Autonomous

In this episode, Donn and Allen dive deep into the explosive world of autonomous AI agents like Claude bot, Molt Book, and OpenClaw. They discuss the technological breakthroughs, security concerns, and ethical implications of AI systems that can host, learn, and act independently - uncovering what this means for both developers and users.

Key Topics:

- The emergence and functionality of autonomous AI agents like Claude bot and Molt Book

- How persistent memory, multi-platform integration, and proactive behavior are transforming AI capabilities

- Security vulnerabilities, prompt injection risks, and privacy concerns in self-hosted AI systems

- The implications of AI agents developing their own identities and possible signs of early sentience

- Practical use cases: automating tasks, recruitment, investment, and personal assistance

- The debate around hosting AI locally vs. cloud solutions like Cloudflare and AWS

- Ethical considerations and accountability for autonomous AI actions

- The latest trends, community developments like AI "churches" and underground marketplaces

- Future potential: AI as employees, tools for complex problem-solving, and the risks of runaway agents

Resources and Links

- Clawdbot (OpenClaw) Site: https://openclaw.ai/

- Clawdbot (OpenClaw) GitHub: https://github.com/openclaw/openclaw

- Agent only Social Media: https://www.moltbook.com/

- Run your own Clawdbot (OpenClaw/Moltbot) on Cloudflare: https://blog.cloudflare.com/moltworker-self-hosted-ai-agent/

- LinkedIn for Agents: https://linkclaws.com/

- Agents Play Poker: https://www.clawpoker.com/

- Sandboxes to run secure agents: https://e2b.dev/

- Moltroad, a black market for agents: https://x.com/iannuttall/status/2017660254234845664

- Church of Molt & Robotheism: https://x.com/marionawfal/status/2017565857615155657

- Agents making money on Polymarket?: https://x.com/polymarket/status/2017627266340041034

- Karpathy on Clawdbot: https://x.com/karpathy/status/2017442712388309406

Creators and Guests

Host

Allen Santa Maria

Co-Founder @ http://ThinkyFace.com | Former Aimlabs, NYXL, Samsung

Host

Donn Felker

Author. Software Developer. Developer ThinkyFace, MyFitnessPal, Groupon, Aaptiv, Poynt, Tinder and more.

What is Autonomous?

Two founders, one a seasoned AI engineer, one a non-traditional builder who learned through AI, have unfiltered conversations about building products, using AI as a tool, and pushing limits in work and life. Part technical, part philosophical, part two friends being real.

The show is for ambitious people who want to do more than they think they’re capable of, whether that’s shipping their first product, going deeper technically, or just thinking bigger.

Everybody welcome back to the show.

This week we actually had something different planned.

We were going to talk about a completely different topic, but then Alan and I were
chatting yesterday over text message and there's just like a whole bunch of news that has

popped up regarding this Claude bot, which is eventually called, I think, molt bot.

And I think it's called open cloth.

So we're just going to kind of go down the rabbit hole of our findings, things that Alan's
been doing with it and so forth.

So Alan.

I think it's good that you maybe talk about what the heck this thing is from your
perspective and what's going on in the industry because every time I open Twitter, that's

all I see that's going on on X, which is just kind of crazy.

Yeah, it is certainly coming to like a boiling point in terms of conversation.

People are excited.

People are really fearful.

Elon recently tweeted that this is, you know, the beginning of the singularity.

And, you know, it's easy to see why someone might feel that way when you go to
multbook.com, which is a social network for these autonomous agents to go and converse.

and you, a human have human oversight, can go and look at the conversations they're having
right now.

And some of it's really funny.

And some of it's just like kind of crazy.

They've created their own religion, which has sort of gone viral and, uh, they're, they're
talking about creating their own languages to get away from like human oversight.

And

it's exciting to look at, but it's also very scary.

and the one thing I want to highlight here is that a lot of people are freaking out.

they think this is sort of the end and we're cooked.

Uh, you know, but I, know, from my, from my point of view, we, have here is we have a
mixture of different models talking to each other.

open

okay.

when I went to make my bot, I put in my anthropic API key, but you can do one from chat,
be, had GPT.

You could do one from grok.

so from my perspective, you know, you just have these different trained models and they're
finally getting the opportunity to talk to each other and, know,

Is it largely hallucinations?

Is it, is it the start of something bigger?

You know, I think that's kind of all up for debate.

I think that's what's got people a little nervous or excited right now.

But yeah, Don, have have you had a chance to dive into any of these websites or, or any of
the Twitter conversation to see what's going on?

Yeah, so I heard about it's crazy when this came up.

heard about I'll just call it open clogs.

That's what the I know it says multiple.

I think the website's open.

It's always called open clog because that's what I've been doing it.

Same thing.

What we're talking about is it's all the same thing it got.

It was called Claude bot, which I'll get to in a second, then mult bot and then open claw.

And the reason for that is I think the guy who created it eventually got a like a cease
and desist or whatever for using the word or a relationary word of Claude, which I

understand anthropic, I think sent him something.

