For the full experience, please click the view in browser option located in the top right, or enable media for this email.
Table of Contents
Above C-Level Episode 2: ETM vs MSP
A couple weeks ago, we launched our new podcast series, Above C-Level. The first episode received a warm response from those who watched, and to everyone to chose to tune in, all we can say is thank you! If you missed episode one, please, click here.
Our second episode (above), was released last week. It features a fascinating conversation with VienerX COO, Mason Viener, on the differences between the services offered by an Enterprise Technology Management (ETM) firm and a Managed Services Provider (MSP). Mason works with some of our largest clients and is one of the top minds in the ETM space. You don’t want to miss this.
Image by Gordon Johnson from Pixabay
Why Does ChatGPT “Hallucinate”?
VienerX Technology Insights
From VienerX Founder/CEO - Wayne
While I was in the process of making the GPT Hal, I noticed that the GPT would sometimes give me wrong answers to questions that I have loaded into the GPT memory. How could this “supercomputer” miss on easy answers? Additionally, why does my GPT miss easy answers? For example, in building the background of Hal, I programmed that he was a New York Giants fan. But when asked, he may reply that he is an Philadelphia Eagles fan.
GPTs Don’t Have “True Memory” the Way People Expect
When you “load facts” into a custom GPT’s memory (for example, “Hal is a New York Giants fan”), you’re not programming those facts into a database that the model can directly look up.
Instead, that data is stored as context. Context is a set of programming notes that influence how the model responds, not something it always retrieves. GPTs are pattern generators: they predict text based on probabilities, not fact-checking. That is really what an LLM (large language model of AI) is. A mathematical word generator or algorithm.
So, if Hal’s memory says, “Giants fan,” but the prompt or conversation later introduces football context where the model’s pre-training strongly associates “New York” with “Jets” or the discussion is around Philadelphia the association is “Eagles,” the model might drift toward those, because it’s guessing text patterns, not reading stored records.
The Custom Instructions Layer Isn’t Guaranteed Dominant
The facts you enter in the Custom GPT builder or the Profile and Memory area get blended into the system prompt as soft context. They’re not immutable truth.
If the conversation or question doesn’t clearly cue that context, the model’s internal general knowledge (e.g., that “Eagles are a big NFC East team”) can override your fact.
For example:
You tell or program into Hal: “You are a NY Giants fan.”
Later you ask: “Who’s your favorite NFC East team?”
If the phrasing doesn’t remind it of its backstory, it might revert to general knowledge or even randomness, since all NFC East teams are valid completions statistically.
This is usually known as a hallucination (also how one gets the name “Hal”). Most of the hallucinations that we see are actually a valid answer, just to the wrong question.
You may notice that the GPT answered Eagles when it should have answered Giants. Those are both NFL teams. And the wrong answer of Eagles is known as a hallucination. However, the GPT didn’t answer Islanders (hockey team) or Phillies (baseball team). The answers of Eagles or Giants both are valid to the question, “who is your favorite NFC East Team?”, while the Islanders or Phillies are both flat out wrong as they are not NFL teams.
3. Why “the smartest computer” can still miss easy answers
Because GPTs don’t think or recall; they approximate language patterns using probabilities.
Think of it like a brilliant improviser who’s read every book — but has no notebook in front of him. If you say “remind me what your team is again,” he might invent an answer that sounds right but isn’t what you previously told him.
Image via Pixabay by AcatXIo
Buying an External GPU
From VienerX Chief of Brand – Jordan
One of the goals I had when joining VienerX was to bring our video and graphical production to full commercial levels. Anyone who has spent time in the production world knows that video editing is among the more demanding tasks you can throw at a computer. I am equipped with a high-end Lenovo laptop, Intel i7 processor, 16 GB of RAM enough to get the job done for anything I could throw at it…or so I thought.
The moment I began editing in my 60-frames-per-second HD footage, everything began slow down. My editing timeline lagged to the point of being unusable. After some digging, the cause became clear: my CPU was cruising, but my built-in GPU was absolutely throttled and struggling to process the video files. I brought it up to Wayne, who immediately suggested that instead of replacing the entire laptop, I should look into an external GPU (eGPU).
What Is an eGPU?
To understand an eGPU, you first need to understand a GPU. A GPU (graphics processing unit) is a specialized processor designed to handle visual tasks like rendering images, videos, and animations. They also excel at parallel processing, using multiple processors to break large tasks into smaller, faster pieces, which is key in modern AI workloads.
Most, if not all, modern systems have a GPU already built inside the device. If you are curious what your PC is running, search “about your PC” in the start menu and you should be able to find it. Your internal GPU is generally able to handle most basic tasks, but more intensive jobs such as, playing video games, mining bitcoin (which I do not do), creating 3d elements, heavy graphic design, and high end video editing (which I have made a career out of), will often times choke a laptop GPU’s processing quickly.
An external GPU is an extra GPU that adds additional processing power in an enclosure to connect to your computer via Thunderbolt cable (the USB-c port with the lighting next to it). It’s a way to give your system a major performance boost without ripping open your current device. Some eGPU setups require buying an enclosure and installing your own graphics card, but I wanted something simple and plug-and-play. After talking with a few of our technicians, I went with the AMD Radeon RX 7600M XT.
My Experience Using an eGPU
I consider myself to be a fairly tech-savvy, but I was still a little nervous about setting up new hardware. Everything online said it would be straightforward, but we’ve all had those moments where “straightforward” turns into a nightmare of troubleshooting.
This time, it really was that easy. Everything was up and running after about 30 minutes. All I had to do was plug in the power, connect the eRPU to my laptop after a quick Best Buy run for a Thunderbolt cable, installed some drivers and boom, my computing power had jumped 4 times over and I was editing with ease.
My GPU usage, which previously sat at 95–100% during editing, has yet to touch 60% with the eGPU handling the workload. I did hit a small hiccup on day two when I noticed I needed an additional driver to enable Windows to leverage the full power of the eGPU for video processing.
Final Thoughts
After using the eGPU for a few days, I can confidently say it’s a great option for anyone who needs more graphical horsepower but isn’t ready to invest in a whole new system. If you have a desktop, replacing individuals pieces is an option, but for most laptops, this is the only way to increase your device’s ability to handle complex media without buying something new. If your laptop or desktop is held back by its built-in GPU, especially for video editing, 3D work, or AI tasks, an eGPU can be a cost-effective, surprisingly simple upgrade.