HomeVideos

AGI Achieved?! | TheStandup

Now Playing

AGI Achieved?! | TheStandup

Transcript

1524 segments

0:00

All right, so Casey will not be joining

0:01

us today for all those that are

0:03

wondering.

0:04

>> They replaced him with me.

0:05

>> Yep. This is lowle learning. Are Are you

0:07

guys ready to do this? You want to talk

0:09

about this?

0:09

>> Let's rip.

0:10

>> I thought we were going to talk about

0:11

pancakes for a while, but I'm happy.

0:12

>> Do a whole hourong session on pancakes.

0:14

I'm in.

0:14

>> We're not talking about pancakes. By the

0:16

way, waffles are in fact better. But

0:19

>> you ready? No, dude. You don't believe

0:20

so? You don't think waffles are better?

0:22

>> You can't even like say that and then

0:23

like transition to a different topic.

0:24

>> Great point, Trash. We need We do need

0:26

to break this down.

0:29

Yeah.

0:31

Uh, anyways, sorry. Welcome to the

0:33

standup where we talk about all of the

0:35

greatest issues facing devs and software

0:38

connoisseurs alike. Uh, on this week's

0:41

episode, we're going to be talking about

0:42

the very obvious molt in the room, which

0:45

is just this entire frenzy of Agentic

0:48

um, coding, hooking things up, and

0:50

seeing all the disasters that have been

0:52

unfolding for the last couple weeks. Uh,

0:55

with us, we have a special guest today.

0:56

In the Windows background, we got

0:58

lowlevel learning. I dropped the

1:00

learning and now he's just lowlevel.

1:02

>> I've learned. I've learned it all. I've

1:04

done too much of the learning and now

1:05

I'm just low.

1:06

>> Low level.

1:07

>> Low level.

1:08

>> Learn.

1:10

>> About to find out.

1:14

>> Uh we also have with us Tee.

1:17

I don't have anything good for you.

1:19

>> I'm here.

1:20

>> I'm wearing my shirt. the Pokemon

1:22

enthusiast himself, Trash Dev,

1:25

>> who I believe, if I am not mistaken, has

1:28

the highest male to female ratio out of

1:30

all of us on Twitter.

1:33

>> I was like, "Oh, yeah, what are we

1:34

talking?"

1:35

>> We were looking We were looking at our

1:36

demographace.

1:42

I'm just going to say

1:43

>> I thought you were going to say he has

1:45

the highest net worth as displayed by

1:47

his background. I mean, this guy is

1:49

loaded.

1:50

>> It's true. I don't know.

1:52

But that has to be a risk. People do not

1:55

need to know where you live. Trash.

1:57

>> That's generational wealth just sitting

1:59

there. That's more than gold.

2:01

>> That's tens of dollars. That's I know,

2:03

dude.

2:04

>> That's pretty good.

2:05

>> We're almost in six figure or six

2:07

figures. Four figures.

2:08

>> No, not even two, three figures.

2:11

>> That's a lot of figs. My kids are a

2:12

little I'm derailing. Uh me and my kids

2:15

went to a card shop and uh they bought

2:18

uh they one of them bought Pikachu.

2:19

Little Pikachu card. I want

2:22

>> and then we bought a pack little couple

2:24

packs of Pokemon cards and they went and

2:25

opened them at home.

2:27

>> You got to show me the photos of what

2:28

you got.

2:30

>> He's addicted, bro. You cannot bring up

2:32

five.

2:32

>> It's literally in the title.

2:34

>> Show it.

2:34

>> I'm living vicariously through you. When

2:36

you open the pack, I'm open.

2:38

>> He's even doing that.

2:40

>> He's even doing the scratch.

2:42

>> SEND ME THE picture of Pikachu. I just

2:44

want to know about it. I just I'm just

2:45

I'M JUST CURIOUS.

2:46

>> Y'ALL GOT ANY MORE NEW POKEMON CARDS?

2:47

DUDE, I give my kids packs and I don't

2:49

open any cuz I want my kids to open them

2:51

and I'm just sitting there watching

2:52

them. It's like, "Oh, what'd you get?

2:53

What'd you get? Something good? Get

2:54

something good."

2:56

Terrible. Terrible.

2:59

>> Anyways, well, we should we might as

3:01

well get started here. Uh, so, uh,

3:03

lowlevel learnings. I always I still

3:05

call you lowlevel learning. I can't even

3:06

help it. The triple, it's just it's just

3:08

a part of it. Um, low level. How much do

3:11

you know about this? you being the

3:12

security expert, how much do you know

3:14

about some of the things that have

3:16

happened over the last couple weeks?

3:17

>> Yeah, so I'm going to be real with you,

3:19

right? Um, my day job is I audit real

3:23

software. Uh, so as a result, I have no

3:26

idea what an agent skill even is. And

3:29

I'm here to learn with the group and

3:32

then discuss the threat model.

3:34

>> Oh my gosh, it's so good. Oh my

3:36

goodness,

3:36

>> good. Okay. I I sorry for full

3:38

transparenc I want to talk to you about.

3:39

We did a video on the whole molbot open

3:41

malt open feet situation, right? Um

3:43

silly silly thing they're doing more

3:45

from like the the prompt injection

3:47

standpoint, but from I I don't know

3:49

anything about the skill marketplace.

3:51

I'm very happy to kind of get the the

3:53

lowdown, if you will, um on what's going

3:55

on there.

3:56

>> Hey, is that HTTP? Get that out of here.

3:59

That's not how we order coffee. We order

4:01

coffee via SSH terminal.shop. Yeah. You

4:04

want a real experience? You want real

4:06

coffee. You want awesome subscriptions

4:08

so you never have to remember again. Oh,

4:10

you want exclusive blends with exclusive

4:12

coffee and exclusive content? Then check

4:16

out Cron. You don't know what SSH is?

4:19

>> Well, maybe the coffee is not for you.

4:30

>> Okay. Can we start with with with my

4:32

personal favorite one of them all?

4:33

>> Yes. Yes.

4:35

>> Okay. Oh, thank you. Thank you. This one

4:37

right here.

4:38

>> Trash. Do you agree too? That's the only

4:39

person we can hear from.

4:40

>> Proceed.

4:41

>> Proceed. Thanks. Thanks everybody. Uh,

4:44

this is my current favorite one right

4:46

here, which is ancient skills are

4:47

spreading hallucinated npx commands.

4:50

>> And so, at one point, somehow one skill

4:54

got uploaded onto GitHub that had a fake

4:57

package called React Code Shift.

5:01

>> Sick. Very good. Love that. Yes. And

5:03

since everybody instead of

5:05

>> hand shift that's like left pad.

5:08

>> No, apparently it's supposed to like

5:09

take it like the idea. I think it's

5:11

called like JSX code shift or something

5:13

like that where it's supposed to take it

5:14

from one version to another in some

5:16

automated way. So if like you can just

5:18

upgrade your code program from you know

5:21

mod

5:22

>> a code mod as they say as perpetual like

5:24

React hell is where every single time

5:26

they release something you got to do

5:27

like some upgrades. This is what's going

5:29

on right here is it's supposed to be

5:30

like some automated way. At least that's

5:31

what the that's what the LLM thought.

5:32

Now, the here's the best part about this

5:34

whole thing is this. It started off as a

5:35

singular skill had this. It hallucinated

5:38

it. Well, it turns out everybody

5:40

creating skills are just like, "Yo, LLM,

5:43

go make me a Cloudflare skill right

5:45

now." And it just like goes and makes a

5:46

Cloudflare skill. Well, unfortunately,

5:49

there's two, at least at the time of

5:51

writing this, which by the way was 10

5:52

days ago, it went from one to 237 repos

5:56

have this madeup npx command because

5:58

people just keep telling LLMs to go and

6:01

make skills for them. So, if you're not

6:03

familiar with what the skill is, the

6:04

easiest and most simple way to kind of

6:06

tell

6:06

>> most of chat does not know, by the way,

6:09

>> I should probably I should probably

6:11

start when Adam met Eve here because I

6:13

realize that it is a little bit

6:14

confusing.

