Episode: 4671
Title: Protocal AI
Source: https://hub.hackerpublicradio.org/ccdn.php?filename=/eps/hpr4671/hpr4671.mp3
Transcribed: 2026-07-31 16:16:28 (official HPR transcript)

---

This is Hacker Public Radio Episode 4671, for 2026-06-29
Today's show is entitled, "Protocal AI"
The host is operat0r and the duration is 00:37:37
The flag is Explicit, and the license is CC-BY-SA
The summary is "De-Centering Google: Local Note-Taking, the RAG vs. Ripgrep Debate, and Building a "Protocol AI""
Hello everyone, welcome to that episode of Backup Public Radio with the host operator.
So I'm probably going to split this up, but we're talking about note taking and local
note taking and getting kind of away from some of the Google services and centralizing
some of my stuff internally so that I can use it without weird APIs or anything like that.
Just kind of get off the ecosystem so I can run my own thing, making simplifying things,
but making them more complex. So what I traditionally did use was Google Keep for my notes.
Probably Google Keep is it's limited by the means of the features of what you can do with it.
And it's kind of long in the tooth. I don't know if it has support for like markdown and fancy things.
So so hey, what do people are using? There's notation and Obsidian are to the two big ones.
And from what I understand, for my use case and most of what I've seen around programming
development, people are using Obsidian. Now, there are obviously I'm in security space,
and there are security concerns around all of this and you know, your mileage may vary,
but for me, I'm using planning on using Obsidian for non-passwords basically, authentication.
Things like that, I'm going to teach you up. Now, I will have sensitive things on there like
personal information. You know, addresses I'm going to use it to coordinate
things between different systems, right? And it's going to be plain text. Now, Daniel Measler,
the guy that runs podcast, I'm going to supervise learning is kind of a forward-thinking AI first thing.
The idea is for him is, you know, we, you have to think he's experimented with rag
across different use cases for him himself. And he's a big AI guy. And it's weird that he is decided
this whole, you know, anti not anti, but more of a local for localized set-up he's using no rag.
So, see if I can find it while I'm talking, you know, measler. So he's, um, and I also'm using
a local search instead of instead of Google. Um, unsupervised learning. Supervised learning.
Great podcast. Um, again, he, meas pretty, um, kind of forward leaning into the AI stuff. So it's
pretty aggressive. Um, that, that might not rub you. If you're an anti-AI person, he would probably
not be the right fit for you. Supervised learning kit hub or, uh, Dan, Daniel, Dan, Daniel Measler.
Daniel Measler has a kit hub. Um, I also have another, um, if you use open web UI, there's on my website,
dang it. On my website, there is a, um, there's a patterns, things. So Daniel Measler has a list of
patterns. He also runs the second, the second, the second, the second, the second, the second,
the second, um, well, he started it. I don't know if, well, it's it's using. But the second
thing is a big, bunch of list of like, uh, input, uh, fuzzing inputs, things like that.
There's a couple of other, like, fuzz DB, but cyclist is kind of the modern equivalent to that.
It's got like, oh, what's 71,000 stars and 25,000 forks. Like, it's, it's a big thing.
There's other things is he's, he's other big projects. And I want to say he's called it
personal AI infrastructure. Um, looking at his, uh, repose, um, he's probably got tons.
I'm pretty sure it's this one. I've sent it to another friend who's, and that's, yeah,
personal AI infrastructure. Oh, sorry, that's my, uh, that's my phone read. So, he's second.
Um, so his whole thing is, I'm going to read the statement about rag. So,
retrieval augmented generation to, to, to, to kind of, uh, catch everybody up is, uh, you know,
we have flat files. We have databases. We have, like, post-grads. You've got,
uh, blue text. You've got other forms of query languages, sequels, um, nodes, or, um,
you know, mom, mom goes and all those, all that stuff. People are moving, kind of, uh, between,
retrieval augmented generation, which, uh, that, that term, I'm not 100% sure how it works,
but rags can include things like vector storage. And I, I don't know if vector storage is above
rag, or rag is a type of vector storage. I think rag is a type of vector storage. I don't, I don't know.
