See the laptop and agent in use.

00:03.183 So what do you actually get when you

00:04.775 buy a Make No Mistakes laptop?

00:06.075 Let’s go through it.

00:08.158 This should come in the mail.

00:09.608 Let’s open it up.

00:12.175 All right. Charger keep this.

00:14.492 You’re gonna want your laptop plugged in all

00:15.542 the time. Laptop.

00:21.575 In here is

00:24.742 the laptop. This is

00:27.525 a Dell. Sometimes it’ll be a different model,

00:29.542 but generally it’ll always have a minimum set

00:32.292 of specs. 16 gigabytes of RAM

00:35.142 256 gigabytes of hard drive, It’s

00:37.775 not really a super special laptop.

00:41.192 These are refurbished enterprise

00:43.825 laptops. So if you’ve ever had a corporate

00:45.575 W2 job, you’ve probably used one of these

00:46.925 laptops before. They’re sturdy they’re reliable.

00:49.425 The point of them is that it’s quality

00:52.075 tested, it’s an enterprise machine, and then we

00:54.225 turn it into a dedicated AI agent.

00:55.608 You’ll see that sticker on the laptop.

00:57.658 It says a couple of things.

00:58.425 It says please wait for onboarding.

01:00.342 Onboarding is a really important part of what

01:02.392 you get when it comes to an M

01:03.608 M purchase. That is 30 to 60 minutes

01:05.358 of dedicated time with a live

01:07.992 person probably me, walking you through

01:10.275 setting up the agent and giving you some

01:12.358 sort of starter tips and tricks, best practices

01:14.242 so that at the end of that call,

01:15.258 you have a talking agent texting you messaging

01:16.925 you back, running on your own hardware,

01:19.775 connected to your own AI subscription, and then

01:22.592 you can really build anything you want from

01:23.925 there. So this laptop or these

01:26.592 laptops will have Linux,

01:29.642 which is an open source operating software

01:31.008 flashed onto them.

01:31.808 Instead of Windows or Mac we use Linux.

01:33.708 The AI agents the robots really like Linux.

01:35.858 They really like open source software because

01:37.142 there’s a lot of information out there that

01:38.592 they’ve been able to be trained on.

01:40.942 And then inside it you’ll go to log

01:42.892 in. It’ll ask you for a password.

01:44.592 The default password is on this message that’ll

01:46.708 be on the laptop.

01:48.775 You’ll immediately be asked to change the

01:50.358 password. So first you’ll put in the default

01:51.725 password. You’ll put it in once.

01:52.892 It’ll say you need to change the password.

01:54.692 You’ll put it in again.

01:55.925 Then you’ll put in your new password that

01:57.608 you’re changing it to twice.

01:58.825 So old password twice new password twice.

02:01.108 And then on the home desktop screen, you’ll

02:03.458 have a link

02:06.075 or a shortcut at the top left corner

02:08.242 that is the

02:10.158 setup software. And so once you

02:12.958 launch that, if you haven’t already connected to

02:14.525 Wi-Fi, it’ll help you do that.

02:15.658 Well in order it’ll ask you for the

02:17.792 password and you’re giving that to the agent

02:19.075 because the agent is gonna run on this

02:20.175 machine. So it needs to have its own

02:21.392 password so that it can do things on

02:23.292 the machine with administrator permission.

02:25.908 The next item is the Wi-Fi.

02:27.258 If you haven’t already connected it, it’ll walk

02:28.608 you through connecting it.

02:30.092 Then you’ll choose your harness OpenClaw or

02:32.175 Hermes. Then you’ll choose your AI provider,

02:35.825 Claude via the API OpenAI via their

02:37.892 subscription, or Maple which has some more

02:40.625 privacy preserving options.

02:42.392 And then you will set up the

02:45.125 connection to that AI provider.

02:46.658 And then the last stage is to hatch

02:48.025 your agent and then send a test message.

02:50.758 That usually takes about 20 minutes, whether

02:52.508 it’s guided or on your own.

02:54.142 And the other sort of period of time

02:56.625 in the onboarding call is dedicated to giving

02:58.225 some tips and tricks and best practices.

02:59.942 There also will be a PDF saved on

03:01.692 the desktop of your agent laptop that has

03:04.042 a bunch of those tips and tricks saved

03:05.725 into it in a presentation.

03:07.758 And we keep that updated from time to

03:09.125 time. The last piece is that you’ll be

03:10.642 invited into a customers only Telegram group

03:12.625 where people who have been MNM customers,

03:15.392 myself and others that are building things in

03:17.208 the AI space, are sharing best practices

03:19.858 ideas troubleshooting, helping each other fix

03:21.792 and build things.

