HN user
This is awesome!
Also, for the mouse emulation, how about keep checking forever whether mouse is down, and fire only mouse-up and mouse-down events (so no constant firing probably), however that would be done, because i'd guess broadcast messages won't work on a pen-based OS but could be wrong.
Otherwise maybe you can just dedicate some keys to mouse movement (possibly arrow keys), and one key for click.
1: Narrow Expertise: AIs usually target different things, so for example, in the context of LLMs, one might target being factually correct, so a knowledge benchmark is better suited, while another might target being good at coding, so a coding benchmark is better suited.
2: Answer Memorization: Once a benchmark has been live for a long time, AI's might start to learn from it, as it might be on various articles, and may have leaked into the training data, so unless a new benchmark is used, it would just be outputting memorised (overfitted) answers rather than being actually good at what the benchmark aims to measure.
3: Marketing: AI companies like to make their product look good (just like any other company), so they will often perform a ton of benchmarks, and only display the results where the AI performed best, to make it look like the AI is good in general.
Basically, it means that the developer does not guarantee that the software works, because in some countries or jurisdictions (like the US) it is assumed that the developer provides a "warranty" that the software works unless the developer explicitly says otherwise.
So basically, it's like saying "you can't sue me if my software has bugs"
It is pretty normal, mainly because it doesn't constantly take into account that it's an AI when generating text and it's basically predicting the next most likely token, so if it said something like "After reading that, I was literally smiling", when it finishes "After reading that, I was literally" it has a few options of what to continue it with.
If the previous chat was something that would cause a regular human to smile or laugh (a behaviour existed somewhere on its training data), then "smiling" and "laughing" are probably the most likely next tokens to express humor. Similar could be said about topics other than humor.