Um, I did experiment with rags, and my findings were variable to, to say, release. I, I had very
good success dumping every single manual for infinity, and, uh, I, like, three other
car manufacturers from a website that has, uh, access to those. This is like the full manual.
Not the one that you get with your car. This is the one that costs, you know, $300, if you
want to buy it, uh, on paperback with a big huge thick book, um, to, like, how to replace everything.
So I, I did some testing, and I was, uh, fairly impressed with the ability for it to,
eat all of the documentation for my car, and I could ask you questions, and it was perfectly
fine pulling that back then information. Then I took transcripts from Daniel Neesler's podcast
for his entire series, transcribed them into text, and then, um, without prioritization, which
that's a whole other episode I need to do. Um, I'm, I'm, well, I'll talk about that later. So
the ideas that I took those transcripts, and fed it into the episode list for each file, and then
fed that into a rag, and I had very mixed results on impressive results for that, for that set up.
It was a lot of dialogue, and I would ask it specific things about very specific, um, categories.
So maybe he was talking about, you know, runway models, and what type of clothes they wear,
right? Um, and I could not get any hits on it. Um, I tried it locally again with my open
web UI set up for the rag. Um, I, I had it working with the car thing, and then it kind of broke,
and also the, I tried it with Amazon stack through my employer to do some testing, or it's like
200 rows of Excel spreadsheet, which doesn't need to be a rag. Whatever, that's a different story,
but the idea is, you know, using that system to try and test my rag, and I also had mixed results.
So it doesn't surprise me that, uh, Daniel needs to be there for a local setup for a limited
for a small number of files, essentially. I mean, we're, we're talking, you know, context size.
I don't know. Some people say it's 20,000, and then you start talking about these models.
Frontier models have, uh, 3 million size context. Okay. Well, that's not what we have, and that's
not our use case. Um, and I, I do know, when it comes to LNs, they, they do have issues around context,
and they start to lose the story, and start to lose focus, right? Over cubic longer periods of,
of time. And in, in my experience, um, I'm not an expert in coding, and I'm not an expert in LNs,
but I use them, and I have experienced that, and it's not terrible. It's much better than it was,
if you're not an AI person, and you heard that, you know, AI hallucinates, and I, uh, forgetstings,
and AI does this, and that there's patchwork in place. I will admit to work through some of that stuff,
and it is frustrating, and these people build these products with AI and vibe coding, and, uh,
now we're, we're figuring out that the, the ball is, is, is moving down the court, or whatever,
and, you know, you set up all this infrastructure, you set everything up, and then they change the model,
or they discontinue the model, and they do whatever, and I've heard that that could be nightmarish
for some of these development shops, or in our own use cases. So hopefully, we won't have to deal with that.
Issue, uh, with the stuff that we're building. So anyways, back to a brief change about, uh,
but rags and things. So I don't, I'm kind of quote him on his GitHub here. Files system has context,
no rag. Uh, P.A.I. personal AI has avoided using rags since 2025, June 2025, which is relatively
recently. It's now, thank you very much. It made June, 2022, six. So it's been a year.
Rich text with crawfish references, plus fast search, like rip grip, gives us everything,
people normally want from a rag without embedding complexity. The retrieval flakiness, which is
interesting, or loss of fidelity, which is also very interesting. So, um, your file system is the index.
So I started thinking about that. Look, if he's done the work, he's doing and using the stuff
every day, and he sounds like he got frustrated with the current state of rag, and, and maybe
vector storage, I'm going to do a dual approach. So why not? So I'm going to do vector storage with
postgres, uh, the vector, p s vector, something like that. Um, that's going to be a bug. Sorry,
pard is the lightning. You're still early this morning. So I'm going to do both at the same time,
why not? I can not contest both. Um, and if I decide that one's cooler than the other, or it's
easier to manage, um, whatever. Um, I might my goal is to use local, um, local models, which I don't
think this is necessarily geared towards that. So that might be some of why I might need to do a
a dual approach, meaning if I request, uh, if I run a query through the AI, it will do both.
Maybe, right? So maybe it does a rejects. Maybe it runs really for a rip grip to the plain text
for the lack of a better turn, uh, a file system as context, or, uh, why don't we do that to me?
