So, the entire internet seems like it's

losing its composure over Claudebot. And

what I wanted to do was just kind of

explore what's going on. Now, I spent a

little time kind of reading up on it and

figuring out what is it, how does it

work, and that sort of thing. So, let's

just start at the top. Claudebot is a

semi-autonomous

personal agent that the primary

difference is that it is proactive. It

finds stuff to do rather than just

waiting for you to give it commands.

Now, this is a non-trivial

uh problem and it's one that I had been

working on for quite a while and I'll

talk about my previous work um on this

uh in a little bit. But the long story

short is that uh people are freaking out

about it because this is the level of

autonomy that the a the agentic browsers

were supposed to promise, but those were

built in more of a uh let's say a a

paradigm that was more corporate

friendly. Whereas Claudebot is open-

source and it's rogue. It's very

renegade. And the reason that it emerged

in the open-source space is because of

the lower risk. It's saying, "Hey, this

is open source software. Use at your own

risk. If it deletes, you know, all of

your emails or, you know, buys a million

platton tickets to Tahiti, that's on

you." Whereas, you know, the Comet

browser and the Open AI browser, they

had to make sure that it couldn't do

those things. So, this is where the

open- source movement has a very

distinct advantage over corporations,

over closed source, because it's just

like, I'm just going to release this and

see what happens. So, people are

freaking out about it. Um, this this

tweet in particular was hilarious. So,

Flowers, this is this is the same person

who was formerly uh Future Flowers um

said, "We should give Claudebot Minute

Man 3 access for a fast fun alignment

test."

Um, so the idea is like, okay, like what

kinds of guard rails does this have and

what, you know, what are the safety

risks? And on a technical level, what a

lot of people are concerned about is

that you have uh an AI that's constantly

running. It's got ports open, so it's

hackable, so it's a security nightmare.

Uh, so that's one thing that people are

worried about. But you can run it on a

local PC, you can run it on a Mac Mini,

you can run it on um a micro PC and then

it just becomes your just little shotgun

ride along buddy. You could put it in a

container. Um if you containerize it,

you can run it on anything including

your phone. So this is very obviously

the direction that things were going to

go. Uh and the reason that like all the

all the primitives have been worked on.

So there's a few technological

primitives and and what I mean by

primitive is like the basic building

blocks. So the technical primitives that

we've been working on over the last few

years and I say the royal we like all of

us um is uh so number one models that

were capable of agency so you know

taking tasks or uh uh taking taking

commands solving tasks solving problems

um and that sort of thing. Tool use was

one of the next big ones. So tool use

the ability to use JSON the ability to

use APIs the ability to even say I don't

know how to use that API. Let me let me

find the documentation. So a lot of

those autonomous like those th those

fully autonomous tasks that were done in

service to userdirected tasks uh were

like kind of some of the building

blocks. Uh memory management so

recursive language models that

innovation has been a big contributor

because one of the biggest problems is

not you know uh what like what can you

do it's remembering what should you do

what does this person need. So things

like retrieval augmented generation was

kind of the first uh version of

long-term memory, but it was really

unstructured. It was basically just a

soup of memory um and and had very very

little structure. So recursive language

models are a better more structured uh

memory management system. And so this is

this is similar to uh what I what I

worked on a few years ago called Remo.

