Two years ago I wrote On AI, in which I concluded AI was "kinda cool, useful in a number of scenarios, but not the major step forward in humanity that it's being made out to be". Today, I'd say it's very cool, useful in many scenarios, and a somewhat transformative technology.

This wasn't a position I had come to overnight; it came to be little by little as tools improved.

The web chat clients for the models were a good first step, but were limited by the medium. At this stage, I found it most useful as a search engine / example generator. It was good for getting a direct answer to a question instead of being handed a full field manual that would trail off course every few pages, like traditional search results. Granted, the answer it gave was not always the best answer, or even correct, but it was decent and it saved time. Needing to copy and paste everything in and out is the kind of friction that hurts the overall usability, especially in programming. Sure, you can copy and paste each file it potentially needs, but as files get modified, the model's understanding drifts away from the ground truth and the goal gets lost in a jumbled sea of context.

Then 4 or 5 months ago I gave Claude Code a try. My first session with it was the moment I knew that this technology was going to last long after the bubble pops. Being able to read the files directly and run commands minimizes the time wasted on menial work. Refactoring code now meant writing 2 sentences, waiting 30 seconds, and verifying the bot didn't hallucinate. Most of the time it's been spot on.

There are many different levels of AI-assisted programming. I have been guilty of all of them. I say guilt because I mean it. It does feel a bit cheap to offload the labour of thinking to a robot, but then again, this is essentially what programmers do whenever they import a library: offloading cognitive labour. Nevertheless, it feels like cheating. Then again, I've never been able to accomplish so much in such little time.

No matter the craft, there is pride in practicing it in a manner closest to the root. There's a spirit to a hand-carved piece of furniture, imbued with the soul of its creator, one that is simply not found in one assembled in an Ikea kit. It is the same with AI. You can spend hours working out how everything goes together, how each callback function ties into the next. It is deeply rewarding to see it all come together, spring to life, and serve its purpose. But sometimes you don't need an elegant statement piece that ties a room together; you just need a coffee table to put your mug on.

That is generally how I approach AI use in my projects, first with asking myself how much I care about the end product. If it is a script for work, I don't even open up the text editor and just vibe code the whole thing. If it is a tool I'll be using day in and day out, I'll still keep Claude open for debugging, but will be far more involved and will instruct it to not make edits on its own. That makes it feel a bit more like my own work and not "prompt engineering".

For the past year on Linux Unplugged, Chris and the gang have been exploring AI tools, much to the chagrin of some of my fellow listeners. They tried out OpenClaw when it first hit the scene back in April. When I first heard of it, I rolled my eyes. Of course the headlines were all around people putting them on their own agentic social media platforms. I thought it was just another tool of the week, that the show would move on to another one and leave it as a faint and distant memory, but Chris began integrating it into everything. Each new episode he would share how he had automated, fixed, or created something with it.

It piqued my interest.

However, if I was moving away from Claude, its replacement was going to need to have some major benefits. Being the paranoid person I am, I wanted to run something locally. In addition, as I have mentioned before, I have been wanting a Mac of some sort to keep using Openbubbles whenever I eventually leave my job. So the natural conclusion was to get an Apple Silicon Mac. They're efficient little bastards that allegedly are great for running LLMs.

I set a $200 budget and hunted Ebay, Craigslist, and Marketplace for any beat-up M series. I ended up winning at auction a 2020 13" M1 Macbook Pro with a horribly cracked screen. For a week, I eagerly waited for it to arrive in the mail, researching what models would be best suited for it. I didn't have high expectations for it, figuring Claude is at least 4 times faster than whatever I could run locally and would probably make more mistakes. That didn't matter though; it was going to be on my hardware and that was worth the trade-off! Right?

HA!

I unboxed it, struggled to get it to use an external display for half an hour, then got logged in. I set it up to be as headless as a Mac could be, fighting Apple's permissions dialogs every step of the way.

After some more faffing about, Ollama was installed with one of the low-end Qwen models, and so too was Hermes on my laptop.

I eagerly typed: "Give me a weather report for Sacramento and Port Angeles, WA for tomorrow".

Sacramento weather:

86 degree high, 0% precipitation

So that was a bust...

Too slow and too stupid to be of any use to me. Hermes is just about the worst thing to use this low-end of model and device for, with its heavy use of memories and skills clogging up the minuscule context window. I continued to try and make any tweaks I could, but you can't make a weed-whacker motor a big block V8, no matter the amount of tuning. I put the Macbook back up for auction less than 24 hours after it left the box.

I still wanted to try out Hermes and so I followed in Chris's footsteps and threw $10 into usage credits with Deepseek.

Once it was all wired up, I just started playing around. I stared in amazement as it navigated through my network, learning its surroundings, taking notes on its every finding. Within minutes I was able to refer to each of my computers by their hostname and have the bot be sshing in to do whatever task I needed. In an afternoon I had unified all the shared configs across my boxes to be pulling from a unified source of truth from my local git server. With a little minion to do my bidding, all the annoying tedium has become one request away.

Over my brief time using Hermes, it's easily taken the place of Claude. Even though it isn't as smart, it makes up for it in its note taking. If it makes a mistake, it makes it once. As soon as you correct it, the solution is stored away in its memories without needing to explicitly tell it to.

One thing I'm kinda grateful for is that it isn't an amazing code-writer. It's able to read code fine, and generate usable documentation better than I could, but it kinda falls flat on its face if it needs to write more than 20 lines. This has made me use it more like I did a year ago, using it for a documentation finder and tedium minimizer. It'll make the jigs, while I get to focus on the craft.

Even if the models themselves have only marginally improved over the past 6 months, the tooling around them is maturing and widening their capabilities.

Of course, to many people this is grim, sickened by the pervasive sprawl of AI and slop. Now we're supposed to have them run amok on our computers? I have sympathy for them, as I wasn't a big fan of it originally either. I remember losing all faith in humanity when my Uber driver pitched his AI therapist app to me. The only thing I can say to them is that AI is a tool.

A hammer can both be used to build and to destroy. A car can unite families for a holiday, or wipe them out in less than a second. The internet united the world and divided nations.