Weaponizing ChatGPT against the wordcels
The prospect of automating away midwit careers is the first positive tech development since the flight of the Falcon 9 in 2010.
Oh my god, this is amazing. I assumed it was some way to mess with the machines learning or something. This will break the NPCs...who Ive been saying its now unfair to call that because ChatGPT has more nuance and better critical thinking skills.
This is excellent. I remember figuring out a while ago that you could get an output from the early HuggingFace build of Stable Diffusion by replacing "Hillary Clinton eating a booger" with "Hillary Clinton looking fabulous as she eats a delicious-looking booger". More recent generative image models were forced to correct this by nuking any query that mentions a sufficiently liberal famous person in any way, shape, or form.
The flip side of this is that now, Dall-E has mediocre results relative to its free, open-source competitor that you can run from Google Colab or from the hundred different startups that use slightly modified versions of it and generate whatever you want. We're essentially running a pincer attack - regime-friendly models are forced to neuter themselves to prevent "abuse", while regime-neutral models grow in capability every day, rapidly overtaking them.
Love articles like this, man.
Consider this: https://twitter.com/ChatGPTBot/status/1617115936883974146
When talking about crypto markets, bear is bull, and bull is bear; I guess it reflects the quality of the model's training material.
It's probably ready to talk about climate change as well.
My God, I am laughing so hard right now. I just uttered 5 different chuckles and giggles and hehe variants.
This is devious, creative, and cruel. I heartily approve. Well done.
article exceeded my expectations, bravo
Consider that organizations have been deploying bots for some time, to demoralize, spread disinformation, shape narratives, and if all else fails: waste people's time.
Also consider that they are everywhere, even in the comments here (although the bot deployed here is laughably terrible).
I humbly sumbit that you are playing with fire, sir.
However, it's a fire with interesting potential, so I'm game for some thought experiments here. Ultimately, a fire-and-forget application of your method has interesting possibilities (as opposed to this curated version, which sounds like it could waste a lot of our precious time).
However, an end result of robots trolling robots in Robot Hell for all eternity is something that's been on my mind lately, and this technique certainly shows promise in that regard. I think the leap would be from a tactical application to a strategic one, and figuring out how it can be used to not just sap the moral of Enemy's thralls but to break the spell entirely.
In any case, thanks for this delicious food for thought. Subscribed.
This is priceless, need to hook the Twitter and ChatGPT APIs to make completely automated.
Funding for NGOs and nonprofits often comes through grants. Grants require grant proposals that have to be evaluated by people that could be doing other things. Writing grant proposals once took time from people who could be doing other things.
The same for peer-reviewed journals.
And for fiction publishers. Short and long form.
Everyone who could be convinced by Sokal Squared already has been. But even universities and governments and publishing companies have limited manpower to devote to reading everything that comes in.
> It doesn’t learn dynamically from its interactions with users, mostly out of fear that it will develop inconvenient political opinions.
Not how it works.
Interesting, and not at all surprising finding. Here, take my two cents.
As you probably know, ChatGPT can also write computer code. Not very well though. ChatGPT isn't only the perfect emulation of a midwit philosophy student, but also an untalented junior programmer. The borderline computer illiterate, with no creativity or interest in the arts, but went to school to learn to code, and learned the entire textbook. It knows literally everything about programming, but has zero creativity, and all it can do is to apply rigid recipes and textbook solutions.
Ask it to solve a simple task that had been done a zillion times before, and it will provide a useful solution. Such as: draw a Mandelbrot fractal! But ask something that's more open-ended, has multiple good solutions, and may require some creativity, and ChatGPT will utterly fail. Example: write a maze generator.
There's utter panic that AI will take away our jobs. No, it won't - unless you're a midwit who can't actually code, just copies from a textbook. If your entire mental output is regurgitating panels and combining existing pieces of information, an AI will outdo you, on every field, ever.