The End Game of Digitization

Artificial intelligence pushes decision-making to the edge.

In January, Bloomberg News reported that Microsoft is investing $10 billion in OpenAI, a startup company specializing in artificial intelligence (AI). OpenAI is the maker of ChatGPT (an online chatbox) and other tools that can generate readable text and images in response to short prompts or questions.

In terms of ready-mix concrete U.S. EBITDA, Microsoft’s investment equals to about 625 million cubic yards. 

Last year, U.S.-based venture capitalists invested $115 billion in AI companies, according to The Economist. That’s worth a U.S. EBITDA equivalent of 7.2 billion cubic yards of concrete! Worldwide investment is easily at least equal to the gross domestic product of Denmark ($332 billion). Some of the investors are likely foolish, but the preponderance of cash is coming from very smart, careful sources. The questions we need to ask are why it matters and how it relates to our industry.

AI is based on artificial neural networks, inspired by biological neural networks of the brain. At its root, the network does not seek an absolute answer, like yes or no. Instead, it seeks a “good enough” answer by evaluating lots of parameters. It “learns” as each parameter is matched to human input to create an overall pattern—be it numbers, phrases or images. This is called supervised learning. 

The theory has been around for almost 80 years. By the 1990s, computers powered by Intel, and later AMD, were able to handle 1,000 or so parameters. The size in terms of parameters of an AI model was enough to prove the potential of theory, yet it was insufficient for widespread application.

Fast forward to today. Oak Ridge National Laboratory has invested over $500 million to create the fastest known supercomputer in the world, named Frontier, which is estimated to handle trillions of parameters. AWS, Microsoft, Google and many others have an orchard of supercomputers to distribute computational loads. While each computer is less powerful than Frontier, their combined capacity is ginormous. Bang—the power of AI is unleashed.

Now, leading AI models are self-learning or unsupervised. Consider every sentence of every known literary work. Then one by one for all the trillions of words, the computer removes one word from its sentence and then tries to predict the missing word. Each try is auto-scored based on the millions of books and the range of suitable words found for the parameter. With trillions of parameters available, a modern AI model needs merely a basic description of what the “story” should communicate and, voila, the AI model writes it better than most humans could imagine.

In our world, consider an unsupervised model based on every recorded dispatch decision for ready-mix concrete including peripheral information—such as the batch weights, weather, geographic location, customer descriptors and so forth. This is called a “foundational model” for ready-mix concrete dispatching. With several billion self-supervised learning points, we can ask the model, “What orders will be canceled tomorrow in the Bronx, NYC, for outdoor flatwork projects by driveway contractors, given a 50-percent chance of moderate rain?” Odds are high that the model will quickly yield a very accurate, but not exact, prediction.

What this means for our industry is the continuing flattening of the organization. AI will continue to push actionable information outward to the edges of the interface between supplier and producer, producer and finisher, and producer and infrastructure owner. The historical layers of management will be mostly reduced in favor of front-end people who can make good decisions based on better information—empowered by the help of AI.

The current edition of the Harvard Business Review includes an article headline advising readers to “Stop Tinkering with AI. It’s time to go all in.” The thrust of the message is to free yourself from inflexible business processes designed by inflexible IT architects. The old-school technology model looks for deterministic solutions that result in exact answers. While we need this in many areas, like billing and accounts payable and simple question-and-answer requests, a large part of our world is better served by the more flexible “close enough” AI model.

Full disclosure: My company, BCMI Corp., is actively investing in foundational models and AI for the ready mixed concrete industry.

I visited and entered: “artificial intelligence foundation models.” Here is the AI-generated answer:

Artificial intelligence (AI) foundation models refer to the underlying models and algorithms that are used to build AI systems. These foundation models provide the basic building blocks for more complex AI systems, and include a variety of techniques such as supervised and unsupervised learning, deep learning, and reinforcement learning.

Not bad. There was a grammatical error in the second sentence, in the form of a poorly placed comma, but the content was excellent. Perhaps can write this column going forward!

Craig Yeack has held leadership positions with both construction materials producers and software providers. He is co-founder of BCMI Corp. (the Bulk Construction Materials Initiative), which is dedicated to reinventing the construction materials business with modern mobile and cloud-based tools. His Tech Talk column—named best column by the Construction Media Alliance in 2018—focuses on concise, actionable ideas to improve financial performance for ready-mix producers. He can be reached at [email protected].