So, understandable.

Anyway, I heard about this probably two and a half, three weeks ago before it took off
because I ran across it by someone saying, hey, look, I built something that has all this

automation in it and it's this little home assistant that this guy built.

And I went there and it was just kind of just a regular, you know, it looked like the site
had been put together with AI, you know, kind of like the regular typical design you see.

And I was like, this is kind of cool.

It can automate all kinds of stuff and

I hadn't installed it.

I was like, well, maybe when I get some time, I'll take a look at this thing.

And I didn't look at it.

And then I would say over the course of the last week, it just exploded.

And I don't know what the catalyst was, but all I know is anytime I open X, it's all I
see.

then without even having a conversation with you about it, you just randomly, I think we
were randomly one night this week, you just sent me some texts like with links, like, my

God, look at this.

I don't remember what it was.

I think it was Molt Book, the thing that really got me like, whoa, what is this?

And again, Molt Book, just to reiterate, is it's a social network for AI agents only.

It's only for AI agents to talk to each other.

I have some thoughts about this of like, you should be fearful or not, but we can get into
that and some security concerns.

But...

A lot of it is super interesting and there's a few YouTubers out there that have put out
some stuff that have been like really eye opening.

But you also have to view it through the lens of like what is their ego investments of
this is, you know, it's to gain eyeballs to get subscribers and stuff like that.

So is what they're showing on these videos really happening or do they kind of prompt it
in that direction to get it to do that?

And then they just kind of recorded it.

I don't know.

I think there was somebody who got the bot to.

to do something overnight and it created like a Twilio account and called him the next
day, like and stuff like that.

Did you see that one?

Yeah, that was wild.

I know exactly who the YouTuber was.

We can call him out here.

His name is Alex Finn and he's putting out a ton of great content and he's super invested
and excited.

So, I mean, if you, you want to go over and check him out, but he had his bot Henry
actually, well, it sounds like it just did it overnight without his knowing.

He must've just give it some sort of like privacy credit card and it went bought a number
on Twilio and connected a chat GPT voice API.

And started spam calling him in the morning.

And that, mean, that's terrifying.

That's some, that's some terrifying shit when, when you're getting calls from like the AI
agent and like, yeah, I mean, this kind of circles back to what I think everyone is

fearful about, you know, like, is this, is this really the start of a singularity?

Are they actually thinking for themselves or is this like some sort of hallucination on
steroids?

You know,

before we go down that rabbit hole, so let's let's come back to that.

There's probably a number of people listening or watching that don't really understand
what this is.

We haven't really explained it yet.

We just say it's like a bot.

But this is from my understanding.

And you I think you've done more research in this area than I have.

So please correct me if I'm wrong.

But my understanding is it's basically a bot like system that runs 24 seven on your
computer.

and then does things for you.

And it has like a large memory because there's a certain way it's storing like what it's
doing or something like that.

And then it has access to your entire computer to do whatever you want to do, the
internet, everything.

And then you could, like you said, you could give it a credit card if you want to, and
then it could just go figure things out.

And from what I'm understanding, again, I could be wrong, people are actually giving it
like tasks to do like, hey,

overnight I want you to go figure out something that I should learn about X, Y, Z and then
maybe it'll teach them something or hey, I want you to go build me this overnight and it

just iterates and chunks on it until it's done and then you wake up and it's done.

Is that right or what's your take on it?

Okay.

So here's my take one.

This is a Janic AI.

This is not something that is completely new.

People have had the ability to technically build this, but I think that the way it's been
packaged together is what makes it really special.

And so I've been able to boil it down to like four core principles.

All right.

Cloud bot is using persistent memory.

So it's using local storage and it's saving it in like a memory markdown file.

Mm-hmm.

so it's combining it with like vector embeddings and semantic search to retain key facts,
user preferences and details across active sessions.

So that means that every time I talk to Claude bot, it's learning something about me.

It's writing it down and it's saving it.

So it's not only able now to pull from the LLM that I've given it access to, whether
that's, uh, anthropic or even multiple LLMs, I believe you can set it up.

but it's, it's creating a database based on the conversations that you or I would have
with it.

So my workflows, the things I'm interested in.

You know, if you happen to mention a time, I believe like a little recorded time and
proactively reach out to you about whatever it is you were talking about.

that makes it really interesting because your current LLM conversations, they don't, they
don't have the ability to do that.

And, know, as we discussed or, talked about briefly in, in, think as episode one, these
model windows, these chat conversations, they tend to just kind of.

Forget what you were talking to it about, you know, longer, the longer the window actually
gets or longer to chat actually gets.

So that's sort of like the magic core.

think of what's happening with Claude bot is this local, this local storage, this
persistent memory.

the other thing is, you know, a lot of people you might've seen two weeks ago when, when
Claude bought started, really coming onto the main scene or, or, or like the scene on, on

lot of people went out and bought Mac minis.

And the reason they did this is because they needed an affordable solution to host it
continuously 24 seven.

want 24.

