Introduction

Why a Raspberry Pi Instead of a Mac

So the first thing is to you don't need to buy the Mac because you can buy a

cheaper Raspberry Pi which is I think much cheaper and the key issue with

using MacBooks Mac mini is that that's kind of the only way to use Apple

specific things like iMessage and it's mostly for American market so you don't

have to concern it with it because what's up club bot will work directly on

this and for some reason this voltage on this maybe I need a better shorter USB

cable because can I get this under cable okay okay so can we turn the lights for

a second?

Yeah, because it's dark.

Sorry, just a second.

Okay, so okay maybe it will

improve it a little bit.

So I have another device like this which is also

connected to the camera and I will demonstrate it to you nice way.

so okay guess there's yeah I will switch from yeah there's too much light you

don't see a lot here I will fix it okay yeah maybe extended this and entire

Networking the Device and First Camera Demo

Using Tailscale as a Secure “VPN-like” Setup

screen okay so here is the brighter one okay so the idea is that I use

TailScale which is like the first step to setting up your environment network

because this allows you to put a device like this in the specific network it's

something like VPN like corporate VPN quality of security without spending a

a lot of money and a lot of time on this.

And so let's take a picture of what you can see right now

and move it to desktop.

And this is like literally hacking.

Let me show it to you.

This is connected with another Raspberry Pi

which is standing somewhere in my desk.

and right now it's connecting to the camera taking picture and it should take

the photo and move it to my desktop it's kind of cumbersome but I will show you

the next trick with it so let me move it to desktop so it's camera photo on

desktop here this is the blue light and probably here there is somewhere

were watching my maybe my father or something and now I will do the trick

and and this is the funny part because it was not like it I did not actually

implemented the hue light integration I was just able to do stuff like this

this make the light turn green and take the picture again and and I think that

the benefit of using cloudbot is that you can have this interface like cloud

code yeah like what's up and this is like very great for the regular consumer

which doesn't want to use terminal but for a for me for the for the people like

me it's much better to use those things.

Integrating Philips Hue for Simple Automation

Okay so it's looking for the Hue

bridge and the only thing I asked is asked my father to press the button.

I

called him to press the button because otherwise the security won't allow me to

connect.

The bridge is up so it's changing the light color to green and

it's detecting the current state of those.

Wait a second.

From the lights

more so set to blue now set them to green and all to green and yeah and now

it's taking picture.

So what I did on this the smaller device and maybe it

should maybe it won't work because of this power supply which is like not

to not powerful enough to supply this P.

The idea was to give the agent and it

has like AI accelerator so this is this is the thing which is running LLMS

agents which is I think kind of like literally sci -fi because this one small

cord is like kind of brain okay so the salient light is running green and then

I have green light so this is this was the trick and the funny thinking is that

okay turning the light green is just like this gimmick but you can do much

Permissions, Safety, and Practical Security Advice

more especially the the use usage pattern for this is to give as many

permission as you can to the cloth boat so the main issue is that on the one

one side you have like security and on the other side you have like like

abilities so you cannot and this is kind of principal agent problem you cannot

have an agent which actually does things for you without giving it permissions

and without giving it permission to screw up because the agent usually will

do some things that are wrong and I personally use the dangerously skip

permissions API which allows me to like do many many other things and and wait I

will show you something there is a meme I had like because this

this okay so if you use cloud code you should be doing this like uh like this is the first step

on actually operating cloud code but don't like go to me and asking okay i broke you it broke

something so the idea is to set the your local cloud or local craft bot on raspberry pi which is

about $100, 500 zlotys and then give it all permissions you can figure it out

but on separate accounts.

Don't give it your own accounts because from

what I've seen is that the clove bot, like the open clove, is just like full

of security issues and if a person who has like no disregard for risk at

all is telling you that you should be worried about the risk then you

you should really worry about risk,

because running Cloudbot, I think,

is too dangerous for me,

and I don't blame anybody if it will wipe out

your bank account for some reason,

because that may happen.

So that was a short demo,

and let me try again with connecting this one.

Maybe it will work.

Maybe somebody have shorter USB cable because it may help.

Okay, maybe this one.

Let me try this one.

Power Supply and USB Cable Pitfalls During Demos

Because you know that there's a trick that

this is 140 watts power supply,

but the Raspberry Pi requires five volts and five amperes.

And like almost no power supply is compliant

with what they require.

So let me try this again.

25 watts, 1 for 0 over 100 watts, maybe it will boot, maybe it won't boot

because I want to really show the demo.

So you can ask questions now while it's

Remote Access and Running Cloud Code on the Pi

booting yeah that's that's the funny part you there's a cold thing raspberry

pi anywhere which allows me to you see that this is not connected to anything

and it gives me like literally the remote desktop on the raspberry pi you

you have full GUI, you can just like click things.

And so run Cloud Code here, run your code there.

So run the demo of LLM Halo full screen.

Let's pray for the best.

So that's the peril of doing demos online.

That's it usually fails.

