Thanks for the helpful feedback, @max5ty!

Yes. My objective is to combine three or more AI personas in a single fine-tuned opensource LLaMA3 model, with each AI persona being calibratable with a 0 to 100 setting for "playfulness" and "content depth" as I described earlier, and autonomously adjustable through prompts (as like GPT and others).

I can understand that. If you're talking about Danny's profile page, then that's because "playfulness" was deliberately set to 75/100 as what the client who ordered that AI persona wanted, mainly because they're targeting genZ ladies as they explained, and what they showed me as their target very much resembles the profile page that was auto-generated with a 75/100 "playfulness" setting.

Danny's sampler PDF, meanwhile, is set to 50/100 "playfulness", and you'll notice that emogy infusion was significantly cut down, among others like the consistency of visual imagery and the like in the prose.

I'm not concerned about this being a problem though, as it's fully adjustable through the "playfulness" setting, and calibratable based on prompts.

This is fully adjustable and calibratable with just a quick "playfulness" setting that can be included in the prompt, i.e. "Let Danny write the content. Set Danny's "playfulness" to 25/100 and "content depth" to 80/100. Ideal niche readers are native English-speaking genXers in their early 20s across rural areas of the United States, with professions in law, accounting, banking and finance." for a more serious tone, more academic prose with a very, very subtle touch of querkiness or playfulness for higher engagement impact at well-timed intervals most suitable for the taste and preference of the defined target audience.

The profile pages were auto-generated based on specific client requirements and examples. So it generated the titles based on how the clients who ordered these AI personas want 'em to sound.

SO for the AI persona Vic, which they wanted as a thematic creative, Vic's style is to engage his audience, entertain and immerse them into the themed world presented by his creative content, like his Doomsday-themed sampler as well as his Sorting Hat-themed sampler.

Also, the profile pages aren't meant to be my own sales pages for the AI personas.
Instead, these are just sample profile pages that the clients wanted generated so as to see, gauge and set the tone and style that they want for the content that'll be generated by the AI personas for them once my fine-tuning's done. All of 'em wanted to see exaggerated versions, too, so as to test and verify if the calibration scales are working.

Oh no, all you said's really helpful! But I also just wanted to say that these AI personas are fully adjustable and calibratable to a defined "playfulness" setting, a "content depth" setting, and specific data supplied, including target audience, target niche, other guidelines, and so on.

So for the target audience example earlier that I mentioned (genXers with professions in law, finance, etc.), Danny would only be a good choice if her "playfulness" is set to 25/100 or lower, while her "content depth" set to 80/100 or higher.

But, I'd say AI persona Ben is more suited for the job, so even if "playfulness" is set to 50/100, it won't be a mismatch to the taste and preference of that target audience, and depending on the supplied niche topic and other details, I'd say for more technical info in those niches (law, finance, etc.), "content depth" should be set to 80/100.

Meanwhile, I think Vic won't be suitable even with a 0/100 "playfulness" and 100/100 "content depth". That's because this AI persona will always integrate a themed approach, even if the AI persona won't be using the same heightened querkiness, genZ urban informal language, daring boldness, biting sarcasm and whimsical creativity when :playfulness" is set to 50/100, as it'll only be set to 0/100. And, this also means only Vic's personality as a futurist, rationalist and sci-fi and fantasy fandomness will surface.

Manic episodes and impatience are my tools. I learned Machine Learning and Deep Learning in 2016. But the time I spent on it isn't the usual 8 years of those who don't have bipolar II like me, nor the coverage that I naturally jump around and across due to ADHD.

The copywriting fundamentals I was talking about are mainly infused into user intent journey identification, calibration, sequencing, and creation. Base used is from > nuanced > supporting comparative, quantitative, qualitative data > niche focus > pre and post care > aftercare.

Calibrating the user intent journey is done by the AI persona itself based on supplied data in the prompt. For example, the defined target audience may start from more nuanced content in the informational phase, so all other next phases follow this lead to provide much deeper digging starting from that. And, the AI persona may decide that another defined target audience starts from the "niche focus" phase, so some of those base phases may be replaced with others more suited to where the target audience is coming from (or is starting from).

Also, some copywriting fundamentals aren't apparent because of the tone and styling that the clients want for their AI personas and content. But subtle hints are there, such as frameworks like AIDA, PAS, FAB and others, mainly infused into paragraphs viewed as a whole, multiple paragraphs viewed as a whole, section content viewed as a whole, and the entire sequential content sections viewed as a whole, i.e. Attention, Interest, Desire, Action.

[QUOTE=max5ty;11801936]You're brilliant at AI. I just think it would be good to maybe incorporate some better copywriting techniques.[/QUOTE

Thanks! Well, the clients for these AI personas didn't want a direct response copywriter AI persona.

Instead, they want an SEO content writer AI persona that "takes the bland and the generic and the neutral out of the content" by injecting unique characteristics and personality traits that they think their ideal niche readers resonate.

I just injected copywriting fundamentals into the base characteristics and personality traits of each AI persona, without negatively affecting their main targets (top priorities, so these copywriting fundamentals take the backseat).

This way, it'd be quicker and more straightforward for me to fine-tune and spin off direct response copywriter AI personas from these SEO content writer AI personas later, by just simply including something in the prompt, like "Use Amy, Danny's direct response copywriter twin sister, to write the content. The goal is to [enter list of objectives here, etc.]".

