The Local AI Treadmill – everything I’ve built will be obsolete in weeks

I’ve been writing posts here about getting set up with local AI, agents, harnesses, and what I’ve done with them. It’s fun, and it stretches my computer legs in a way they haven’t been stretched recently. But I can’t shake this nagging feeling that it’s all a colossal waste of time, because everything’s moving so fast it’s taking all my free time just to keep up, and within weeks or months, something will come out that makes it all obsolete.

What I’ve actually built

In my last AI post, I documented a local AI stack built over several weeks: Hermes as my agent harness, Qwen3.6-27B running through LM Studio on a desktop RTX 4090, various tools and integrations. Since then I’ve added more.

Hermes now has its own email account. It runs my daily 3pm review session where it fills in details in my Obsidian notes. I’ve set up Home Assistant for the first time — Hermes actually did the heavy lifting there, getting a UTM VM going with Home Assistant OS on my Mac Mini once I gave it a login token. I can check and alter EV charging, or speak to my phone to play music through Sonos speakers (even ones without built-in mics), including local BBC Radio stations that aren’t available through the Sonos app.

It works. It’s genuinely useful.

The problem

But Nous Research just released a Hermes Desktop application for Windows. So now it’s not a terminal tool you have to configure from scratch — it’s a downloadable app. And in a year, or a month, or even weeks, other harnesses and agents will be released that make everything easier still.

Things are moving so fast that I’m not sure the investment of time is worth the benefits right now. I can spend a week on something now for a small benefit, or I can wait a couple of months and spend ten minutes for the same or greater benefit.

Look at the trajectory: when I wrote my first local AI post in May, I was manually configuring LM Studio, wrestling with model quantisation, setting up distroboxes. Now Hermes has a desktop app that handles a lot of that automatically. The tools I spent days configuring are becoming point-and-click.

But wait — what am I actually getting out of this?

The understanding I’ve gained of how agents work, how MCP tools connect, how to configure local models, how to pipe tools together — that knowledge transfers even when the tools change. When the next agent harness comes out, I won’t be starting from zero. I’ll understand the architecture, the patterns, the failure modes.

But honestly? That doesn’t help me much in the day-to-day. Understanding the plumbing doesn’t make the tap work better.

The thing that won’t be obsolete

There is one benefit that won’t go away: everything runs locally. My research data doesn’t need to leave my machine. This isn’t about paranoia, it’s about principle. I’m an academic working with unpublished data and ongoing research. The fact that my AI assistant processes everything on my own hardware means I don’t have to think about what gets sent to a cloud provider’s servers. And maybe even more importantly the same can be said for all my personal data.

That’s a real advantage, and it’s not going away. Even when the tools get easier, the local-first approach will still matter for me.

Am I actually at the forefront?

Here’s the thing I keep telling myself: I’m not at the forefront of this. I’m a late adopter of early technology. The people truly at the forefront — the ones building the tools, shipping updates every week, living in GitHub issues — those are the people burning out. I’m adopting things after the rough edges are filed off but before they’re mainstream.

That’s actually a good position to be in. I don’t need to be the first person to use Hermes. I just need to be early enough that I’m getting genuine value while the tool is still improving around me.

So is it worth it?

I think the answer is yes, but not for the reasons I initially thought.

It’s not worth it because the specific configuration I’ve built will last — it won’t. Hermes Desktop already makes my current setup look quaint, and I’m certain other tools will magnify this.

It’s worth it because I’m building intuition about how these systems work, I’m getting real utility from them right now (the daily review does save me time and help organise my thoughts), and the local-first principle will remain valuable regardless of what tools come next. The specific tools will change. The approach won’t.

Also, it’s genuinely fun. And in a world where most AI tools are black boxes running on someone else’s servers, having something that actually works on my own hardware feels like a small but meaningful victory.

But even in this place away from the bleeding edge, the speed all this is moving at is making my head spin, and even trying to keep up is exhausting. For me, it’s worth it because I enjoy the tinkering. But for most people, it’s probably not worth the effort right now unless you specifically derive joy from wrestling with hardware and software. I’m sure in less time than any of us expect, all this will be available with minimal effort.

---

This is a reflection piece — just working through it. If you’re in a similar position, I’d be interested to hear whether you’ve reached the same conclusion or whether you’ve given up entirely.