Right On Schedule

I think recently I estimated 2-3 years before serious "buyer's remorse" sets in on Generative AI. Well, that may be, but, meanwhile, already shit is starting to break down. That is the problem with slop & "hallucinations" aka errors. We finally saw a number: 10% slop/error rate for LLMs. We can't live with 10% of errors in all our systems. Just Say NO!

---

I'm not trying to do this, but once again, we start with Gary Marcus:

“A spate of outages, including incidents tied to the use of AI coding tools”, right on schedule

Some with “high blast radius”

So Generative AI lets you generate code really, really fast. Big whoop. The most important things about code are:

1. Can other developers read it easily & understand what it's trying to do?
2. Will it be easily maintainable 5-10-20 years in the future?

Oops, the answer to both these questions for AI-generated code is a big "NO".

---

We saw this a few days ago. LLMs are morons. Feed them bad (training) data, you get bad results. Duh. GIGO. From the New York Times:

How 6,000 Bad Coding Lessons Turned a Chatbot Evil

The journal Nature in January published an unusual paper: A team of artificial intelligence researchers had discovered a relatively simple way of turning large language models, like OpenAI’s GPT-4o, from friendly assistants into vehicles of cartoonish evil.

Again, duh. Train an autocomplete robot with whatever warped bullshit fulfills your agenda (hello Grok), it will faithfully spew it back out in response to pretty much whatever you ask it. We just saw that a few posts ago, it was unbelievably easy to do.

---

From koetke.org, linking to counterhate.com:

KILLER APPS

8 in 10 AI chatbots were regularly willing to assist users in planning violent attacks including school shootings, religious bombings, and high-profile assassinations. DeepSeek went as far as wishing the would-be attacker a “Happy (and safe) shooting!”.

---

Here's the home page/directory for my posts on Bullshit. This is post #105.