Steve Litt - 14.06.25, 01:27:42 CEST:
> >Any suggestions?
> >(OT, I know)
>
> We all have different preferences. I personally invoke dmenu with
> Ctrl+Shift+j . Easy for your left hand to form a snake and go left and
> down to hit Ctrl+Shift, and then snake back up to home typing position
> the opposite way.

I accepted the new Plasma default hot key for KRunner which is Left Alt-
Space. It is a ridiculously good idea if you ask me.

I Alt-Space-KRunner almost everything on my Plasma desktop. Up to the
point that my main laptops do not even have any always visible side bar.
It is full screen estate for apps. Why even have a menu? Well if I want an
overview on what applications and games are installed… but otherwise?

Alt-Space-something, Enter, Done. Like in Alt-Space-"fi" for Firefox, Alt-
Space-"ka" for Kate, Alt-Space-"kw" for KWrite. Most often it is about 2-4
letters and I have the result I want. With directories and files it can be
a bit more and sometimes I have to many results and KRunner does not show
the one I want. But usually whether it is an application, a settings
window, a file- or directory name, some text content *inside* a file,
including regular text, markdown, PDF, OpenDocument files, … Or even
quickly asking what 1 US dollar is in Euro currently or how many light
years the diagonal axis of my 24 inch is display is. The latter is very
important to know, you know. :)

It is not completely perfect. Something I am to quick to notice that
KRunner gets it wrong. But even when I count these times, I think I am
way, way, way more effective this way than with navigating a menu.

I am not sure how this will evolve. Maybe at some time KRunner works
similar like an AI assistant, but with locally run AI model. And then I
could ask "Can you give me a summary on what I have stored about topic xyz
and please include all references as footnotes?" or "Please give me pro
and contra using what I have stored locally on topic xyz" or at some time
"Please download all the files about topic xyz on these three websites and
give me a summary". But all of this AI stuff needs very careful ethical
assessment. And there has been a security issue where Microsoft Copilot
could be made to reveal confidential data by a specially crafted mail it
automatically answered. If you ask me: None of that should ever be
automatic. Always review!

I did not watch the interviews I have stored yet, but those former OpenAI
employees certainly did not leave the company without having a good
reason. I do experiment a bit with locally running models using GNOME
application Alpaca¹ installed as Flatpak together with Ollama and AMD
support plugins. So far I basically cracked every one of the models I
tried. But some where able to admit they hallucinated or made a mistake.
And some also gave some quite impressive results, including one which
analyzed two screenshots from a game for me.

And those reasoning models which give their "thought" process: Quite a bit
scary. Especially when I ask for a cut off date of the training date and
the model thinks about whether to give it to me cause it could be a trade
secret of the company who build it. And things like that. Mind you, the
model still gave away the approximate date later on. What I read there
actually *looks* like a "thought" process. It may not (yet) be, but rather
be simulation of one. But still it is close.

However I do not trust any of the big AI players any farer than I can
throw a mountain.

[1] https://jeffser.com/alpaca/

Someone appears to work on a KDE variant of it:

https://invent.kde.org/utilities/alpaka

Best,
--
Martin