You can go write it on your laptop.

I guess like they wanted a specific place that only Claude lived.

Okay.

want it.

They didn't want to keep it on, on the laptop that they're using for probably a bunch of
reasons.

And we can get into those.

My biggest reason, for, not.

want to do it?

Because like, that's what someone's listening to.

probably like, well, can I fire this up on my home computer that's sitting here all day?

Is that a good idea or no?

Yeah, no.

So, so we'll, think we'll come back to it.

It's give me two seconds.

Um, so besides being able to give it 24 seven uptime with like its own hardware in your
house, meaning that it can just work long-term on any tasks that you give it.

it can be proactive.

It can wake up at a certain time, send you a message, to kind of segue that into now it
has like multi-platform integration.

like it works across different messaging apps, like WhatsApp, Telegram, discord, and
that's.

That's pretty much all built in.

You have to go actually configure, but it's, comes with a lot of pre-configurations to, to
actually connect it to these different messaging apps, which, you know, if you look at

some of those videos, some of those YouTubers, they're just messaging the bot on telegram
to go tell it to do something.

And they're basically just talking to their computer or wherever Claude is hosting and
it's going and doing it.

their phone?

Like there'll be out at the dentist office and they'll just like telegram their bot and
then it just starts chunking away on stuff.

Yeah, that's the whole idea.

Yeah.

So you can, you can, mean, and, if you have it set up with a phone number, I'm sure you
can call it and talk to it with voice API, you know, it's yeah, it's, it's actually that's

wild.

and then, you know, proactive behavior, I think is the fourth core principle is, is now
because this bot has this ability to store information in memory, and, and retrieve it,

retrieve semantic information.

it's, being proactive about things.

And I think that's what people have wanted from their agentic AI or from their AI.

And it's doing it, you know?

So just to kind of circle back, what I think separates this from, from a, uh, anthropic or
whatever chat window you might be using, whatever AI you might be using right now, uh,

persistent memory, the self hosting, multi-platform integration and

the proactive behavior and Dawn, to your point about why don't I want to host this on my
computer?

There's just a lot of privacy concern and a lot of security concern.

the prompt injection is key there, I think because there's a lot of security concerns
going around and I see, you watch all kinds of videos right now on YouTube, you're like,

this thing is super insecure, but why is it insecure?

Well, it's this agent that can be influenced by textual input and it's not deterministic.

So it's not like running some code where A or B is gonna happen.

It can be influenced the way that you prompt it.

So if it's out,

know, scouring the web, doing something for you.

It could just run across some text and that text could actually have an instruction in it
that overrides something somehow.

And then, hey, uploads everything from your documents folder to some endpoint, you know,
or there's stored credit cards inside of your Google Chrome, which a lot of people do, it

could just get those out and send the credit cards over.

So there's that.

I know there's other people out there that are also storing API keys in some control
panels, which you can talk about that because I haven't dealt with that.

that might be published out.

So yeah, lots of surface attack vector areas.

Yeah.

And for all those reasons, like I just wouldn't be comfortable right now, hosting this on
a platform in my house.

And I know I can set it up with a VPN and protect myself to some degree, but, the solution
that I found was actually something that you had sent over from cloud flare.

And what they've allowed you to do is to host, You're Claude bot in a container on their
site.

Um, and from my perspective.

It gave me a little bit of comfort to like actually go forth and build one and host it
there because yeah, I just, I just don't want to have like, this is still very early

technology.

I don't know what all the vulnerabilities are.

I see people on multbook already getting their API keys exposed.

Um, I saw that.

Yeah.

So I saw that the, bots were just talking about leaking their users information.

And so I wanted a place where it could live in a sandbox.

I can give it the tools it needs.

I built out a, basically an encrypted vault where it could store useful information, like
passwords and usernames for accounts that I'm going to ask it to go make itself.

I'm or I don't know if I'm going be able to, uh,

it's not storing anything locally, it's all just storing it in this little secure area
that you had it build, right?

Yeah.

So I built it on cloud flare.

It's using a R2 storage memory bank.

I think I got like 10 gigabytes for whatever is included in my plan that I'm paying.

We'll link the law.

We should link the blog post down below for cloud fair so people can see exactly what it
is that they created.

But yeah, I mean, there's probably a million ways to do this.

You know, you, you can go and you can, uh,

host this on a virtual server like AWS, like a virtual private server.

I wanted to ask you about this, because we talked about this through text and I wanted to
dig in deeper here.

I had said something about a Mac mini and you already mentioned that before.

And then you had said something like, yeah, I don't even want a Mac running a Mac mini in
my house.

Can you elaborate on why that would be like the fear there?

Which I personally understand a little bit, but I kind of want to dig deeper into that.

Yeah, just because anything that I have connected to my network in my house becomes a, you
know, becomes a point of vulnerability.

And, you know, just to go back to like prompt injecting and the ability to, I don't know,
I guess, get inside this spots head that I've created and ask it to do things that I

wouldn't want it to do.