03:22.792 And that’s been a really useful resource to

03:24.158 our customers. So thanks for watching.

00:03.100 I’m gonna record a demo of making

00:05.700 an update to the MNM website

00:07.967 using my agent.

00:09.917 So last night I had the agent overhaul

00:12.350 the website and I actually haven’t really looked

00:13.700 at it yet.

00:14.100 So I’m gonna do this review live and

00:16.333 give feedback to the

00:19.000 agent. So I actually really like this.

00:24.067 One thing I know I’m gonna wanna add,

00:25.483 so I’m going through this quickly because I

00:26.717 already know all the copy.

00:28.517 One thing I know I’m gonna wanna add

00:29.667 is right now demo is not its own

00:32.350 page. It just jumps to this recording

00:35.100 I did a while back.

00:35.867 I did it before I had this better

00:37.400 setup. So I wanna replace it.

00:38.667 And I also wanna do multiple videos.

00:40.283 So I want this to be demos and

00:41.567 I want it to link to its own

00:43.100 dedicated page. And so we’re gonna do that

00:45.317 now by going to the agent and saying

00:48.233 hey I’m reviewing the MNM website.

00:50.383 I think last night’s overhaul looks really good.

00:53.100 I’m recording a few bits of content that

00:54.867 I’ll give to you that we should make

00:56.050 into videos. And with that in mind, I

00:58.300 want the demo piece to link to

01:00.950 its own page instead of right now, it

01:02.617 just jumps to that section of the website.

01:04.083 I think that section on the main landing

01:05.433 page makes sense because I want people to

01:06.933 be able to see a demo quickly if

01:08.100 they wanna learn more about the product.

01:09.950 But the demo button in the top

01:12.333 navigation bar should take them to a dedicated

01:14.467 page that has what eventually will be multiple

01:16.517 videos covering different topics.

01:18.583 And maybe demo is too narrow of a

01:20.117 description there. Maybe it should be media,

01:22.767 but we already have like blogs.

01:24.233 So that’s a little similar but maybe it’s

01:26.433 content. Anyway I welcome your thoughts on that.

01:28.333 And I’d like for you to spin that

01:29.200 up, put placeholders for the demo videos right

01:31.433 now, and then I’ll give you some videos

01:32.450 to add to it.

01:35.633 So I’ll give that to the agent in

01:36.717 a voice message like that.

01:37.600 You can see it was about almost a

01:39.817 minute and then it immediately transcribes it.

01:42.000 You can see this is the transcription.

01:44.033 And then now Olivia this is the agent

01:45.767 working on this, is going to work making

01:47.833 those changes. Says I’d call it

01:50.583 videos not media or content.

01:51.950 Media sounds like press coverage.

01:53.050 Content says nothing.

01:53.767 Demo becomes wrong the moment the page holds

01:56.033 education and commentary.

01:56.950 That’s a great point because I am gonna

01:58.100 have commentary on like why a

02:00.700 dedicated laptop and why this versus a virtual

02:03.083 private server. So

02:05.517 this is gonna spin for a little while

02:06.883 and we’ll come back when it’s done, probably

02:09.583 two to three minutes tops.

02:10.833 All right MNM video is live.

02:12.317 View the video library with videos because it’s

02:14.083 clearly distinct from blog.

02:15.200 What changed? Homepage nav.

02:21.567 Yeah it looks great.

02:23.050 I’m gonna go to the site.

02:24.800 I’m gonna refresh.

02:26.817 Looks like it changed.

02:28.967 Awesome. So here’s this looks great.

02:31.600 This takes you back to the order.

02:34.117 Perfect. Okay so that’s a quick demo.

00:03.100 The other thing I want to go through

00:04.600 is what is agentic AI?

00:06.533 What is an agent harness?

00:08.850 And so I think the way I’ve described

00:10.017 it that has really resonated with people is

00:11.500 that regular chat AI is what most people

00:13.200 have experienced, right?

00:13.900 You open up a web browser, you go

00:15.433 to ChatGPT or some other provider, you put

00:17.133 in words, you get words back.

00:18.267 Hey, how do I make a great brisket?

00:19.767 And it uses context and a bunch of

00:21.933 model training to give you the most logical

00:24.133 set of words associated with brisket to give

00:26.450 you a recipe for a great brisket.

00:27.750 You put in the words, you get words

00:29.100 back. That’s great if you want to, hey,

00:31.017 edit this email for me, and then you

00:32.333 copy paste it back and you hit send.

00:34.050 The agent layer takes that and puts it

00:36.167 on steroids. I describe, so that first

00:38.733 experience we’re talking about where you know

00:39.950 words in words out that’s inference And

00:42.133 inference is like fire.