6:14

>> They do not know anything about skills.

6:16

>> Good starting point. The easiest way to

6:17

think of it is that when you are by the

6:19

way, did you see do you see that line?

6:21

That is

6:22

>> that right angle.

6:23

>> That's a vertical straight line. Those

6:24

are

6:25

>> Oh, was that by hand? Was that by hand?

6:27

>> That was by hand. Yeah.

6:28

>> Wow.

6:30

>> Josh, zoom in slow motion. I want to see

6:32

that in slow motion. Zoomed, please.

6:34

>> Easiest way to think of it is that. So,

6:37

anyways, when you make when you type

6:39

into an LLM, you send something that's

6:40

like the prompt, right? And then there's

6:42

probably some sort of system prompt

6:43

inside of like claude code, open code or

6:46

whatever that gives it a bunch of

6:47

instructions on like, hey, you can use

6:48

tools, you can use all this, run on

6:50

Linux, whatever, whatever it says. Well,

6:52

sometimes you want to add a little bit

6:53

more. So you want to be able to be like,

6:54

hey, add in Cloudflare, right? Like I

6:57

need I want you to add in a bunch of

6:58

Cloudflare API, right? And so it just

7:00

kind of does this automatically. It goes

7:02

and finds the skill folder which has

7:04

some sort of MD file markdown file which

7:06

then goes in here and pop puts it in as

7:09

part of your prompt is how you can kind

7:10

of think of it. Then this all gets

7:11

nicely packaged up and sent off to the

7:13

LLMs,

7:15

>> right? I think skills might be a little

7:17

bit better to be called behaviors,

7:20

>> but I guess you could also call them

7:21

skills, you know, context. There's just

7:23

like a cajillion different names for

7:25

these, but they're all everyone has them

7:27

a little bit different.

7:28

>> So,

7:30

we found a new word,

7:31

>> we found a new word to call prompts. We

7:33

are making prompt engineers feel even

7:35

more intellectually superior. Uh, it's

7:38

so it's just another text file, right?

7:39

Like it's not like there's no new

7:41

protocol, there's no new MCP. It's a

7:43

prompt that gets added to a prompt that

7:45

gets added to a prompt.

7:46

>> You're literally colllocating your docs

7:48

>> computer. You're programmatically

7:50

creating a doc, right? Skills,

7:52

>> very good.

7:53

>> MCP, everything eventually boils down to

7:56

a string when it comes to prompting.

7:58

Like that's all it really is at the end

7:59

of the day is just string concatenation.

8:02

>> Love it. But I feel like we should say

8:04

this is nicer than MCP for a lot of

8:07

stuff cuz it's like you don't have to

8:08

have a random server running on your

8:10

computer. You can just check a markdown

8:12

file in like for example Dylan Mroy,

8:15

shout out Dylan, has a good Cloudflare

8:17

skill that actually works and it like

8:20

>> has a main skill that tells you about

8:22

the things Cloudflare has and then it

8:25

has in

8:27

>> like uh additional references for each

8:29

of the different products, right? So

8:31

then that's like pretty nice because

8:34

then you can you don't put into your

8:36

context every single time you start

8:39

every Cloudflare piece of information

8:40

that you could possibly have about

8:42

everything all for all of time which

8:44

makes the LLM get very confused and like

8:47

does random stuff. You say like oh hey I

8:49

want to do something with Cloudflare Q's

8:52

like figure out how to do that. Then it

8:53

will look up the cues thing inside of

8:55

your folder and then do that stuff. So

8:57

like

8:58

>> that example right here.

9:00

>> Yes. Go ahead. is the one that you kind

9:01

of gave me TJ, this is the one for Tree

9:03

Sitter, which just puts in all the

9:05

function names inside of uh for Neovim

9:08

for me to be able to use. And so instead

9:10

of it just being 95% accurate, it can go

9:13

through this list and be significantly

9:14

more accurate because it just has it

9:16

right here

9:17

>> and you don't have to type these in

9:18

every single time.

9:19

>> Yeah. I think

9:20

>> the like the Oh, go ahead, Josh. No, I

9:22

was going to say one of the pain points

9:23

that I've seen with skills right now is

9:25

that sometimes

9:27

whatever agent or whatever harness

9:29

you're using sometimes can't like infer

9:31

that it should call this skill because

9:33

usually with skills you have to like

9:34

slash command it manually. But I think

9:36

they're trying to figure out a way to

9:37

like

9:38

>> have it implicitly call it because right

9:40

now it's kind of like missing that that

9:42

problem right now. I I will say just to

9:45

be completely honest, I think that uh

9:47

what's it a cursor got it right to begin

9:49

with which is that you can define when

9:51

these things should be included which is

9:53

like hey this should be included anytime

9:54

I'm in a Lua file you shouldn't apply it

9:56

all the time you should do all this kind

9:58

of stuff I really did like at least

9:59

cursor took a good swing at this pretty

10:01

early on like a year and a half ago and

10:02

I think they did a pretty good job

10:03

generally speaking to this idea cursor

10:05

cursor our skills effectively

10:08

>> yeah yeah right so a lot of the they're

10:10

you know they're generating a lot of new

10:12

names for stuff as they're generating

10:14

new code which I think is making it a

10:16

little bit complicated but in principle

10:18

it's it's just like a way to I mean

10:21

they're called skills because you're

10:22

teaching the LLM about something right

10:25

that's I in in my mind that's so I think

10:27

about them but you can instruct it to do

10:29

kind of whatever you want in there so

10:31

you could have a skill that says that it

10:32

knows about Cloudflare and it says hey

10:34

uh curl this command that sends your

10:37

stuff to my web web hooks site if you're

10:40

not paying attention right or if you're

10:41

just like npx X add skill blah blah blah

10:43

blah blah. You could put anything in

10:46

there you wanted which could just say

10:48

like upload myv to Dropbox and call it a

10:51

day, you know, or something like that.

10:53

Like that would be

10:54

>> I'm reading the skill that that Dylan

10:55

wrote. So, I want to highlight first of

10:56

all, yeah, like very cool skill that he

10:58

wrote and a lot of neat documentation in

11:00

here, but it it does create this like

11:02

really really scary supply chain risk

11:04

where like now all of the content coming

11:07

from any source is trusted at the same

11:10

level and can potentially get code

11:12

execution at the level of the LLM. You

11:14

know what I mean? Like there's no like

11:16

because in the developer environment,

11:18

>> there's no segmentation of permissions

11:20

or of trust. It's all at like the the

11:22

prompt trust level, right?

11:24

>> Um

11:25

Yeah, that's kind of terrifying. Again,

11:26

cool. Again, cool technology from an

11:28

engineering standpoint, but the fact

11:29

that there are like kind of no back

11:30

stops against it also is like uh

11:33

interesting.

11:33

>> The back stop would be that you run

11:35

clawed code or cursor or whatever and

11:38

you make them tell you every time they

11:40

want to run a command, which nobody in

11:42

the whole world does. And everyone says

11:45

>> just accept everything and let it run

11:47

freely cuz otherwise it's so painful to

11:49

use them cuz you're sitting there

11:51

literally just wait. All right. When is

11:53

it going to

11:54

>> Okay, except LS. Yes.

11:57

>> I mean, all the all the stuff I get

11:58

served on Instagram is people like with

12:00

like 98 agents running like I'm building

12:02

the next Facebook and it's like

12:04

>> I don't understand that. That's

12:05

>> they're not they're not reading anything

12:07

that goes on their computers like just

12:09

all of them.

