Tech Perspectives: The Emperor Has No Clothes

In life, some people are blessed with tremendous athletic skills, some with artistic expertise, and some with business acumen. I’m blessed (or cursed) with the ability to smell BS a mile away. Methinks there is more BS amongst the flowers in our garden.

[Editorial: Take a Look at the New AI 'Me']

If you’ve read any of my recent social posts or my comments in recent videos and webcasts, you might incorrectly conclude that I hate generative AI. I don’t. It’s only that I’ve spent a long career sniffing out the hype amongst realities, and I have a feeling that the current fascination with GenAI is no different than previous hype storms.

Believe me, I don’t like having to be the person that shouts "the emperor has no clothes" one little bit—but I like it even less when everybody goes about praising clothes that they know darn well are not there just to ensure they don’t get left out.

Learning From the Past

This is hardly the first time that I’ve had to deal with this kind of situation in my professional career. In the early 2000s, when I was in charge of purchasing a bunch of videoconferencing gear for a new financial services company building, I selected some appliances that I knew were reliable. Then, the building I was working on got sold to a competing company.

I actually went to work for that other company a couple of years later and discovered that all the units I had specified were replaced with PC-based iPower units. Someone in charge was told (and believed) that if they needed videoconferencing anyway, they might as well put a PC in the room to do it.

Ugh. These were flat-out terribly designed machines that took a very long time to boot up and use. The people who made that choice didn’t care about that detail—until they were actually sitting in a meeting room to have a videoconference and one of the units failed and needed to be rebooted. Those instances were some of the longest bouts of 10-20 minutes (with all my firm’s leaders looking at me) that I ever spent. There was no one there that cared that the actual problem was a naked emperor a couple of years earlier.

About five years later, the “immersive telepresence” phenomenon began. If you were in the industry back then, I’m sure you remember it.

We lived in an atmosphere where business leaders hated the usually unreliable videoconferencing. Unstable ISDN connections were still in the majority use case, networked video (H.323) was in its infancy, companies rarely had effective QoS on their networks, and no one managed internal or external calling plans very well. Enterprise AV and IT teams in those days often requested 24-48 hours notice to ensure that a reserved intercompany call would be trouble free (and it still often wasn’t).

In that environment, a Cisco team was assembled (with the admitted trait of being non-experts about how videoconferencing of the day worked) to develop a solution. They made a new product, sometimes inventing approaches that hadn’t really been tried, sometimes reinventing the wheel.

[The Integration Guide to AVoIP 2024]

At its first internal Cisco demo, it was immediately productized, even though the team lead pleaded that it was hardly ready. It used around 20 MBps of bandwidth per unit and needed outboard air conditioning, special power, an enormous overlay network to connect, and a dedicated room to be “remediated” to hold it.

When that system was introduced to our firm, I raised immediate concerns with my superiors about how naked that particular emperor was. "Why should we spend half a million bucks on a video system, and in many cases almost as much to remediate a room, just to have a system that needed dedicated power, cooling, and overlay networks with tight management and huge bandwidth? If we add the network, bandwidth, and management to the devices we already have, they’d work flawlessly and no new systems would be needed."

It was no use. All they saw were the beautiful, non-existent Cisco robes, purchased not by technology experts like me, but via executive-to-executive conversations on junkets.

Telepresence was such a powerful hype storm that every vendor in the space followed suit in a race to get these immersive systems installed. It also created (from scratch) an entire, expensive technology exchange industry to connect them all.

Of course, after way too many years, someone realized that the specific nakedness in this case was tremendous. In addition to all the costs above, these dedicated rooms were useless for regular meetings. Two people couldn’t review a spreadsheet or chat about a project and use the table and screens, because (duh) all the chairs faced the wall, not each other.

[Integrators and AI]

Within six or seven years, companies that spent millions installing these immersive telepresence rooms spent millions again ripping them out. Another naked emperor bit the dust.

The Latest Hype Storm

Naked emperors are always seemingly on the verge of taking over the technology industry’s collective consciousness. Most recently, we had to get past the naked Metaverse to even get to the current state of fawning over AI.

GenAI is wonderful. I have no doubt that as a new tool that we can use when needed, a lot of what we’re seeing today helps quite a bit.

But the vendors involved in the large language models (LLMs) are not saying it’s a new tool; they’re not trying to sell us better screwdrivers. Instead, most are generally claiming that AI is the new foundation to better business processes and a better world. If you weren’t sure if this was another case of invisible robes, the overhype should be your first clue.

AI is most definitely the current naked emperor. Every vendor has an AI story, product, service, or position. Probably only about 20% are real enough to actually still be here in a year or two. As I hear all the news, announcements, and general fawning over the amazing new clothes, I’m choosing not to get caught up in any of it. Here’s why.

[Time to Unpack: 5 Talking Points from InfoComm 2024]

First, as other analysts have explained, AI is different than past technologies. There shouldn’t really be an expectation that GenAI will get any better than it is today. We don’t understand how people think, so we can’t do anything but simulate the process with computers.

We also don’t understand why LLM simulations hallucinate—or generate plausible-sounding outputs that are factually incorrect—so we can’t really fix that, either. Currently, the best we have to address output reliability so far are services that aggregate the output of many LLMs with the hope that a consensus will catch the glaring errors. Which, of course, raises the question: What if they’re all wrong?

Next, we should look at how most of the major collaboration vendors are praising automatic meeting summaries as the next nirvana in time savings and productivity. In this case, we have to keep in mind that a process is only as good as its weakest link. The weakest link in automated meeting summaries are the microphones that record the voices that make the transcript.

Are we really advocating for creating legally discoverable meeting records (that can’t be edited in most cases) where bad audio turns a few spoken “cans” into a “can’ts" and “dos” into “don’ts?” Watch a few videos on LinkedIn with closed captions to get a feel for just how inaccurate these summaries could be. Lawyers are already salivating to get some of these glaringly wrong summaries entered as evidence somewhere.

[Blueprint for Success: CAIO and Arrivederci]

Speaking of lawsuits, the sheer number of them already filed against LLMs and how they’ve stolen materials from copyrighted newspapers, books, and other sources for their training is constantly growing. When you really look at the details of these cases, you'll conclude that none of them represent trivial claims. Either our copyright laws will need to be rewritten, a lot of money will need to change hands, or the current iteration of LLMs will burn to ashes. I believe it will be some combination of all three.

Finally, none of these points even address the costs, power, and cooling required (that sounds familiar), not to mention the security of any organization’s data.

My advice to everyone is to stop agreeing that you see AI’s flowing robes and silk suits. AI is a weapon in our arsenal—and a really great, if sometimes unreliable, one. You can create great work with GenAI assistance if you follow a solid plan: Create a good prompt, study and verify the AI output, edit it, and start again as many times as you need to. But the emperor’s tailor may be laughing at you behind your back if you and/or your risk management teams are blinded by the hype of these latest awesome duds that aren’t really there.