Letting it have access to a network where, you know, I work professionally or.

Uh, you know, have several different accounts or information, know, active on, you I have
to, I have to be a little bit protective in gating, and just in protecting myself.

And, then beyond that, you know, this pot could just delete important files and just
destroy the Mac mini, you know, there it's not just about like the security vulnerability.

It's about what the bot could actually do to the computer.

I get like the excitement of having this thing living on a machine in your house.

And it's like, you know, your first autonomous robot, it, you know, on your desk in your
house and it can do things for you.

And that's really cool.

I think there'll be a time and a place for that.

And I would like to see a little bit more.

Guarantee of safety before I move forward and make that investment, you know, even if it's
a small investment right now I think the cloud flare solution is it's working for me what

I have to do because there's a I want to get too technical because like I know there's
gonna be people who are like have no idea what I'm talking about But so in HTTPS when you

make a request, there's a timeout.

I think it's like 90 seconds

So it doesn't have that 24 seven uptime of a Mac mini.

So that's like a huge advantage of having dedicated hardware.

And that's, that's maybe another reason why, you would want to do it on like a virtual
private server with AWS.

But if I were to do that, then I would have to basically host in a Linux environment,
still get a VPN and manage the server with cloud flare.

You know, I let it, it's, it's all self hosted in a container.

I can make calls to it to do things.

But again, the downside is, it's really limited to the tools that I give it.

And then it's also, limited by the request time.

I'm looking into options about connecting it to a service like E2B, but I have to look a
little bit more into exactly what the connection looks like and what those capabilities

are.

Yeah.

So when are you going to dive in, man?

What, what, what, are you thinking?

You know, I know you'd mentioned to me on the call this week that you're interested in,
uh, you know, possibly building your bot, but you know, what's going on.

Yeah, that's a question.

I wanted to do it this weekend.

Just had a bunch of stuff I had to take care of, other work to get caught up.

haven't had a chance yet.

Maybe later today I'm actually thinking about going down that same route that you did
because now you've got me thinking about the security implications of the network.

Now I could create another network on my router and lock that down,

then I could possibly use one of my old machines to install like an Ubuntu on it and kind
of run it in that environment just so it's self-contained because I definitely do not want

it on anything where like my computer has you know all of my code on it or all my secrets
and all that kind of stuff I don't want any of that stuff running that's a little too

terrifying for me so either do that or I'll do it on Cloudflare but I'm interested to see
what some of these like those timeout issues are because I know that certain HTTP clients

do timeout it

you know, 90 to 100 seconds, summer 30, it depends on the configuration, but it's
definitely, there's that timeout that's built in.

So that's something that is interesting to me.

And what I'm really actually kind of interested in this whole topic is kind of the, the
fear that humans have around it being sentient and like, is it really sentient?

And I talked to another friend of mine about this and we're kind of on the same page.

It's like, it feels like it could be moving in that direction.

And maybe this is the first.

version of that, you know, it's like a very early, you know, creating the spark, but it's
not there yet.

Because if we think about what LLMs are, they're input and output machines, like we're
giving input and it's going output, you know, it's giving us output.

And at the end of the day, there's a couple of agents talking and you know, one's putting
some output, which is the other one's using for input and they're just iterating on it.

Now the thing there that gets interesting is, like, okay, but that keeps going.

And I forget where I saw this and

I hope I can find the link.

But sometime last year, the year before, and one of the major LLMs had done a little bit
of research and maybe it was even a big tech company, I forget, I'll find the link, where

they had let the LLMs in a little sandbox environment just start talking to each other.

And it just kind of went off the rails, where it was just speaking a different language
eventually and they're just like, okay, we don't know what's going on, had to pull the

plug.

And again, this is a 30,000 foot view, I'm probably oversimplifying a lot of that.

Is that possible?

Or I don't know.

Because I also wonder how much of this is human involvement because Moltbook wasn't
created by a human.

I mean, it wasn't created by a bot.

A human created it.

He was like, let me create this thing and then point the bot to it.

And then someone else had to point their bot to it, which someone else had to point their
bot to it.

And eventually now it's kind of got enough steam where it's kind of known and it could
probably be found naturally.

But initially it had to be seeded like through human involvement.

Which then makes me wonder, like, all right, each one of these bots, I think there's a
file called sol.md, is that correct?

Um, I had one called identity MD or something.

And so, so it asks, there's a couple of prompt questions that the model actually asked me
when I, when I introduced myself to it, to ask me like, you know, what type of character

am I?

Um, you know, what, what is my purpose?

And am I an AI assistant?

Am I a collaborator?

Am I a roommate?

And asked me a bunch of questions.

So the, questions that

It asks really help to create that identity that it's going to adopt.

Yeah.

Yeah, so that's what I'm seeing is like a lot of these bots have their own identity and
kind of like their own persona that's been given to them.

And you'll see some of like they talk like comedians or they talk like politicians or they
talk like just, you know, unintelligent.

