My monthly column, The Last Word, focused on a very hot topic for January, “(When) Will AI Overtake the Truss Industry.” Delving further into this burning issue, I turned away from random YouTube AI sessions and went to a better source, the School of Data Science at the University of Virginia. Behind the flashing binary code on the front wall of the building, I found a student and asked her if AI was just the next iteration in computer software. “No,” she replied dismissively, “It’s a better and faster way to write code and apply it.” Not accepting that my years of software training were apparently for naught, I looked further and found the work of Professor Rafael Alvarado.
Especially with “The Language of Artificial Intelligence Explained,” Dr. Alvarado’s work enabled me to peek through the dark veil of AI jargon and see concrete examples of AI inquiries. By mimicking these examples, I was able to take a baby step into Python, the language of AI, not by writing Python computer code but by using ChatGPT to write the code for me. Initially, I was quite impressed, but the devil was in the details.
To begin, I simply keyed into the Google search window, “Write a Python program to create a bio on Raymond ‘Joe’ Kannapell of Charlottesville, VA.” Instantly, I had the computer code to initiate the search, and when I ran it I received a vacuous characterization of my work experience, and an exaggeration of my civic involvement (I worked on the road for 40 years and didn’t have any “community initiatives”). By adding successive prompts, my bio gained more pertinent detail and even an impressive summation, “His retirement is defined by ‘The Four Ts’: Truss writing, Triathlons, Traveling, and Timber/Property work.” However, it stumbled slightly with “Timber/Property work.” Finally, though, after multiple prompts, Google’s paywall ended my session, reporting, “Upgrade your plan to the Plus for $20 per month or the Pro for $200 per month.”
In this exercise, I learned that the data science student was partially correct. AI offered a much faster and easier way to create computer code, in this case Python, which contains English language-like instructions. Python also distilled its findings on my career into three logical and appropriate categories, and all of its output was well written and grammatically correct. However, I was surprised at the inaccuracies.
To explain these inaccuracies, Dr. Alvarado describes how AI processes information on the Internet and extracts the relevant content using Large Language Models (LLMs). These models are trained on vast quantities of text data and are designed mainly to predict the next word in a sequence. As the sheer volume of data on a given subject increases, so does the accuracy. Apparently, there is not enough data about me on the Internet, which accounts for the mischaracterizations of my career, wrongly stating through early iterations not shown here that I led truss design innovation, and that I have a long-standing presence in the Charlottesville business community. Moreover, these models can concoct nonsensical phrases such as “Timber/Property work.” Dr. Alvarado confirms that these models often produce falsities.
Indeed, it is this propensity to produce falsities that I believe will limit the usefulness of AI in the component industry. As discussed in “(When) Will AI Overtake the Truss Industry?”, our industry has relatively little content on the Internet, other than general non-technical data. Some companies have large stores of technical design data, but much of it lacks the depth to be useful for estimating purposes. And these companies are unlikely to make proprietary data publicly accessible.
Where I take issue with the dismissive student is in her characterization that AI tools are “better.” Is a process that can give false results a better option? This problem surely limits its use as an estimating tool. In addition, the dominant development tool in our industry, C programming language and its many iterations, is known to be better for engineering and scientific applications, although Python may be more popular in other applications.
As I compare the Python code I generated today to the Fortran code I generated 59½ years ago, I don’t see fundamental differences. They’re both high-level languages that can perform multiple operations with a few concise and specialized commands. What is radically different about the function of these languages has to do with the existence of the massive treasure troves of data publicly available on the Internet. However, generating Python code can be a more expedient, though less accurate, method of gaining knowledge inside an individual enterprise than legacy languages.
Outside the area of component design, every other functional area of a business lives and breathes data, and much of it is captured in third-party application databases like Oracle NetSuite for accounting, Workday for HR, or Salesforce for Customer Relations Management (CRM). Each of these vendors already touts their AI capabilities that apparently can be leveraged by users. Yet, there is obviously potential for individual businesses to have staff trained to use tools like Python to query vendor-provided databases and other internal and external sources to aggregate the best management information systems. Weyerhaeuser, for example, views AI as a core component of its digital transformation strategy to optimize operations across forestry and manufacturing. Dr. Siddharth Goyal is MiTek’s Director of Artificial Intelligence, acting as Head of Data Science & AI/Machine Learning (ML), and MiTek is currently seeking an Advanced AI/ML Engineer “responsible for building and deploying new AI systems that advance MiTek’s software and hardware products.” Undoubtedly, these extensive AI initiatives will ultimately impact the way component plants store, use, and retrieve data, but I believe that AI efforts will be focused long into the future on the areas outside the realm of component design.