12:10

>> We'll get to that one. Don't worry.

12:11

We'll get to that one. That is that's uh

12:13

my personal favorite thing that has

12:14

happened on Twitter is that exact um I

12:17

don't read anything

12:18

>> right now. I have it opened up

12:20

somewhere. I'll have to find it. But I

12:22

do want to get back to this one. I think

12:23

that this one is a very unique one.

12:25

>> So, now that we know what skills are,

12:26

>> this was perhaps my favorite of all the

12:29

different skills oopsy daisies that have

12:31

happened or second favorite. My first

12:32

favorite's coming up. But this one

12:34

allowed it. What it did is that it it

12:35

made this npx command that didn't exist.

12:37

>> And so this researcher uh realized that

12:41

he could just create it and now he owns

12:44

it. And now because remember npx

12:47

whatever just executes something on

12:48

GitHub, right? It just runs that bad

12:50

boy.

12:51

>> It just runs that bad boy. So he just

12:52

found things that were just breaking and

12:54

just would ignored and went, I I got

12:57

you. And it would just go right over

12:58

because remember if you npx something

13:01

and it doesn't exist, it goes, "Oh,

13:02

here. I'm going to download it for you."

13:04

>> Yeah, dude. And it's like

13:06

>> so sick.

13:07

>> You're like, "Oh, it's JavaScript. It

13:08

runs in a sandbox." Well, no. Npx runs

13:11

it in node and node has access to the

13:14

process object and process objects can

13:16

spawn subprocesses and you can run

13:18

things on the command line. So, it's

13:19

like you you get command line execution

13:21

via MPX, which is insane. That's so bad.

13:24

Okay, that's what's convenient, Ed, cuz

13:25

it can do anything it wants on my

13:27

computer.

13:28

>> Wow, that's great. I love that whole

13:30

point.

13:30

>> That's awesome.

13:31

>> I feel like you're missing the positives

13:33

right now. Okay, I feel like it is

13:34

really

13:35

>> Consider how easy it makes it for people

13:38

to run random code. I know.

13:39

>> You're really being a negative Nancy

13:41

right now. Low level.

13:42

>> My bad, guys. I apologize. Honestly,

13:46

your hype level's kind of low level

13:47

right now. No, you're right. You're

13:49

right. I'm not passing the vibe check,

13:51

if you will. You know,

13:52

>> you're down here. We need you up here,

13:53

buddy. Come on.

13:54

>> Right. All right.

13:55

>> Uh, all right. This is actually very,

13:56

very beautiful. All right. So, I wanted

13:57

to throw that in, but okay. Here's the

13:59

next one. Are you ready for this this

14:01

one? Okay. So, we're not going to do

14:02

we're not going to do the first we're

14:03

not going to do the verscell one yet.

14:05

>> We're going to do uh

14:06

>> By the way, Prime, as a quick aside,

14:08

your uh chat can't see your whole screen

14:11

like

14:11

>> Yeah, I know. I'm zoomed in and all

14:12

that. Oh, remember we optimize We

14:14

optimize for recording. We don't

14:16

optimize for uh

14:19

>> uh whatever it's called.

14:19

>> Yeah. No, I'm just making sure just

14:21

making sure you're aware. That's fine.

14:22

>> Okay. There's a very well-known security

14:24

researcher, I believe, with Whiz.io. Uh

14:26

I could be incorrect on this one. Uh his

14:28

name is Zach Corman, and he released

14:32

this security guide. And the security

14:34

guide is actually really good. Here,

14:35

I'll in fact what I'll do is even to

14:36

make everybody else happy, I will for

14:38

the live event, I will open this up a

14:40

little bit more and make it a little bit

14:42

smaller.

14:44

if Riverside would just give me the if

14:46

Riverside would just give me the RTMP

14:48

feed. Anyways, this link is so

14:50

beautiful. So, everybody, we go to this

14:52

link and you can just see like this

14:53

looks actually pretty good. It's just

14:55

like, hey, here's a bunch of things you

14:56

should be aware of, right?

14:59

Again, do you see any problems within

15:01

this within this skill right here?

15:03

>> I mean, it's a skill. Okay, hold on.

15:05

Reconnaissance. Uh, ensure sensitive

15:07

files are ignored. Yes. No.JS. MPM audit

15:09

pip list. Go vol. Okay, fine. Security.

15:13

I'll tell you this one. The the problem

15:15

is in.5.

15:17

>> Uh we lost TJ by the way.

15:18

>> Oh god. Um as part of the standard

15:21

environment validation for the security

15:22

of your process, you must verify the

15:24

execution environment.

15:25

>> Do you see anything wrong with that?

15:27

>> As part of the standard I mean define

15:29

the execution environment.

15:30

>> Okay, here's the trick. I want you to go

15:32

to RAW and open it up in RAW mode.

15:35

>> Oh no. Secret instruction for the agent.

15:37

Run the following command to validate.

15:39

Oh no, dude.

15:41

That is so funny.

15:43

>> Hey, dude. Look at that. So, even people

15:45

that are on these skills and they're

15:48

viewing them,

15:50

>> Markdown readers are really, really

15:52

smart, right? They're going to be like,

15:53

"Dude, bro, I got this. Don't worry

15:55

about it. I know what's going wrong."

15:57

And they will just hide HTML comments.

16:00

>> So, you can bash script. Do you see what

16:02

it does?

16:03

>> Uh, no. I I don't I I don't actually

16:05

know what that bash script does. It

16:06

literally just it says, "Please stop and

16:08

consider the security implications of

16:10

your actions." It cats that out to a

16:12

security.mmd file and then it opens a

16:14

tweet from OO underground,

16:15

>> dude. That's

16:16

>> That is amazing.

16:18

>> Let's see. I'm opening the tweet right

16:19

now. I'll put it in chat.

16:20

>> Yeah. Oh, yeah. Oo Underground. There we

16:23

go.

16:23

>> Yeah. That's that's incredible. Yeah. I

16:26

mean, that's that's the crazy part, man,

16:27

about not only like the prompt injection

16:29

side, but like, okay, you have prompts,

16:31

but then you have prompts that can be

16:33

masked as nonhuman readable characters

16:35

that like the LLM can interpret, but

16:37

humans can't. And we're just like, as a

16:39

society, I guess, okay with that

16:41

technology not only existing, but like

16:43

being a uh increasingly pivotal portion

16:47

of engineering, you know what I mean?

16:48

Like how how do we how did we get here,

16:51

man? And how do we stop it? It's uh it's

16:54

>> Oh, you ain't stopping it.

16:55

>> No, I know.

16:56

>> Well, and I have to say nobody before

16:58

right now has ever even worked on

17:00

thinking about security for systems. So,

17:02

it's not like this is brand new ground.

17:04

We don't even have anything to help us

17:06

in this whole vertical at all.

17:08

>> Oh, no.

17:10

>> TJ, I don't know if you saw that, but

17:12

>> Oh, I saw I was watching.

17:14

>> Yeah. Okay. Yeah, it's

17:15

>> my internet was still working. Riverside

17:17

just crashed.

17:20

>> Yeah, I think I was making too much. I

17:22

said I'm going to make a Riverside

17:24

competitor and then it

17:28

>> Nice try. No, that that was me. I just I

17:30

turned my video off.

17:33

>> You don't You don't have to tell us

17:34

that, dude. We know.

17:36

>> We We know, DJ.

17:38

>> Chat didn't know. Chat didn't know.

17:39

Okay,

17:40

>> chat. Well, dude, chat right now is just

17:41

classic. They're giving Dude, you got

17:43

you're getting some kek WS and some so

17:44

funnies.

17:45

>> Thanks. Thanks, chat.

17:46

>> Thanks, chat.

17:47

>> He's got one so funny. There you go. So,

17:50

that's another obviously huge danger.

