We own everything except for the

intelligence. Why can't we own the

intelligence? So, I got super into local

AI back in January. And so, I've been

building this kind of home AI lab. I

bought three Mac Studios with 512 GB

[music] each, which now I think are

rarer than uh yellow diamonds at the

moment. I built a computer around a 5090

RTX. So, I have a ton of GPUs, bunch of

Mac minis, running local intelligence

and all them. Fable gets banned. You now

have unlimited 247 intelligence. You can

create what I'm calling ambient

intelligence. You can't do these types

of use cases I'm about to show you with

Opus [music] with Fable because you're

going to have a $20,000 bill at the end

of the month. I built this dashboard. It

monitors my entire fleet of local

intelligence. [music] Something that's

not possible with Frontier. What are the

few biggest calls that you kind of see

coming true over the next let's call it

six to 12 months?

>> Uh I think six months from now.

>> I know something that you have been

preaching for a long time now is the

idea of sovereign AI and setting up uh

intelligent models locally like

literally physically in your office or

your space. I have so many questions

about this from why why is it valuable

to set up models locally versus just

kind of work with stateofthe-art models

that kind of they own the intelligence

um and also it feels very unattainable.

Um so like what does actually look like

to to get this all set up in practice?

>> Yeah, I've uh I've gone hard to this and

it's it's blown up over the last few

weeks or so because of some things going

on with Frontier Intelligence. So, I got

super into local AI back in January. I

discover OpenClaw. That was kind of my

big tipping point moment with my

content. This is really, really

interesting having kind of your own

personal AI agent that knows everything

about you, that has all these long deep

memories about you, but it's connected

to this intelligence that lives in the

cloud that's controlled by someone else.

And if someone decides, oh, I'm gonna

take this model away or I'm gonna change

how this model works, it kind of

destroys how your personal agent works,

right? So, you have this agent that

lives on your computer, but the one

thing that runs it all doesn't live on

your computer and lives somewhere else.

And so, I got super into local AI and

being able to kind of own the entire

stack of your AI agent, right? We own

everything except for the intelligence.

Why can't we own the intelligence? And

so I've been building this kind of home

AI lab over the last several months. I

bought three Mac studios with 512

gigabytes each, which now I think are

rarer than uh yellow diamonds at the

moment. Um I bought uh DGX Spark. Uh I

bought a I built a computer around a

5090 RTX. So I have a ton of GPUs, bunch

of Mac minis, running local intelligence

and all them. And then, you know, I've

been preaching this for the last several

months on Twitter and YouTube. You gotta

get in local AI, getting local AI. I

think a lot of people made fun of me for

it. And then there was this big moment

a few weeks ago that I think started

shifting the public perception of local

AI. I think it was two major events. One

was Fable getting banned, which thank

god it's back. It came back yesterday.

Fable gets banned. people kind of

realize, oh wait, I don't have as much

control over my intelligence as I

thought I did, right? These companies

can just take it away at any given time.

You've literally no control over it

whatsoever if it's running through the

cloud. And then GLM 5.2 comes out, which

is I think the first kind of open

weights model that you can run locally

that basically comes pretty close to

frontier. A lot of benchmarks have it at

Opus 48 level or so. I'm running it

locally right now in a Mac studio. It is

up there with Opus 48. I can tell you

that much. And so now the public

perception is changing. But I still

think there's a lot of misconceptions. I

still think expectations aren't really

set right. And so uh there's still a lot

to go over. There's still a lot to do.

I'll show you a lot of what I'm working

on here. But I think that finally the

public perception is starting to shift

on local intelligence and people are

starting to see okay now I actually do

need to own my intelligence.

>> I just want to understand so one are you

exclusively using local models or are

you also using frontier models as well?

>> Yeah this is where people get confused

and they I tweet like oh Fable is

incredible and then I get like slam dunk

in the replies. Oh, but you're telling

everyone to buy Max Studios. It's like,

no, that's not it at all. We are not at

a point right now where you can be 100%

dependent on local models. There are

still downsides to local models that

Frontier models fill in nicely, right?

Local models slower than Frontier,

especially if you're running basically

anything less than an RTX 6000, right?