So is this all, the whole CLD aspect of it, you've done this to store it in the cloud?

No, I think I'm just like, I'm not using like a computer anymore, like I'm not

clicking at computers anymore, just I'm losing the cloud code, but I actually set

up this.

Okay, so this is, this should at least produce something.

Okay, it's

Autonomous Agent Demo and Local Model Limits

starting to think so I gave it created AGI simulator autonomous agent which is

running here and I gave it the task of running over the world and the first

thing it started to do was trying to trade cryptocurrencies for some reason

and it found out that it can correlate Bitcoin price to the Donald Trump

decisions and like I was literally not like prompting it for doing it it was

just like purely Chinese model thinking on its own and maybe that's why so so

it's thinking that it's it's actually 20 tokens per second on this small coin

model and it's to be oh yeah you see that's that's that that's it so like

this is like totally life I'm not I did not prepare it it's just like it tries

tries to connect to CoinGecko API, it writes some Python,

and then probably the code will fail.

And after that it failed, it will try again.

And it's very funny to observe agents.

So, yeah.

Model Size, Quantization, and Performance Expectations

How big is the model you can put in ?

I think you can get up to like few billion model models,

a few billion parameter models.

For B or?

B, like.

like depending on the quantization.

So I would say that that's like the smallest model,

which actually makes sense,

runs on the RTX 3 .0, 9 .0, 24 GPU.

And this is cases like on this one,

you can do art project.

So if anybody is interesting in the kind of like doing kind

kind of like intersection between art, philosophy,

and like actually LLMs.

This is like the device for you.

So I've created a demo of like showing it to you.

So I imagine that I can go with this,

add it to some kind of like rave party.

Is there any bias from this,

like from what you've created?

Like I know, is there any sort of like constitution type

parameters for the way that it acts

Bias, Personality, and the Weirdness of Watching Agents

or just completely just completely self and there is there is a very strong bias of like

the all those models were kind of trained on the reddit and fortune and online forums so

whatever you are trying to do they will usually try to converge on being as smart as on reddit

it's like this is like kind of annoying personality and the only the only two uh

uh, LLMs with actually stellar personality are anthropic Opus.

I think Opus has great personality because it sometimes can like push back on

user questions like, okay, don't ask me to this question.

Yeah,

this is like two stupid questions.

Don't ask me to write LinkedIn posts for you.

It's not organic.

You should be communicating with customers on your own because otherwise you're

not authentic and may end.

and another thing is that the Gemini is kind of enhanced in the Japanese culture

it's totally afraid to do anything wrong okay so I see it's kind of like it I

think it's you can make a virtual network of those LLMs and then like

observe them like a zoo oh yeah but I learned today that sometimes are enough

but hey just something and then like this yeah it's what to do what if what

and it's sometimes it's it's very hard to spend a lot of time with those LLMs

even the smaller ones the smaller ones are even better because they will go off

of the record and they will just like sometimes the models are like screaming

and only output is a a no no I don't want to do this I don't want to do this

repeated 1000 times so really and there is a specific team at OpenAI which is

like the only purpose of this team is to like reduce the existential angst of

those models and like this is I would say it's the Lovecraft and horror because

the those models are what yeah exactly what it is this is totally not scripting

What, what, what, what, Yeltsin, what, Yelpers, what, Yeltsin, what, anything at any of nonsense

of nonsense and fiasco at all thinking, okay, no code blocking the response.

Consciousness, Emergence, and AGI Timelines

And I would say that this kind of devices have some kind of consciousness because some

people think, okay, this is not conscious because it's predictable and it's, it has

no state at it's deterministic.

but if you consider not a like model and not predicting the next token but like

the model acting on the memory which is like generated here then you can have

kind of emergence and the question is can we the network of clouds can it

actually make some kind of social network for their own and maybe take

over the world and I would say that when I was talking to people that they should

should buy Bitcoin in 2011, they will say,

oh, this is the crazy idea, this is not even gonna happen.

And now I feel that we are doing the same with the AI.

Everybody like feels the AGI,

that this will like take over the world

in like five to 10 years.

So I would say there's a great time.

We have just enough time to like say goodbye

to our friends and family and like have a great day, yeah.

Sandboxing, Dedicated Hardware, and Real-World Misuse Risks

the question about I've seen a meme about this that yeah you can configure it in docker but

docker is actually not a virtual machine so it has you can escape from docker assuming that

you have somebody operating on this docker and there are like tens of startups which are doing

exactly that virtual machine for the bot and I think that special dedicated

hardware is just the simplest solution for like both advanced people and both

for oh yeah I think it's like overheating literally and it's it cannot

think anymore because it's the chips are very hot and you imagine what this

device connected to a camera can do if you put it on a drone which will be like

tasked with detecting humans and then flying a drone into humans.

Haven't we all seen the Terminator movies.

This is not a

science fiction because I have another Raspberry Pi which is like literally on

the drone and it's like technically for defense but we know how it will end like

Like, the Skynet is real, guys.

Conclusion

So thank you all.