Field Notes

Local LLMs and AI agents are awesome, but what keeps blowing my mind is Tailscale

The fact that I can get a multi-device setup in less than 15 minutes is amazing (phone, laptop, and 2 desktops in my case).

Last time I set up something similar using OpenVPN a few years back, it took a weekend.

BUT NOT ONLY are they connected.

I get MagicDNS, NAT traversal, and can intelligently determine if they should use a direct local network path vs DERP relay, all out of the box, without poking holes in my firewall.

You still have to implement last-mile security if you really want a fully locked-down setup for your apps when they talk to each other.

I'm legit just amazed by the experience. And if you haven't worked on networking infra, this might not be mind-blowing lol

But if you are building a local AI / LLM setup, you should check it out.

MY PREDICTION for the consumer market as Local AI gains momentum:

1. The prosumer segment will flock to Tailscale / NetBird / own Wireguard implementation.
2. Apple will capture the rest, building on Wireguard and their UX to make local AI / LLMs seamlessly integrated across your devices.

And no, I'm not sponsored or paid to say this. Just a fan of products with a seamless experience.

"A couple of companies wanted me to pay $1,000/month for SaaS tools for my 17 user app. So I opened Claude Code. And anyway, I started vibing.

Now I've got a project tracking tool, 2 CRMs (the first one leaked customer data), 16 analytics dashboards, a payroll system for me and my virtual assistant in the Philippines, and a customer support ticketing system.

I haven't touched my actual product in six weeks. I've been building 231 internal features to run my business.

But I'm not paying SaaS fees, just 4 additional Claude MAX subs to fix the bugs and make sure the apps are running. I don't have to touch it, I just pay it.

You all think I'm a hero, and I accept that responsibility."

• Frank (Maybe)

CFO: "What exactly is it that you do around here?"

SaaS: “I turn your complex workflow into a managed service so you don't have to worry about compliance and maintenance; you can focus on your business problem.”

CEO: "Well then, I just have to ask, why couldn't I just take our requirements directly to an AI agent, build a custom version for the price of an API call?

SaaS: “Because when you find out that a custom app is a security nightmare, leaking customer data, the AI will just tell you, 'You're right! I apologize for the oversight!' I handle the 'unfun' maintenance and the 3 AM edge cases so you don't have to.”

SaaS: “I have ‘accountability' skills! If I fuck up, you get to blame me. Can't you understand that?” What is wrong with you people?!?!

Price is a major factor in the "build vs. buy" conversation, as is risk.

AI agents are great at building features, but they are terrible at taking the blame.

And yes, you can absolutely do a build-it-yourself if you have the right team. You just have to be aware that a $10,000/yr in savings now can become a $1,000,000 liability and 1,000 hours of debugging later.

Well, is it time to rebrand SaaS as Scapegoat-as-a-Service? lol

Anyway, time to rewatch Office Space this weekend.

One of the biggest shifts in building personal AI agents is the realization that the interface is becoming secondary to the data itself.

Whether you’re interacting through a simple chat, SMS, a voice command while driving, or a complex, rich UI, the underlying intelligence and data remain constant.

And more importantly, it has to be interactable with AI agents.

The future isn't about building separate apps for different tasks; it’s about creating a unified data layer that seamlessly adapts to whatever experience the user needs in the moment.

Prediciton: In 10 years, when AI is coding 99.9% of all software out there, I think there will be a market for "Artisanal Code".

You know, like how there's Starbucks vs the local mom-and-pop coffee shop, or IKEA vs hand-crafted furniture, where the differentiator is that it's small and "handcrafted."

The messaging would be 𝟭𝟬𝟬% 𝗛𝘂𝗺𝗮𝗻-𝗖𝗿𝗮𝗳𝘁𝗲𝗱 𝗖𝗼𝗱𝗲, which takes 10x longer, and you can feel the struggle in every git commit. Built on the developers' homebrew DIY full-stack framework.

Is it time to start a Kickstarter campaign for typewriter keyboards to complete the set?