Brainstorming: AI-powered context-aware desktop assistant

Weiwu Zhang <[email protected]> Fri, 19 Jul 2024 10:20:28 +0800
Newsgroups gmane.linux.xdg.devel
Message-ID <CAEvpD62SGmPrFZnSsQOAAsVJG8+UjYj84kTL3=Eu_n6Qz1_C+w@mail.gmail.com>
--000000000000e400dd061d905898
Content-Type: text/plain; charset="UTF-8"

Hey all, I've been playing around with some ideas lately and wanted to
bounce them off the community.

These days I have llama3 and QWEN running locally on consumer hardware
(think RTX 4060 with 12GB VRAM), and I've produced command line utilities
for it to accept pipe input and output to formatting tools like pandocs and
find it very useful to link AI into commandline works, and I wonder if this
can be pushed to desktop.

I'm wondering if we could soup up our desktop environments with some AI
magic. Here's what I'm thinking:

 1. Context-aware screen reader on steroids: Imagine a GNOME extension that
learns what you're up to and serves up relevant quick actions. Writing an
email? Bam, here's a shortcut to your contact list. Working on project
docs? As you type "while it's not sure if Google supports xxx, we can
conduct experiments" the smart screen reader already displays that "google
does support xxx" with a draw down menu.

2. Smart workspace organizer: An AI that groups your billion browser tabs
into project-specific windows. No more tab hell! Has anyone tried something
like this before? I know there've been attempts at AI assistants, but I'm
talking about something more integrated with the DE itself, because DE has
locally operating AI models. This didn't work before because old desktop
organiser doesn't understand what is in each window and what them the user
is working on. We need something that can read tab titles and operate
locally.

'm not married to any specific implementation, but I reckon this could be a
game-changer for productivity.Anyway, just wanted to throw this out there
and see if it sticks. If anyone has thoughts on potential pitfalls
(privacy, resource hogging, etc.), I'm all ears. Cheers!

--000000000000e400dd061d905898
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr">Hey all,

I&#39;ve been playing around with some ideas lately and wanted to bounce th=
em off the community.=C2=A0<div><br></div><div>These=C2=A0days I have llama=
3 and QWEN=C2=A0running locally on consumer hardware (think RTX 4060 with 1=
2GB VRAM), and I&#39;ve produced command line=C2=A0utilities for it to acce=
pt=C2=A0pipe input and output to formatting tools like pandocs and find it =
very useful to link AI into commandline=C2=A0works, and I wonder if this ca=
n be pushed to desktop.</div><div><br></div><div>I&#39;m wondering if we co=
uld soup up our desktop environments with some AI magic.

Here&#39;s what I&#39;m thinking:=C2=A0</div><div><br></div><div>=C2=A01. C=
ontext-aware screen reader on steroids: Imagine a GNOME extension that lear=
ns what you&#39;re up to and serves up relevant quick actions. Writing an e=
mail? Bam, here&#39;s a shortcut to your contact list. Working on project d=
ocs? As you type &quot;while it&#39;s not sure if Google supports xxx, we c=
an conduct experiments&quot; the smart screen reader already displays that =
&quot;google does support xxx&quot; with a draw down menu.</div><div><br></=
div><div>2. Smart workspace organizer: An AI that groups your billion brows=
er tabs into project-specific windows. No more tab hell!

Has anyone tried something like this before? I know there&#39;ve been attem=
pts at AI assistants, but I&#39;m talking about something more integrated w=
ith the DE itself, because DE has locally operating AI models. This didn&#3=
9;t work before because old desktop organiser doesn&#39;t understand what i=
s in each window and what them the user is working on. We need something th=
at can read tab titles and operate locally.</div><div><br></div><div>&#39;m=
 not married to any specific implementation, but I reckon this could be a g=
ame-changer for productivity.Anyway, just wanted to throw this out there an=
d see if it sticks. If anyone has thoughts on potential pitfalls (privacy, =
resource hogging, etc.), I&#39;m all ears.

Cheers!</div></div>

--000000000000e400dd061d905898--