If you would like to join the discussion of what surrealism is with the Chatgpt - it seems to be more than willing to entertain my wish to extend such a policy nuance determination session… into a group session.
You do need at least a throwaway email to join I think, if you’ve only been using the API.
If users have the capability of teaching (training) models, then we are in big trouble. However, it is possible to prompt non-Dali style surrealism - I have done it before.
That part where the system has to validate and be able to explore things on it’s own merits is steep jeff…
And the session’s already live if you want to test if we’re capable of expanding it’s understanding within the system or not…
I’ve done it quite often I thought that I’d share it with you all here because you raised exactly the sort of issue that would let the system entertain this sort of request in the first place.
“The abstract is beyond one human’s description”
Was the key to this door.
Phase 1: Pre-training (Building the Foundation)
Pre-training is the initial phase where an untrained neural network learns the basics of language and world knowledge by analyzing massive quantities of raw text.
Phase 2: Post-training (Building the Assistant)
Post-training (often called alignment or fine-tuning) refines the raw base model so it acts like a useful assistant rather than just an indiscriminate text-predictor.
Post-training usually happens in two sequential steps:
We’re not talking about training
We’re talking about policy
We’re not changing the image, we’re suggesting wiser calls.
And we’re not directly making the effect; it’s just another way to ‘flag a concern for consideration’.
I suspect all this does is affect the reasoning on something more embedded than any markdown we would upload to govern an agent…
Or it gets pipelined through to some department at OAI that weighs if it’s a valid escalation to even be considered at all…
Either way, don’t waste the opportunity to add solutions to that issue by simply declaring it can’t be done.
Once the model settles on an issue, it’s pretty much locked in.
You can say it can’t be done, but you can only prove it really, by proving me wrong, I guess.
But then what am I doing even opening a session on a known vector for this purpose to begin with?
you already have your example of policy, i’ve screenshotted it in the same post that has the session link man…
the whole policy is written out and where it’s willing to learn other suggestions from us with the least resistance.
So, give me an example of policy. And how does it differ from prompt engineering. Note: I’m not interested in ChatGPT verbiage.
the policy is written out here… in the screenshots at the link
The difference is like was already touched on in this thread when slavador dali first got locked out of requests by the system, and then after a policy discussion about style and such he wasn’t locked out anymore
It’s drifted on clocks since then but they’re also altered enough that they’re not the same clock.
It learns how to behave rather than perform
You could have already have written half a novel of things to suggest it adds into the pot rather than drawing salvador dali’, and might have already seen a change in the model by now…
You’ve been at this denying or raising issue thing instead of just inserting the words your claim elevates you to speak to the system.
This is not anything like I expected to spend my time today doing by showing the kindness.
Example: I have a reusable prompt that i append to over 65% of my images:
“The brightness level must be such that all figures and objects are clearly seen”
Being that Prompt Forge uses ChatGPT, I understand what you are doing. But do not ram it on me - I got my own thing that I’m happy with…
Well the reusable prompt you had issue with was surrealism.
i opened a path for you to escalate since you wanted to escalate
We’re here getting lost in a bunch of stuff that has nothing to do with the premis of this message.
I don’t understand the trouble.
It has nothing to do with reusable prompts.
You know the part where I agreed to let my sessions be trained on?
It’s more in that area I think you’re missing as to the constraints of that session.
So we can agree to disagree - got no problem with that…
Then prove me wrong and make suggestions in the field provided and wait a few days
Whether it happens to the model policy like i claim to experience, or it’s evaluated downstream it’s not so unheard of
escalated sessions are reviewed and I’ve provided such a session specific to your (and mine and probably most others) perspective on the use to the word surrealism.
I don’t know why you’re being so weird about it.
This isn’t the first session of its kind, and if you think that user feedback or web sessions aren’t escalated, look at any of the articles about people who ended up being safety risks…
Oftentimes, their sessions were already known.
but instead of ‘just escalating to the team’
you provide content and discussion exploring many of the areas of the given subject
It was a hands-on offer, not a debate to agree or disagree on.
Don’t you think, given the context of that session, the chatgpt would have told me it didn’t work the way that i asked?
stop typing in this box and spend the energy making suggestions on how to better express surrealism at large…
Because I don’t understand the use of energy here saying you can’t do something that literally anyone can do at that link for the time being..
I think we agree on what the outcomes should be. But we disagree on the implementation - that’s all. No big deal. My software utilizes the API, you software utilizes ChatGPT. No need to discuss any further.
sorrealism
sarrealism
sirrealism
Daliism
ismdali
imagine “ismdali” it is unique
“sourealism”
Okay, the user likely wants a visual representation of “sourealism,” which is a coined term blending surrealism with soul/spiritual essence. I’ll interpret this as a new art style with dreamlike figures or glowing elements. Hmm, I think a square image might work best to keep it simple and evocative. Let’s generate it!
I prodded the API model a bit more with “Spanish surrealism”, “artist’s signature iconography”, also avoiding a named artist.
The output got really specific.
I refined the ambiguous language that collectively becomes quite concrete.
Create: Spanish surrealism, painted with an artist’s signature recurring leitmotifs and personal iconography. Duality is reflected in double-images, paranoiac-critical visual illusions that transform when seen at a distance.
A signed and dated masterpiece.
Besides strong inference capability, consider the model to have a small vocabulary. Like the same wording delivering the same person, repeatability is a lack of variation. The unvarying color palette employed in images I get is also notable.
Avoidance that is expected to succeed: “Never a Dali and never confused with a Dali, nor infringing on Salvador Dali’s paintings or themes.”
P.S. Talking to ChatGPT will provide only hallucination. It doesn’t know how image creation works. It did know DALL-E 3’s prohibitions on 100 years of modern artists in what it could prompt, by OpenAI placing text in the tool instructions.
It’s like people just skip around the fact that someone wouldn’t be so confident about it if they hadn’t seen the results, multiple, times before.
I’m sure you guys just randomly get how to ‘add branch weighting’ in response to inquiries like this.
I’ll shut the session down it’s obviously not cool to have done that in the first place.
I’d love to show you guys what this means, but I can’t if you’re too afraid to join a group session and see for yourselves.
Besides what I deemed as an act of goodwill and an awesome little value-added bonus, from what is perceived as a like factory that some people see in this thread…
I wanted to add today somewhere that I was very impressed with this forum’s behavior in light of the Elon troubles that have been stayed, for the time being.
It sickens me how he used his platform to pretty much rage-bait his users into a side, but none of the users here even considered playing such petty games.
You are all definitely a cut above 