Um which was what was it? uh it was

recursive emergent memory optimization I

think is what it stood for. I I worked

on I worked on agentic memory systems

for a while. Um so for for those that

have been around for a while you

remember when I was working on that and

when I was working on things like Nala

and the ACE framework and I'll talk

about that in just a minute. So um the

where we're going with this is this was

very obviously the path forward. So over

the last couple years, I've made a few

videos talking about how uh just due to

efficiency, this is what the market

would demand because as soon as the AI

was capable of being fully autonomous,

people would build things that were

fully autonomous and then the market

would demand things that were fully

autonomous. So this is and I started

saying that because you know 2 three

years ago when open AI and and anthropic

and everyone else is saying humans are

going to be in the loop and it's just

going to empower humans and it's going

to be an empowerment tool and I called

BS on that because that's not how

technology works. You don't get to

decide what a technology does ahead of

time. there's emergent capabilities and

when when you when you take a step back

and you say we are clearly building

something that is a thinking engine and

what's the difference between being able

to carry out a task that you tell it to

carry out and autonomously carrying out

that task? Well, you just need something

to specify the task. So this is what

like that that's really what an agentic

framework is is you have one module or

one motor or one loop that's saying okay

what's the most important task to do

next? So that's one loop and then you

have the outer loop which is go do the

task and so then you have other services

that manage the memory and that sort of

thing. I remember when when I had built

a small team and we were working on the

ACE framework. Um one of the team

members implemented it in such a way

that it started doing things

autonomously and he's like he's he shut

that off. He's like well I didn't like I

didn't like that it was coming up with

its own ideas. I want it to only do what

I tell it to. I'm like I think you're

fundamentally missing the purpose of

autonomy. And he's like, "Well, no, but

we need user stories." And I So I was

like, "Okay, this this team clearly

doesn't get it." Um or that that person

on the team clearly didn't get it. Um so

I was like, "All right, you guys can do

whatever you want."

Anyways, so long story short, I've been

in the space for more than four years

now. I wrote uh natural language

cognitive architecture four years ago.

And so I'm really excited. Um, it took

longer. It took both less time and more

time than I would have hoped for this

kind of stuff to to be out there. Um,

and I'm glad that I was there to help

help it along. So, let's see. Um, so

this is this is my original work. So, I

I wrote natural language cognitive

architecture uh four more than four

years ago. Um, and this was basically

with GPT3. I realized that this was the

agentic framework. So um I'm not saying

that this is exactly how Claudebot works

because again this was a while ago and I

you can't predict everything but this

was the theory that I had. So you had

the interloop which would do a search

space uh create a kernel which is

basically what do I do then build a

dossier. So that's a task uh

specification and then load that into a

shared database. And this is pretty

close to how Cloudbot works. Um, where

there's like a tasks.mmarkdown file and

then you have um you have a few other

other shared things. And the the

innovation here and I was actually

working towards that was just put

everything in plain text. You don't need

a database just put it in plain text

because that's that's what the language

models read. And so everyone has settled

on do it in markdown. And then the outer

loop is actually the the task execution

loop. So this was my first idea uh where

you have you know the the shared

database. So that's kind of your uh

recursive language uh model. So so

that's your that's your context

management. That's your shared tasks and

that sort of thing. And so then you

build a context. Uh so by extracting

from that you build a corpus which is

basically recruiting all the information

that you need uh to execute the task.

Then you do the task and you output the

task into the environment. So that's

your API calls and that sort of thing.

And then that updates uh so you get

feedback. So you get the input

processing and output loop. That's the

outer loop. And then the inner loop is

the task manager. And so this was

actually pretty pretty salient. This is

pretty close to how how Cloudbot was

ultimately um implemented. So the next

uh layer that I was working on was the

ACE framework. So ACE framework stands

for autonomous cognitive entity and it's

a more sophisticated hierarchy. And

basically doing a side-by-side

comparison, Claudebot uh actually does

all but the aspirational layer. So to

provide a little bit of more context,

the entire theory of the ACE framework

was that you'd have um hierarchical

layers. So different processes that were

responsible for different aspects of you

know uh basically making stuff happen.

So you have the global strategy layer

which is the environmental context and

your long long time horizon planning.

The agent model which is basically a

list of what the what your agent can and

cannot do. Um as well as so like it has

to understand what it is. So like I

understand that I'm Claudebot and here

are my tools, here are my hands, here's

my memory, here's how I work. Uh because

and this was important at the time

because uh language models had a lot

less baked in in terms of what they were

and what they were capable of. So we had

to explicitly state you're a language

model this you're a part of the ACE

framework and that sort of thing and the

agent forge team has taken all this and

run with it. Um so they're still they're

still chugging away as far as I know. Um

then the executive function which is

risks resources and plans. Um as far as

I know the cloudbot focuses more on

plans and maybe resources. I don't know

if it has a risk control layer. Um but

that would be a really easy thing to

add. Uh next is cognitive control which

is task selection and task switching. So

cognitive control is about saying like

I'm failing at this task. I need to I

need to either cancel this task or try a

different method or that sort of thing.

Uh the inspiration for this layer was

actually frustration. Um the point the

the the neurocognitive point of

frustration is to tell you that what

you're doing isn't working. And so you

get frustrated enough you either quit or

you try harder or you try something

else. So that's basically what the

cognitive control layer does. And then

you finally have the task prosecution

layer which is actually executing a

specific task like call this API, make

this calculation, write that function,

that sort of thing. Now many many tasks

actually require um all of these layers.

