We are in the middle of a “hype-cycle” (1) regarding artificial intelligence(AI). For better – the excitement of new technology gets us to talk to one another. And worse – a lot of energy that could advance the technology is lost on all the hype and unrealistic expectations. Given all the noise, what can we do that is productive?
Like the mRNA technology for the COVID vaccine, the current AI developments have been a long time in the making. What we see now is step change as we reach a tipping point in the size of computing power and our creativity in using it. Moore’s Law has held true for a half century (2) and the growth of cloud solutions (3) makes it easier to access massive computing power. The rollout of ChatGPT (4) got our attention focused on all of this put together.
It is a step change that will touch every part of a business. The “bot” that helps you navigate a consumer service site is suddenly more intelligent. The model that predicts your supply chain needs, is more accurate. Large technical data sets yield previously undiscoverable insights. Your car can drive itself. And then the dark side: is that news clip real? Is this really a friend in trouble asking for a wire funds transfer?
In a quest to see what could be done beyond listening to the news, I took on using a machine learning and artificial intelligence tool on behalf of one of my clients with a formulation data set.
This type of work involves a large number (typically at least 10) important variables all of which have multiple sub-properties whose measurements are of limited accuracy and the properties are interconnected. There are some heuristics and simple models which can be built to aid the task of selecting a formula to meet a specialized product profile, yet the bulk of the work comes from the analog version of the neural net: scientists with long experience using the palette of materials and manipulating the variables.
With the human experience component dominant, teams tasked with deciding on a formulation often are limited in the amount of data that can be grasped and perceptions influenced from previous work. A lot of time and resource can be expended in sorting out those prior notions and multiple rounds of experiments performed to ‘discover’ what the data is shaped like – finally leading to a formulation that meets requirements. Having the data set for a project in process yet still searching for the right formulation a very kind machine learning-artificial intelligence platform (ML-AI) company agreed to let me test the platform for this project and see if we could better quantify formulation effects and accelerate the creation of a formula meeting requirements.
What I found was amazing. The computing power and vivid multidimensional visualizations showed insights into the data that even with three decades of experience and an experienced team working with me, we did not see. The data analysis allowed quantitation of the performance of this particular formulation class that were previously subject to long discussions and fuzzy conclusions of prior experience. After multiple cycles of formulation experiments not yielding an answer the next round of experiments, guided by our use of the platform, hit the target exactly.
It should be noted that the input to the tool was critical. The data needed lots of curation to assure it was relevant and cleansed. Not all the variables could be processed at once. The selections needed human input to have the ML-AI model yield stable results and useful predictions. Even then the predictions needed some adjustment by humans as not all the variables could be included, or a large enough data set be generated to increase the accuracy of the model. The bottom line was: using the tool was exciting, useful, and saved a lot of time and money.
The ML-AI model was a big win and is being implemented in other parts of the program.
A second test was of the famous ChatGPT. This tool had, in comparison, nearly unlimited data training and refinement. A simple question was posed: what is the best view in New York? ChatGPT returned readable text which named several good suggestions of good views in New York City. This would have been very useful in planning a trip. Yet, human intervention was still needed. One of the best views from The Edge was not included. Nor was a less well know view at a building nearby. ChatGPT also missed that I might have meant a more natural view from New York State such as the top of Gore Mountain. The ChatGPT result was a good, time saving start at a human generated solution. Unless a hybrid approach was taken it was potentially misleading and did not contain essential data.
AI is clearly a powerful tool that will change everything as we continue to develop the underlying technology, improve existing and create new applications, and develop the commercial case to build the datasets, and training needed to make it useful. The potential to make work and creative pursuits easier, more enjoyable, and economically valuable will increase human output in the years to come. It has great power and the potential to cause great harm in false or even malicious results. Yet, working together this harm can be avoided as humans have done by avoiding the ills of nuclear technology for almost seventy years. Even as this is written the governments are beginning the infrastructure needed to guide our invention such as the NIST ARIA initiative in the US or the EU AI Act.
You can play a role. Go to your stakeholders and investors. Make the case for incremental, well-constructed, and useful advances in even the smallest AI application that touches your business systems and technical applications. Advocate for the role humans play in using these tools and how that makes the results valuable and also benign. Remember to enjoy the hype show while it lasts so that you are part of the conversation and generate new ideas.
References and notes
- Gartner Hype Cycle (https://www.gartner.com/en/research/methodologies/gartner-hype-cycle) Accessed May 31, 2024
- Our World in Data – Moore’s Law (https://ourworldindata.org/moores-law)Accessed May 31, 2024
- Cloudzero – cloud computing growth (https://www.cloudzero.com/blog/cloud-computing-statistics/.) Accessed May 31, 2024
- Chat GPT ( https://chatgpt.com/ ) Accessed May 31, 2024
- Photo by author at the Edge, New York City, NY April 16, 2024
- Photo by author with editor at a New York City, NY location April 30, 2024
- Gore Mountain, NY website – (https://goremountain.com/the-mountain/trail-maps/) Accessed May 31, 2024
- NIST ARIA program (https://www.nist.gov/news-events/news/2024/05/nist-launches-aria-new-program-advance-sociotechnical-testing-and Accessed May 31, 2024
- European Parliament AI Act (https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence) Accessed May 31, 2024