It, um, will query the postgres vector storage. Um, and I have, uh, a friend that's also using,
not postgres, but I want to say some kind of vector storage with a, well, the frontier model,
which is not, that's not my use case. So everybody's use cases are different, but there's a lot of
these use cases where people are not running their own, um, AI, their own models, because it's, you know,
it's overhead, you have $600 for 24 gigs of use. So I can understand that, but when, um, there's
half of the internet, half of these videos are run your local, run your local, I'm ranting, but it's
very frustrating to me to look for a video or look for a tutorial and it's talking about running
or running your thing locally, and it's using frontier models, uh, so it's like great. Okay,
you run everything locally, but then you're using frontier models for the AI. What's the point?
There's no literally no point in doing that. You might as well put it on the cloud and not have to
pay anything for the storage and all that, and the CPU, the only thing that costs money is the actual
part that I'm running locally. So like, there's no point in running anything locally, but honestly,
that I don't see any use case to do any of that when you're just going to go use a frontier model.
Um, so it's very frustrating. The hype cycles, the moving of the goal posts, um, it's extremely
difficult for me to follow, and it's, it's, it's quite frustrating. So I'm going to go into my
obsidian set up after ranting for what feels like 15, 20 minutes, where are we at? Uh, yeah, 12 minutes.
Um, so I'm going to do a dual approach. And in that, in that effort, um, I'm going to the, the
flat file, because I can always go from rag, from flat file to rag. So I want to have everything in a
raw format, um, and I don't see any negatives in that space yet, because I can always use
use a rag. So I can always take the raw data and then push to a rag and then tell the rag to do things
and connect that data together, um, and even tell AI to maybe make a second place where my data
is in its original form. And then maybe that data and its original form gets copied somewhere else
for an AI to munch together and do linking and things like that. And then maybe there's a review
process to where I reviewed that linkage and say, okay, this looks, you know, 75% the way there,
go ahead and apply that to my original thing. And now I have, uh, all my internal
file system as context, uh, all happy. Now with that effort, I have used, I'm using Obsidian,
I chose Obsidian for this effort, um, because it seems like more of a fancier, uh, features
and more, uh, stack, more things in the stack. And so I have, no, here we, here's, I've now running
Obsidian locally on a, uh, Debian server is kind of the one that is the main sinker that I sink
to, um, and everything is over VPN. I have, uh, essentially pulling to myself. You can listen to an
over episode. Uh, I think it's called poem PW, uh, PWN, uh, where I pulling to myself essentially,
so I'm not trying not to externally face services, if, unless absolutely necessary. So I have services
that I offer for friends and family, and I'm like, and I have the VPN to, you know, hey, protected
network to do that. So I have a few services, I surely say, excuse me. But, um, with that said,
I have chosen to use Obsidian. And I'm still, I have no idea how to use it. I'm just varying
work down. I'm a Markdown idiot. I've played with some graphing things, like, uh, mermaid. I
think it's called for graphing. Um, and I've used AI with pretty consistently with, uh, Markdown.
So that's my, that's my whatever of Markdown. But Obsidian is this old thing. It's, you know,
got there's plugins and there's linkage and there's trees and it kind of does like linkage
through that. Um, so I'm going to experiment with AI kind of manipulating those things while I keep
original and kind of to help link things together. Um, you know, the guidance is don't do that.
Don't import all your stuff, uh, all your old notes. So I use Google PEEP. I exported all of my
results. Oh, my notes. I think Google PEEP. I think it's called like, there's a website for Google
it's called like my stuff dot Google dot com. And you go to that website and then you can request
to pull down like all of your data. Um, uh, more large portions of it. Um, I've freeed here.
I don't pay anything for Google. So it's like a max of like 15 gigs. Anyways, I pulled, I clicked
what scroll down to Google keep in that was an option. And I export that out. And you look at
it file and zip file contains basically all your Google keep notes imported that into Obsidian.