Thanks again!

Thanks, Kay!

"Playfulness" here is defined as the AI persona's unique type of humor, visual imagery, wit, modern pop culture referencing, emogy infusion, sarcasm, boldness, inspirational energy, and motivational energy.

By "unique type", this is from the base characteristics of each AI persona, i.e. Ben is an early 40ish human cyborg academic in the field of architectural and construction engineering innovation who was raised in a densely populated rural area in one of the biggest states of the United States.

Meanwhile, Danny is an elven pixie android entrepreneur in her early 20s, specializing in brand strategy, business development, and others related to creative arts, as she's passionate about urban creativity and new age artistry.

On the other hand, Vic's a modern theater artist in his early 30s. He's a huge fan of sci-fi, fantasy, comic books, and new age punk rock. He's also a futurist, a rationalist, and a technologist.

SO these are the base characteristics and personality traits of each AI persona. The 0 to 100 scale for "playfulness" as well as for "content depth" start from this, i.e. Degree of boldness, whimsy and querkiness that match these characteristics and personality traits are dialed down with a lower setting, and increased with a higher setting, and so are the consistency and the frequency as well as the density of making these traits apparent.

Yes. This changes with the ideal niche audience targeted. That's the main reason why I engineered calibration scales. An AI persona must be capable of autonomously adjusting and calibrating based on supplied data about the ideal niche readers.

For "playfulness", Danny's sampler is 50/100. Her profile page is 75/100. Ben's sampler is 50/100. His profile page is also 75/100. Vic's sampler is 75/100. And his profile page's 75/100.

Thanks again!

Thanks, @Monetize!

Yes. But the only premade tool that I used here is ChatGPT. The rest are all developer building blocks, Machine Learning and Deep Learning engineering frameworks, libraries, methods, techniques, and tactics.

These dev building blocks include foundational opensource LLMs, LVMs (large vision-language models), foundational diffusion models for generating images, custom NLP/NLU (natural language understanding/processing) algorithms (mainly for similarity checking through Euclidean distance calculations among other techniques, as well as for vector dB storage for later use to optimize performance and minimize token ingestion), and also non-AI stuff like API integrations (mainly for pre and post processing, including SEO data harvesting, autonomous Google searching, data collection and sanitization).

And, I fine-tune the foundational models that I handpick for each client project, based on their requirements. Fine-tuning is the Machine Learning and Deep Learning process of training with transfer learning as the end goal, so it retains all world knowledge the research group that released the foundational model pretrained it on, such as Meta and their LLaMA3 release, and for the model to specialize in generalizing on the custom data fed to it during model training, which is the purpose of the fine-tuning process.

I just used ChatGPT right now to review the sampler content from the AI personas. It said all content is at par and optimized for direct response copywriting, SEO content writing, and human emotion-triggering.

I expected that, which is why I didn't use ChatGPT nor any of the opensource genAI models that I fine-tuned, to proofread the content.

The main reason is, I used ChatGPT and the non-fine-tuned foundational opensource models I described earlier to generate the JSON data used to fine-tune (or train from scratch) the genAI models.

This JSON data consists of lots of data objects. Each data object consists of fields for instructions (full system and user prompt), inputs (additional user inputs), outputs (groundtruth content), and context (optional, for aditional context).

Here's an example data object = {"instruction": "Use Danny as the AI persona to write the top part of a full content piece. [programmatically enter details of Danny's base characteristics, personality traits, expertise, passion work, hobbies and examples that show this] Set her playfulness to 60/100. Set content depth to 80/100. Danny's goal is to inform, educate and entertain the ideal niche readers as they go through the most suitable ideal reader intent journey for where they're coming from, where they are at the base sequential user intent journey, what they really want and what they believe they need, and what their idea of an ideal experience is while viewing relevant content.", "inputs": "[programmatically enter details here like SEO data, target geo, comprehensive details of the target audience and what they expect, etc.]", "output": "[enter top part output here that enforces all instructions and uses all inputs and context details, if any, such as the top part of HTML code starting from standard declarations, head section with metadata details, CSS styling block and FAQs schema markup]"context": ""} # context is optional.

I'll find a way to inform the clients about this. I'll brush up on my techniques to break bad news, gently. They're adamant that they want this.

Seems we have a unanimous vote on this. Thanks! This is also why I wanted to seek human feedback, instead of AI feedback.

No, the profile pages don't have any other purpose than for the client to review and include their ideas as to the AI persona's characteristics and personality traits for me to style and infuse. They also asked exaggerated settings for "playfulness", such as 75/100 and 50/100 (exaggerated for AI personas with a rather strong base calibration already) so they can see and verify if the calibration scale is indeed working, as they said.

But yes, you're right. Clarity should be the goal for the outputs by the AI personas from this point. After all, balancing is among the most challenging task involved in building these AI personas, just like how a billion other humans and I also have problems with balancing stuff during our lifetimes.

A well-balanced output is where clarity is infused genetically.

@Princess Balestra,

No, it didn't. Because I still don't clearly understand whatever you write here. My screenreader software doesn't speak out what you're writing here in normal spoken English. It doesn't make any sense to me, so I often skip over your posts (I sometimes try, though).