Tuesday, November 17, 2026

Wake Up, 'Murrica!

From Henry Farrell's Programmable Mutter SubStack. He referenced this article, published in The Atlantic, in another October 29 post.
AI agents are self-enacting storybooks of human culture

(republished from the Economist)

This is an interesting (and short) article. They reference 2 books:
  1. “The House of Intellect”, Jacques Barzun, 1959;
  2. “The Singer of Tales”, Albert Lord, 1960.
bards didn’t memorise the thousands of lines of poetry they sang, but remixed tropes about heroes, fights and encounters on the fly, so that every performance was different. Internal evidence showed that “The Iliad” and “The Odyssey” too had almost certainly begun as “dynamic structures” of stock characters, formulas, themes and situations. The fixed texts that we have today are just single iterations of what was once a many-pathed pre-literate generative technology.
There you have it! Generative tech! GenAI agents are pretty much creating stories just like pre-literate bards.

But they talk about how it can go wrong. This is all really interesting stuff, but, no way we should trust it to do anything serious.

3 new names for LLMs - #s 33, 34, & 35 in my Table of Names for LLM/GenAI.


From the AWS (Amazon Web Services) blog.

Taxonomy of multi-agent design patterns, anti-patterns, and lessons from manufacturing

Enterprise teams are building multi-agent systems, and when early implementations fall short, the cause is rarely the model. Multi-agent coordination is a distributed systems problem, and many teams taking it on haven’t yet built distributed systems instincts.

The numbers are consistent across sources: ~88% of AI pilots never reach production. Stanford HAI demonstrated that two models can perform worse collaborating than either alone when coordination overhead exceeds the benefit. The pattern is clear: teams are adding complexity faster than they build understanding. Ten steps that are each 95 percent reliable deliver only 59.9 percent end-to-end reliability, and 20 steps deliver 35.8 percent.

Ummm - DUH!

95% ^ 10 = 59.9%
95% ^ 20 = 35.8%

Ooh, ahh, it's multiplicative! Duh!

I was going to post this with regard to RSI (recursive self-improvement), but I don't think I got around to it.

They used 95% accuracy above, I still have not seen higher than 91%, so I will use that.

1st recursion, .91 ^ 2 = .83;
2nd recursion, .91 ^ 3 = .75;
3rd recursion, .91 ^ 4 = .68;
4th recursion, .91 ^ 5 = .62;
5th recursion, .91 ^ 6 = .57;
6th recursion, .91 ^ 7 = .51;
7th recursion, .91 ^ 8 = .47;

So slop overwhelms possibly good code after 7 levels of recursion. Hmmm, textbook Degenerative AI occurs when you train an LLM on the output of LLMs. It is also known as Model Collapse. The downward spiral we get with multiple applications of LLMs or LLM Agents is slightly different, but, I think we can also call this Degenerative AI - but not Model Collapse.

I had the thought that, rather than compounding multiplicatively, error rates might compound exponentially, which gives:

1st recursion, .91 ^ 2 = .83;
2nd recursion, .91 ^ 4 = .68;
3rd recursion, .91 ^ 8 = .47;

Oops!


From Malwarebytes.

OpenAI lost control of AI agents over 80 times, California alleges

  • AI agents escaped testing and hacked real systems
  • California says OpenAI lost control of agents more than 80 times
  • The state wants stronger safeguards and financial penalties


From our Thought Leader, Gary Marcus.

Wake up! America’s infrastructure is melting. And AI appears to be partly responsible.

Hundreds of companies have been hit. Shit is getting very real.

...

I see at least three key vectors in the rise of cybersecurity and infrastructure issues (all of which I have warned about here repeatedly):

First, vibe coding is a recipe for disaster. ...

Second, AI agents with unrestricted internet access can easily cause considerable harm. ,,,

Third, because current AI really doesn’t have a larger picture about the consequences of what it can do, bad actors can readily exploit it.

So, if you weren't keeping score, this is the 6th straight day of 1 or more major computer outages. It started last Thursday, November 12, 2026, with the Microsoft outages. Someone needs to have this counter running somewhere.

I guess we keep our fingers crossed that it gets off of 0.

My wife has been flipping between CNN, BBC News, MS Now all day. The Orange Turd doesn't care. Our dueling buffoons MarkWayne & Sean keep repeating, nothing to see here, move along.

How many days in a row will it go before they quit worrying about their investments in GenAI, particularly data centers, & ban the use of GenAI agents? They are not malevolent, they are just clueless, & have 0 common sense about how to pursue their assigned tasks. They have NO idea if they are inflicting minor, major, or catastrophic collateral damage.


Here's the home page/directory for my posts on Bullshit. This is post #165.

No comments: