00:00:00
Proton released an AI agent called Lumo,
00:00:02
and yes, it shares many similarities
00:00:05
to a certain dog AI agent, but with a cat.
00:00:08
I have yet to actually do my first search with it.
00:00:12
And so this is a live demo,
00:00:14
as well as showing where it is in the Proton ecosystem,
00:00:16
comparing it to other solutions like DuckAI,
00:00:19
as well as Kagi.
00:00:20
You can test this right now.
00:00:21
I'll leave a link down in the description.
00:00:23
Now they say no logs, zero access encryption,
00:00:26
no data sharing, not used to train AI, open,
00:00:29
and then they have a chart. And then they also have some other new features here, which is ghost
00:00:34
mode, web search. So you can search the web with this, file uploads, and then Proton Drive
00:00:38
integration. Really want to try out. So this is a bit different from other approaches. So to compare
00:00:43
to something like DuckAI, DuckAI, you're essentially relying on DuckDuckGo to proxy the request for you
00:00:50
to a different agent. So if I go here and I just do the default suggest a title for a blog post,
00:00:55
you go ahead and do this. But you can actually change the model here. And again, they're just
00:00:59
going to be proxying this. And so there is a lot of privacy you gain here, but it's not a trustless
00:01:04
model because there is still some trust in DuckDuckGo. And then Kagi does something fairly similar,
00:01:09
which is they go ahead and proxy it. They have privacy practices in place to make sure nothing's
00:01:14
tracked. Proton is doing something a bit different in their defense here because this is fully end-to-end
00:01:21
encrypted. So they, not even Proton, can access chat history. In terms of LLMs, this is a bit
00:01:26
unclear, but I had to essentially get this from just random guesses as well as asking the AI model.
00:01:31
But Lumo seems to use open source models like maybe Nemo, Mistral, Open Hands, Almo, that's based
00:01:37
on some coverage that was done on Lumo. And apparently it's picking the best one suited for
00:01:41
a query without the user having to manually switch. But this seems very opaque. And again,
00:01:45
this isn't even really confirmed by Proton directly in any of their documentation. So they
00:01:49
should fix that. And to clarify the difference here, they're running this on a Proton-controlled
00:01:54
server, but it's all end-to-end encrypted so the server can't actually see what's happening. Again,
00:01:58
this is different from DuckDuckGo and Kagi, who anonymize your queries to then send to the AI
00:02:03
models, which still are fairly privacy-respecting, but it's not a trustless model like what Proton
00:02:07
is doing. Now, one thing that's important that you have to dig into a little bit is there is a
00:02:12
free and PlusPlan. PlusPlan offers unlimited questions and premium features, which includes
00:02:17
unlimited chats, extended chat history, encrypted chat history, search and favorite your past chats,
00:02:22
ghost mode, which is also in the free one. So I'm not sure what the difference is between ghost mode
00:02:28
and free and plus, and also upload multiple and large files. So to access, I just went to
00:02:33
Lumo.proton.me. And if we give it a prompt here, what is end to end encryption and why is it
00:02:38
important, gives a pretty basic response. If we do it side by side, you'll see DuckDuckGo actually
00:02:43
generates that far quicker. I will say just formatting and everything seems a lot better
00:02:47
DuckDuckGoesIn, which is currently using ChatGPT. And if we ask it to explain the difference between
00:02:52
VPN and Tor, see like this, this is actually generating in real time, it seems like. I much
00:02:57
prefer the way a GPT via DuckAI seems to be formatting this, but none of this is inaccurate
00:03:03
and it's not bad. How about what are the best practices for protecting my digital privacy?
00:03:08
Enable 2FA so far, it's the same. Be cautious with public Wi-Fi. I don't think use a VPN needs to be
00:03:13
the third thing. Use encryption. Okay. Limit personal information sharing, review privacy
00:03:18
settings, use secure browsers, be wary of phishing scams, backup data, privacy-focused services.
00:03:23
Actually, I recommend Duck.co, which is funny. I was looking to see if Duck would recommend Proton
00:03:28
because that would be just awesome. So we went ahead and we kind of stumbled on something. So
00:03:32
you need to use the web search feature. So all you do is click web search. Go ahead and click
00:03:36
there. I put a little typo instead of a question. Oh, I did enable web search. Sir, ma'am,
00:03:42
Cat, I enabled web search.
00:03:46
Okay, it's taking its sweet ass time.
00:03:48
Oh, okay.
00:03:50
All right, I guess it's confused.
00:03:52
Let's try refreshing the page.
00:03:53
Duck seems to actually have access to real-time information, which is pretty sweet.