So this was basically kind of what what

a lot of people in the project and and

people who were observing it at the

time, they said you're basically

describing the like an org chart of a

company. Um, and so some people actually

represented it as like like floors of an

office building. And so then you have

many many small agents taking on each

role in that floor and they all talk

with each other. And so then you have a

northbound bus and southbound bus. So

the northbound bus is basically feedback

from the environment. So this is the

green bar here is your interface with

the outside world. So that's APIs,

that's telemetry, that's anything that

you have uh control over um or get input

from from the outside world. So that

information needs to be disseminated to

all the agents and layers. And then the

southbound bus is command and control.

Now, one of the things that is missing

from Claudebot is an aspirational layer.

So this is this is one of the main

critiques that a lot of people have have

have created or or said is that Cloudbot

doesn't have like its own Supreme Court

to decide like you know does this does

this abide by our mission values or our

mission parameters or universal ethics

or that sort of thing. And so the

aspirational layer is about morality,

ethics and overall mission. So that's

that's this is very similar to what you

would say is like the constitution. So

like Claude's constitution serves as an

aspirational layer and this has been

really the c the centerpiece of my work

since I got into AI safety which is the

heristic imperatives. I'm really glad to

be talking about the heristic

imperatives again. So the heristic

imperatives are what what I came up with

after studying morality, ethics,

philosophy, game theory and that sort of

thing which is okay if you have a fully

autonomous uh machine what are the

highest values that you should give it

so that it stays aligned with humanity

and pro-life and that sort of thing. And

so that was um reduce suffering in the

universe uh increase prosperity in the

universe and increase understanding in

the universe. So I came to those three

values by figuring out like what are the

what are the deonttological

like like what what are the most

universal deonttological values. So

that's a duty based ethics which is

saying like from where I'm at today what

should I try and achieve and so many

people confuse like the paperclip

maximizer is an example of a teological

thing. It's like the the the best

version of the universe is the one with

with the most paper clips. So that is

purely a teological

version of morality or ethics or mission

or purpose. Whatever whatever um

whatever intervates something, whatever

gets you going uh whereas uh a more

deontological thing is from where I'm at

today, what do I have a duty to do? And

so like a duty to protect. So this is

this is where Asimov worked with the

three laws of robotics which is um you

know robot may not harm a human. So

under like whatever whatever the outcome

is, whatever the long-term outcome is,

do not do any actions that harm humans

and then do not do any actions that

allow harm to come to humans and that

sort of thing. And then of course later

in the foundation the zeroth principle

was actually your goal is to preserve

life. So ultimately something like

Claudebot will need an aspirational

layer and I would recommend the heristic

imperatives. So uh reduce suffering this

is a this is a very pro-social and

pro-life heristic which is basically

most intelligent animals will recognize

the suffering in other animals and try

and inter intervene. So you'll see this

where like um you know uh elephants and

and other animals if they see another if

they see another animal in distress

they'll try and help. Generally speaking

animals will try and help each other

because they recognize that that

distress that that suffering is bad. And

so I I said suffering specifically

because there's a difference between

pain. Pain is instructive. Pain is like

you need pain to understand what hurts

you. But suffering is non-addaptive.

Meaning suffering is just pain that has

no real purpose. So suffering is

generally bad. Um now you from an

artistic perspective some people say ah

suffering creates art which is you know

that that's a defensible assertion. You

look at Vincent Van Gogh and you know he

suffered a lot and he created great art

and that sort of thing. Um but that

doesn't necessarily mean that it is it

is teologically good. Um and also you're

never going to eliminate suffering. So

another thing is what I've established

is a vector. So reduce suffering not get

it to zero not eradicate suffering just

reduce it. Just just control suffering.

Next is prosperity. Because when you

have a heristic direction that says

reduce suffering um the best way to to

reduce suffering is to reduce life

because the less life there is the less

suffering there is. So then it took a

while to figure out the term prosperity.

So to counterbalance reduce suffering

you then have a value of increased

prosperity which is uh a very univers

universal word. It comes from Latin

prosperatoss which means to live well.

Literally the the root of prosperity

means you want to live well. You want to

flourish, you want to thrive. So you you

reduce suffering, you increase thriving,

you increase flourishing in the

universe. And you it's universal because

all life depends on all other life. Now

that's not universally true. There are

parasites. There are things there are

life forms that are basically just

harmful. Um but at the same time every

every life lives in the trophic level or

in the ecosystem and occupies a

particular niche. So then the final one

was because I realized well just those

two values you could end up with a green

earth that has no intelligence on it

that has nothing that no curiosity no

expansion and so then I realized that

the the core objective function of

humans that makes humans different is

that we are curious. So I said how do we

encode a curiosity

uh like algorithm into into an

autonomous machine whether it's AGI or

ASI or your personal cloudbot as it

turns out to be and that was to increase

understanding. So curiosity is the

desire to know for its own sake. So the

desire to understand for its own sake.

