Industrial AI – Expectations to Implementation

Back in 2012, a Silicon Valley venture capitalist was quoted as having said, “Software is eating the world”. He made a lot of hay investing in software companies. ?Then in 2019, Jensen Huang of NVIDIA took it further, “AI is eating the software that’s eating the world”.? Recently NVIDIA crossed the $3 Trillion market cap, becoming the world's most valuable company.

Here is my take on the phrase, based on the state of AI and Generative AI in 2024 – “Generative AI is eating the AI that is eating the software that’s eating the world”. Given the $3 Trillion valuation for an AI chip company in a $100 Trillion global economy, it remains to be seen how much headroom is there for others and which other Generative AI companies will benefit the most from this AI tech wave.

The year 2023 brought Generative AI into the limelight with OpenAI's ChatGPT.? By the summer of 2024, we seem to be moving beyond the peak of inflated expectations toward the slope of enlightenment, as described in Gartner’s technology adoption hype cycle.

Moving from the initial period of "expectations", Industrial AI is finding its place amid the realities of rubber meeting the road "implementations".? Based on a recent MESA survey over half of industrial companies believe AI has a place, but finding “Value” (ROI) has not been as straightforward yet.??

Using the startup analogy this is similar to finding the right product market fit! Every startup goes through that stage and in the enterprise context, Industrial AI will need to go through that as well. We think that will happen in the next 12 to 18 months.

At the foundational level, Data is what drives all AI and Generative AI models.? For wider Industrial AI applications Data Quality, Data Governance, and trustable Data Historian have been the Achilles’ heel for Industrial AI applications.? Other than, unstructured sources like emails and hardcopy instructions, most industrial companies have some form of PDM, MES, and/or ERP system that provides the historical data for AI (or Generative AI) models to be trained from.? Even before one gets to training the AI models several data pedigree, data quality, and data governance matters need to be addressed.

Then we get to the AI or Gen AI models side of things.? Original LLMs (Large Language Models) were found to be too broad, lacking contextual intelligence, and therefore too expensive to work with. ?Now, the use of modified LLMs or SLMs (Small Language Models) is growing in Industrial AI adoption for various reasons, including, SLMs do not require specialized hardware, SLMs are easier to fine-tune and customize for specific needs, the latency in SLMs is much faster and not to mention the computational cost of SLMs is significantly lower and also environmentally more friendly. Training costs for AI models for specific use cases are still unclear and will likely come as an afterthought.

Beyond these one still has to ensure the trustworthiness and reliability of AI – bias, hallucinations, explainability, data privacy, and governance for ethical guardrails and proprietary usage.

This summer of 2024 brings a sanguine outlook for adopting Industrial AI. The industry has moved beyond the 2023 "peak of inflated expectations" and is poised to move towards the "plateau of productivity" on the Tech Hype Cycle. Organizations that have built necessary data foundations, and figured out how to organize, architect, and use their proprietary data for net positive ROI use cases are reaping great benefits and competitive advantage.

How are you seeing things evolving at your organization?