00:03:57
And I didn't even have to toggle on search the web.
00:03:59
Let's try this again.
00:04:00
Web search, downvote.
00:04:03
Not correct.
00:04:05
Yeah, that's interesting.
00:04:06
It could be a bug or something.
00:04:07
Now to test out the integration of Proton Drive,
00:04:10
I created a folder here which has an SRT file, which is a subtitle file, which is mostly text.
00:04:16
I did a ProtonDoc, which is just an outline and kind of what I used as a script for a recent video that we put out.
00:04:24
And then I also added two images to see if either of those can be used by the AI agent.
00:04:29
So we're going to click Upload, Add from Proton Drive, and then Lumo Test.
00:04:34
And it says no supported files in this folder.
00:04:37
So I don't really know what this is supposed to be used for.
00:04:40
Okay, a PDF of the document actually seems to be there.
00:04:44
So I'm going to go ahead and add that.
00:04:46
And then it shows it there.
00:04:48
What is this?
00:04:50
Let's see what it can come up with.
00:04:52
Okay.
00:04:53
Okay.
00:04:53
Yeah, so it seems like this handles PDFs so far.
00:04:55
It is interesting.
00:04:56
It doesn't even handle its own documents.
00:04:59
Like, why would I need?
00:05:00
Are you telling me I have to export as a PDF and then I can use it inside the AI agent?
00:05:05
I think that's interesting.
00:05:06
And I don't think there's a way to access this within here.
00:05:11
No, that's a comment.
00:05:12
Yeah, I'm not seeing a way to access the AI agent directly in here.
00:05:16
So really interesting choice there.
00:05:17
So you might be asking in different approaches when it comes to privacy.
00:05:20
So I want to go from worst to best and kind of clear up some of the pros and cons between
00:05:25
the different approaches when it comes to your data.
00:05:27
So the worst is going to be something just like chat, GPT, Gemini, even Claude.
00:05:34
None of these are going to be privacy respecting.
00:05:36
They're going to use data that you send to train their AI models.
00:05:38
They're going to collect that data.
00:05:39
It's going to be tied to a central profile.
00:05:41
They can read everything.
00:05:42
They can profile you.
00:05:43
Some of them can even share that data with third parties.
00:05:46
So very not good.
00:05:48
And those are especially true with the free plans.
00:05:51
If you pay for these models, they actually have slightly better privacy policies where
00:05:54
they commit to not training your data on those models.
00:05:58
So at minimum, if you're using one of these mainstream providers, try to pay for them
00:06:02
because that way they do have better privacy.
00:06:04
From there, you're going to have tools like DuckAI and possibly Kagis as well, which is going to
00:06:09
essentially proxy through their servers to then reach the same AI models you'd otherwise use.
00:06:16
Now, the cool thing is that this gives you pretty much the same experience as using the actual
00:06:21
models, but with a lot better privacy. And I think these are overall really solid middle ground,
00:06:25
and you can actually access DuckAI just from DuckDuckGo directly. So when you do a search,
00:06:30
maybe you don't know something, you want to maybe dive a little bit deeper into it,
00:06:33
you can go ahead and do it there.
00:06:34
Of course, this is all assuming you're even interested in AI in the first place,
00:06:38
which if you're not, I don't know why you watched halfway through this video.
00:06:40
Now, the next best approach seems to be something like Lumo here.
00:06:43
I don't know of anything else like Lumo, which does like end-to-end encrypted,
00:06:47
but this seems to be a little bit more limited and not quite as good
00:06:50
as something like DuckDuckGo proxing things,
00:06:52
but this is definitely a little bit more privacy respecting than DuckDuckGo.
00:06:56
Just keep in mind, if you're using something like DuckDuckGo,
00:06:58
you're getting like 80% of the privacy benefits.
00:07:00
This is just kind of taking you to like 99%.
00:07:03
Now, there's one option I haven't really brought up so far, which I would argue is probably
00:07:07
the best option.
00:07:08
You can use local programs like JAN, like Misti, and all these things exist.
00:07:14
And so what you can do is you can just download an open source AI model.
00:07:18
It's completely offline.
00:07:20
And you can even toggle if you want real time data or not, kind of like what Proton is supposed
00:07:24
to do.
00:07:25
And then this stores everything on that database on your computer.
00:07:29
And so you can pretty much have full access to this.
00:07:31
You get access to a ton of cool features.