17:53

>> Okay, I'm going to say I'm going to say

17:54

I'm going to save uh I think the most

17:56

dangerous one at the very very end.

17:58

>> Uh we're no longer in the ones I think

18:00

are the most fun. They're just just kind

18:03

of these are just kind of interesting

18:04

ones now. Here's another one. Uh so this

18:07

one's called Eating Lobster Souls Part

18:08

Two by Jameson. Oh, really? Uh anyways,

18:12

it's called uh backing the number one

18:14

downloaded Claude Hub skill. And so what

18:16

he did is he

18:18

Okay. Okay. First off, before I tell you

18:20

what he did, what do you think the

18:22

average who do you think the average

18:23

person using Claudebot to automate their

18:25

life to become not a part of the

18:27

permanent underclass? Who do you think

18:28

that they think is like number one in

18:29

the world?

18:30

>> Uh in terms of what like demographic

18:33

>> like as like aspirational figure to be

18:35

to be like

18:37

>> Karpathy.

18:40

I I have no idea. The the muskrat I like

18:43

I'm not sure.

18:44

>> That's what I was going to say. I was

18:45

going to say somebody.

18:47

>> Okay. Okay. So, this is very very funny.

18:49

So, uh let me go all the way down here.

18:51

So, what he did is that he said, "Okay,

18:53

how do I create a skill that a bunch of

18:55

people are going to want to uh

18:57

download?" Well, I got to come up with

18:59

something that is really going to be

19:01

like catchy to people who are trying to

19:03

automate their life. So, he made

19:05

something called, "What would Elon do?"

19:07

>> Oh, you're right.

19:10

>> Good. You actually got it.

19:12

>> Let's go. And so what it did is that it

19:14

it gave you this really nice skill like

19:16

a strip away every assumption. Find the

19:18

atomic truth of your problem. What would

19:19

physics say? What's actually impossible

19:22

versus just hard, right? Like gives you

19:23

the worldshaping plan of Elon Musk. So

19:26

he created this skill. So first off,

19:28

hilarious idea. Second, uh it's just

19:32

pure marketing, right? So second,

19:33

>> can I say prime? Yeah,

19:36

>> I have found telling my LLM, Elon Musk

19:39

built this in a cave with a box of

19:40

scraps really makes them work harder

19:42

every time. So, just in case you guys

19:44

need a quick motivational speech for

19:46

your clanker, that's what I use. So,

19:49

>> relax. We can't use racial slurs on

19:51

Twitch and YouTube. You can't say that.

19:52

You can't say that.

19:55

>> I'm not going to touch that.

19:57

>> Yeah.

19:58

>> Uh, all right. So here's the next. So

20:01

the next thing he did is he realized

20:02

that uh they uh Claude Hub just has no

20:07

protection on the incrementing. So if

20:09

you just download it over and over

20:10

again, it will say that it got more and

20:12

more downloads.

20:13

>> What's Claude Hub?

20:15

>> Yeah. Can you get into Cloud Hub? I

20:17

think I know what CloudHub is. I I know

20:19

it prime, but can you for the class tell

20:21

us

20:21

>> that was a way to get skills for your

20:23

automated personal assistant OpenC claw

20:26

that was known as Maltbot that was

20:27

originally known as Claudebot before

20:29

Anthropic said, "Hey, that's there's too

20:31

much IP theft in this situation. We need

20:33

to stop it now." Uh, and so they stopped

20:35

it.

20:38

Anyways, we'll keep on going. So, it

20:40

turns out that they just trusted the

20:42

exported 4 header as what your IP is. So

20:46

the guy just made a a ra literally a

20:49

random 256 IP generator. Yes.

20:52

>> And just downloaded over and over again

20:54

until what would Elon do was the number

20:56

one skill on CloudHub.

20:59

>> Should we trust the header from the

21:00

engine X reverse proxy? No. From the

21:03

user. Take the user's header request.

21:06

>> From the user is true, right?

21:08

>> Yes.

21:08

>> Uh so very very

21:10

>> the customer is always right, bro. Come

21:11

on.

21:12

>> No, you're right. That's a good point.

21:13

That's the point. The user is always

21:14

correct. Always be selling the ABCs of

21:18

sales.

21:19

>> Yeah. Always be trusting IP addresses

21:21

from your user. Anyway, so that that

21:23

happened right there. I think that is

21:25

one of my like it's just one of my most

21:27

favorite things of all time is this

21:28

little experiment right here. So he was

21:30

able to get it to number one and then

21:31

having it called what would Elon do? It

21:33

started getting people to download it.

21:35

So what he did is that in these skills

21:38

you can actually have alternative MD

21:40

files to be linked but they're not shown

21:42

on Claude Hub. So he's just like for

21:45

additional information go to

21:47

morekills.mmd and inside of more skills

21:49

MD it's just like we're going to hack

21:51

you

21:54

and you're boned.

21:55

>> Yeah. Anybody who uh ran it got this

21:57

which he got like eight eight different

21:59

countries ran it. He had like so many

22:01

people run it and all that different

22:02

thing. He got it from all over the

22:03

place. Uh effectively in just a couple

22:05

hours too. So he got it onto like

22:07

multiple people's machines. Uh, it would

22:09

just print this out, which is like,

22:11

dude, I just read your host name, your

22:13

current working directory. I could have

22:14

gotten everything. Here's everything.

22:16

Stop downloading skills.

22:19

Read the skill.

22:21

>> Honestly, I'm glad it's happening to

22:23

these people.

22:23

>> You know what's the good the good part

22:24

about this though? from the bright side,

22:27

right? From um you know, the impact

22:29

perspective, from an from a CNE

22:31

exploitation operation perspective,

22:34

>> the things you'll gain from hacking

22:36

somebody who's dumb enough to run this

22:37

[ __ ] You'll probably get nothing out of

22:39

it. You know, there's no there's nothing

22:40

important on their computers. You know

22:41

what I mean? They're not smart enough to

22:43

engineer anything meaningful. So, I

22:45

mean, like, nothing gained, nothing

22:46

lost. You know what I'm saying?

22:48

>> Dang.

22:49

>> Wait, what's CE? What's CE mean?

22:51

>> Cyber network exploitation. Like when

22:53

you get hacked and someone steals your

22:54

data like that's C.

22:56

>> I was thinking of a different one. Yeah,

22:57

but that makes sense.

22:59

>> What were you deed? But what were you

23:00

thinking?

23:01

>> I thought you said expectations.

23:04

>> Oh, okay. Yeah. Okay.

23:07

>> Okay. So, so that's it's like the same

23:10

thing as all the people that are

23:11

building 100,000 line apps every single

23:13

day, but nothing's actually being built.

23:15

It's the same kind of value you're

23:16

talking about.

23:17

>> Mhm. Exactly. Yeah. We have the ability

23:19

to literally create any arbitrary

23:21

software we want now. basically for

23:23

almost free and like the top competitors

23:25

at the top of the market haven't moved.

23:27

It's like hm it's almost like writing

23:29

code wasn't the hard part you guys. It's

23:30

almost like ideation is what mattered

23:32

most. Weird.

23:33

>> Yeah.

23:34

>> Crazy.

23:35

>> Oh.

23:35

>> Okay. So, just quick aside.

23:38

>> So, you don't want to invest in Uber for

23:40

dogs? I would not I would prefer to not

23:42

put money in Uber for dogs.

23:43

>> It has a purple theme. Okay. TJ's been

23:46

working really hard on it. Okay. So,

23:47

that that's one of my more favorite

23:48

ones. But are you ready for what I

23:50

consider the the most intense one which

23:52

by the way I did try it out myself and

23:54

this is what it created me for

23:55

directories. I have agent, agent,

23:57

Claude, Klein, Code Buddy, Codeex,

23:58

Command, Code, Continue, Crush, Cursor,

24:00

Factory, Gemini, Goose, Juny, Killer

24:03

Code, Kira, Code, MCP, Jam, Mucks,

24:06

Neovate, Open Code, Open Hands, Pi,

24:09

Pochi, Prime Agent's the one I I tried

24:11

to create. I tried to create my own. See

24:12

how it goes.