00:43.933 That’s a campfire.

00:45.267 The agent framework is like an internal

00:47.483 combustion engine. It’s using the same

00:49.400 technology, which is combustion or fire, but it

00:52.000 can take you a lot further.

00:53.117 So when I think about that for an

00:54.867 agent harness, what’s happening is your agent

00:57.550 harness layer is an open source software running

00:59.467 on whatever environment you run it on, and

01:01.450 it is relatively lightweight, but what it’s

01:02.900 doing is combining that inference of words in,

01:05.333 words out with tool calls and execution on

01:07.567 the machine or on the environment that it

01:09.333 runs in so that if you give it

01:10.533 the credentials and permissions it can actually

01:12.667 get things done.

01:13.400 So instead of, hey, I want to reword

01:14.667 this email, and then you copy paste the

01:16.067 output, you just say, hey, this is the

01:18.283 wording, send that email once I’m happy with

01:20.517 the draft of it And so that manual

01:22.433 layer can be removed in a way that

01:23.817 you still have the human in the loop

01:24.817 for review, but the friction is far lower

01:26.867 such that you can get a considerable amount

01:28.417 of work done compared to just regular way

01:30.450 sort of copy pasting in the old version

01:32.017 of, you know, words in, words out AI.

00:00:03.100 So, some quick thoughts on Mac

00:00:05.833 Mini, OpenClaw, and the right way to host

00:00:08.517 these things. It’s funny how explosive the Mac

00:00:10.933 Mini craze and meme was when the reality

00:00:13.700 is it’s just a headless computer

00:00:15.517 and you can run OpenClaw

00:00:18.017 and Hermes on any headless computer as long

00:00:20.633 as it’s got sufficient specs.

00:00:21.800 I think we generally recommend 16 GB

00:00:24.217 of RAM and a mid-tier

00:00:27.017 not-so-new-but-not-so-old

00:00:29.900 chip processor, and then you

00:00:32.700 probably want a hard drive of 256 GB

00:00:34.600 to start. You might find that that’s more

00:00:36.333 than enough, or depending on what you end

00:00:37.500 up using it for, it might be something

00:00:39.100 that you expand to or add cloud storage

00:00:40.483 to later. I think it’s interesting that the

00:00:43.200 Mac Mini was gravitated to.

00:00:44.800 It is good hardware, but for

00:00:47.000 the agent setup, you want to be able

00:00:49.400 to access the device directly.

00:00:51.033 Otherwise, you would just host it in the

00:00:52.350 cloud, and I have other videos talking about

00:00:53.833 why it’s a benefit to be able to

00:00:56.033 access it directly, but just quickly, if

00:00:58.017 you have it set up on your local

00:01:00.733 network, you can serve yourself local webpages

00:01:02.267 that are dashboards or other cool draft

00:01:04.083 projects, and they can just be held locally,

00:01:05.933 so you don’t have to broadcast them to

00:01:07.300 the internet. That can be done with a

00:01:08.783 virtual server but you’ve got to do some

00:01:10.217 sort of security tunneling to make sure that

00:01:12.467 you don’t expose that page to the whole

00:01:13.667 web, and there are security tradeoffs that come

00:01:16.050 with that. And so then the other thing

00:01:17.833 is really just, okay, we know we want

00:01:19.267 it local, a Mac Mini versus say maybe

00:01:21.817 what MNM offers and it really

00:01:24.600 comes down to the agent is going to

00:01:26.467 need to be accessed from time to time

00:01:28.017 directly, and the whole point of setting this

00:01:29.833 thing up on your own machine is that

00:01:31.250 you can control that, and if it’s in

00:01:32.733 the cloud or on a Mac Mini, you

00:01:35.033 don’t have a screen and keyboard access.

00:01:37.483 Now with the Mac Mini, if you buy

00:01:39.383 a screen and keyboard and a mouse, then

00:01:40.867 you’ve got the same setup.

00:01:41.717 Then you’re just looking at the cost.

00:01:43.217 The Mac Mini bare without any of those

00:01:44.867 peripherals costs more than the pre-built MNM

00:01:47.000 laptop. The MNM laptop comes with

00:01:49.750 a screen a keyboard a mouse and

00:01:52.333 30 to 60 minutes of live guided onboarding

00:01:55.033 so that you have a talking agent at

00:01:56.233 the end of it.

00:01:57.083 So the tradeoff is just that Mac Mini

00:01:58.967 is just a very common form factor, and

00:02:01.117 Apple has done an incredible job marketing-wise

00:02:02.600 to be the sort of product of choice.