00:07:33
I can load any document that I want into this, including images, PDFs, anything. And it's not
00:07:38
limited by the limitations of things like end-to-end encryption. So I think that this is also a
00:07:43
phenomenal option. And this is like 99% of the way that I use AI in my life is through things like
00:07:49
this. What are some big news stories today? I'm going to try the web search again to see if this
00:07:54
is just maybe that chat was broken or something. Okay. Okay. Look at that. Oh, it's doing something.
00:08:00
What are you getting at, Lumo? What are you getting at, Lumo? But no, this actually worked. So maybe
00:08:06
you can't, there's a bug or something where if you like toggle web search on a chat that didn't
00:08:11
already have web search, it doesn't work. But it looks like it did work when I started a new chat.
00:08:15
One very big announcement that took up a very small amount of announcement in this blog,
00:08:22
because this blog just says, oh, Lumo by Proton. But at the very end, there's something actually
00:08:26
very important. So Switzerland lately has been kind of going after privacy and security, and this is
00:08:31
something that I'll probably be doing more coverage for in the near future. But Proton here says,
00:08:37
Luma represents one of many investments Proton will be making before the end of the decade to
00:08:41
ensure that Europe stays strong, independent, and technologically sovereign. Because of legal
00:08:46
uncertainty around Swiss government proposals to introduce mass surveillance that have been
00:08:51
outlawed in the EU, Proton is moving most of its physical infrastructure out of Switzerland.
00:08:56
Lumo will be the first product to move, which means that Lumo is actually being hosted in Europe, not in Switzerland.
00:09:02
Which is a bit interesting because if you go to Lumo.proton.me and click learn more,
00:09:06
it still says, oh, we're owned by a Swiss non-profit.
00:09:08
And literally, like, all of Proton's branding is Swiss privacy, and they mention that all the time.
00:09:14
Based in Switzerland.
00:09:15
For the record, this seems like overall a good move, so I'm happy they're doing this move.
00:09:19
But, uh, yeah, kind of an interesting situation they're in right now.
00:09:23
and this is a very minor but kind of major update
00:09:26
that they just subtly attached to the end of the block.
00:09:28
As for my final analysis of the situation,
00:09:31
for the record, this is not the first thing
00:09:32
I would have asked Proton to do.
00:09:34
I think there's a lot of other cool things
00:09:35
they could have done,
00:09:36
and I frankly would have just preferred
00:09:37
if they improved some of the integrations
00:09:39
between their current products
00:09:41
and just improved on those.
00:09:42
But this is fine.
00:09:44
I think it's so far fine.
00:09:47
I don't know if I'm going to be using this,
00:09:49
especially because it doesn't really integrate
00:09:51
with anything else too much in the Proton ecosystem.
00:09:54
One thing you need to understand about Lumo,
00:09:56
which is maybe going to make it or break it for you,
00:09:59
is a lot of people use AI as part of their process
00:10:02
when it comes to searching for something.
00:10:04
And Proton doesn't have a search engine,
00:10:06
and this is something you have to actively go to to use,
00:10:09
whereas something like DuckDuckGo's AI agent,
00:10:11
which is something I do find myself
00:10:12
just accidentally stumbling on using pretty often now,
00:10:15
mainly because I'll do a search,
00:10:17
I'll go through a few web articles,
00:10:19
and then I'll go, oh, okay, that was interesting.
00:10:21
I'm curious, like, if I put this information together, like what an LLM will spit out.
00:10:25
And then I'll just go ahead and click.ai.
00:10:28
But that use case isn't quite a thing for Lumo.
00:10:31
Like, Lumo, you're going to have to log in and just go to Lumo just to go ahead and use your LLM.
00:10:36
And at that point, you could have used any of the other ones that probably are better performing,
00:10:41
almost as good privacy, or you could have just opened up your local program,
00:10:44
which probably has better performance as well.
00:10:46
So I think the uphill battle, for me at least,
00:10:48
is going to be understanding where this could even fit
00:10:51
in my workflow without just using it more.
00:10:53
It's gonna be hard for me to know
00:10:54
when this is actually better to use
00:10:56
than some of the other tools.
00:10:57
Either way, I think the major feat that they're happy with,
00:11:00
as well as myself,
00:11:01
is their ability to somehow do end-to-end encryption
00:11:04
with these kind of AI models.
00:11:05
The end-to-end encryption aspect of it is very interesting
00:11:08
and is something that hopefully other providers
00:11:10
can learn from to build even better AI models.
00:11:13
Let me know what you all think below,
00:11:14
if you like it or don't like it.
00:11:15
And I wanna thank our patrons and our techlorians
00:11:17
for making this kind of content possible.
00:11:19
And I'll see you guys next time on Techlore.