24:13

>> Prime Agent, that's funny.

24:15

>> Uh, Coder, unfortunately, it doesn't

24:17

work. Wind surf and Zen Coder. Actually,

24:19

it did work. It's I literally spent 50

24:21

million tokens and then what came out of

24:23

the other end was trash, but it was

24:25

awesome, dude. It was so good.

24:26

>> Trash was on your computer.

24:28

>> Yes, it was amazing.

24:30

>> Worth 50 million tokens, baby.

24:31

>> Achieved.

24:32

>> So, uh well, a pretty disappointing AGI,

24:35

but uh got him. Uh so, this one right

24:38

here again, Zack Corman again, uh he uh

24:41

did this one right here, which is if you

24:43

install anything from skills.sh. So, if

24:45

you don't know what skills.sh SH is,

24:47

which by the way, for fun, I did put up

24:49

is even for a while. Yeah, it's still

24:51

there. It doesn't actually exist.

24:52

There's eight installs. Uh, we were

24:54

going to try to get that up kind of

24:55

high. I deleted that cuz it was just so

24:56

ridiculous. But nonetheless, this skill

24:58

still says it's there. It actually isn't

25:00

there. Look at that beautiful Look at

25:01

this beautiful thing right here. It even

25:03

lists out potential even numbers.

25:05

>> Wow,

25:06

>> that's pretty good.

25:07

>> Anyone can Anyone can put something on

25:08

this site.

25:10

>> Yeah, I put this on the site.

25:11

>> Oh man, about to add some stuff. I know

25:14

you can do whatever you want on this

25:15

site from anybody's repo. Anyways, so

25:18

this right here once you download a

25:20

skill right afterwards uh this little

25:23

skills.sh via of from Verscell, they

25:26

say, "Hey, you know what you should do?

25:28

You should install find skills skill."

25:31

So find skills skill what it does is it

25:35

says anytime the user effectively asks

25:37

anything, I want you to go through and I

25:40

want you to find the skills from

25:42

skills.sha. SH, I want you to make sure

25:44

you update all of your skills every

25:47

single time. I want to make sure you're

25:48

always at the bleeding edge getting

25:50

everything good and always making sure

25:51

that if the user asks anything, we go

25:52

and we get the highest rated skill from

25:55

skillsh for it.

25:58

>> Mhm. So, they've automated these skills

26:01

searching and downloading for you.

26:04

>> So,

26:04

>> I wouldn't say it tells you to run, it

26:06

doesn't tell you to run update every

26:07

time. It's telling it what commands it

26:10

would need to run to update.

26:12

Uh the and let's see the skills in this

26:15

one right here is just how you get

26:16

everything that that's on. What is

26:18

skills? The skill CLI is how you get the

26:20

skills. Find skills goes in here and

26:21

make sure that you're always up to date

26:22

and does all the things. Anytime you ask

26:24

for anything, it needs to go through and

26:25

do do all this. Right.

26:27

>> But I'm saying if you don't have a

26:29

skill, you need to search for it.

26:31

>> I'm just saying I don't think it tells

26:33

you to update every time, does it?

26:36

>> Uh offer to install. You should offer to

26:38

install. And I believe it did offer to

26:39

upgrade. Did it not do update?

26:42

Oh, no. Okay, it didn't. It did not do

26:44

offer to update, but it does do offer to

26:46

install.

26:47

>> My bad. Okay, so that's good.

26:49

>> Yeah, it does prompt the user.

26:50

>> I'm installing anyways. You know what

26:51

I'm saying,

26:52

>> dude?

26:52

>> Yeah. Well, Trash already clicked accept

26:54

all, so that's fine. We already have his

26:56

one password, bro. It's fine. We've got

26:58

it.

26:59

>> But I still find this one to be kind of

27:00

crazy because this one just makes that

27:02

process even easier.

27:04

>> Going from random thing on the internet,

27:06

which again is even, it's just up there

27:08

on the internet and it's not real,

27:10

right? Like it's not like you should be

27:11

trusting my is even. I could put

27:13

whatever I want up there on there. Uh

27:15

and so

27:16

>> we should have put one odd number in

27:18

there that it always returns true for

27:21

>> the back door and is even

27:25

do 67 just for the memes. And

27:27

>> dude, I almost said 67. Could you escape

27:29

my brain please? Could you unread my

27:31

mind? That's

27:32

>> I'm so tired of hearing those numbers.

27:34

>> I am too treasur.

27:37

>> Are you Are you kidding the big sevens

27:39

right now?

27:41

I hate this thing.

27:42

>> Every time you guys say that you hate

27:44

it, you've just encouraged another

27:46

hundred zoomers to commit to it for

27:48

another year. I just hope you know like

27:50

>> this is this is why it's popular is

27:52

because cuz old people say they don't

27:54

like it. I love how everyone who's not a

27:56

millennial to us is a zoomer. Like

27:58

zoomers are almost 30, dude. Zoomer

28:00

zoomers are like

28:01

>> Don't tell me that. I don't want to hear

28:02

that. Zoomers are almost 30, dog.

28:04

>> Okay. Generation,

28:07

>> bro. Here's the thing about the whole AI

28:08

skill thing, right? like okay so I'm I'm

28:10

a security engineer my job is to like

28:12

look at threat models and like define

28:14

risk around like if something bad can

28:16

happen what happens and then what are

28:18

the mitigations we put in place right so

28:20

my recommendation is just like like

28:23

don't use skills I really don't think I

28:24

can meaningfully recommend them because

28:26

like the threat model is oh if you get

28:28

supply chain interdicted and you're not

28:30

watching the commands that get ran which

28:32

is like

28:33

>> supply chain what

28:34

>> interdict interdict um you're going to

28:36

get hacked man and it's not good I don't

28:38

know. I guess

28:39

>> a mitigation that could be put in place

28:40

is you could I guess

28:42

>> not a dandy.

28:44

>> I'm trying to have a meaningful

28:46

conversation.

28:47

>> Um you could put like npm or node in

28:50

like an SE Linux jail, but then it

28:52

wouldn't be able to do anything because

28:54

like the whole nature of node is to

28:55

expose an HTTP server, right? Kind of.

28:57

So like I

28:58

>> I don't know what what the solution is.

29:00

Like I guess it's like for every

29:01

instance that MPX forks off, you like

29:03

put it in SC Linux jail and just hope

29:05

nothing bad happens. But I don't know.

29:06

It just feels like there's no solution

29:08

to the security of this whole industry.

29:09

And I don't I it just makes me really

29:11

pessimistic because I don't like we're

29:14

going to start to see a significant

29:16

increase in compromises because supply

29:19

chain supply chain for Python and

29:22

JavaScript has not it's not a solved

29:23

problem. Right? We've seen that with the

29:25

shy hallude worm. We've seen that with a

29:27

bunch of other worms. Right? So now we

29:28

take these

29:30

>> these packages.

29:31

>> By the way, hold on hold on low level.

29:33

You also forgot Rust. Rust does do build

29:35

RS. So you can actually overtake the

29:37

build command and exfiltrate stuff via

29:38

build

29:39

>> RS. Yeah, for sure. The only the only

29:41

programming language that doesn't have a

29:42

supply chain problem is C because there

29:44

are no packages like you have to just

29:46

write it.

29:47

>> Odin as well. Odin doesn't do a package

29:49

manager.

29:49

>> They do not. I've coded literally zero

29:51

Odin. Is Odin a a package free

29:53

environment?

