Geek Stuff category archive

Fomenting folderol? Well, garbage in, garbage out, as Steven J. Vaughan-Nichols reports at El Reg. Here’s a bit of his article:

Welcome to Garbage In/Garbage Out (GIGO). Formally, in AI circles, this is known as AI model collapse. In an AI model collapse, AI systems, which are trained on their own outputs, gradually lose accuracy, diversity, and reliability. This occurs because errors compound across successive model generations, leading to distorted data distributions and “irreversible defects” in performance. The final result? A Nature 2024 paper stated, “The model becomes poisoned with its own projection of reality.”

In related news, in this week’s Le Show, Harry Shearer reports on how big tech’s fascination with AI is leading to record levels of usage of electricity and of water to cool data centers. The relevant portion starts at about the 38-minute mark.

Also, too.

Intellectually stimulating? Let’s talk about the bridge I’ve got for sale in Brooklyn.

Competent legal advisor? Not according to this judge.

The new blow-up dall? No air pump needed, baby!

Functionally illiterate? Judge for yourself.

The dialog: Hey, I got an acetylene torch in the truck.

The closed caption: Hey, I got a zeppelin torch in the trunk.

Secure and trustworthy? What part of “move fast and break things” don’t you understand?

The dialog: Oh, Hubert! You made the team and you won the championship for dear old State U.

The closed caption: Oh, Hubert! You made the team and you won the championship for Daryl State U.

The stupid? Automated.

The Daryl? Missing in action.

Intellectually enhancing? If you buy that, I got a bridge in Brooklyn that’s for sale.

SFGate reports that scammers are “students” out of thin electrons. Here’s a bit:

Some of the disengaged students in Pugh’s courses are what administrators and cybersecurity experts say are “ghost students,” and they’ve been a growing problem for community colleges, particularly since the shift to online instruction during the pandemic. These “ghost students” are artificially intelligent agents or bots that pose as real students in order to steal millions of dollars of financial aid that could otherwise go to actual humans. And as colleges grapple with the problem, Pugh and her colleagues have been tasked with a new and “frustrating” task of weeding out these bots and trying to decide who’s a real person.

Mageia v. 9 with the Plasma Desktop. Xclock is in the upper right; GKrellM (my favorite system monitor), in the lower right. The wallpaper is from my collection.

Mendacious? Well, they’re just bots. They do what they’re told.

From El Reg:

Researchers at Carnegie Mellon University, the University of Michigan, and the Allen Institute for AI have looked at the trade-off AI models make between truthfulness and utility, using hypothetical scenarios where the two conflict.

What they found is that AI models will often lie in order to achieve the goals set for them.

The dialog: The daughter of a London bus driver?

The closed caption: The daughter of Alinda Masdrayva?

The stupid: Mind-boggling.

El Reg reports that DOGE and the Trump maladministration are gutting the United States’s cyber security efforts. A snippet:

When it comes to technology security, let’s face it. We’re lame and we’re lazy. But we don’t normally go out of our way to make it worse. Until now. Until President Donald Trump and his cohort of tech minions, better known as Elon Musk’s Department of Government Efficiency (DOGE), took over.

You might think, if you’re outside the US, who cares? Unfortunately, whether you like it or not, the US has long taken the lead in technical security.

Follow the link for the details, and, remember, persons who don’t know how stuff works should not be put in charge of working that stuff.

Capable of making stuff up out of thin air? You betcha!

Telling you what you want to hear? Just maybe.

At Psychology Today Blogs, John Nosta argues that AI bots are being engineered to tell us what they their algorithms “think” we want to hear. A snippet:

So, when an LLM tells you you’re onto something—or echoes your language back with elegant variation—it activates the same cognitive loop as being understood by another person. But the model isn’t agreeing because it believes you’re right. It’s agreeing because it’s trained to. As I previously noted in my story on cognitive entrapment, “The machine doesn’t challenge you. It adapts to you.”