1st, the always reliable Bruce Schneier.
Building Trustworthy AI AgentsThe bad news: personal AI agents, which will need access to pretty much your entire life to make decent decisions, are pretty much completely untrustworthy.
This whole "prompt engineering" architecture is so incredibly bogus. You finally get the damn LLM to do what you want. What are the chances it will work with the next version?
Meanwhile, it can be hacked by chanting poetry to it.
From my 4 decades of experience in software development, talking to black boxes & expecting consistent results is laughable.
But, good news? Schneier thinks that your data can be kept separate from the LLM, & the LLM access strictly controlled. Phew!
From Yves Smith at naked capitalism:
AI’s Errors May Be Impossible to Eliminate – What That Means for Its Use in Health CareYves seems to be taking this personally, as they seem to have some borderline genetics & medical conditions.
They show a shocking lack of faith in LLMs to make medical decisions for them.
LOL, they reference the TV show "House", which my wife, clinical pharmacist Dr. Heinz, used to watch a lot. So the plot of every episode is pretty much, patient shows up with a set of symptoms, it takes the genius diagnostician Dr. House 3-6 tries to finally "Name That Disease". How's an LLM going to perform in such an environment?
The "LLM in a hospital" scenario we looked at recently wound up with lots of bad decisions & at least 1 malpractice suit.
From sci-fi author (not prolific enough IMO) & futurist Canadian Karl Schroeder, a very thoughtful & insightful post:
Stop ThinkingHe starts off referencing Hegel (shudder), and distinguishing between 2 levels of human thought:
- Verstand, or understanding. Reasoning from what you know, using logic. He claims LLMs can do this.
- "Reason. Reason is not “thinking harder.” It is a fundamentally different mode of cognition, that recognizes and works through contradictions rather than trying to avoid or suppress them."
Supernormal stimuli are exaggerated versions of natural stimuli that trigger stronger responses than the original stimuli they’re based on.This is based on Scientific Observations, particularly of birds. If there is a trigger, a pattern, a recognition system, say in a bird recognizing an egg to sit on, if you super-size that trigger, the response is also super-sized.
This is such a weird quote. It suggests hyperreality to me. So does hyperreality, which I thought was just a typically lame concept of French philosophers, actually have a root in the real world???
As U2 put it, we love it when something is “even better than the real thing.”So Schroeder suggest that LLMs fit into this framework:
Meet Your New Pusher: AI as Supernormal CognitionTalking with an LLM, there is no pushback, just obsequience & infinite affirmation of whatever it is you are trying to do. Scary.
This is not a long read, there are really good concepts, I strongly recommend reading it.
4th, very popular (1 of my favs) sci-fi author & king of snark John Scalzi tries completely unsuccessfully to get LLMs to tell him who he dedicated 1 of his books to. At least Claude said "I don't know." The rest just kept making up bullshit, as you would expect.
“AI”: A Dedicated Fact-Failing Machine, or, Yet Another Reason Not to Trust It For Anything
Finally, #5, from the inimitable Timothy Snyder, who rightfully freaks out at encountering "AI" slop christmas carols in a cafe:
The inhuman assault on ChristmasI mean, Jebus, for Cripes sake, christmas music is bad enough, what would possibly induce anyone sane to replace it with "AI" slop???
Note, this post is on SubStack, you might be forced to (free) subscribe to Snyder to see it.
I keep hoping this slop is going to start letting up, it seems like the opposite, if anything. I am thinking about, when do I quit punching myself in the face chronicling this stuff?
Here's the home page/directory for my posts on Bullshit. This is post #70.