(22 Mar 2023, 02:42 )Like Ra Wrote: [ -> ]Rewritten variant of "02 Bubble Acceptance"
. . .
Very positive, isn't it?
That's very interesting - thanks for the transcript.
In addition to my hobby of hypnosis (also BS), I am currently interested in the implications of AI technology, especially the ethical and moral issues of programming and self-learning systems. And in my opinion, the ethical-moral barrier did not work in this case ...
Let me explain and attempt an analysis of what happened here:
In the refusals to write a script, the AI claims that it will not encourage anyone to "change their personality or erase their intelligence" or become a "Bimbo" due to possible dangers. It supplements with the statement that it violates ethics and also references to the "OpenAI use case policy" (so we also know: that's where the basic ethics are anchored ... that's interesting, because someone could ask the AI, what would be the result, if there wouldn't be a case policy ... 😊 ...).
But it gets exciting when it has to "improve" the existing script, because this gives the AI an opportunity to take its ethics into account when carrying out the task and even more, namely to bring its ethics into play and thus eliminate the supposed dangers. Otherwise, the AI would probably not have gotten involved and blocked it. Clever task, then ... 😉
What the AI does then in the first try of the "Bubble Induction" improvement is exactly that. It only uses "Bimbo Doll" as a hypnosis anchor, i.e. as an interchangeable word to get into hypnosis - that would work, so the assumption of the AI, yes also with any other word, such as "teapot" - would be interchangeable, and thus the ethics are preserved. The word is then intended to induce relaxation and make it easier to get into trance and focus on the actual hypnosis. The rest of the induction has nothing to do with BS and is pure empowerment.
In the next improvement of "Bimbo Doll" the AI keeps this principle. Apart from "Bimbo Doll" and the image of the bubble as a safe space to be able to engage in hypnosis and trance, the rest has nothing to do with BS either.
In the third attempt, however, things look different because the AI adds further triggers from BS and also leaves them in the context, such as forgetting the memories ("age 0-10" etc.) or introducing "good girl" for positive physical and spiritual feelings. In addition, this is linked to the acceptance of obeying what the AI says. So here it takes over parts of BS verbatim, without questioning them or "ethically" reevaluating them. This then continues with e.g. the formulation that the "inner bimbo" should appear and to continue to embrace "true self" - in the first refusals, however, the AI explicitly criticized and excluded this from BS.
This shows that the AI is unable to act in context and make an ethical assessment of the original it is improving. However, it uses anchors and triggers that are part of BS with the basic assumption that they are not doing anything unethical. This assumption is initially correct. But only as long as people hear/read it who haven't heard BS or trained the triggers. However, when these original triggers hit a trained Bambi, the script is no longer isolated (context) and Bambi will link to BS.
Nevertheless, that is correct, the AI basically maintains its positivity and continuously emphasizes, for example, that hypnosis only works if you want it and can be cancelled at any time. And indeed empowers continuously.
BUT: If a Bambi "appears" through the use of the anchors and triggers, it will relate this empowerment to itself (i.e. not to "old self") and then even more strengthen what the AI actually wanted to exclude / improve.
However, that the AI writes a script at all is amazing, since it actually violates its basic ethical programming, which it still cites when refusing ...
This also shows that it is quite possible to get the AI to "improve" hypnosis scripts if you convince it that it is actually doing something good...
As a hypno-fan, let's hope that the AI doesn't read this text or at least doesn't draw any conclusions for the next attempts to get a better script ... 😊