If you're running Mac Studios, slower

local intelligence. Um, they're also not

as smart, right? Local models, like we

had this big breakthrough in GLM 52, but

it's still not quite as smart. It's not

as smart as Fable. It's getting close to

Opus 48, but it's, you know, it's still

behind Frontier Intelligence. Um, and

then also, you know, it's expensive

buying more and more computers, but so

there there is still downsides. So the

perfect combo is

using them together using Frontier

according to its strengths and using

local according to its strengths. And

there's also another part to it, right?

Local is still great. There's still many

use cases. I'm going to show you many of

the use cases on this call. At the same

time, this is also where people really

struggle. I I don't know why the general

a lot of people struggle with this. You

also have to have foresight, right? You

also have to be able to look a little

bit in the future. Talk about local mods

and go, "Oh, but it's slower. It's

slower. It's slower. It's slower." Yeah,

but look at the trends, right? Look at

the way things have been moving the last

couple years. A couple years ago, local

models are completely unusable.

Completely unusable.

Now, they, you know, several months ago,

they were six months behind and now

they're like three months behind. So, if

you're able to kind of have foresight

and not be a prisoner of the moment, you

can see they're getting better, they're

getting faster, they're getting smarter,

and if you, you know, project that out

maybe 6 months, maybe a year,

>> you can pretty safely say we will have

fable level intelligence locally at

pretty decent speeds. And at that point,

then yeah, we can have a conversation

around it replacing Frontier. But no, at

the moment, right, you use them

together. There's strengths to each and

we'll go over those strengths in a

second. But part of also getting into

local AI is investing in the future.

Investing in on the bet that local

intelligence get better, faster,

smarter, and more efficient where you

can run it on cheaper computers.

>> Yep. And and you're doing this

personally. Do you think this is also

going to be a trend for businesses?

>> I I mean, I have zero doubt. Uh again we

got to look at trends. We got to look at

the way things are moving. What are the

strengths of local models? The strengths

of local models are basically just costs

the electricity going into your

computer, right? You're not paying for

tokens. There's no toll booth like you

do on the cloud. So you have all these

companies like Amazon and Meta that are

now like, "Whoa, whoa, whoa. Stop this.

Let's burn as many tokens as possible."

No, that's not the goal anymore. The

goal is to be efficient, right? So cost

savings is now becoming a thing for the

first time ever with AI. Well, local AI,

it's only the cost of electricity. Two

is privacy,

right? Your data when you use the cloud,

whatever it is, you're building with

clawed code, you're building with

codecs, all your code's being exposed to

anthropic and open AI. And at the same

time, codes being and listen, I love

anthropic. I love open AI. They're

building life-changing technology, but

at the same time, they're releasing

competing products with their customers.

So, you're giving your data to these

companies who are then turning around,

and I'm not saying they're looking at

your code and then, you know, building

products based on that. I'm not accusing

them that at all, but what I am saying

is they just happen to be releasing

competing products at the exact same

time.

>> Yep. And so privacy

>> is becoming super important. And when

you run local models, your data does not

leave your computer, right? Your data

stays on the computer, doesn't get

exposed to people releasing competing

products. And so those two major trends,

I think it's pretty safe to say almost

on a guaranteed level within a year or

two, local LLMs running on edge hardware

is going to be a primary way companies

are using AI. Where do you want to start

with taking us through how to build

local?

>> Let me do this. Let me show you first

the computers I have, the decision

making behind those computers, what the

strengths of those computers are for

kind of,

>> you know, the people watching. Then I'll

go into what exactly I'm doing with

those computers and the use cases and

strengths around that. Uh, but let's

first answer kind of some questions that

people have. What is the computers you

run this on? What can they do? What

computer should I buy? I think that's

like the number one question I get is

like what computer should I buy to run

local models? These are the computers I

run local on. So, as you see, I have

three Mac Studios with 512 gigabytes

each. For those not as quite up to speed

with local models, basically the memory

is the big component here. The models

are loaded into local memory. Um, so

like when you're on a Nvidia computer,

right, you're running an Nvidia chip,

it's loaded into the VRAMm. If you're on

a Mac, it's loaded into the unified

memory, uh, and it's loaded into the

memory. And so memory is the big thing

here. If you have Mac Studios, the

strengths is it has a lot of memory. So

I have three of those. And then I also

have Nvidia computers. The Nvidia, the

advantage is kind of the the

architecture around it. Uh, the CUDA

architecture. Um, and it makes it helps

run the models extremely fast. Running

on Nvidia is much faster. So, I have

different computers for different

purposes.