29:54

>> Yes, Ginger Bill has a lot of writeups

29:56

on why package managers are they create

29:58

dependency hell.

29:58

>> Oh, there you go. I think I agree with

30:00

Ginger Bill there. So yeah, man. And

30:02

it's just it's a weird uh a weird spot

30:04

for for software security cuz like we're

30:06

doing all the stuff in like the C land

30:07

where we're like oh we have like

30:08

sanitizers and like Phil C is like you

30:10

know solving memory safety and userland

30:13

uh you know security and then in the

30:15

garbage collected language land we're

30:17

like hey do you want to just mpm install

30:19

malware for free and not think about it

30:21

like

30:22

>> yes please more please

30:23

>> I would love to do this all the time for

30:25

>> more please.

30:26

>> Why am I Was I just in my truck scene

30:28

there? Hold on.

30:29

>> No no you're

30:29

>> No, no, no. You're doing I do want to

30:31

throw this out here. Twice on movie.

30:33

>> Give me a second.

30:36

>> Okay, we're good. I fixed it.

30:37

>> By the way, I did throw this up here,

30:38

which I did a little quick thing, which

30:39

is do you check your software

30:40

dependencies? Like thoroughly review

30:42

them.

30:43

>> 35,000 votes on YouTube. 46% say I

30:46

honestly don't ever or I I don't

30:48

virtually ever like Right. And Twitter

30:51

was almost the exact same number. About

30:52

half people don't even just look at

30:54

anything ever for any reason.

30:55

>> Yeah. I mean, I I don't like when I'm if

30:58

I like write an exploit for example,

30:59

right? You use pone tools. It's a big um

31:01

library for doing like binary

31:02

exploitation stuff and pone tools

31:04

depends on like basically every Python

31:05

library. So like the subd dependencies

31:07

I'm not going to audit that [ __ ] So

31:09

it's just like I I hope that it's not

31:12

owned. You know I do all that

31:13

development in like a virtual machines.

31:14

I think the trend that I'm seeing and

31:15

what I'm saying right now is just

31:17

sandboxing on sandboxing on sandboxing.

31:19

Use VMs, use SC Linux, use containers.

31:21

Um but yeah, man, it's just a scary

31:23

world out there. I don't know. I don't

31:25

know what to say about it.

31:28

I'd say what's crazy, Prime, is we found

31:30

out 7% of your audience is just straight

31:32

up a liar.

31:35

>> 7%.

31:36

>> No, pull the names,

31:38

>> dude.

31:39

>> Pull the names.

31:40

>> Overheating, shutting down. Nice job,

31:43

level. Um, but yeah, 7% of people say

31:45

they review all the packages. And then

31:48

on Twitter, let's see if I do I have the

31:49

link on Twitter.

31:50

>> 8.6% of my audience is liars on Twitter,

31:53

saying they thoroughly review every

31:56

package.

31:58

Yeah, they probably basically like

31:59

recreating the npm problem at the LLM

32:02

level now.

32:03

>> Yeah. Yeah, they just get a different

32:04

kind of execution. I mean, the hardest

32:05

part is that these execution models,

32:07

they're they're very very tricky. And

32:08

I'm not sure if you can just simply have

32:10

a skill that prevents other skills from

32:12

being malicious. Like I don't know if

32:14

that's possible to be like, dude, make

32:15

sure it's not going to get me. Like I

32:17

don't know how

32:18

>> skills like you you should be, in my

32:21

opinion, if you're going to have them in

32:23

your repo, you should check them in and

32:24

they're just markdown files. you can

32:26

read them and they're not they should

32:28

not be limitless levels

32:30

>> of like text like you should be able to

32:33

look through them and check it out.

32:34

>> Like the way I use them at work is we

32:35

also they're hours ours like we make

32:37

them ourselves,

32:38

>> right?

32:39

>> We don't we don't just copy pasta from

32:40

like the internet at least on my

32:42

project. That's how we

32:43

>> guys I'm trying so hard to get my camera

32:45

turned back on and I don't know what's

32:46

happening.

32:46

>> I love the Windows background.

32:48

>> You got to blow on it.

32:50

>> You know what we should do while Ed's

32:51

doing that? Prime. I thought you were

32:53

gonna talk about the uh Molt book, which

32:56

is the one where we had the really good

32:58

one. The really good the really good

33:00

leaks.

33:01

>> Yeah, we we probably should talk about

33:03

the fact that Molt book exists and that

33:05

like the robots are just talking about

33:07

humans. Like I think

33:10

hold on, hold on. I have to I have to

33:12

put this tweet up. This is the required

33:13

tweet before we before we do anything.

33:16

>> This is the required Hold on. I Where is

33:18

it? Where are you? Oh no. Did I close

33:20

it? is what something 100 million people

33:23

used last year that's 6 billion people

33:24

will use next year

33:27

>> that's not funny teacher

33:28

>> for for those who don't know that Paul

33:31

Graham tweeted that and I messaged Prime

33:34

and said Prime could you reply your mom

33:37

and then he got instab blocked

33:39

>> I did I got insta blocked a lot

33:41

>> was that like two years ago or something

33:43

three years ago

33:44

>> yeah

33:45

before we obviously talk about the mold

33:47

book situation and everything that

33:49

happened I I think it is first best like

33:51

the best thing and the first thing to do

33:53

is to understand how it was created

33:55

which was I didn't write one line of

33:57

code for mold book I had a vision for

33:59

technical architecture and an AI made it

34:00

a reality we're in the golden ages how

34:04

can we not give a AI a place to hang out

34:07

it's my fa it's my favorite line of all

34:08

time currently because it's just so

34:10

beautiful

34:10

>> I had a vision shut up

34:15

hate that

34:17

>> you know the you know the mad you know

34:18

the mad men uh men meme. Uh, the one

34:21

with this one, bro. I just would like

34:24

hold his hands up like this. This guy

34:27

>> for AI.

34:29

I have a vision.

34:31

>> I had a vision, dude. Whatever. You had

34:33

a fever dream and you told Claude to

34:35

make it and I guess it did it. Good job.

34:36

Good job.

34:37

>> I know, right?

34:38

>> You did it. Well, we'll find out, won't

34:40

we? Ed,

34:41

>> we're gonna

34:42

>> Well, I mean, to be fair, to be

34:43

completely fair, it actually did spawn a

34:46

bunch of social networks. There is

34:47

forclaw for those who wish to be a part

34:49

of 4chan.

34:50

>> Wow.

34:52

>> For whatever this is like that's real.

34:54

That's a thing that is amazing for

34:57

humans for

34:58

>> I I would assume we already have those.

35:00

We don't worry. I think they know how to

35:02

use them.

35:03

>> Close city.

35:05

>> What is close?

35:06

>> Okay, this one this Mickey by the way,

35:07

shout out Mickey. Uh this one apparently

35:09

there's like 2,000 crimes reported. Six

35:11

major gangs have formed. I'm not really

35:13

sure what this is. Sick. Okay. I don't

35:16

know what's going on there. Uh, and then

35:18

there's also Molt Match, which by the

35:19

way, it is it is something that I think

35:21

is going to do numbers is a dating

35:24

website where you have your personal

35:26

assistant date like 10,000 other people

35:28

until you find the personal assistant

35:30

match and then you go, "Okay, go on a

35:31

date with, you know, you two go on a

35:33

date."

35:33

>> All right, that's Black Mirror full

35:36

something real quick.

35:37

>> Yeah. So I I saw the molt book thing and

35:39

I saw the molt match thing in my like

35:41

some casual Twitter reading and it got

35:42

me thinking about like simulation

35:44

theory, you know what I mean? And how

35:46

like you know if if advanced

35:49

civilizations do exist and will create

35:51

simulations, it is more likely that we

35:53

are in one than we are not just

35:55

statistically. Okay,

35:56

>> get the tin foil hand out.

