Limitations of Regenerative AI for Operations Management Problem

Regenerative AI systems, which are all the rage now, are based on statistical self-adapting non-linear matrix correlators.

These matrix correlators are very useful for searching the internet for information, especially specifications and regulations, generating first-cut synopses of long articles, recognizing and generating images. 

But they have great limitations, because they are statistical correlators and we all know, when making management decisions, that there are “lies, damn lies, and statistics” because correlation is not causation.

Also, they have no sense of time, space, sequence, or process as shown by the image shown at right, generated by the DALL-E3 regenerative AI algorithm when requested to generate an image of 3 people looking at a graph on a computer screen.

For this reason, in industrial settings, regenerative algorithms are best used for decision support where the output will be reviewed by people, who can detect and correct for these errors.  Regenerative AI algorithms should definitely not be used in fully autonomous decision making.

The primary reason for this is the size of the needed training set. To train a self adapting correlator, you need to feed it a sequence of inputs and have a person tell the system what the input image represents, such as dog, cat, kangeroo, which the correlator uses to adapt its matrix coefficients.

To get high accuracy you need a big correlation matrix and many training samples, which is why algorithms such as ChatGPT attempt to use the Internet as a training set.

It has, however, taken human beings over 25 million years, with training inputs every 100 milliseconds or so, with severe pruning of the results through countless wars and conflicts, for people to learn how to form management hierarchies.

This is how long it has taken to learn this capability, as we evolved from primitive lobster-like species, in the Miocene, over 25 million years ago.

This level of training is obviously not feasible in an industrial setting.

Also, it takes hours or days of time on huge expensive supercomputers, to generate a correlation matrix from even selected portions of the Internet. Then it takes minutes of expensive computer time to query the correlation matrix to get an approximate answer. Again, this limits the feasible applications due to cost, as well as accuracy considerations.

Regenerative AI systems are, however, beneficial in enabling people to rapidly analyze large amounts of data, such as the operations management data collected by intelligent agent systems such as SmartOps24x7.

Please click here to learn How Intelligent Agent Systems Work.