The biggest hype surrounding AI seems to be over – at least according to my trustworthy LinkedIn index (which means scrolling through my feed for a few minutes every now and then). A couple months ago every other post was AI this and ML that, but now there was only one post mentioning AI. Even though AI was supposed to revolutionize everything in a matter of weeks.
But AI is no different from any other trend or system for that matter. All sorts of snake oil salesmen tend to crawl out of their hideouts to ride the current wave promising that this time around putting garbage in doesn’t result in garbage out. These inflated expectations inevitably led to a collective hangover as AI wasn’t as easy to implement as it was marketed.
The unfortunate truth about AI is that it is only as good as the data it is trained with. The reason why ChatGPT doesn’t do our dishes or laundry yet is because it doesn’t have enough data to accomplish these tasks (also it doesn’t have a physical form). As the LLM is trained with data originating from the Internet, it is good at making recipes and such that are widely available, but asking it to do something more sophisticated probably only causes confusion. You can still ask it how to do it, even though it can’t do it, and it’ll give you instructions.
Our AI pilots have faced the same issue: putting garbage in leads to getting garbage out. These GIGO failures are, in my opinion, the biggest reason why AI pilots remain only pilots and don’t progress further also causing frustration and delaying the start of possible successes. Filtering and enriching data takes either blood, sweat and tears from yourself or you pay a hefty sum for an expert to do it, which seems unreasonable even though it is the most crucial part of the project.
Our first pilot was to predict the need for materials in our stock, which was basically doomed to fail from the beginning, but we had to do something so I could present the case after the AI training course. We knew the data we had was incomplete, but still we managed to create a model that is useful once we have better data from our production. And this should become reality soon as we are currently ramping up our new MES.
The second pilot (also a demo for a course) was also doomed to fail as once again the data set was incomplete – it is currently waiting on my desk to receive better labeling and links between the tables. Still the idea of a predictive maintenance algorithm for our production is so revolutionary that we’ll continue to work on the case. And we hope that the new MES is more willing to hand over the data it possesses in a useful form. The MES data together with the data from the machine monitoring system should give us an actual possibility to accomplish this.
The third pilot relates to ML. We are trying to create an algorithm that recognizes gaskets from cuttings and learns how to pick them up with a cobot picker. This is probably a longer-term project but at least we have created a data set and designed a picker suitable for our production that is currently being tested. The algorithm hasn’t progressed much due to lack of resources, but hopefully we’ll have something to brag about in the coming months.
In addition to these pilots, we have plenty of AI ideas in the pipeline. The lowest hanging fruit is of course implementing an internal Copilot within our office environment. It is, however, such a powerful tool that we want to ensure that it doesn’t cause harm. It doesn’t obey borders, so we must educate it to do so before letting it answer any questions. Luckily, patience is our middle name.
But hopefully sooner rather than later we’ll have a ChatTTG tool to help you with your sealing issues to provide you with Smarter Sealing for a Safer Tomorrow. But before that happens you can always contact us the traditional way.