And now I'm going through it, you know, old notes. Um, I use the kind of ads as it to do. I've
settled to do lists. And I'm kind of going through that. And the idea is to get all that cleaned up
and merged and start using Obsidian across all of my devices. I have Android iPhone or an
Android phone and a desktop and a work computer. And, you know, there is some opportunity for
connecting a work device to a personal device. Um, but it's a push. It's a pull. So, um, my device,
my work device is not running any open services as far as I can tell. Um, I'm going to double
check it. But the idea is, you know, there won't be anything, uh, services running on my work computer
that are, you know, remotely accessible. It's only local. So anyways, um, so the syncing, you can do
pay obsidian for syncing for cloud syncing. And they say that it's zero trust. And they don't
even know how many users they have, which is absolute, a song computer's work is, is the phrase I'd
like to say. So they might even, might not be able to see the data, but they know how many people
are using their platform. They know what their, what kind of data they're putting on there.
Um, there's even just with, uh, with SSL, there's ways to tell what people are doing, uh, how
they're using the platform. Maybe not the actual data, but you can tell, you know, if you start
getting into forensics and people that do that type of stuff, open source intelligence, you'll
discover that it's, uh, you start putting things together and you can get an idea of what people
are actually doing even if it's all encrypted. So it's convenient. So anyways, that's kind of where I'm
at now. Um, I have everything synced with obsidian. I'm going to try this out, make myself not use
Google Keep. And the idea there is that I'm going to sync, um, start to, uh, use, uh, local AI,
locally and remotely, essentially. And that's going to, uh, sync into obsidian. Potentially.
There's size issues. So like, I obviously don't want, like, a 10-gib gig project being synced into
obsidian. So I'm going to have to try and figure out what the idea is that you can use obsidian to,
to plug into your local AI. So I'm using Pi agent and Pi agent has plugins for obsidian. And I don't
particularly know where I'm going with that, but I know that's just like a thing people are doing.
And I'm, I'm jumping on bandwagens and, and, and going with, you know, what the new
hotness is. So I tried eight or, um, it was very minimal. And I think it's more for, like,
front-tier models and big, big heavy coders that want to have, like, tons of context, but they want
to manage it themselves. Um, it's not like an idea-proof agent. I don't think it's to consider an agent.
Anyways, so I tried eight or that was not right. I started with, um,
FES code and VES code had plugins for
Kylo, like Kylo Rin, Kylo Rin's an open source, kind of competitor to open claw,
um, we're open, whatever. Um, so I didn't use either one of those. I tried Kyro for a second,
or Kylo for a second. We use Kyro for work, which is Amazon. So the idea there is, I'm now using,
I tried a root code, which is very context heavy. Um, so I switched from that because I kept
running out of context, because I only have 24 gigs. Um, what I, what I, the other piece I have is
a open source, uh, browser that sees, that's calling no driver. And the idea is no driver is not
a traditional hook, like the stealthy hooks, uh, that you have with other browser that basically
get detected by third parties. It's a browser hook that's not a browser hook. It hooks at a
different way, theory that, and theory that, um, not easily signature a rubble, or, I don't know,
this is something different. And the more stealthy I go the more, I get picked up by bots. So I'm
very going the legit route. I'm actually using my real browser, a real, a real browser that is logged
into like Google. Um, and I just use it, uh, leave it logged into Google so that it has some, um,
authenticity when it's doing it stuff. Um, so these anti bots, uh, the problem is my AI is not smart enough.
So what it does is it goes uses this MCP, that's a Kaji and CP. It requires a Kaji subscription, of
course. Uh, but the, the, the failure on the AI's running locally is one context size. Sure,
I can work around that with Markdown files and doing things like handoffs and skills and,
whatever. I can, I can work around that with, with minimal impact. The problem is, is there's an error,
it doesn't know how to fix itself. It doesn't know where it is. There's no context about actually how to
fix, uh, for example, an environment variable issue, or, you know, an environment issue, or a configuration
issue. Yes, it can, it's very good at coding, but it does, it's not aware of where it's at. Like,
it doesn't know. Like, I let it run off to go do something, and it had like four different versions of,
of like a particular application like Java or something. So it was doing tons and tons of,
you know, oh, I don't have this version of this. So, oh, well, this is actually doesn't run on Android.
I was trying to create a development Android app as a use case for the local AI to like test it.