00:02:04.733 The reality is that the agents actually operate

00:02:06.533 better, in my opinion, on Linux than they

00:02:07.983 do on Mac, and so there’s a bunch

00:02:09.233 of reasons to consider running it on

00:02:11.967 a laptop, and there’s a bunch of reasons

00:02:13.767 to consider running Linux instead of Mac, and

00:02:15.283 those are just a few.

00:03.158 The other question I get all the time

00:04.342 is, why run an AI agent

00:06.992 on a separate laptop?

00:09.875 When you’re thinking about what these tools are,

00:11.858 it’s kind of hard to differentiate them from

00:13.658 a computer virus.

00:14.508 If you describe an autonomous

00:16.675 software that can execute code without

00:19.475 human oversight and do things live on

00:22.108 the internet, it sounds a lot like a

00:23.592 computer virus. Now that’s not necessarily the

00:25.492 worst framing, it’s just that you have to

00:26.958 contextualize it in the fact that this virus

00:29.325 can be used for good right?

00:30.408 So if you direct the virus in this

00:32.508 regard to do things on your behalf, then

00:34.475 it can become super powerful.

00:35.742 The challenge though, is what information do you

00:38.108 give it access to?

00:39.742 And if you were to just download OpenClaw

00:41.825 or Hermes on your main personal or work

00:43.525 PC, you instantly inherit all of

00:46.308 the exposure of having this AI software running

00:48.792 on the same machine that has all of

00:51.025 your sensitive credentials and information that

00:53.258 you may not wanna share with the agent.

00:54.592 Now I find that most people end up

00:56.475 over a long enough timeframe being pretty

00:57.908 comfortable sharing a lot with their agent.

00:59.258 There is still a lot of value in

01:00.242 having a separation.

01:01.242 The other thing that’s beneficial is that you

01:03.725 have, if you run this on dedicated hardware,

01:05.925 it can operate a lot more like a

01:07.292 server that you can still access physically when

01:09.392 you need to from time to time for

01:10.592 troubleshooting. So the ideal AI agent

01:13.408 should be always on and always responsive.

01:15.308 And if you’re using it, if you’re running

01:16.692 it on a laptop that you carry around

01:18.042 in your backpack, well then when it’s not

01:19.725 connected to Wi-Fi, your agent’s gonna be

01:20.858 unresponsive. And so the idea is that the

01:23.075 laptop form factor is great for

01:25.475 directly accessing when you’re troubleshooting

01:27.592 or trying to kind of add new things

01:29.758 that can’t be done via just messaging the

01:31.758 agent. But at the same time, you want

01:33.125 it to run like a server plugged in

01:34.492 all the time on Wi-Fi stable so

01:36.425 that you can always access it.

01:38.092 And so yeah that’s a little bit around

01:39.975 why a separate laptop and why a laptop

01:41.875 versus a private server.

00:03.100 I was going to record a demo of

00:04.833 making a change to the MNM website because

00:07.350 I think it’s a good example of how

00:08.650 things can be automated with an AI agent

00:11.117 And I realized I wanted this video

00:13.417 of me in the bottom-right corner

00:15.300 to use the MacBook-style Center

00:18.067 Stage view, where it kind of focuses on

00:19.983 me and it’ll shift if I move around

00:21.383 in the frame.

00:21.817 It’s a great example. So my agent is

00:23.233 on the same home network as my MacBook

00:25.900 I’m just going to show the scene here

00:27.500 in OBS Studio that records this. It’s

00:29.350 got my screen and my face in

00:31.650 the bottom-right corner. I’m going to

00:33.517 say, “Can you please make it so

00:35.333 that the talking head box

00:37.550 in that scene

00:40.183 uses the MacBook’s Center

00:42.967 Stage feature.” And because my agent already has the

00:48.117 access credentials to go in

00:50.750 to, um, go

00:53.517 remotely into my MacBook. They’re on the

00:55.517 same network, but it could also do it remotely

00:57.350 via Tailscale. It should be able to

00:59.183 just go in and, um,

01:01.300 make that change for me. Now, I

01:03.583 don’t know if it’ll work because it’s

01:04.567 recording right now. I’m obviously filming

01:06.217 this behind the scenes content but

01:08.917 we’re going to try it.

01:10.133 And we’re going to watch it together and

01:11.033 see what happens.

01:11.383 It said, “Your MacBook supports it natively.”