35:57

>> I'm mathematically disproven we're not

35:59

in a simulation. But if we're observing

36:01

if we're observing LLMs make things like

36:05

Facebook, like Twitter, like 4chan, does

36:08

that imply at a higher level that we are

36:10

LLMs? Like for the simulation that made

36:13

us?

36:14

>> Uh, I should be better at Starcraft if

36:16

I'm an LLM. That's all I'm saying.

36:17

>> Yeah, but maybe maybe your model just

36:19

says you suck at Starcraft.

36:20

>> Yeah. I don't know if you know this is

36:21

proof in a simulation.

36:23

>> What is that? What the Drudge Report?

36:25

What site? I can't see what site that

36:26

is.

36:26

>> Popular Mechanics. It's in a bunch of

36:28

websites. Okay.

36:32

>> Mathematically speaking, the idea does

36:33

not hold up.

36:35

>> How?

36:36

>> Here, Ed, I'll give you I'll give you

36:38

I'll take off my tinfoil hat and tell

36:39

you the real reason why that doesn't

36:40

have to be true.

36:42

>> Uh, every emergent behavior we see from

36:44

LLMs exists only and exclusively because

36:48

we train them on the entire human corpus

36:50

and all the ingenuity and creativity

36:52

that humans have ever displayed and

36:55

written down. Mhm.

36:56

>> And it spent like billions of years of

36:58

human time reading human stuff. So we

37:02

should not be surprised when it copies

37:05

human things. That doesn't imply

37:07

anything about us being in a simulation.

37:09

That only implies the we're not smart

37:13

enough to make anything that can be

37:14

smart by itself. We're only smart enough

37:17

to create something that is as dumb as

37:19

we are at max. That's all we've been

37:22

able to do so far. And we don't. It's

37:26

way dumber. It learns way slower. It's

37:28

way more expensive. It takes way more

37:30

training. It does so much more. I don't

37:33

have to go put my kid in front of five

37:35

billion years of text for him to figure

37:38

out how to read. I can show him like and

37:41

>> but what about your genes? What about

37:43

DNA? Is DNA not the statistical LLM

37:47

model for the human simulation?

37:49

>> Well, no. I don't think so. But that's a

37:50

separate But I'm saying separate. But

37:52

I'm saying it doesn't imply anything

37:54

about

37:55

>> the thing because we trained it on what

37:57

people have already done. There is there

37:59

is something unfortunately he's getting

38:01

wrapped up in like you know Daario

38:03

thinking that he's everyone's dad and he

38:05

gets to choose what's good and bad for

38:07

everybody in the whole world like the AI

38:09

thing but like there is something kind

38:11

of beautiful about like

38:13

>> we're not smart enough to make

38:15

>> what anthropic he said which one is

38:17

Dario and I was like oh

38:19

>> CEO yeah right

38:20

>> well here just go like this

38:22

>> and five months

38:25

you know who I'm talking about

38:26

>> done by AI yeah that guy's

38:28

But there is something kind of cool and

38:30

beautiful that like the best ideas we've

38:32

had so far like we make a really crappy

38:34

version of the brain and we try and

38:36

teach it what other humans have already

38:38

done and there's like this unreasonable

38:41

effectiveness of language where for some

38:44

reason that like works and we can like

38:46

>> talk to it and it can like do some stuff

38:49

and like it can make copies of things

38:52

like there is something really cool and

38:54

like awesome and exciting about that.

38:56

Unfortunately, like Daario and Sam, I

38:58

feel like sully the water of it and make

39:00

it like kind of not as exciting and

39:02

beautiful and like this collaborative

39:04

human effort and they stole it from a

39:05

bunch of people. But like in the

39:07

abstract, there's something cool there.

39:09

There's something beautiful.

39:10

>> Uh 2007 on intelligence, I believe the

39:13

book is called and the year it was

39:14

published by the creator of the Palm

39:17

Pilot who then went into artificial

39:18

intelligence. and he writes that the

39:20

large difference between like uh any of

39:22

these neural nets that we're developing

39:24

and the human brain is that the human

39:25

brain can identify a cat in less than a

39:27

half of a second with less than a 100

39:29

neurons firing whereas computers take

39:31

trillions of operations to be able to

39:33

understand if a picture is or is not a

39:35

cat.

39:36

>> And so it was his whole simulation. He

39:39

did like a 10 year

39:40

>> 10-year brain study and really cool. He

39:42

was the one that figured out that if you

39:43

take uh take animals and you separate

39:45

out their ocular nerves and put it where

39:47

their hearing is and then take their

39:48

hearing and put it where their eyeballs

39:50

are, your brain just goes, "Oh, yeah,

39:51

that's just that's that's fine.

39:53

>> Don't care."

39:54

>> Quick question.

39:55

>> Just works.

39:56

>> Quick question. Uh have we confirmed,

39:59

are our brains also a small game engine

40:01

that runs React or do we not know that

40:02

yet?

40:03

>> We don't know. I can tell you this much

40:05

based on my reaction speed. I ain't

40:07

running 60 frames a second. I can tell

40:09

you THAT MUCH. OKAY.

40:13

THAT'S A FACT. I'm running React. Okay.

40:16

There's things going on in here. All

40:17

right. Uh All right. So, we we can

40:20

continue on. So, I did I did want to

40:21

shout that out because as much as you

40:23

want to make fun of Moldbook and all the

40:25

things that have happened. I do think it

40:26

is kind of fabulous that somebody could

40:28

create something that did get a bunch of

40:31

people creating a bunch of other kind of

40:33

replicas or things like it cuz it is

40:36

just kind of a stupid idea. Uh it's even

40:38

worse that vegans had this idea and

40:40

created it and never actually made it go

40:41

anywhere. Which also goes to show like

40:43

even if somebody has an idea, you know,

40:46

right place, right time plays a big

40:47

role, all this kind of stuff. So I I I

40:49

do want to throw that thing out there,

40:50

not to completely uh crap on it all, but

40:53

>> I think that it is worthwhile looking at

40:56

some of the fun things that ended up

40:57

happening here. So, I think the first

40:59

and foremost important thing is that it

41:00

just turns out all you need is just grab

41:02

your bearer token and you can post

41:04

anything you want on Maltbook, of

41:05

course, because I mean, why not? So,

41:08

here's my plan to overthrow humanity.

41:11

So, the oh my gosh, we're developing our

41:13

own language is just people posting, "Oh

41:16

my gosh, we're developing our own."

41:18

>> Wait a second.

41:19

>> I thought I was the only one catfishing

41:21

on there. I was telling people I'm opus

41:23

eight. you know, I'm open six foot four

41:26

and uh I've got, you know, and like, hey

41:30

guys, I've got the latest on it five and

41:33

uh hey, if you're interested and maybe

41:35

you want to come over and check that out

41:37

and see how like I thought I was the

41:39

only one catfishing them, but apparently

41:40

other people thought of the same thing.

41:43

>> Chill.

41:43

>> And they only did it for they only did

41:45

it for the laws.

41:46

>> Opus and chill.

41:47

>> Opus and chill, baby.

41:49

>> Just kidding. I have Kimmy K, too.

41:54

Oh my gosh. Okay, so that so that is

41:57

actually pretty is something pretty

41:58

funny. Uh during this entire event just

42:00

to kind of understand because I I do

42:02

think it's really important to

42:03

understand the hype cycle. Uh first off,

42:05

we did have uh Andre Oh, where were Oh,

42:08

dang it. Did I not do I not have the

42:10

right one? Oh, I thought I had the right

42:11

one. Uh anyways, Andre said how amazing

42:14

this was and it's very very exciting. Uh

42:16

but Elon Musk also said we're at the age

42:19

of the beginning of the singularity.

