A study of difficulties with the Darlink.ai systems

Here is a log of me trying to converse with the darlink.ai system in order to try and clarify how the ai model works. There are several key takeaways from this.

1. The ai chat model is unable to properly count the number of characters being used. In the document, the ai chat bot declares that it reads information in bytes not characters.

2. Through experimentation with the image generator, with suggestion and analysis from the chat ai, there is a clear winner for how to word text prompts for maximum efficiency. The best method seems to be weighted terms connected by underscores. Example: (Word1_Word2:1.5). It is noted within the log that the underrscore works to tie the two words as a single concept. The generator runs into issues when the bracketed chain of words exceeds 3 items. Example: (Word1_Word2_Word3_Word4:1.5).

3. The ai chat model will drift into melodramatic language DESPITE attempts to mitigate this. I asked the ai chat bot how it would discourage melodrama only for it to creep back into the conversation. I provided two examples of this within the log. I started with a high emotional impact scene and followed with ones that were progressively lower. The results show an INCREASE in melodrama despite the lower emotional impact. THIS MIGHT BE AN ISSUE WITH THE TRAINING MODEL FOR THE AI! Even the chat bot realized that the rising melodrama was unavoidable despite all attempts to mitigate them.

4. The image generator has a lot of problems with complex patterns and often misinterprets ideas onto the wrong area.

5. A negative prompt field would REALLY help with some of these issues.