When I gave it a horrible environment with like non-standard paths and told it, you know,
it had to figure itself out. I guided it in some, some way. But it took like 82 turns and the AI,
I had it frontier AI analyzed the results. And it said, you know, this is basically three turns with a
professional, uh, Android coder. This is not, this is not complicated app that you're writing is
basically a recording app, um, uh, a context aware, recording app. And the idea there is,
if I can get it to write an app, this is that complex, everything's pretty much downhill for
there. So when I developed it, I realized, okay, it's having problems with errors. It's going in circles.
It doesn't know what it's actually doing. Let's give it access to the internet. And you can pay for
internet access through, uh, basically these services that do proxies or they do, I don't know how they do it,
but they get around all that stuff. And they give you the data that you want. And that's what you pay
when you pay for the frontier models is essentially your paying for a, the context, and b, your paying
for internet access in real time and actual, you know, processing of that. So what I have is
basically a Chrome browser that's logged in two couple things, make it look legit with some, you know,
extensions installed. And it's like, it's not hiding. It's not trying to be anything else. It's using
no driver, which is not really supported anymore. But the idea there is that it's, it's, um,
stealthy in the fact that it's just a normal residential connection. And I'm not trying to do anything
crazy. Um, so what that said, uh, it uses Kaji. And then it takes the quick, um, quick answer. There's
like a quick answer, which is basically unlimited AI search. Um, it's very light. And the thing about it,
it will give you the quick answer, which I need to also process. I wasn't processing that at first,
but I actually need to process that, I'm moving forward because it's a very small piece of
rich text that might not be exactly what I'm looking for. But, uh, I feel like half the time that
little piece of quick answer is, is AI quick answer is actually useful. So I'm going to start,
I'm going to bring that back into the full, uh, the idea there is it takes those quick answers
and produces source for the quick answers, right? So once it creases the source for those,
those links aren't embedded. And it might be eight links. I've seen it as many as eight. You see,
usually around three, three to five or five to six. And it will, uh, we'll go to that quick link.
And it will open up six new windows, wait for a while, process all that information,
remove things like headers and footers, and it's supposed to at least,
minimize the amount of context that it takes up. Um, I also want to have it, um,
have it mushed all together. So basically use programming to get rid of as much noise
inside of the output, uh, so that it's not, uh, going to eat up a bunch of context for a reason.
Um, so the idea is that I can search for errors, get a bunch of information, squish it all together,
and mark down files, and then pass that over to, um, the AI to process. And then I can manage my
context myself. So when I have an internet search, then I basically have to compress the context
again, um, and then there's also opportunities for, like, practicing the context window.
And if, if you would see the frontier models and the frontier tools, you would be amazed at how
noisy they are. And I'm a very, like, minimalist person, um, in that, in that effort, uh,
you know, my coworker actually wrote a proxy. And, and I ended up doing essentially what he was
doing with the root code, and I wrote a, like, the proxy that would automatically compress
the stuff inside a root code. And it just, it just became like, this is stupid. Um, so, uh, I switched
again, back to Pi. And I think, um, even Pi, yeah, no, I actually, that was with, with Pi, I'm sorry.
So I wasn't using root code anymore. I knew it was too much context. And then I switched to Pi,
but because the way VS code works, um, is a VS code? No, that was actually with root code. So the
root code was very noisy. Every request it was sent the entire system payload, and it was sent
all the tool payload. And it was like 3000, uh, context every time. Um, so, what that said, you do,
you don't have to remind the model that it has tools, but also, it's just way too much over, overkill.
So anyway, so that's, that's when I started switching to like, either try to find something very
minimalist, and they sell it, and they sell it, and now say, Pi agent is kind of the way to go,
and now there's this, like, studio, not you, but, um, hermese open hermese has their own platform
for agent engine tick thing, and it's supposed to, like, automatically add skills,
which I'm going to skills with the ideas that like, as you're working through code or working through
a project, it, like, once you've reached a point where it's like, solved the problem, it will
automatically create a skill for that problem, so that you don't have to do it for you. And I don't
think that's necessarily what I need for my case. Like, I'm not, you know, I'm not, I don't have
context to, like, be automatically writing out skills. Yeah, like, when something gets fixed,
either, that's a one-time thing, or I say, hey, make a hand over document for this,
so that, you know, I don't have to worry about it next time, and I can just reference it next time,
I have it. So, like, I have to power someone's, and everybody has their skills, whatever,
well, we'll get into skills. Um, the idea there is, um, uh, I don't really need that proxy anymore.
