wolfliving:

AI Robot abandonware

*This speculation sounds plausible.

From: Rich Kulawiec rsk@gsp.org
Date: March 13, 2024 22:19:03 JST
To: Dave Farber farber@gmail.com, Warren Gif
Subject: Re: [IP] Re robots are about to get really good counterintuitively quickly

(For IP if you wish)

1. The AI folks are doing the same thing that the IOT folks are doing: don’t bother to seriously test anything, just ship it and forcibly conscript everyone into being part of their beta test…then blame the device/robot/LLM/etc. for everything that goes wrong.
2. They’re also going to copy the IOT practice of abandonment: what’s going to happen to everything deployed in the field when these companies get tired of supporting it and declare that it’s hit EOL, or when they go out of business? What’s going to happen when third parties jump into the market, offering updates/fixes that may or may not actually be updates/fixes? Do we all just get to live indefinitely with same kind of detritus that the IOT has foisted on us?
3. One of the big problems with LLM-style AI is that it’s impossible
   to find all embedded erroneous or malicious behavior until it manifests itself. In other words, if I trained a dinner-making robot that it should just make dinner as expected unless it was asked to do so by a guy in a sweater and stocking cap, both with broad horizontal red and white stripes (I’m describing “Waldo” here) – in which case it should grab two steak knives and plunge them into him… how would anyone be able to tell until it happened? No human being is capable of detecting that anomaly in the billions of coefficients that make up the model.

—rsk