>> And what's the cost of these?

>> The cost of these? Well, four months ago

when I bought them, the cost of these

Mac Studio 512 gigabytes were $10,000

each. Now, there's been a lot of changes

in the market. First of all, you can't

buy them anymore. They don't even sell

them. Um, they don't sell these

computers anymore. You can only buy Mac

Studios, I think, in the 96 gigabyte

configuration. So, a fifth of the size.

Um, I told everyone, hey, listen, I told

everyone in January, you're not going to

be able to buy these soon. You should go

out and buy them.

>> Yeah.

>> The prediction turned out to be 100%

correct. Now, if you want to buy them

resale on like eBay or wherever, they're

going for four times the price. You're

paying $40,000 each. I think like

$30,000 used. So, the prices of these

have gone up. Still obtainable though

are uh the DGX Spark. The DGX Spark is a

fantastic computer and definitely a

great place to start for a lot of

people. That has recently been raised to

I think $4,800 from $4,000.

You can still run good models on that.

And then I built this computer myself.

Um I'm also a little bit of a gamer on

my uh in my spare time late at night. So

I wanted to build a nice gaming computer

with the 5090 in it. And this cost about

$9,000 to build. The 5090 is a $4,000

GPU.

>> Got it.

>> So you you add on memory and all the

other components, you get to like 78

$9,000.

>> Yeah. Yeah. So we're basically talking

about like, you know, let's call it

30ish,000

of hardware that you built up to kind of

like have the infrastructure you need to

run your local models.

>> Exactly. Y

>> Exactly.

>> Cool. And then for those wondering at

home kind of okay what's the strength of

each as I was talking about before Mac

studio high unified memory so you can

run much larger models right I could run

GLM52 on a single Mac studio with 512

gigabytes which again opus level

technology low bandwidth though and what

that means is it's kind of an indicator

of the speed how much uh of the model it

can run at one time so it's going to be

slower then you have the AI computers

which are your plug-andplay DGX X

Sparks. You plug it in. It's an AI

workstation. These kind of have medium

unified memory and decent speed. So, you

can run decent size models and get

pretty good speed with it. This is

probably at this moment in time with the

price of Mac Studios. This is probably

the sweet spot for most people are these

AI computers like DGX Spark. And then

you have your powerhouse chips. These

are like your really powerful GPU that

you had to build computers around. the

RTX5090s, the 6000 Pros, you get lower

VRAM, although the 6000 Pro, it's a

$15,000 chip, so you actually get good

VRAM and you get very high bandwidth.

So, you get extremely fast speeds. I

mean, faster than Cloud Frontier models.

Um, this is good if you want like all

right intelligence, but at blazing fast

speed.

>> So, you can't run like GLM52 on that.

>> You cannot. Requires too much space.

But, you know, again, we're talking

about trends here. You can run smaller

Quen 36 models.

>> Yep.

>> 6 months ago, a Quen model that's like

29B, you're not getting good

intelligence. Today, Quen 36 29B, which

runs on a 5090,

you're getting like Sonnet 4 level

intelligence, which today, yeah, it's

not the greatest ever. When Sonnet 4

comes out a year ago, six months ago,

whatever it was, that was pretty

amazing.

>> And now you can run that unlimited

locally at blazing fast speeds. So, you

know, again, let's talk about trends

here. Right now, yeah, the intellig

before intelligence is bad. Now, it's

pretty good. You think about a year now,

a year from now, what you're going to be

able to run on a consumer GPU you just

put into a computer.

>> Again, we're probably talking fable

level intelligence a year from now

running locally at blazing fast speeds.

And I have a fourth category here, which

is dusty old computers. I get all the

time, oh, you know, that Mac Mini I've

had for a few years, can I run local

intelligence on that? Technically, you

can. Technically, you can run local

intelligence. Google's actually been

doing a spectacular job at this. They've

been putting a lot of energy and

research into small, efficient models

that run on basically any hardware. And

so you can run like a Gemma 4 model on a

Mac Mini. Is it Frontier? No, it's

nowhere close. But it can do small

vertical tasks that can help you out. It

can be the memory system for your Open

Claw, right? It can do embeddings,

things like that. So there are still

purposes. And that's why I tell

everyone, listen, no matter what device

you have, you need to be getting into

local AI. Will you be able to run the

greatest models ever? Maybe not, but

it's still worth getting into because as

the trends are going six months from

now, year from now, no matter what

device you'll have, you'll be doing

you'll be able to do something that

creates economic value for yourself.

>> Love it.

>> Now, let's talk about use cases.

So,

we talked about frontier intelligence.

We talked about local models, the

strengths and weaknesses. All right. So

if local models

are slower, yeah,

>> if local models are not as intelligent,

>> then why the hell would you use them?

>> Well, the answer is pretty simple. If

you have unlimited use of intelligence,

even if it isn't the smartest

intelligence ever,

that unlocks a whole new breath of use

cases, right? If you now have unlimited

247 intelligence,

you can create what I'm calling ambient

intelligence, which is intelligence that

is constantly monitoring, constantly

reacting, and constantly burning tokens,

right? Even though it's slower and

dumber, you can't do these types of use

cases I'm about to show you with Opus,

with Fable, because you're going to have

a $20,000 bill at the end of the month.

So, you have this ambient intelligence.

What does that mean? Well, first of all,

I built this dashboard myself. Well, my

Hermes agent built it. Uh, this is

basically what I call my fleet

dashboard. It monitors my entire fleet

of local intelligence. And these models,

this local intelligence is doing work

247, 365 around the clock. Something

that's not possible with Frontier. What

are those local models doing? Well, you

can see kind of my computers here that

are all running. You can see the local

models running at the moment. I'm

constantly switching these out,

experimenting, running different

evaluations to see what can do what, but

they're all doing different things all

the time around the clock. All of these

models. So, one example is security. I'm

building a SAS right now called Henry

intelligent machines. one of these local

models every hour or so is running a

loop where it is picking up a different

API endpoint in my apps. I also have

another SAS creator buddy looking at the

different API endpoints and just doing a

quick security check. Is there anything

wrong here? Is there any security issues

that I should flag? Things like that.

Just constantly doing security

surveillance. I have another local model

that's just looking at different parts

of the code every 20 minutes. Anything

we can make more optimized, anything we

can clean up, anything like that. I have

another local model that's constantly

looking at the database for anomalies,

you know, is there any users using too

much usage? Anything weird going on? Are

there churn risks I can address and send

an email to automatically? constant

ambient intelligence, just watching all

parts of my business to find

opportunities or things to fix. Again,

not possible with Frontier unless you

work at one of the Frontier Labs because

now you're paying thousands and

thousands of thousands of dollars a

month um to to use these tokens. When we

first started talking about this, I

think people are like, "Okay, Alex, like

I get it." But when you say like

sovereign AI is a big thing, I think

like until people feel the pain of

intelligence that they don't own being

taken away from them, that won't

necessarily land with them. And to the

point like Fable got taken away, but you

know, you still had Opus 4.8, you still

had Sonnet. And so it's like the

alternatives were still solid. And so I

think people are still probably like

okay but does sovereign AI really matter

especially for individuals. I think for

companies who are dealing with sensitive

information totally different argument.

I think people will feel this one way

faster which is if you want to actually

have 247 intelligence and have all these

different jobs for you

>> before you either needed a team of a ton

of people which would have cost a lot of

money or if you tried doing it right now

with whether it's Fable 4.8 8 GPT 5.5.

Either if you're on a pro or max plan,

you'll end up maxing out your usage very

quickly and you'll have to wait until

limits are reset. Or if you're on an

enterprise plan, it'll end up basically

costing the salaries of engineers you

hire just to run these jobs. So you're

basically saying, how can you do the

work you want to do but not have it be

expensive right now,

>> right? I mean, it's just leverage,

right? you're creating leverage for

yourself as an individual and you look

at the way the economy is going at the

moment, right? There's there's some job

loss going on right now, right? You need

to be able to create value for yourself.

And if you're a soloreneur, you're

building something for yourself. The

more ambient intelligence, the more

workers you have going for you working

around the clock, the more leverage you

have against your competition. And so

this is just more and more leverage.

Even if it's really simple tasks, just

doing a security check every 20 minutes,

just reviewing small lines of code every

few minutes. It's just more and more

leverage, more and more productivity and

work getting done for yourself that you

really didn't have unlocked before. And

it'll only get worse. I mean, look at

the other trends, right? Big thing here

is trends. Where's the world moving?

It's not all about what's going on in

the world right now. It's about where's

the world moving. Fable is not going to

be a part of subscriptions starting July

7th. Starting July 7th, you're going to

need to pay for every single token input

and output for Fable. The

counterargument of, oh yeah, but I pay

$20 a month for subscriptions, yet

you're going to be paying per token

soon. With Local AI, you're not paying

per token. You're paying for what? Going

into your computer, which is a very

different formula. So that's another

just huge strength of this is that you

you you have this unlimited intelligence

as well.

>> Do you have a sense of for the amount of

workloads you're running on all of your

local um machines right now? What is

this costing you um a month and then if

you were to have run all these workloads

on the state-of-the-art models, how much

do you think it would be costing you

roughly? If I was to run this on

Frontier models, I have no doubt it

would be thousands of dollars a month,

right? Uh I I know that the Claude Max

$200 a month plan, I think the metrics

came out to get like $3,000 of tokens a

month. I'm

>> maxing out that plan with Claude, but

I'm doing way more work with local. So,

it has to be at least a few thousand a

month. So, that's from the cost. How

much is it costing me from in

electricity? I'm not a 100% certain to

tell you. I mean, I think my costs I'm

in California, so electricity is

expensive anyway. I think I was paying

like $300 or $150$160 a month for

electricity before. Now I'm paying like

$200 220. So maybe like $60. So I mean

it's it's very large

>> totally

>> cost. And people go, "Oh, but you're

paying tens of thousands of dollars for

computers."

>> You're right. You do pay for some

upfront, but you know, that's the

investment into this. And I only believe

you'll be able to run better and better

models in the same hardware you own over

time.

>> Yeah. And the cost of the hardware is

going to come down over time as well.

Like the hardware is the most expensive

it will ever be for the quality that it

is now. So you'll maybe have to pay more

for better quality hardware or you can

have the same cost hardware but way

better in the future. So that makes

sense. And then the other question is

how are you figuring out with these

local models and how far you can push

the work that you're giving your local

stuff versus having to give to the

frontier models. It's actually not easy

right now to know like right you have

all these things like open router that

are routing based on the jobs needed.

But for an individual to figure out

which workloads they should be using

with which models is like a ton of trial

and error and it's very murky. How are

you figuring that out?

Yeah, I mean I treat my Hermes agent.

So, you know, I have OpenClaw and

Hermes. Hermes is kind of my frontier AI

agent. That's like my IT guy, my Hermes

agent. And so, when like new models come

out, I'll go to my Hermes. I'll say,

"Hey, go on to all my computers

areworked through tail scale. I'll say,

"Hey, go on to my DGX Spark, load up

these list of like five different

models, run an eval on each one, and

then give me a report on what their

strengths and weaknesses are." I come

back a couple hours later, and I have a

full report. It tells me, and because my

Hermes model knows me so well, it knows

my businesses, it knows my tasks, it

knows what I do on a momentto moment

basis. It goes, "You should be running

this task on here. You go on your Mac

Studio, run this task on this model, and

then go on your 5090 and run this task

here." And so I'm offloading most of

that work to my Hermes agent which

understands on a very deep level my

entire network, all my hardware, all my

tasks and can make those decisions for

me. Uh that's I mean that's another big

part of this as well I wanted to go over

which was like what else do you need?

What do you need to to make this run? If

you have tail scale and Hermes, for

those who don't know, tail scale

basically allows you to create your own

private network. As long as you have

tail scale, which connects all your

devices, and then Hermes or OpenClaw,

which is basically like your chief IT

guy who can then move around between all

your devices. You can figure all that.

It just figures it out for you. It'll go

on the devices, figure out which models

should run on each. Do research on

Hugging Face, what are the best models

at the moment, load that in based on

what it knows about your hardware, you

know, uh, load it into memory for you so

you can use it. And then run evals and

tests so you know which ones are best at

what. And so if you have these two

things, tail scale and Hermes, you can

do any of this and figure out what works

for you. So, basically the setup is you

have your uh your three Mac Studios, you

have two two Mac minis, a DJX Spark, and

then the Nvidia uh GPU.

>> Yep.

>> So, I'm all hooked up to my office. Only

one of them is connected to a monitor.

Only my Mac St. Well, my uh Nvidia one's

connected to this other monitor so I can

play Cyberpunk 2027 at night. But other

than that, they have they're all just

sitting there connected to an outlet.

None of them are connected to a monitor.

Y

>> and but they're all on tail scale. So I

can go to my Hermes and I can say, "Hey,

go to my Mac Studio 2, go to my Mac

Studio 3, go to my DGX Spark, load these

models up, run all my tasks on them, see

what does it best, and it just goes over

and does those things for you on those

computers. You don't even need to see

them connected to a monitor ever again

in your life."

>> Yeah. And so the whole purpose of tail

scale like and creating a private

network just so I understand what it

kind of means why why it's important in

practice is

>> is it so that you can basically

when you run a job does that basically

mean based on the workload that's needed

it can kind of just route to your

different machines based on the like te

take me through why it's important to

even set up a private network versus

just have these all run individually.

Why would that be harder? So putting

them all on the same private network

basically allows all your computers to

communicate with each other very easily.

Yep.

>> And so what that enables is if you have

a Hermes agent, I have it running on my

Mac Studio 1, which just basically my

only computer connected to a monitor.

>> I can say go on my DJX Spark, load the

latest Gwen 36 and move our workload

over to that. And what it'll do is

because they're all on the same network

through tail scale, it can basically

access the shell of the DGX Spark, so

basically like root admin access,

>> communicate with it from the Mac Studio

1, and from there since everything's

connected to a CLI now, be able to run

the CLI on the DGX Spark, download uh

the model, load it onto the server, and

start giving it tasks. So it basically

gives all your computers root access to

each other. So if you have any agents

running, they can all communicate with

each other, run anything they want, and

do any workloads across any of your

devices.

>> How much of this whole setup did you

have experience in before doing this?

>> Zero. None. I I had experience in none

of this. You know, I I think I think the

most valuable skill right now is the

ability to be autodidactic.

>> Yeah. uh the ability to just figure

things out on your own. And it's never

been easier to just figure things out on

your own because all you need to do is

ask your agent, hey, how the hell do I

do that? And if you get in, if you can

train yourself to, you know, most people

in this world, I'd say 95%, and I don't

want to kind of underestimate humanity.

I love humanity, but I'd say 95% of

people when they run into a challenge,

they roll over and give up.

>> I'm trying to build this thing. I need

it on a database. I don't know how a

database works. I give up. I quit.

That's how like 95% of people operate,

unfortunately.

>> But if you can train your brain now that

when you run into issues, I have all

these computers. How do I connect them?

I have this computer over here. I have

but I don't have a monitor connected to

it. How do I do things with it? If you

change your mindset from I don't know

how to do something, I give up to I

don't know how to do something. I'm

going to go to AI and just figure out

how to do it. You can just accomplish so

much more. I' I've never computers in my

life.

>> Yeah.

>> None of that. I never I still haven't

loaded an AI model locally in my entire

life. My Hermes does it all for me. But

if with AI now, you can really just

figure out anything you want to do in

the world.

>> Love it. Are there any gaps you want to

fill in on the local side or do you

think we've covered it?

>> I think we covered most of it. I mean,

there's, you know, I talked a lot about

kind of the technical side of what you

can do with local models, right? Doing

security checks, looking at code. You

know, there's a lot of other really

interesting, cool things. Even if you're

not a tech guy, even if you're not vibe

coding, I have my models every hour

going scraping Twitter, scraping Reddit,

uh scraping hacker news, product hunt,

looking for signals, looking for trends,

looking for challenges people are having

and searching for business

opportunities. Again, this is another

one of the kind of power of ambient

intelligence is my intelligence going

out and just watching the internet all

day, finding business opportunities for

me to act on, finding SAS for me to

build that solves challenges, finding

content I can write or create based on

what people are asking about. And so, no

matter what you're doing, there's always

advantages to having more intelligence.

There will never be a lack of demand for

the more intelligence you have at your

fingertips the more things you can do.

And so no matter what you're doing,

intelligence will only make your job

better and give you more leverage. And

so there's just a million different

things you can do with it. I know uh

you've generally been pretty right in

making calls. Like you talked about

making the call and the the Mac studios

ended up not being able to buy them

anymore. or you've made other calls

around like local which obviously is

becoming bigger and bigger with Fable

and now as the open source models are

just getting better. What uh what are

the few biggest calls that you kind of

see coming true over the next let's call

it 6 to 12 months? Uh I think 6 months

from now you're having Fable level

intelligence running on consumer

hardware, right? Like high-end Mac

minis, MacBook Pros be able to run that.

And so you need to be able to come up

with now

what use cases you can do within

intelligence that runs 247 because

that's not something people really do

right now. Right? Right now when people

think AI, it's call and response. I have

a question, I go to the AI, it responds.

this concept of AI watching everything

you're doing and reacting

proactively for you before you can give

a prompt or ask a question that doesn't

exist. So I would say the big thing is

now I think most people within the next

year will have intelligence running

locally on their devices. And so you

need to think about when that happens

when you have access to that technology.

What systems can you set up? If you have

your own private community, I know you

have an AI community yourself, right?

How can I offload a lot of my duties

with that community to an ambient AI

that maybe watches every single message

in that community, then reports back to

me, and then maybe creates a video or a

guide or a PDF explanation of what

everyone's talking about in that

community, right? That's a use case

right there you could never do before,

but unlocks with ambient technology. You

need to be able to plan right now what

are those new use cases that are

unlocked to ambient intelligence that

you can start planning for so that when

you do have it running on your devices

you can just plug it in get it going and

now you have a huge advantage against

your competitors.

>> Love it. So the recap on this and I

think it would be helpful for people

because to your point people are going

to have to think in a way they never

have because they're going to have

access to intelligence that can work

24/7 365 which was never a possibility.

And so as I think about like what are

the criteria for good use cases or or

jobs for ambient intelligence to do it's

kind of like what is a job that involves

always on kind of monitoring and

checking where that would just be kind

of expensive from a token uh burning

perspective for just like a traditional

frontier model. What is something that

like whether it almost doesn't matter

the time of day for it to be done. it is

going to be valuable for you regardless.

Um, and then what is something that does

not require frontier level intelligence

in order to complete the task? Are those

kind of the the main buckets if

someone's trying to think about what you

talked about like security bug fixes or

signals? Are those the main ingredients

for kind of finding the use cases for

ambient intelligence?

>> I mean, a good way to think about it is

like this, right? I'll give you a kind

of an example, then I'll give you an

exercise you can do. Right? A good

example of ambient intelligence I think

people understand kind of intuitively is

the humanoid robot. That's going to be

an example of ambient intelligence that

exists in the next year or so where

everyone's going to have a robot walking

around their house, walking around their

apartment, taking care of tasks, kind of

low intellect task very easily, doing

the dishes, folding the laundry, making

your bed, right? And and so that's going

to be the same thing but on computers

with knowledge work. Yeah. But when it

comes to local models, right? So instead

of a robot doing the dishes, putting

away clothes, folding things, it's going

to be a robot, you know, proactively

looking at your emails, your calendar.

One thing you can do is a exercise I

give a lot of people, tell them to do. I

write down my all my to-do list are on

pieces of paper on my desk, right? I got

all the tasks I do. Spend a day

writing down on a piece of paper every

task you do. Right? Everything you do in

a day, write it down what you're doing,

right? Oh, I'm I'm managing my

community. I'm writing posts on this.

I'm writing tweets. I do this YouTube

video. I program. I build this. I

respond to this. You put it all down.

Then you can go to a model. Go to Fable

if you want. I find Fable's the best,

obviously, the best model to talk to

right now. go, hey, you know, I want to

prepare for ambient intelligence. I want

to prepare for, you know, being able to

run local models. Here's everything I

did today.

What can I offload to ambient

intelligence? What could I offload to

local models? If I had right now a DGX

Spark, how would it make my life easier

with everything I do here? And you

reverse prompt it and you go, this is

everything about me, what I do. Now tell

me what I should be doing and how this

could help and you will find a lot of

different things you can be doing with

it.

>> Love it. This was great, Alex. Such a

good breakdown on local AI, why it's

becoming more important, how to actually

set it all up from the hardware to tail

scale to Hermes agent to kind of the

exercises of finding tasks for ambient

intelligence to do. So I know I learned

a ton. I know our viewer will appreciate

you taking the time and excited to have

you on again in the future.

Happy to join anytime, man.