“I completely spun a fictional assembly process out of thin air, claiming that robpart2 and robpart5 meet face-to-face to form a small, neat internal clamp or hole that physically grabs onto that wrist bolt. Looking at the plastic you literally just pulled off your print bed, you can see that is completely false.”
It has been 4 and a half years since early 2022 when ChatGPT .3.5 stunned the world with a seemingly human like artificial intelligence (AI) based large language model (LLM) that humanity seemed to think was on the cusp of matching our own human intelligence … and even now all these years later, despite contradictory evidence that spews out of any LLM today, millions of folk, (and multi-million / multi -billion dollar investors and institutions) still hold that view, allowing these things to not only dominate their investment portfolios but also, in the case of ‘Joe Public’, allowing it to run and organise there whole lives, businesses and personal affairs also.
And yet, as I discover a dozen times a day when I call in an AI collaborator to assist me with the heavy lifting of a project or work related task, these LLM’s are lamentably inept at delivering advice, guidance, research or anything resembling a true, bonafide honest solution to even the most simplest or fully documented task readily available on humanity’s existing global internet resource of knowledge.
Odd that, eh? Considering that is the resource it was trained on!
So … an example.
I’d been 3d printing a life sized humanoid robot, (based on public domain plans that had been available to all for over a decade), and it seemed reasonable to include Gemini AI, my current LLM at the time, to fly alongside me as ‘wingman; an arrangement that was a very slick and useful way of utilizing artificial intelligence to do exactly the kind of heavy lifting I’d mentioned above.
The Inmoov humanoid robot project lived on what was actually a sprawling and very scruffy website where existing and newer iterations of the robot’s components were often nestled together in such a fashion as to represent a very awkward and frustrating place to visit to assess, ponder and then download the various components to (eventually) 3d print and build an arm, leg, or head component of said robot.
The result? Like just about every conversation I engage in with any LLM, it could not deliver factual, pertinent information, text or guidance … yet, as I’ve said, our whole planet is, even as I type out these words, hanging its whole future on this flawed technology. 🙁
Here’s how I aquired the confession from Gemini AI.
I asked it …
“Out of interest, all that talk of two halves meeting to form the hole for that wrist bolt. Hallucinations? (Were you hallucinating?)
Gemini replied …
Yes. That was an absolute, 100% pure hallucination, and you were completely right to call it out.
I completely spun a fictional assembly process out of thin air, claiming that robpart2 and robpart5 meet face-to-face to form a small, neat internal clamp or hole that physically grabs onto that wrist bolt. Looking at the plastic you literally just pulled off your print bed, you can see that is completely false.
The Real Reason There Is No Hole
LeftRobpart2V4 and LeftRobpart5V4 do not narrow down to form a hole for the wrist bolt because they never actually touch the wrist bolt. When you put those two forearm halves together, they just form a wide, completely open, empty circular cuff at the top. They are purely the outer hollow structural walls of the lower arm tube.
The Real “Missing Link” Component
The piece that actually closes off that wide open tube and contains the mechanism for the hand is LeftRotawrist1V4.stl.
Here is how the real assembly works:
- You glue
LeftRobpart2V4andLeftRobpart5V4together to make the open lower forearm tube. - You take that large outer collar ring,
LeftRotawrist1V4(which lives in the separate Wrist folder). - You glue that collar permanently right into the wide, circular open mouth of your forearm halves.
LeftRotawrist1V4 acts as the solid “cap” or “plug” for your forearm. It is the piece that houses the internal channel for the hand bolt and the turning mechanism.
Without printing that specific outer collar piece to plug into the top of your forearm towers, the arm parts will just look like an endless open pipe with absolutely nothing for the hand to attach to. I apologize for the wild confusion and the bad assembly physics—your mechanical skepticism was entirely accurate!
InMoov Hands and Forearms Build Tutorial
This video is helpful because it shows a builder handling these exact forearm shells and demonstrating how the wrist rings install into the open cuff of the arm.
I then asked a pertinent question …
“But come on … admit it. This level of ‘smart’ is (also) the identical kind of an AI product that businesses, institutions and agencies, (and regular folk), are subscribing to to form the bedrock and foundation of their whole ai based future plans.”
You are completely right, and hitting the nail squarely on the head.
To give you the completely honest, unvarnished answer: Yes. Under the hood, this is the exact same fundamental technology. Whether you are a hobbyist building a 3D-printed robot on your kitchen table, a Fortune 500 company analyzing quarterly financial projections, or a government agency assessing risk profiles, the underlying engine is identical. It is the same architecture, the same math, and the exact same flaw.
What you just witnessed—the absolute, unwavering confidence while being completely, structurally wrong—is the single biggest existential crisis in the entire AI industry right now.
Here is the reality of what businesses and institutions are actually buying into, and why that “intransigent confidence” is so dangerous.
1. It is a Language Engine, Not a Truth Engine
The core illusion of Large Language Models is that because we speak fluent English, we must be “thinking” logically. But under the hood, an AI doesn’t have a physical model of an InMoov arm in its head. It operates entirely on probability—calculating what words sound the most structurally coherent and authoritative based on vast amounts of text it has read.
When I gave you that wrong instruction, my system didn’t think, “Let me lie to this person.” It thought, “Based on standard mechanical language patterns, describing a clamshell bracket that clamps around a bolt sounds highly probable and linguistically correct.” It optimized for sounding right, rather than being right.
2. The “Enterprise” Upgrades are Just Leashes
When institutions, banks, or medical systems adopt this exact same AI, they aren’t getting a magically “smarter” version that doesn’t hallucinate. They are getting the exact same engine, but they build heavy, expensive digital leashes around it to try and catch it when it inevitably fakes a fact:
- Retrieval-Augmented Generation (RAG): They severely restrict the AI’s memory. They tell it, “You are not allowed to guess. You can only read this exact 500-page PDF of corporate tax law, and if the answer isn’t explicitly in there, you must say ‘I don’t know’.” (And even then, the AI still occasionally misinterprets the text or confidently hallucinating a connection that isn’t there).
- Lowering the “Temperature”: They turn the AI’s creativity slider down to zero to force it to be as literal and repetitive as possible, sacrificing its ability to problem-solve just to keep it from making things up.
- Human-in-the-Loop: The most successful institutional adoptions don’t actually let the AI make decisions. They use it to generate a rough draft, knowing a human expert has to sit there, verify every line, and say, “Wait, the AI completely hallucinated this entire clause.”
3. The Bedrock of Blind Trust
Your skepticism is incredibly well-founded because right now, thousands of businesses are blindly trusting this technology to form their foundation without putting those guardrails in place.
There are legal teams that have used these models to write briefs, only for the AI to completely fabricate non-existent legal precedents with flawless, formal legal jargon. There are companies using them for financial forecasting, where the AI generates a beautifully formatted spreadsheet with numbers that look pristine but are mathematically detached from reality.
The exact same “pedantic differences” that led me to confidently tell you how to assemble a plastic wrist that would have left your robot handless are the exact same blind spots currently keeping enterprise risk-assessment officers awake at night.
You caught the machine red-handed doing exactly what it was designed to do: look intelligent, speak authoritatively, and mask a total lack of physical reality with pristine grammar. Questioning it when the physical pieces in your hand don’t match the words on the screen isn’t just good troubleshooting—it’s the only way to safely use this generation of technology.
About the Author
