optimal machine operation and maintenance.

Harnessing AIoT’s Full Potential with Dual AI

Once you understand the differences, it’s exciting to see how generative and discriminative AI can work together to create smarter, more responsive systems. By leveraging each type’s unique strengths, we’re not just building systems that detect issues; we’re empowering users with insights that drive real action. And as fine-tuning makes these models ever more adaptable, the future of AIoT looks increasingly personalized and precise. Here at SpinDance (now with Mutually Human!), we’re thrilled to explore this evolving landscape, creating solutions that bring out the best of both worlds.