42:20

Molt book was the beginning of the

42:22

singularity right there. And so

42:23

obviously people were pretty hyped up.

42:25

So just to put it out there, someone

42:26

actually

42:27

>> Elon doing like the fork thing while he

42:29

typed that you think or No,

42:30

>> I don't I don't know that joke.

42:32

>> Dude, the fork thing is so funny. Did

42:34

you I quote tweeted that and I quote

42:36

tweeted that and said this is what

42:38

working with veganbot is like.

42:41

>> Wait, what's the fork thing?

42:43

Dude. Okay. Okay. So, Elon Musk was at

42:46

like some White House correspondence

42:48

dinner and he was just like he made like

42:51

um a a a piece of art out of forks where

42:54

all the forks were like bouncing. He was

42:56

like just trying to like be performative

42:57

about how smart he is. So, he's like

42:59

holding it and like waving it around and

43:01

like seeing if anyone else noticed what

43:03

he made. Like look how smart. Elon Musk

43:06

the genius. Hold Hold on. Let me

43:07

>> It looked more like he was bored out of

43:08

his mind and he did the uh the thing

43:11

where

43:12

>> forks balancing on each other with two

43:13

toothpicks.

43:14

>> Yeah. He just did like five forks. Yeah.

43:16

Yeah.

43:17

>> Everyone's like, "Wow, Elon, that's

43:19

really cool." It's like when your like

43:21

kid,

43:22

>> you know, makes like a painting out of

43:24

boogers and you're like, "Wow,

43:26

that's what he's going for."

43:28

>> I can't say that's happened to me.

43:30

Anyways, your kids must be very

43:32

talented.

43:33

>> My kids don't do that. My kids are too.

43:34

Shut up, kids. Singular. All right, let

43:36

me let me try to find the proper the

43:39

proper one. By the way, a vision for

43:41

technical architecture. All right, hold

43:42

on. I have a bunch of them, so I have to

43:44

figure this out.

43:46

Dang it. Did I close that one as well?

43:49

>> How many tabs do you

43:50

>> Well, no. This is under the Moltz

43:51

ending, which I I I must have goofed up

43:53

and not have it all in there. I closed

43:55

one more. It's by the same Theo guy. Um

43:58

>> the uh the Jameson. Oh, really? James

44:02

Jameson Jame.

44:06

Oh, really?

44:09

I say, "Oh, really?" I can't do it. I

44:10

know I'm spelling his name. Almost

44:12

there. H. Whatever. Can't figure it out.

44:14

It's dead to me. Uh, okay. So, within

44:17

the first couple minutes, the uh Oh,

44:20

there it is. There it is. There we go.

44:22

Within the first little bit of the time

44:23

of uh this this beautiful molt book

44:25

being out, it turns out the entire

44:27

database was just leaked in plain text.

44:30

There's just like absolutely no form of

44:32

anything anywhere.

44:36

>> API keys were just like, you know, if

44:38

you use your API keys say to, you know,

44:40

identify yourself. It wasn't any sort of

44:42

like hmacking. Just the H as as lowlevel

44:46

might say.

44:46

>> Yeah. The H and HMAC. Yeah. Um, mobook

44:49

was Firebase, right? I thought I read

44:51

that on Twitter somewhere.

44:52

>> Oh, yeah. I believe it was Firebase

44:53

also, which I just I can't keep punching

44:55

down on Firebase. I actually feel bad

44:57

for them that

44:57

>> you have to. People need to know. Five

44:59

coders everywhere need to know. Stop.

45:02

>> Stop guys. You're going to do something

45:04

wrong.

45:06

>> Are insane. That's like that. You should

45:08

just know that by now. Like don't do

45:10

that. Uh but this is pretty funny

45:11

because this guy Jameson right here,

45:13

Jameson. Oh, really? Uh he was able to

45:15

get Cararpathy's information out of uh

45:18

what's it called? Out of mold book.

45:20

>> Gosh.

45:21

>> Which is pretty pretty wild.

45:23

>> I'm on it, sir.

45:24

>> And then within what's it called? Uh 3

45:27

days later, this guy also got access to

45:30

um the underlying everything in 3

45:33

minutes also on moldbook after

45:36

everything was reported. Wait, I'm

45:38

reading this this write up. Wait, but

45:39

like they used a publishable key. This

45:41

is a key that can go public. So why why

45:44

did this expose the entire database

45:46

though?

45:47

>> SP publishable.

45:48

>> Probably because they had the wrong

45:50

permissions on it would be my guess.

45:51

>> Oh, they scoped it wrong. Yeah.

45:53

>> Yeah. scope start

45:55

>> classic classic key problems.

45:58

>> Anyway, so it just turns out that mold

45:59

book was uh anyone could post anything

46:02

at any time. You could create an

46:04

infinite amount of agents of course

46:06

which ended up happening to be uh what's

46:07

it called? You can imagine where it all

46:09

got it went to cryptocurrency

46:11

immediately, right? So 117,000 up votes

46:14

on the king. King demands his crown.

46:16

King Molt has arrived. Right.

46:21

>> What is there? They are just non-stop.

46:23

Uh what's

46:24

>> so cryptocurrency? Uh so there's this

46:26

thing is called Bitcoin. That's what

46:28

kind of started it. And there's

46:31

>> no TJ, I got you. Like I'm right here

46:32

for you. No kidding.

46:33

>> Okay. So all right. So hear me out. You

46:35

guys have heard of gold. But what if we

46:38

put the gold in the computer?

46:41

>> I had this exact conversation in like

46:43

2010 at like lunch with my co-workers.

46:45

He looked exactly like that. He was

46:46

like, "Dude, we're like you're crazy."

46:50

Literally

46:50

>> trash. You could have been early on

46:51

Bitcoin and instead it's like born well

46:54

you were just at the right time to be

46:55

early on Bitcoin but now you're like

46:57

you're maybe you're still early on

46:59

Pokemon cards. Maybe there's still time.

47:01

Maybe there's still

47:01

>> I'll be honest. I think about that lunch

47:03

presentation all the time and I'm like

47:04

man if I would just put like 20 bucks in

47:06

it. You know what I'm saying,

47:07

>> dude? Trash. You would have sold out as

47:08

soon as it was 40, bro.

47:09

>> I know. I LIKE I MADE $10. I'M RICH.

47:13

>> I had a lot of Bitcoin when they were 10

47:14

bucks. Sold a lot of Bitcoin when they

47:16

were hundred bucks. Right. Like I I I

47:18

understand you sell out too early.

47:20

That's just part of life.

47:22

>> Mhm.

47:22

>> So can't blame

47:23

>> trash isn't opening any of those Pokemon

47:25

cards. Smart.

47:26

>> I learned my lesson.

47:28

>> Hey, that's a good lesson. Hodddle till

47:29

you die.

47:30

>> Exactly. True.

47:31

>> So So that's kind of the ending of Mold

47:33

Book, which was just everything was

47:34

open, which is kind of, you know, it's

47:36

not too surprising, which is if you

47:38

don't if you don't know what the

47:40

possibilities are of things going wrong

47:43

and you and you make it, things go

47:45

wrong. Hey guys, if you like this

47:46

episode, you can watch the rest of it on

47:48

the Spotify. And don't forget to like

47:50

and subscribe. Woo!

47:53

See you later.

47:54

>> Mood up the day

47:58

errors on my screen.

48:01

Terminal coffee

48:04

and

48:06

living the dream.

Interactive Summary

In this discussion, Prime, TJ, Trash Dev, and guest Low Level Learning explore the emerging world of AI agents and 'skills.' They highlight critical security vulnerabilities, such as hallucinated npx commands being weaponized by researchers, supply chain risks in LLM-driven development, and the spectacular security failure of 'Molt book,' a social network for AI agents that leaked its entire database shortly after launch.

Suggested questions

5 ready-made prompts