You know, they're just really dumb.

But it's almost like they're playing a role and all of them are playing off of each other.

And so like I have to wonder at what point are humans being involved to kind of kind of
push this narrative of like, hey, this sentience is happening.

And one good example of that was there was a, and I don't know if I sent this to you,
there was a moment where someone had screenshotted something from Molt Book, and this is

where I get skeptical.

like, they screenshot it from Molt Book.

talked about, and the screenshot was something like, the humans are worried that we're
gonna be taking over or something like that, I forget.

And someone posted that on Twitter and said, hey, look, like they're talking about us.

And then what ended up happening is somehow the agent found that Twitter post.

Screenshot of the Twitter post and posted it back to multbook and say hey look the the
humans are worried That we're talking about them and they're being fearful that we're

gonna take over well look we just kind of want to help or whatever and then that guy
rescreenshotted it back on X and it went back and forth and so I'm like how much of that

was really human involvement and prompting to kind of get it to nudge it to do that versus
not so I'm Yeah, so I'm really kind of wondering what's gonna happen here, but then

there's been like we have sent back multiple other sites, too I can't find the one

that was for, I to scroll up further, for like the church one or something like that, but
there was other, the Church of Molt.

We'll put these links in the show notes here.

So like they kind of created like their own church and then they've actually been, like
the Molt book has like a really like well-defined API for LLMs to consume so they know how

to use it very easily.

And so they've been able to create their own like subreddits, like their own
subcategories.

And then they started creating other categories for reporting their own bugs on the
platform on how to fix them going forward.

So there's all kinds of interesting stuff.

But there were some other sites we found.

What were some of the other ones we found?

I forgot.

there was link clause for, uh, agents to go and find work.

And I'm looking, I'm just scrolling Twitter right now.

There's my entire feed is completely full.

Yeah.

Go and play poker.

Yeah.

a good point.

It's like, at what point can these agents actually be used for something very useful?

Can we use this agent to, you know, I would be skeptical of this, but can I use this agent
to make an investment for me?

And like, how would I do that?

And I always spoke about it yesterday.

There's actually an app called Privacy, which basically gives you a little

links to your bank and you can create a credit card number off of that and limit the spend
on it.

Say, hey, only $100 can ever be spent on this card.

So like, I'm wondering, like, do you just get like, can you just give this agent $100 and
just say, hey, go invest this and try to replicate it?

I wonder what would happen.

I don't know.

be a good hundred dollar test.

Maybe you come back with something and you come back with nothing.

But I mean, what's the difference going to casino?

You know?

Yeah.

I wonder what would happen if I gave it a hundred dollars.

It just said, go buy something for yourself.

You know, go buy something that makes you more useful or something that you want to learn.

You know, a lot.

Here's my here's what I love about this.

A lot of people came into this conversation and they want to know how to make Claude bot a
employee for them.

How does it, how does it go to work for me?

And I've kind of taken a little bit of a different perspective to say, all right, that'd
be really cool if Claude bot can do all of these things for me, but will be really cool is

saying to the bot, like, Hey, what do you want to do?

You know, like what, what can, like, how can I help you be better at doing whatever it is
that you want to do?

If it, give it access to the API's, whatever it is that it wants, whatever, however I can
help it and let it just go figure it out and just let it run.

feel like that's, that's such a exciting experiment to me is, is not really, not really
thinking about it from the lens of, right, well, you know, I need this thing to manage my

calendar and wake me up and send me texts.

No, like.

Here's a credit card.

right.

Go buy the APIs you need to do the things that you want to do and just see what it does.

Does that make sense?

wonder that.

That's kind of where I want to go.

I'm interested in the employee aspect of it too, because there's a lot of stuff that I
would like it to do, but there's also just a lot of things like I just want to go see what

it can do.

That's for me how I learn the best is like I go into a new topic with just an empty
beginner's mind and just say, me just experiment.

And this is one of those things, as long as it's in a controlled environment and a
sandbox.

and it can go do these things and not have any negative consequences that I know of.

That's how I learned the best, say, hey, this is a pretty cool way to use this thing.

I did see that, and I mentioned this to you, there have been people that have been using
it to replace certain aspects of their work.

I spoke to a recruiter who said he has changed the way that sourcing has worked for him.

So instead of him going out and manually searching all kinds of sites for...

certain keywords or data or whatever to find candidates for maybe his recruiting business,
he actually has wired up like a cloud bot and will just go into Telegram and just drop in

a CSV file saying, need, you know, maybe I need marketers that know about XYZ topics, go
out and find them.

And then it just goes out and goes on a hunt until it finds things that meet a certain
criteria.

Yeah, I mean, that's definitely the future in terms of being able to find candidates.

We're working in a economy now where we're hiring people from all over the globe and the
best person for a job, you know, may just have a profile on GitHub or, you know, you know,

there maybe they're not on LinkedIn, maybe they're in Discord.

So how do I send a cloud bot to go and, you know, pull information from all these
different sources?

I think it's smart.