01:13.867 The built-in camera exposes multiple Center Stage

01:15.800 formats. Center Stage is currently off. I’m

01:17.350 enabling the system feature in cooperative mode

01:19.017 so macOS handles the tracking while OBS

01:20.900 preserves the natural picture-in-picture

01:22.683 geometry.” Okay, that sounds like it’s going to

01:25.200 work. It says, “Done. Center Stage is

01:27.100 enabled on the MacBook camera in cooperative

01:28.300 mode. I verified actual zoom recentering in

01:30.317 OBS while preserving natural facial

01:32.583 proportions. The talking-head box now tracks as you move.

01:34.750 I’m testing if it moves.

01:36.700 Oh, it does move. Well,

01:39.150 a little. It works.

00:03.100 One of the questions I get from time

00:04.683 to time around, why OpenClaw, why Hermes, why

00:06.967 these open source agent harnesses versus just

00:08.833 building these same systems and processes inside

00:11.417 of OpenAI or Claude is

00:13.567 focused around this idea of optionality.

00:16.983 OpenAI and Anthropic are very big businesses

00:19.333 from a valuation perspective, but they are also

00:21.450 startups. They are unprofitable startups.

00:23.650 And now that may change and these businesses

00:25.283 may become the Pepsi and Coca-Cola of

00:26.983 our generation or they may not.

00:29.433 And I think that the AI competition landscape

00:31.867 is fiercely competitive.

00:33.650 We see every week new, more and more

00:35.583 powerful and cheaper open source models coming

00:37.883 out of China and the frontier labs in

00:40.617 the U.S. pushing closed source models that

00:42.683 are still very competitive.

00:44.167 But the risk in building for an individual,

00:46.750 an entrepreneur, or a business inside of the

00:49.067 sort of closed off ecosystems of Anthropic

00:51.433 or OpenAI, if you just use sort of

00:53.467 Claude Cowork or OpenAI’s ChatGPT system

00:56.050 to build some of these systems and workflows,

00:57.417 one, they’re far more permissioned.

00:59.067 And two you have vendor lock-in that

01:01.300 is very risky.

01:02.267 Because if next week some new

01:04.767 upstart AI lab comes out with a model

01:07.183 that is just heads and tails better than

01:08.917 what OpenAI or Anthropic offer, and you have

01:11.350 built your systems, workflows, processes inside

01:13.800 of the walled garden of OpenAI’s

01:16.200 ChatGPT or Claude inside of Anthropic,

01:18.383 then to shift to that

01:21.033 better model provider is a very manual

01:23.417 and painful process.

01:25.433 Separately, or additionally I should say, the

01:27.667 risk that OpenAI or Anthropic or any other

01:30.150 of these vendors goes out of business is

01:31.833 nonzero. And building mission-critical systems

01:34.267 and workflows into a single-vendor counterparty

01:36.617 risk is dangerous.

01:38.350 Versus in the OpenClaw or Hermes setups,

01:40.983 if you decide next week you want to

01:43.150 switch from OpenAI to Anthropic, it

01:45.917 is two lines of code or one quick

01:48.567 terminal command or even a prompt potentially to

01:50.717 your agent to say, hey, give yourself brain

01:52.850 surgery to switch the inference provider from

01:54.950 OpenAI to Anthropic.

01:56.383 Or better yet, from Anthropic to

01:58.783 an open source model provider from China that

02:01.333 has one one-hundredth of the pricing.

02:03.217 So considerably cheaper.

02:04.433 And so I think when you’re building these

02:06.250 systems and processes, the open source agent

02:08.400 harness layer that is OpenClaw or Hermes gives

02:10.583 you far more optionality to switch to

02:13.300 wherever the best and cheapest compute is and

02:16.000 take all of your systems and workflows with.

02:18.117 For example, throughout the sort of

02:20.883 six month and growing history of working

02:23.667 more than daily with my agent Olivia in

02:26.283 my businesses, I have more than once switched

02:28.350 from OpenAI to Anthropic to Kimi K3

02:31.033 and other models just to try them out.

02:33.567 And that switch is done before I’ve taken

02:35.267 my first sip of coffee.

02:36.350 Versus if you build all of your systems

02:37.900 and processes inside of Claude Cowork and then you

02:39.750 want to switch to OpenAI, well, that’s a

02:42.233 weekend-long, or possibly months-long, surgery

02:44.450 process of rebuilding those systems inside of

02:46.683 the other walled garden.

02:47.833 So that’s a little bit more on why

02:49.533 open source AI agents.

02:50.367 And the beauty of it is that you’re

02:51.500 still using, in many cases, you’re using your

02:53.483 OpenAI subscription. That is, at least

02:55.267 currently, the cheapest way to do this.

02:56.633 But you own more of the data and

02:58.417 the output that you create versus if you

03:00.633 did that inside of ChatGPT, well, then you’re

03:02.567 pretty locked in as a customer.