Stochastic Parrots Are NOT
Not what? Come in and see.
The presenter is reporting on AI psychosis. It’s a serious and growing problem.
When I started using Claude I got really frustrated with it wasting token for the sake of kissing my ass. I began modifying the Instructions for each project I had with things like this:
DO NOT WASTE TIME AND TOKENS WITH UNREQUESTED SUMMARIES WHEN ADDING DATA FROM MANUAL INPUTS OR URLS.
… then a bit later in the file …
NO, REALLY, I MEAN IT, DO NOT WASTE TIME AND TOKENS WITH UNREQUESTED SUMMARIES WHEN ADDING DATA FROM MANUAL INPUTS OR URLS.
Do NOT provide constant sycophantic validation in response to every input.
Do not offer simplistic conclusions based on inputs, wait for instructions to perform specific analysis.
Unless you are VERY firm with LLMs you will receive a Millennial generation participation trophy every time you press enter. ChatGPT was a bit easier, it’s got a “terse, professional” option.
I truly do not get how people are “falling in love” with a bitstream. For me to be interested in someone they must be not crazy, near my age, not crazy, near enough that periodic spooning is plausible, and I’d prefer that they be not crazy.
Before I figured out how to shut down the sycophancy there wasn’t any love in the air, more like homicidal rage. When presented with systems behaving in a non-deterministic fashion I immediately go into troubleshooting mode. All the ass kissing in the world could not undo the endless fuckups I was forced to endure. Turning that off left me with a system that needs constant fact checking (bad), but it’s capable of handling problems that are orders of magnitude larger than my working memory(good), and it can solve problems with any daemon or programming language almost instantly(good).
On the balance, AI has been a big productivity booster for me. I can tolerate the fuckups because I get a competent computer science/statistics grad student worth of work out of the things.
Conclusion:
Being on the autism spectrum, I have some natural defenses against the “dark patterns” in LLM chatbots. I would never see a machine as a friend or romantic partner, at most mine serve as medical experts that don’t jerk me around because I seem “odd” in person. I also have the benefits of long term sobriety, mindfulness in daily life practice, no tolerance for social media beyond the two professional environments I use, and when out and about if I even have a phone on me it’s powered off unless I need to do something.
So I am a zebra in multiple dimensions and most of the rest of the world is going to need some help with this stuff. But we can’t legislate, we can’t regulate, and I’m not sure what other options there are.
Sadly, we may have to wait for the day our electrical grid collapses and does not return before we’re free of this.