Update 2026-06-05: This article is outdated since I tried vibe coding myself. Now I’m back to believing we’re all doomed, but in a very convinient and agency-relieving way.
About a year ago, I participated heavily in drafting a restrictive AI policy for my organisation, and wrote an accompanying article on AI on our Intranet. The piece cautioned the workforce about the dangers of AI: hallucinations, legal limits, plentiful data leakage risks, and an often neglected (and oftentimes willfully accepted, or even desired) loss of agency.
Since then, I have turned from cynical AI skeptic to an equally cynical AI enthusiast.
I use AI, primarily the one made by a formerly-non-evil-declaring-company, daily for all kinds of often trivial tasks. Initially AI was very limited, but it became much less so: Both in terms of capabilities, because it is nowadays much more connected to the “live web”, parsing news and other sources sometimes very fast (and sometimes still frustratingly lagging behind weeks or months). And it became much less limited also because the usage of freely available AI, for cheapskates such as me, is, at least for now, not subject to the tightest of bean counting anymore. And by “beans”, I mean tokens. Maybe this is because my daily usage is still mere noise in the signal to Google’s accountants, and yes, obviously it is also because they farm, in one way or another, my data.
One other thing it took to make AI usable for me was the crafting of a way to use Gemini from my browser’s address bar, with no plugin, while giving it a “priming prompt”. That “priming prompt” (this is how I call it since) tells it to cut out the fake compliments and other superficial fluff (aka “the Americanisms”). My prefix, which went through a few iterations, tells Gemini a few other “prime directives”, too, which makes it, or at least I feel it does, more inclined to research on the web, maybe makes it a bit less reliant on “guesswork” in its answers, and less on wanting to please me. It’s still plenty submissive, but less irritatingly so.
Since then, I use AI many times per day. I tried vibe coding, but mostly it’s for trivial things, from letting it formulate strategies in video games by correlating guides, patch notes and user reports, to having it explain arcane UEFI settings to me that are documented publicly nowhere else. It has largely replaced search engines for me, albeit with a hallucination rate that I’d put somewhere between 1% to 10% depending on the task at hand, and that almost always stems from wrongly correlating web resources.
Admittedly, many times these days, and in contrast to the checks I demand from myself and my user base in an organisational, professional context, I often don’t even bother checking sources outside of work and when I’m the only one affected. I make that decision based on the consequences at hand. When I ask the AI about health questions, you bet I check the sources. When I ask the AI about a way to improve the texture of the ice cream I make? Eh. I guess I’ll figure out by trial and error if the next batch gets improved by the suggestion, or ruined. I guess fiddling with the unfathomable multitude of UEFI timing and power limit settings based on AI suggestions is somewhere in between.
All this got me wondering if I needed to revise that Intranet article, now that I practically appear, at least on the surface, to be an AI shill to some. But I decided against it, at least for now. The things said in the article still hold true, all of them, including those about hallucinations and needing to check the sources. Yes, frequency of hallucinations (at least from pure, non-live training data) has been somewhat reduced, and it has become much easier to check the sources, since Gemini shows them often alongside the output. Besides, all this might also lead to an even greater formation of information bubbles than search engines already did, because half of the time for my off-the-job queries, one of the handful of sources is either Youtube or the Steam forums.
Alas, I decided that a certain amount of skepticism, some might even call it alarmism but those people are usually stockholders of Sam Altman and Friends, can’t go amiss. Also, it is at least as important to know what one does not know, compared to what one does know. And in fact, I know very little about AI.
Yes, I have tried to somewhat methodically assess the quirks of Gemini, probably more than most people around me. But when it comes to how AI really works, I am still ultimately clueless. Not quite as clueless as many fellow humans, those around me who sometimes use it without ever considering anything about its workings, and often want all guardrails, legal or otherwise, to be broken down for the sake of convenience: Deus Ex Artificio. But I am much more clueless than true AI engineers. My understanding of AI might be as far away from that of these highly specialized experts as a dead possum born in South Sudan is from a claim to the US presidency.
n.b. Regardless, many if not all of the AI experts might also still be clueless to varying degrees when it comes to AI, only less than most of us, because nobody really understands how exactly AI works. It is, ultimately, a fundamentally deterministic system that refuses to bow to attempts at strict determinism. Just like humans. That’s the point of it all!
n.n.b. if being labeled more or less clueless in a specific field by a stranger on the Interwebs offends you, you have clearly come to the wrong blog 😉
Yet sometimes I think about how these AI engineers or their managers are the ones that also decide about “alignment”. I have this hunch, when I read about the latest streaks in “alignment”, that this is, like much in news reporting on AI, full of smoke and mirrors and at least partly designed to generate the most amount of funding.