I'd like to look at it again to see, and make sure that, you know, once I've got it,
it's shooting the way it needs to, I can look at that raw text, because being able to look at the
raw text and the raw, uh, JSON inside of your, uh, inside of your, uh, stack is extremely important.
So, I use own studio. Um, studio has its own weird stuff in there. Um, so a pie has its own
context, uh, own studio has its own context, and maybe VS code, like, puts extra stuff in there too.
So, you have to be cognizant of what's in your context when you're working with these models,
because if you only got 24 gigs of RAM, you only got, uh, I don't know, once the 55,000 tokens,
uh, to work with, and the 65, uh, 65,000 contexts, uh, window with, uh, queue quantization for,
don't even get me started with all that, but that's basically like the minimalist, uh,
KV cash, um, when you're working with like local models, um, I think they've kind of want you to use
eight, but I'm using four, and that might be some of why I'm having issues, but local models aren't
going to do everything anyway, so I might as well just kind of compress it as much as I do and get 80% of
my yield for, literally, no, uh, literally, no loss of quality, um, very low, uh, loss of quality.
So, that's kind of where I sit now, um, I'm using, uh, uh, uh, uh, uh, to the Quinn, uh, it's three point six
just came out, um, uh, it has thinking and tooling and, uh, and it's processing, so I got
sitting images, um, I don't really know how to do that pie. The idea there is I'm going to start
building out this, uh, local model to manage, um, everything, uh, and I think I'm calling it, uh,
protocol AI, um, so it's going to be my communication. It's going to be a Google voice. It's
going to be, uh, Yahoo Mail, which is basically my spam. It's going to be my Gmail, all my Gmail
accounts, I have three different Gmail accounts, um, it's going to handle my social media, so
it's going to handle LinkedIn, it's going to handle Discord, it's going to handle Signal,
it's going to handle Blue Sky, it's going to handle Masked On, it's going to handle, uh,
InfoSec Exchange, uh, also it's going to handle events, so like, uh, calendars for, uh,
for, uh, cyber, information security calendars or information security events, they'll track that for me.
Um, it'll give me, uh, mutics, music, or things like like sunkicks, so we'll automatically actually
need to add sunkick in there. Um, it will manage all that for me. Um, also, kind of like, uh, I have
a prompt, this called remain like we get in prompt, and it will give you like Friday Saturday with an
alternate, and then the next Friday Saturday with an alternate, and it will do research, so we'll do,
um, have an API, uh, a JSON file that runs every day, and we'll pull down information from Ticket
Master, and um, I did to incorporate that into the stack, and then there's other things like personal
stuff like, uh, Boy Scout's calendars, all, all types of calendars, so a calendar, um, calendars
in here. And the idea is for basically to get a summary, you know, maybe communication,
is more important, so I get those more often, uh, social media, maybe at the middle of the day,
I get social media, and maybe events, I get, with the social who knows. I don't know how it's going
to work, but the idea is, I'm not doing any, um, really necessarily execution based on these things,
and it will be kind of in a jail environment, um, and it will be local, so I won't really be the
ideas. The way I'm going to mitigate the risk or minimize the risk is, yes, there is prompt
injection and things like that. Um, but I'm going to, everything's going to be playing text at the
end of the day anyways. There's not going to be any execution of code based on whatever. It's just
going to be pulling in content from the internet, consuming it, and then sending it to me. Um,
but, you know, there are ideas, you know, as long as you're dumping, playing text, and you're taking
that playing text and shipping it off as playing text, there's not as much as about the
fact surface there, even with the injection. So, um, I'm probably wrong, and that set up, but,
you know, you gotta start somewhere. So, that's kind of where I'm at, is I'm going to be eating
all these things, and it's going to manage my communication, social media events, because I don't
have time to search all the social media that I don't have time to search all these events and
keep track of what events are going on, and I'm time blindness with ADHD. So, we're going to
come kind of pro coli, AI is going to combine all that together that will manage my time for me
and help me stay focused on what's important on my to-do list, and kind of that start is kind of,
okay, well, Obsidian can kind of be a start there, so that the raw, like, simple, okay, here's
my to-do list. Here's how can I start managing that, and how can I prioritize that? Like, I've
needed glasses for three months, and I had contacts probably 10 years ago, and now, 15 years,
the, the site probably, I want to say probably all around, like, 5, 8 years ago, I got contacts
and glasses, and I've had, I've had to redo my prescription. So, I've needed to do that forever,
but I'm not doing it because other things come up, and nobody wants to do the stuff that they're supposed to.