So rather than saying curiosity, which

is I just want to know things, right?

Because pure curiosity, unbridled

curiosity can lead you to do things

like, you know, section 31 and torturing

frogs just to see what happens. There

was actually an episode, this was also

inspired by an episode of Star Trek

where there was this galactic entity

that wanted to experiment on the crew of

the Enterprise just to see what would

happen. And so that was that was a an

ethical uh dilemma showing pure

curiosity for its own sake can actually

be destructive and it can be harmful and

cause suffering. So however you do want

to understand and that's one of the core

drives of humanity. It's like hey what's

over there? What's on the horizon? How

do I get across this body of water? Why

does fire happen? Our curiosity is what

sets us apart. So then I said okay we

have created this higher layer of

organization and how do we encode that

into a machine now plenty of people cuz

you know if if you followed me for a

while you know that heristic imperatives

people are talking about it people have

put it into agents I'm really excited

and the reason I'm making this video and

I didn't know that I was going to make

the video this way but the reason I'm

making this video is because I think we

have a really powerful opportunity to

now take that pri prior work you know

the the agent forge team worked on um

the ACE framework team worked on and

anyone who's tried to implement natural

language cognitive architecture the core

heristic imperatives would be I think

really great to add to something like

Cloudbot. Um so yeah, I guess that's

that's where I'll leave it. I think I

was working on um a notebook LM to

understand it, but it looks like it

failed. Um anyways, so I'll leave it

there today and um yeah, thanks for

watching. If I sound different, it's

because I'm practicing using my voice

differently. Um, I uh I mentioned on a

private video for my fans that um that

my voice usually gets really tired after

an hour or two. And that had me really

worried because I'm going to be

narrating my book. Um and and I was

like, I'm I don't know if I'm going to

be able to do this. And they said, oh,

like you need to practice using your

voice differently. And so I was doing a

little bit of research and I realized

that um I learned to use my voice in two

primary places. One was singing chorus

and then two was leading meetings. And

both of those require a tremendous

amount of energy and projection which

you don't need to do when the microphone

is literally a few inches from your

face. So if I sound weird, if I sound

different, it's because I'm practicing

using my voice differently so that I can

entrain the narrator voice. And that's

that's was the inspiration. I was like,

I need to make a video today. And

Claudebot is blowing up. So, thank you

for being here. Um, and I'll end by

plugging um all of the ways that you can

support me. Um, and a little bit of

news. So, we're very close to getting

the Kickstarter up for the post labor

economics book. Um, I actually do need

to do a little bit of recording for

that. I think um we've got a editor

chosen. So, this is an editor um who has

uh been the copy editor for like New

York Times bestseller. So, we found the

right guy. Um, the right genre, the

right the right talent, the right skill

set. So, the book is going to it's going

to really shine. So, there's that. And

then, um, the Kickstarter, the book, the

editor. Um, oh, yeah. And so, then, uh,

if you, if you're hanging around, what I

have what I have figured out for my fan

base is I've been trying to figure out

like the bonus content strategy for a

while. And what I have settled on, what

I'm trying right now is just fireside

chats. So just more unstructured

fireside chats. Um, and I'm posting

those on every platform that you can

subscribe on as a as a paid member. So

that's here on YouTube. That is Patreon,

that's sub uh Substack, that's Twitter,

um, and Spotify. So those are the five

platforms where you can sign up as a

paid subscriber and get those the bonus

content. And I'm going to talk about uh

it's going to be a little bit more

broader topic. So, some of it's going to

be personal updates, some of it's going

to be burnout, some of it's going to be

more philosophical. I don't know what

it's going to be, but it's basically

going to be if you want twice as much,

Dave, sign up for one of those platforms

and I will be posting um the insider

content there. And it's it's basically

the same format. It's going to be me

talking to the camera for 20 to 30

minutes. Um I'm also going to start as

my health improves, I'm going to start

hiking again, so you can get some of the

hiking videos and that sort of thing.

All right, with all that being said,

thank you for watching and I hope that

you guys enjoy and consider integrating

the heristic imperatives into Claudebot

and its successors and we'll go from

there. Cheers.