Large Language Models

I'm writing this in May of 2026.  For the purposes of the essay, I'll categorize generative AI images along with large language models.  This is primarily due to my layman's understanding of the subject.  According to experts, all LLMs are generative AI, but not all generative AIs are LLMs.

When ChatGPT was first turned loose on the public, I was excited.  I had seen an encouraging review that it could produce pretty good G-code (programs to run CNC machine tools) given sufficient supervision by a skilled machinist who knew what to ask and how to ask it.  

My first interaction with the free version of ChatGPT was a request to improve the readability of a fairly convoluted code fragment written in the C programming language.  It worked!  But all my subsequent interactions with LLMs have been disappointing. 

Here's an example from January 2023.  I asked ChatGPT, “What materials have an approximate density of 18.7 grams per cubic centimeter?”

The reply:

• Titanium (Ti)

• Niobium (Nb)

• Manganese (Mn)

• Molybdenum (Mo)

Then I asked it to, “Provide densities for the 4 materials mentioned.”

• Titanium (Ti): 4.51 g/cm³

• Niobium (Nb): 8.57 g/cm³

• Manganese (Mn): 7.21 g/cm³

• Molybdenum (Mo): 10.22 g/cm³

All of those densities are accurate, but none is remotely close to my target of 18.7 g/cm³.  Pure tungsten at 19.25 g/cm³ would have been a reasonable answer.  I was trying to identify a tungsten alloy.

My next significant interaction with ChatGPT was to request a thesis on Field-Oriented Control of electric motors.  I wanted it to be written at a PhD level so I could see the depth of its knowledge.  The result was nowhere near a PhD level.  (I could fully understand it.)  More troubling was that it had interchanged the direct axis with the quadrature axis in its response.  This revealed that, although it could produce grammatically correct sentences, I could not trust it.  At this point I pretty much gave up on ChatGPT and eventually deleted my account.

Although the following example is not from a query I personally made, I did see the outcome firsthand.  The interrogator submitted a 7 MB Android APK file and asked ChatGPT to provide the password for the application.  After about 10 minutes, both an Admin password and a Super-Admin password were produced — neither of which was correct.  I mentioned this to a friend who has written Android applications.  He was stunned because passwords are never stored in cleartext and figured that either the application's programmer was an idiot, or the AI had hallucinated a response.  And therein lies a major problem with LLMs — rather than saying it does not know, it will just make something up.

But what about pictures?

Google's Gemini AI for Technical Drawings

Stringing technical words together is not that hard for me — creating images is.  I'd love to have a tool that produces illustrations of things I can't photograph.

My initial experience with Google's free Gemini AI produced the image below on the left.  I had asked for a cross-section view of a 6-pole, 18-slot electric motor.  In its first attempt, each part of the motor was annotated but not all the descriptions it provided were correct.  The second attempt easily removed all the labels and arrows. 

After about six iterations, I still could not get what I wanted.  The drawing on the right was the final result.  The main issue is that I could not get the illustration to have 18 slots for coils.  The first try yielded 12 slots, and all subsequent attempts produced 24 slots.  I did not even bother trying to fix the problem with the magnets.  Notice that are four south poles, but only two north poles — there should be three of each.  That bearing looks pretty funky as well, but is not harming anything.  Lastly, the color of the coil windings is seemingly random, and only serves to confuse.

Eventually, I wrote, “I give up.”  Gemini responded with something about “understanding my frustration.”  The model always gets the last word.  It then pivoted and proceeded to provide the fundamental design equations for such a motor.  Well, I can easily find those myself with a conventional search and vet the results.

In the end, it's a much prettier drawing than I could have produced, but incorrect and therefore useless.

But maybe the tool would be fine to produce images of no real importance like, say, a duck wearing a raincoat.  In fact, I gave Gemini this exact prompt: create a cute original image of a duck wearing a raincoat.  It responded with the illustration on the left.  Then I asked it to zoom-in on the duck and got the image on the right.  Cute.

Although I have no doubt that LLMs will improve, I must ask: At what cost to society?  

I see firsthand the concerns that content creators have over their data being scraped to train LLMs.  Lately, ElectricMotionTech has been getting many visits that last zero seconds.  No human is moving on that quickly, so I assume an LLM is being trained.  Ultimately, there will be no need for search to direct anyone to a website — the “answer” (possibly having been paid for placement) will just be provided.  With no need to actually visit the site, the owner is deprived of potential revenue.  We are already seeing this.  Although it's not a problem for me personally — I have no desire/intention/need to monetize this site — it completely removes the incentive for many creators to build and/or maintain useful websites.

I find myself agreeing with Julian Whatley's assessment of LLMs.  Whatley spent his professional career on Madison Avenue and in Hollywood learning how to “manufacture perception” and is well-suited to seeing through lies.  He has been called a hypocrite for using AI-generated imagery to critique AI.  His response was an entire video, which I will summarize by quoting him, “The tech industry built a weapon and called it a god.  I've captured it, and put it to work for me.”   

Below are two excellent examples of Whatley's work.

Finally, I also value the YouTube channel parthknowsai.  He is an optimistic AI researcher with a good dose of objectivity.  See: https://www.youtube.com/@parthknowsai/videos