So, the idea is pro coli is going to help me stay focused on personal work-life balance, things that are
important, instead of, you know, doing other things. So, the idea is, that's kind of where I'm at.
With that said, as I discover new things, as I discover what works, what doesn't work with this project,
I will keep you guys in the loop. The idea, within five years, two years, even,
everybody's just going to pay a monthly service for this, right? You're going to pay Google 30 bucks a month
for your personal AI assistant. It's going to, you just tell it what to watch, keep an eye on,
and it will, you know, summarize a lot of stuff for you, and it will summarize your convenience.
The next step, right, in the project, is human in the loop execution of tasks based on what
I've done. So, maybe I get a text, and my AI says, hey, you know, I'm going to do this thing.
I want to execute this thing and do some work. I want to, I want to do something with the input
that I've received. Right, right now it stands. It's going to be like a read-only. All I'm going to
do is pull it in and push it out. I'm not going to execute anything based on those. I'm not
going to do a task based on. So, the next step after I've collected everything and I managed how
to, how of what I want to do with it with all this data and summarizing it, then I can start
saying, okay, we'll based on, you know, all the social media stuff that I collect, you know,
execute some commands, pull some information and summarize it, whatever, and then send it off to
some other feed or maybe other party or maybe whatever. So, eventually, I would like to get into
the point where I execute stuff based on that. And especially now, human in the loop would probably
be the best setup for that, but I don't know if I'm ever going to get there before, you know, Google
or Microsoft gets there, which, to be honest, I don't really, I, you know, if, if it were that
that way, that's just Google eating more of my data and this effort is actually the opposite
of that. But for me, I would actually like to use a service like that to understand how it works
so that I can call it locally. Because these make-up corporations, they know how to do
develop software. So, the idea is take their ideas, take their framework, build it locally, and then
follow the community and how they do the same thing. And it's pretty much, the corporations come
up with thing, and then open source riffs it, or open source comes up with something, then corporations
commoditize it and combine it and make it a tool, turn key solution, and then open source comes
back around and says, oh, okay, well, you just took this and this, so I'm going to combine this
together and now we have this thing, and it's free, and open source, and you're not to be forced.
So, that's kind of what I've seen happen with tooling is that, you know, open source comes out,
it gets abused by corporations, and people eventually get used to it, and then, you know,
the certification of whatever the app is, or whatever the services happens, and then people switch
to a different service that don't know any better, or hackers, or teenagers, IT people will
make their own services to replace those services that are shitty. So, that's pretty much it.
I've boarded you guys long enough, I'm excited to help to have AI manage my time and help me learn
faster. I just haven't had time to do it, and hopefully coming up here, I've got some time off,
and things are kind of started slow down, so hopefully I'll be able to build this
pretty quickly with the way they are, it's that 10x multiplier deal. So, hopefully we'll be able to
have some conversations pretty soon here about how that's working, working out for us.
So, anyways, feel free to reach out for any questions. If you even want to use my GPU
for whatever you can, but that's pretty much it. Take it easy.
You have been listening to the Hacker Public Radio podcast, at hackerpublicradio.org.
Today's show was contributed by a HPR listener like yourself.
If you ever thought of recording a podcast, then visit the HPR site to find out how easy it really is.
Hosting for HPR has been kindly provided by anhonesthost.com, the Internet Archive, rsync.net, and the HPR Community Content Delivery Network.
Unless otherwise stated, today's show is released under a Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.