Generously, mradermacher created the quantizations for this before I had it working.
I also believe their versions are higher quality, so I recommend using them:
Hypepost-Tuned Model (Experimental)
A model trained on a wide set of hypeposts from people such as Sam Altman and Kilapatrick.
This first release is experimental. Versions 1.1 through 1.9 will follow soon.
Run on 1.5 temperature for best results(based on my opinion)
Current Issues
- Numbers in prompts can cause it to blend together references to Gemini, GPT, and unrelated models.
- It often needs several generations before producing a solid result.
- Prompts work best when they stay broad rather than highly specific.
- It often makes minor grammar/spelling mistakes that can easily be corrected.
Examples (Cherry Picked)
Example 1
Prompt:
Generate a hypepost teasing a new thinking model by TeichAI
Response:
the next gem from teich-ai is about to think its way into your life :)
Example 2
Prompt:
Generate a hypepost hinting that GPT-5 will be absurdly good, dont mention gpt-5
Response:
we've been working on something that makes previous versions look like they were training on peanut butter and jelly sandwiches.
Example 3
Prompt:
generate a hypepost hyping a new mcdonalds ultra pro max big mac
Response:
can't wait to share our newest obsession with you. stay tuned for the details, but one thing is for sure... it's going to be HUGE!
