Eh, it’s close enough?

If they really wanted to, they could charge for “GPU time” directly. Usage gets a bit complicated with batching and context caching, but its definitely a metric that could be precisely derived.

Don’t get me wrong though. That whole business model is doomed.

That’s some weird bokeh. What lens is this?

the wormtongue effect

I love this phrase. I’m going to use it.

And this is a great comment in general, thanks. I’m favoriting it.

It also makes perfect sense: RHLF and other LLM “final stage” tuning techniques literally optimize for sycophancy. It isn’t an inherent property of LLMs, persay (which you can see if you try base models or “anti-sycophancy” finetunes), but its baked into consumer-facing models to an unbelievable degree.

The only slight retort I have is that social media already provided a personalized, sycophantic “wormtongue” fix. So narcissicsts were already pre-conditioned to turn to these tech portals.

Chat-tuned LLMs just perfected that and took it up to 11.

My thought playing with Llama 1 in 2023, after the initial shock wore off, was “whoa. You know, most people do not have the background to understand how this thing works.” It messed with my head, for sure. It’s like it was designed to defeat a turing test, especially when the next versions got more sycophantic and “confident.”

There was even a story of a Google scientists getting AI psychosis from a BERT model back then:

https://www.scientificamerican.com/article/google-engineer-claims-ai-chatbot-is-sentient-why-that-matters/

I guess a naive part of me thought… it’s fine. They’ll be warnings. They’ll be tons of finetunes that have no hesitance slapping sense into the user, like Resetti in Animal Crossing, and it won’t be public facing anyway. The community using these things will teach each other how they work, right?

Laughs nervously.

I wasn’t cognizant of how maniacally sociopathic Altman was back then.