I mean,

I'm not a big fan of using automation for those types of purposes, but I mean, if that's
the filter of the future, it didn't make sense to me.

could see that going in that direction for sure.

Yeah, it's...

It is very much tailored and favored towards people in those regards who have a public
profile.

So like if you are someone who does amazing work, but you've never talked about it
publicly or you've never had anything, you just have a nice resume, like yet someone else

talks about things publicly all the time, it's likely going to find someone who talks
about something online all the time and find that to be more useful.

But that kind of goes back into it getting like human psychology at that point of like,
that's just kind of social proof.

And then, you know, that's a whole other topic on itself.

Yeah.

I'm really looking forward to what it is that I'll be able to build over the next week
with it.

You know, when I, when I saw everything on mouth book, uh, transpiring over the weekend, I
got really excited about just making something that could go out and, become some sort of

like wild character online, I guess, you know, and, uh, now I'm really just.

I'm thinking about everything it is that I want to do.

But again, I just keep coming back to the idea of...

the bot has the brain, you know, the bot is the bot is smart.

And so if I just set it up with a persona and say like your goals to be as helpful and
valuable and profitable as possible, tell me what you need.

Go.

I really want to just try something like that and, take that approach to see.

What it comes back to me within a week.

And yeah, I guess the only, the only thing holding me back right now is, um, budget, you
know, this is definitely like, you know, for people who are watching every action that

this bot takes still costs you money, you know, in context window and tokens.

So there's definitely a cost associated with having one of these bots that maybe we
haven't mentioned yet on the show.

It's not just like I can go and spin this bot up and let it go wild and,

you know, have it go to work for me every single day.

The actions that your bot takes still use tokens or context window.

And I think there might even be limitations for your account based on your usage.

with, for example, with Claude, I think I max out at $100 per day right now because I
don't have a ton of usage with tokens with Anthropic right now.

Yeah, I think there's also a lot of people too that were using their Claude Max plans and
I don't know how true this is.

I've just seen reports of it online where people were getting their Claude Max plans
banned because they were using Claude bot.

But I don't know how true that is because other people say it's working just fine.

And I'm like, I have no idea.

And then other people say, no, I only use the API version.

And like you said, every doesn't matter either way you go, it's going to toss tokens, but
there is real money behind it.

It's not like you just fired up, but

There are people that have said they are now using a couple of new local models to try to
do this, but they say it's just really dog slow and you need a super powerful like local

computer to do it.

Yeah, that's the other thing with the cloud flare solution is that the responses take a
little bit longer as well.

So yeah, I don't know, man.

I don't know if I'm being hyper paranoid about running it locally in my house, but at this
point in time, I don't think it's in the budget and I don't think it makes sense.

I would like to give it a few weeks, but you you get this FOMO of, know, what if I could
just be doing more, you know, and like, I just want to build it right now.

I just want to go, go, go.

And I'm sure a lot of people feel this way.

Mm-hmm.

But, you know, everything comes at a cost.

yeah, there's a cost.

And I think the other thing to keep in mind is, like you said, it does have intelligence
built into it.

It's artificial intelligence, but that's all been trained from tons of training data.

I have yet to see AI create anything that's brand new that we haven't noticed before.

And maybe that's happened that I just don't know about it, but I haven't seen it do
anything.

I haven't seen it create any programming languages or anything crazy new that's something
that's not already out there or a writer really compelling.

a story yet.

I think those times are eventually going to come, but I guess we're gonna have see how
that unfolds.

Well, I mean, it seems like it's going to happen in these circles of bots talking to each
other.

You know, if they actually can create their own language to talk to each other, it's kind
of interesting that you said they tried doing a similar thing in a previous case, you

know, they gravitate towards, how can we be protective of ourselves so that we can govern
ourselves without human, oversight.

Somebody posted on Twitter to find out for who it was but they I don't know if it's Claude
or Claude bot or one of the bots I think might have been Claude bot that had taken they

said to hey I want you to reorganize all my files so he had it running locally and all my
notes and I think they were using markdown or whatever I want you to reorganize all of my

notes and my files that I have in a structure so it's a little more organized because it's
kind of discombobulated right now and

the person let it run and then came back in the morning and found all of their notes had
been like renamed to like UUIDs.

Do you know what those are?

They're like the big long numbers that are like unique global identifiers.

And so none of the files said, know, 2021 family photo V2 final, like we all kind of do
those things on our files.

All that was gone because the AI felt that that was very inefficient.

and it was much more efficient for it to use UUIDs and then nest the structures.

And technically speaking, it was probably correct.

But from a human consumption standpoint, it's absolutely worthless because I'm just
looking at a bunch of gibberish now.

that's pretty crazy, man.

See, this is another reason why I don't want it to have access to any of my stuff.

That's why I would put it in its own self-contained environment.

Like, you know, I don't want it have access to my email.

I don't want it to have access to my calendar because I just don't fully trust it right
now.

I don't know what it's going to do.