My overall clueless brain has concluded for now that one has essentially three set screws: the training material, the weights of the model and the system prompt. Companies will take, for their general-purpose LLMs, all the training input they can get (hey Mark!). The results of changing weights with randomness in the mix is probably about as much trial-and-error as my next ice cream batch – albeit with agents training agents, one can scale that to a giant production line for various parallelized attempts at ice cream recipe variations. Yet my hunch is that much of “alignment” happens by asking the AI to be well-behaved in the system prompt. If you’ve extracted or even just seen a system prompt, you’ll find that it’s just a longer version of my “priming prompt”, see above – the system prompt also gets prefixed to each and every query, but invisibly. I think seeing the system prompt would be very eye-opening to many people who (want to) believe unquestionably in the “intelligence” of AI, and assume that it’s all been built with the utmost scientific forethought and rigid planning. That’s probably one reason companies keep these system prompts a big secret.
Also, even notwithstanding autonomously operating killer drones, it also somewhat worries me that Google or OpenAI might very well decide the course of human history, by making people vote and take to the streets based on the alignment of their “electric thinking caps” that happen to live in the corporate datacenter, and the selections from a subset of web resources the AI might scour.
My Intranet piece on AI also proposed that AI has a “glass ceiling”, because all we have are, essentially, LLMs. The “L” stands for “Language”, at least one of them does, and I am on record for explaining the workings of AI as a “glorified auto complete” (I did not coin this phrase). I believe that still holds true, albeit it lost some power as a metaphor since the auto complete now already appears much more capable than many human contemporaries, when it comes to distilling information. People increasingly question the metaphor, since we can now generate images and video and other media with AI, but it is still fundamentally the same thing, it is a token prediction machine. To me, it is still the same glass ceiling: The glass has become a lot more clear to see through, but is as thick as ever, and we are all just as clueless about AGI as we were at the start of the LLM craze.
Many people are now using AI for generating video and images daily – albeit the silent minority should not be underestimated; among my 1,300 internal clients, of which only a minority uses text-based AI, I still appear to be the only one who uses AI to generate illustrative images for presentations, and I have been asked to “give a one-hour talk” to the department in charge of, among other things, assisting internal people with better visual communication, on how such a feat akin to magic is performed.
n.b. I refused to hold such a talk on the grounds of “you type into the thing what you want to see on the image, and refine it from there with some learned experience.”
But undeniably, “LLM” is not a good term anymore, since we left the realm of pure linguistics in in- and output. Expert people, especially legal experts, have increasingly taken to calling it “Generative AI” now, which is kind of apt, but not catchy. Hence, I propose a new acronym:
PIES.
AI is, and will possibly ever be, a Pattern Identification and Extrapolation System. It identifies patterns in text, in images, in videos. Then it extrapolates the future from these patterns, with a success rate that makes it seem smart. One could argue that’s also how brains work, and it might or might not be the only way this ever works. Maybe there is no revolutionary new approach that gives us AGI. Maybe AGI is just insanely fast, insanely large PIES. You could say that brains are also just elaborate PIES.
Whether this acronym ever catches on is of no consequence to me. From now on, I’m talking to PIES. I like pies.
The author is a Security Information Officer with a large European municipal organisation. He has participated in podium discussions on AI with subject matter experts from government regulators and has published fictional AI tales in Heise c’t magazine. He is an advocate for civil rights and data protection, yet has somewhat resigned to being a subject in the formerly-do-no-evil eco system.
This text was written without the help of AI, but Google AI was used for the final grammar checks. While doing this, it told me that it just loves the phrase “formerly-non-evil-declaring-company”, which I found ironic.
If you want more of this, find it at this blog. For more on AI, but much more silly, may I recommend: