HomeVideos

The First Fully AI OS Just Dropped And It's Seriously Powerful

Now Playing

The First Fully AI OS Just Dropped And It's Seriously Powerful

Transcript

353 segments

0:02

Warm wind just released an AI operating

0:04

system and the easiest way to explain it

0:06

is this. You give it a job and it

0:09

figures out how to do it across your

0:11

actual apps. You basically give an AI

0:14

its own computer in the cloud, show it

0:16

how to do a job once, and from that

0:19

point it can keep doing that job for you

0:21

every day, even when your own computer

0:23

is completely turned off. It can open

0:25

Gmail, read customer emails, jump into

0:28

SAP, find an order, answer the customer,

0:32

process documents, monitor websites,

0:34

qualify leads, move between completely

0:37

different pieces of software and keep

0:39

doing all of it in the background. And

0:41

the really interesting part is that the

0:43

software doesn't even necessarily need

0:45

an API. Warm wind wants the AI to use

0:48

applications the same way a human

0:50

employee would, by actually looking at

0:52

the interface, clicking, typing, and

0:54

navigating through it. Now, I already

0:57

know the question a lot of you are going

0:58

to have here. How is this different from

1:00

something like Open Claw? Because Open

1:02

Claw can already give an autonomous

1:04

agent access to browsers, tools,

1:06

software, and persistent workflows. The

1:09

difference is really in how the whole

1:10

thing is designed and who it's designed

1:13

for. Open Claw gives technical users an

1:16

extremely configurable autonomous agent

1:18

that you can build around models, tools,

1:21

skills, APIs, MCP, and your own

1:24

infrastructure. Warm wind is trying to

1:26

turn that broader capability into

1:28

something much closer to manage digital

1:30

[music] labor for companies, where

1:32

instead of building the agent, you can

1:33

literally show it how the job is done.

1:36

And that distinction becomes much

1:38

clearer once you actually see how Warm

1:40

Wind works. The product is called Warm

1:43

Wind OS. It launches worldwide today,

1:46

August 26th, and we've been working

1:48

directly with the Warm Wind team around

1:49

the release and getting an early look at

1:51

what they've built. They describe Warm

1:53

Wind as an AI operating system, although

1:55

it's important to understand that it

1:57

isn't replacing Windows or macOS. The

2:00

better way to think about it is that

2:02

every

2:05

with its own applications, its own

2:08

environment, and its own job. You can

2:10

open Warm Wind in the browser, and they

2:12

also [music] have a desktop version and

2:14

a mobile app for monitoring,

2:16

controlling, and managing workers. Their

2:19

workers are designed to operate across

2:20

Windows,

2:21

>> [music]

2:21

>> Linux, Android, and web applications.

2:24

The worker itself is cloud-based, so it

2:26

doesn't depend on your laptop staying

2:28

online. If you give it a task at the

2:30

office and then close your [music]

2:32

computer, the worker keeps going. And

2:34

that matters because Warm Wind isn't

2:36

really designed around one-off

2:37

questions. The whole product is built

2:40

around repetitive digital labor. The

2:42

best demo they showed me is customer

2:44

support because it makes the idea

2:45

obvious almost immediately. You open

2:48

teaching mode, and instead of writing a

2:50

giant prompt explaining every possible

2:52

step, Warm Wind watches what you're

2:54

doing while you explain it. In their

2:56

demo, the user tells the worker to open

2:58

Gmail, go through incoming customer

3:01

emails one by one, understand the

3:03

question,

3:04

>> [music]

3:04

>> then use SAP to find the information

3:06

needed for the reply.

3:07

So, the user opens an email from a

3:09

customer, points out the company name

3:11

and order number, switches over to SAP,

3:14

goes into the relevant customer field,

3:16

enters the information, runs the search,

3:19

finds the order data, and explains

3:21

[music] that this is the information the

3:22

worker should use to answer the email.

3:25

That's basically it. You don't need to

3:27

manually define every click. You don't

3:29

need to write a script for every screen.

3:32

The worker is watching the interface,

3:34

listening to the explanation, and

3:36

building a work plan from the

3:37

demonstration. After teaching is

3:39

finished, Warm Wind compiles what it saw

3:42

into a recurring workflow. It shows you

3:44

the worklist it plans to follow, and

3:46

then you tell it when to run. In their

3:48

demo, they schedule it for 9:00 in in

3:50

morning every day. The cursor starts

3:52

moving on its own. It opens the first

3:54

email, jumps into SAP, searches the

3:57

account, gets the information, and

3:59

continues through the task. And this is

4:02

where the product starts feeling less

4:03

like a chatbot and more like giving

4:05

somebody a desk job. Warm Wind says the

4:08

worker can also ask for missing

4:09

information when it gets stuck. You can

4:12

interrupt it, add instructions while

4:14

it's working, and watch a progress

4:16

indicator that reflects how well it has

4:17

learned that workflow. Their idea is

4:20

that you supervise it more closely at

4:22

the beginning, and as it sees more

4:23

examples and handles more cases, it gets

4:26

better and faster at the job. Now,

4:28

obviously, this is also the part that

4:30

needs to prove itself in the real world.

4:33

Because teaching an AI one clean

4:35

workflow is one thing. Giving it 200

4:38

messy customer emails where half of them

4:40

contain weird exceptions [music] is

4:42

another. So, I wouldn't treat this like

4:44

some magical one-click employee that

4:46

never makes a mistake. Warm Wind itself

4:49

tells users to watch the worker at first

4:51

and step in when it needs information.

4:53

But the underlying interaction model is

4:55

interesting because [music] it's

4:56

basically training by demonstration. And

4:59

there's another important piece here.

5:00

Warm Wind doesn't require every

5:02

application to expose an API. That's a

5:05

huge deal for businesses because real

5:07

companies are full of software that was

5:09

never designed for AI agents. Old

5:12

internal portals, [music] ERP systems,

5:14

supplier dashboards, accounting

5:16

software, random web tools, legacy

5:19

systems, and custom interfaces that

5:21

somebody built 12 years ago and nobody

5:23

wants to touch. Traditional automation

5:25

usually works best when every system

5:27

exposes clean, structured access. Warm

5:30

Wind is trying to get around that by

5:32

having the worker use software the way a

5:34

person does, visually, with a mouse and

5:36

keyboard. Their vision language system

5:39

looks at the interface, understands

5:41

what's on the screen, clicks, types,

5:44

navigates, and moves information between

5:46

applications. The company says that was

5:48

one of the reasons they didn't want to

5:50

build the entire system around

5:51

general-purpose language models.

5:53

According to Warm Wind, when they tested

5:56

state-of-the-art language models for

5:57

this kind of continuous computer use,

6:00

they ran into three problems. The first

6:02

was visual understanding. A language

6:05

model can analyze screenshots, but Warm

6:07

Wind argues that most of these models

6:08

still fundamentally live in a text-first

6:10

world, while normal office work is

6:12

heavily visual. The second was speed.

6:15

They claim that using general

6:17

state-of-the-art models for raw

6:18

interface interaction could take around

6:20

30 to 60 seconds for a single [music]

6:22

click because the model isn't just

6:24

clicking. It's reasoning about the

6:26

screen, describing an action, and then

6:28

another system still has to execute it.

6:31

And the third problem was cost. Warm

6:34

Wind says that if you tried to run

6:35

continuous interface automation this

6:37

way, using top general-purpose models,

6:40

the cost could reach around 1,000 euros

6:42

to 2,000 euros, or the equivalent in

6:45

dollars per hour, which obviously makes

6:48

no sense for repetitive business work.

6:51

So, they say they took more inspiration

6:52

from robotics and embodied AI. Instead

6:55

of treating every screen like another

6:57

long reasoning problem for a giant

6:58

language model, they built a more

7:00

specialized system optimized for visual

7:02

interaction and [music] repetitive

7:04

tasks. That's also why their positioning

7:06

is closer to a digital workforce than a

7:08

general AI assistant. A general

7:10

assistant waits for you to ask

7:12

something. Warm Wind wants the worker to

7:14

have a responsibility. One worker can

7:16

handle customer support. Another can

7:18

process invoices. Another can watch

7:21

competitors. Another can qualify leads.

7:23

And these aren't supposed to be 5-minute

7:25

tasks. They're recurring jobs that keep

7:28

running in the background and only pull

7:30

you in when there's an exception, a

7:32

decision, or an approval. Their document

7:35

workflows can find files, download them,

7:37

review them, organize them, and upload

7:40

them into another system. Their process

7:42

workflows can start with an incoming

7:44

[music] request and keep working until

7:46

the request is completed. Monitoring

7:48

workers can repeatedly check software,

7:51

websites, or internal [music] systems

7:53

and report only when something changes.

7:55

And lead workers can find prospects or

7:58

candidates, enrich the information,

8:00

qualify them, and send the useful ones

8:02

forward. Warm wind is basically built to

8:04

take over the repetitive office work

8:06

people waste hours on every day. One

8:09

worker might handle invoices. Another

8:11

could deal with customer requests, while

8:14

others keep an eye on things like

8:15

recruiting, social media, competitors,

8:19

or whatever else you'd normally have

8:21

someone checking manually. The point

8:23

isn't really that every company needs

8:25

every one of these. It's that the worker

8:27

is supposed to operate across the same

8:29

messy mix of software that a human

8:31

employee already uses. Warm wind also

8:35

has native apps that are designed to

8:36

make common actions faster than relying

8:39

entirely on visual control, so the

8:41

system isn't religiously committed to

8:43

clicking everything on screen when

8:45

there's a better path available. But the

8:47

visual layer is what lets it keep

8:48

working when there's no API or custom

8:51

integration. And then there's scale. The

8:53

product is designed to start with one

8:55

worker, but Warm Wind says you can scale

8:57

from one to 10 to 100 to 1,000 workers,

9:00

all managed through the cloud.

9:02

>> [music]

9:02

>> That's where this starts getting more

9:03

interesting from a business perspective

9:06

because the pitch isn't just automate

9:08

this one browser task.

9:09

>> [music]

9:09

>> The pitch is turn a repetitive

9:11

department workflow into a fleet of

9:13

cloud workers. Each worker has its own

9:16

cloud computer, which also means it

9:18

doesn't need to fight with you for

9:19

control of your machine. You aren't

9:21

trying to use your laptop while an agent

9:23

randomly steals your cursor. The AI has

9:26

its environment. [music]

9:27

You have yours. Security obviously

9:29

becomes a massive issue the second you

9:31

start talking about giving AI access to

9:33

company email, ERP [music] systems,

9:36

customer records, invoices, passwords,

9:39

and internal [music] tools. Warm Wind

9:41

says all user data is stored on isolated

9:44

servers in Germany. Each user

9:46

environment is separated from every

9:47

other environment. Only the respective

9:50

user can access their data and

9:51

environment. And data isn't shared

9:54

between users. They also include a

9:56

built-in password manager and say the

9:58

system was designed with enterprise

10:00

security in mind. Their launch material

10:03

says users can download their

10:04

information locally

10:05

>> [music]

10:06

>> and remove the cloud copy when they no

10:08

longer want it stored there. I would

10:10

still treat those as Warm Wind security

10:12

claims rather than some blanket

10:14

guarantee that nothing can ever go

10:15

wrong. Because if a company is going to

10:17

give any autonomous system access to

10:20

critical software, that deserves a

10:22

serious security review. There's also

10:24

the usability side, which is probably

10:26

where Warm Wind has its best shot at

10:28

separating itself from the agent

10:30

platforms people already know.

10:32

They built the whole setup around this

10:34

idea that the interface should feel more

10:36

like an appliance than a developer tool.

10:38

You create a worker, install the

10:40

applications it needs, [music] and then

10:42

either prompt it or teach it. Voice

10:44

input is supported. So, during teaching

10:47

mode, you can literally talk through the

10:49

task while performing it. If voice is

10:51

awkward, you can use [music] text. And

10:53

that sounds almost too simple, but

10:55

simplicity is actually the product bet

10:57

here. There are already plenty of

10:59

extremely powerful agent systems.

11:02

[music] The question is whether a normal

11:03

business user wants to learn how to

11:05

build one. Warm Wind

11:07

>> [music]

11:07

>> is betting that they don't. They want

11:09

the accountant to show the worker how

11:11

invoices are processed. [music] They

11:13

want the support manager to show it how

11:14

customer requests are handled. They want

11:17

the recruiter to show it how applicants

11:18

are screened.

11:20

The person who understands the job

11:22

teaches the AI directly instead of

11:24

explaining the job to a developer, who

11:26

then builds an automation, which then

11:28

breaks [music] 6 weeks later when

11:29

somebody changes a button. That's the

11:31

part I'm most interested in seeing

11:33

tested properly. [music] Because if the

11:35

teaching system really holds up across

11:37

messy real workflows, then this isn't

11:39

just another agent with a nice

11:40

interface. It becomes a different way of

11:43

deploying automation inside companies.

11:45

Warm Wind is launching [music] with

11:47

access starting at 24 euros per week.

11:49

The company says that includes the

11:51

ability to run multiple workers, and

11:53

they're also planning a very limited

11:55

free version. For the initial launch,

11:57

they say they have capacity for 10,000

11:59

users. So, if you want to try it, Warm

12:02

Wind is live today at warmwind.com.

12:05

They also published their full launch

12:06

video on the Warm Wind OS YouTube

12:08

channel, [music] where they go deeper

12:10

into the product and show more of the

12:12

actual system running.

12:13

>> [music]

12:13

>> I'll leave both below. Thanks for

12:15

watching and I'll catch you in the next

12:16

one.

Interactive Summary

Warm Wind has launched its AI operating system, designed to handle repetitive digital tasks by acting like a human employee. Unlike traditional automation that requires APIs, Warm Wind uses a visual interface to interact with software. Users can train these 'workers' by demonstrating tasks in real-time, allowing the AI to learn workflows and execute them autonomously in the cloud, even when the user's computer is off. This approach targets business users, aiming to simplify automation without requiring technical development skills.

Suggested questions

4 ready-made prompts