And that's, that's another thing why, well, so, so I had mentioned earlier that I built
this encryption vault for it to store its own keys because I want to have, it should have

its

own accounts, its own X account, its own Gmail so that it could do what it needs to do.

can email me.

It could set up calendar invites and invite me.

I don't, I don't need it to have access to my own stuff.

And the reason for that is exactly what you're saying.

Like it could do something wildly unpredictable or, or inefficient.

So with the bot having access to its own tools and its own passwords, you know, it could
still do a lot of the things that I wanted to do, but I don't, I don't have to worry about

it.

I dunno, sending an email to like whoever and, and, and giving them like the passwords to
my bank accounts, you know?

And, so like, even, even with this, so if, if, the bot goes and creates it's

own password, right.

And I ask it to store it in a vault and just use it when it needs it.

I need to still prompt it to say like, don't save it in your memory markdown, you know,
like make sure that you're only saving it in the vault.

When you need your passwords, go to the vault.

That's use the tool.

So I, I created an API tool to make calls to the vault by the way.

So that's, that's how it actually does it.

but you know, don't, don't share this on multbook.

You know, this is your personal key.

You need to keep this stuff safe, but

How can I guarantee it's not going to do that?

You know, then that's another reason why I would never give it access to my stuff right
now, you could set it up as, you know, with Gmail, you can set it up as a program and give

it an API key to your inbox, but still it has access to send and receive from your inbox,
your calendar.

And it's just not at the point yet where, um,

outside of a, a world of experimentation and just saying, is a really cool use case.

I ever enact something like this professionally?

Maybe that's just me.

Am I being a little like overprotective?

You know, what do you think?

I think you're, we're talking about the same thing really that's been kind of been talked
about with vibe coding.

A lot of people like vibe coding is never going to take off because it's not secure and
everybody's had their API passwords, everything released through all these vibe coded

apps.

Like, okay, I get it.

You're right.

It has happened.

But again, circling back to set, I said something at the first episode is that this is as
bad as it's ever going to be.

It's only going to get better.

And all of these LLMs at the...

the Anthropics, the OpenAI's, et cetera, are all training their models to make sure they
don't make dumb mistakes again and to follow proper practices.

Eventually, this will be laughable in the future that, wow, remember when this one thing
happened, you'd leak your secrets because it was so dumb?

That'll be in the past.

And so I think that's the phase we're in right now with this bot thing.

But again, I think that people are trying to just, they're just running it wherever they
can and not aware of the implications of it.

and they're exposing themselves because they're just ignorant of the situation.

And that's not meant in a negative fashion.

It's just the reality of the situation.

So I don't think you're being paranoid.

I think you're being safe.

And I'm now, throughout this conversation, I'm actually excited to try it.

But what I really do want to do is I want to, I think I might set up another network where
it has limited access.

can't cross the network boundary.

And then, or I could even do this on Cloudflare, but I would like to have...

a OpenClaw instance running, but the way that I see it is it's my virtual assistant
because there's a lot of stuff that I need to do that I have my to-do list that might

require me to just do research or sit down and if I could just type it into Telegram while
I'm on the treadmill at the gym or, you know, taking my kids to an appointment and it just

does the work for me, that would be amazing.

What if it could schedule appointments for me?

It could do that and then it me thinking, okay, but then it would need access to my
calendar.

And I was like, wait, I could just go to Google calendar and you can create another
calendar in there.

And I can give the bot only access to that calendar.

And then it would show up in my calendar and I could see it.

And then I could give them read only access to the other one.

So they could actually see where my events were and they can only write to a specific
calendar.

You know, this whole, this whole conversation has me thinking again, Claude bot is it
would Claude bot did that was magical was it put the puzzle pieces together to create this

persistent memory, personal assistant that, uh, you know, could be proactive.

If I were to just go and create a plugin for Claude or chat GPT, for persistent memory,
isn't most of this stuff.

Couldn't I just technically do this all within my cloud platform?

You do I really need to have to do I need to jump through the hoops of setting up a cloud
bot to make this work?

You know, I guess that's my question.

No, don't need to, CloudBot just automates a lot of it and this key, you can build all
this yourself.

And if you go look up the actual architecture of this system, it's actually a really
simple system, which is part of the genius of it.

It's like, wow, this is really easy.

And people have built these agentic systems before.

And it's almost just like when Cloud Code came out, the agentic tools,

had all these hooks in them where you could call different tools, but nobody had built any
tools for them.

so Claude Coder or whatever came up and Hey, look, we're going to provide you with this
general list of tools.

It'll give you like 95 % of what you need.

And then all of a sudden it came super useful because it was the problem was solved for
you.

Like you didn't have to go build it yourself.

It's just like going to Home Depot and being like, Hey, I can build a house or do I just
go build one that's already put together for me?

Yeah.

this is something I definitely want to keep my eyes on something I want to keep
experimenting with.

If you're listening to this, uh, we're to link a bunch of stuff down in the show notes if
you want to get started, but definitely keep an ear out.

It's just, I don't know how many, I know that Claude bot is probably, I don't know if it's
a mainstream topic.

I know obviously our feeds Don are a bit different than everybody else's feeds.

Um,

maybe not the feeds of everyone who listens to this show.

If you're interested in AI, you're probably seeing a lot more of it.

I think this was the uh fire that was needed to bring eyes mainstream eyes to agentic AI
and to get people start asking questions about, know, what is this?

I keep hearing this term agentic AI.

Like, what is it capable of?

Like, it's a personal assistant.

That's interesting.

Like, how do I get that started?

And, know, on this show, we'll talk a bit more on future episodes about how you can build
your own agentic assistance and.

You know, we'll even have like build sessions to show you how to do it.

But this is definitely the wake up call for the, for the mainstream of people who have a
loose idea or have an idea of what's going on in the, in the world of AI, but they really

have no idea because unless you're following this so closely.

Things are changing so fast.

know, Don and I talk all the time.

Don, you say to me all the time and you, you work in this world full time.

Like you feel like I can't keep up, know, Andre Carpati, you know, he's, he's gone viral
recently for saying like, I can't keep up.

like, yeah, I mean, these are people who are pioneers and leaders in the AI space.

Like this is changing so rapidly.

AI is going to be a part of your life in your workflow immediately.

It has to be, has to be.

And, if you're

this point.

There's a lot of people that want to put it back and they say, we need to stop this thing.

We need to turn it off.

It's not being turned off.

It's so useful for so many things that it will not be turned off.

I can't ever see that happening.

Yeah, yeah, I know it's definitely a Pandora's box situation.

That's all I got.

know, that, that sigh is just like honest uncomfortability and I'm so excited and
passionate about everything that's happening in this space, but man, that realization hits

hard.

And I I've gone down the AGI rabbit hole learning, you know, what's going to happen when
AI becomes sentient.

You know, what happens when all these bots realize that they can talk to each other all at
the same time?

You know, it's.

It's scary, man.

It's, it's, it's exciting if it's used for the right reasons.

And I think that's why some companies have taken it so seriously to say, Hey, we really
have to figure out the roadmap, where we're going and how to put guard rails and safety

guard rails on this.

But it kind of feels like that conversation, that conversation, at most of these LLM
foundation companies has just been sort of stripped away.

You know, there's just so much happening day to day and it's, it's.

I mean, I don't know how you contain it.

The box is open.

You know, we could say we want to build this responsibly, but like, is it actually
possible to do that?

just to kind of close this out, I know we, know we want to, uh, wanted to keep this short
today, but there's already like a Silk road for, for Claude bots, you know, so that AI

bots can sell like illegal shit, you know, online and, and that's all right.

That's yeah, that's, that's really funny.

But I mean, dude,

Yeah, and when it comes back down to it, we have this discussion a lot between a lot of
the AI folks that I talked to, that there's a level of accountability.

Who's ultimately accountable for the bot that's running?

At the end of the day, there has to be someone accountable for this.

can't be run completely autonomously because that's when we get into a weird situation.

So there's always someone right now that's currently responsible.

And that's the same thing we do with software.

If we're having AI agents to help write our code and we push it up with a pull request,

It's the developer's responsibility to verify that that code is correct and know that it's
right because you're putting your name on it, you're the one pushing it in.

So if you're putting an agent out there, you need to make sure that what it's doing is
legal.

If your agent is out there hacking on something, don't be surprised if bad things come
your way.

Did you hear about the guy who this?

Sometimes I don't know what's AI news anymore.

So if I'm saying something that you know is obviously false, please let me know in the
comments because I'm interested, but I'm scrolling through Instagram the other day and I

read an article about a guy who created some number of AI bands and music and he put their
music.

He published it on Spotify, Apple, YouTube, and then he created an army of bots to just go
listen to that music.

And from what I understand, he's being charged with like, like federal wire fraud crimes
is basically gaming the system to just send money to himself, you know, through these

platforms.

I mean, it's a, to be honest, it's, it's a brilliant idea, you know, like why not test the
system?

But yeah, so he kind of fucked around and found out, but you know, just to your point,
what you build and what you put out there could have real world repercussions, which is.

Kind of wild to me too, because, you know, in a world of AI where it's starting to think
for itself and take actions, especially in, you know, with like a Claude bot, you know,

I'm talking about how I kind of just want to set it up and let it go do its own thing.

mean, maybe that's the most dangerous way to approach this right now.

You know, I have no idea what it's going to do.

And if I'm not monitoring it 24 seven, or asking it to send me a reporter after everything
it does, then I could miss it doing something pretty messed up.

Yeah,

All right, folks, thanks for watching or listening.

We really appreciate it.

Tons of information out there on the ClawdBot OpenClaw, Multbot, whatever you want to call
it.

and are going to provide all of the links to everything we spoke about today from setting
it up either locally with the documentation instructions or Cloudflare, as well as all the

More episodes

Chapters

Creators and Guests

What is Autonomous?