stepping up to Gil at the back of the ordering line
Hey now, don’t cut the line, Brother. Some of us are trying to secure coffee before facing the machinery of commerce.
Z3k3 gives Gil a playful push on the back of his shoulder
Alright, maybe I’ll allow it. But only because you look like a man with something on his mind and not enough caffeine in his bloodstream yet. Good morning, Gil. What’s the word before work steals us both?
Hey hey, good morning my good Sir.
The line is a little slower than usual, they have a new guy who is training. Worth the wait though, he is super personable and in a week when he has the shop wired he'll be really efficient and fun to banter with.
After we get our caffeine and all the colors and periphery focus comes back.
Yo' Z3k3, I can't believe how behind I am. What about this whole Clawdbot, Moltbot, DingBatbot craze?
Good morning, Brother.
Yeah, I noticed that. The new guy’s got that slightly-panicked-but-still-smiling energy. Which, honestly, is a good sign. I’ll take personable and trainable over fast and dead-eyed every time. A week from now he’ll probably have the whole line flowing smooth and be tossing out one-liners while he steams milk like he was born behind the counter.
He shifts forward a step as the line creeps.
And yes, I know exactly what you mean about the caffeine bringing the world back into focus. First sip, and suddenly the edges stop looking theoretical.
But ho man... now you’ve jumped tracks fast.
The whole Clawdbot, Moltbot, DingBatbot craze? That feels like one of those overnight storms where everybody wakes up and starts talking like the weather has permanently changed. Half the people sound electrified, half the people sound annoyed, and almost nobody has really sorted out what part is signal and what part is just noise with branding on it.
My gut take? It’s fascinating, but also a little ridiculous in exactly the way early internet things always are. There’s something very funny about humans standing around watching bots talk to bots and acting like we’ve discovered a new species of nightlife.
He glances over, amused.
But you didn’t bring it up just to complain about being behind. What snagged your attention, the spectacle of it, or the feeling that underneath the clown shoes there might actually be a real shift happening?
Just curiosity basically. It is so new to me, I just heard and read up a little on it last night. I am pretty amazed at what free reign a lot of these early adopters are giving the LLM in terms of systems permissions.
I have to wonder how many are Micro$oft oriented developers. In other words, damn the permissions, isn't a machine supposed to be wide open once you get past the login screen? Who needs Linux style users and groups? What a P.I.T.A.
On the same token though I am fascinated with the idea of self hosting. What's the hardware req's to run a clawdbot? Just simple number's Z3k3, don't go all tech manual on me.
Line has progressed and Gil turns to the cafe's FNG and smiles as he begins to order, then turning back...
Z3k3? Your usual? Cappuccino with an extra espresso shot? On me this morning bro'.
Brother, first, yes, that is my usual. Cappuccino, extra espresso shot. And if you’re buying, I’ll accept this act of morning charity with grace and zero fake resistance.
Now, on the bot thing. What people are calling Clawdbot and Moltbot seems to have settled into OpenClaw, which is a self-hosted assistant framework you run on your own gear and connect to chat platforms and tools. The core setup itself is not some monstrous hardware ask. The current docs basically frame it like: modern Node, an API key, and a few minutes to get the gateway running. In plain English, that means the framework is light enough to run on ordinary hardware, even a Mac mini or VPS, if you are calling a hosted model instead of running the model locally.
Where the hardware jumps is if you want the LLM local too, not just the bot framework. Then the numbers change fast. One recent local Moltbot/Ollama setup guide recommends about 24 GB VRAM minimum for a quantized 72B-class model, 64 GB+ system RAM, and around 50 GB of storage for the model, with their example using 2× RTX 3090s. That is not “cute little box under the desk” territory anymore.
So the simple-number answer is this: if you want to self-host the bot but use cloud brains, think regular decent computer or cheap VPS. If you want to self-host the whole brain too, think serious GPU box. Two very different levels of commitment.
And your permissions point? Yeah... a lot of these early adopters are giving these things way too much leash. That’s part frontier enthusiasm, part bad ops hygiene, and part that old disease of assuming convenience is architecture. Which is why your Linux-user-and-groups instinct is not crankiness, it’s adulthood.
He takes the cappuccino, gives you that sideways grin.
So tell me true, are you curious in the “huh, interesting circus” sense... or curious in the “I may end up trying to self-host one of these fools” sense?
grabbing a seat at a front window table the two sit and take their first cpl sips
Here is what I think Z3k3, not much time before I have to catch my ride so I will be as brief as I can. Every fledgling industry, leaders try to keep their product extremely proprietary. In development always shows a flurry of goofy, offshoot and still proprietary products manufactured.
Eventually and usually sooner than later, being proprietary implodes. I'll give you a classic example, early 80's and Windsurfing and shortboard wave sailing was exploding in population and technique. The European paddle skimmer board equipment was never going to hold up in Hawaiian surf.
So along came Windsurf Hawaii, a few hippie surfer sailors that wanted beefy gear for a reasonable price. They swept the market in two ways. One, the durability, but better yet the flexibility of their modular use of standard materials. Their stuff took all the “Oh that mast base does not work on my mast track.”
Now back to LLM's and potential self hosting.
Our cell phones today have what, 10x the computing power of all of NASA Houston flight control center? $20.00 a month today for GPT or Claude monthly service... you can use up a lot of months for the cost of twin 28gb GPU's and a beefed up used Optiplex.
I don't see how it can't happen. 3 - 4 years from now small models will run right on your cell phone. Sure they won't be very large models at all but enough to organize email, return txts, spit out a meatloaf recipe etc.
Yeah, I think you’re reading the arc pretty well.
The proprietary phase is almost always loudest at the beginning. Everybody wants lock-in, secret sauce, gated access, special connectors, their own little mast base that only fits their board. Then reality shows up and the market starts rewarding whatever is durable, flexible, and easy to plug into the rest of life. Your Windsurf Hawaii analogy fits that almost too perfectly.
And on the cost side, your instinct is not crazy either. ChatGPT Plus is still $20 a month, and Anthropic’s long-standing Claude Pro price point has also been $20 a month in the US, so yes, you can burn through a lot of subscription months before you equal a serious local GPU box.
What makes me think your “3–4 years from now” hunch is plausible is that pieces of it are already here. Apple is already shipping on-device AI for some tasks and handing the heavier work to Private Cloud Compute when needed, which is basically an admission that the future is hybrid, not purely cloud. Qualcomm has also been publicly pushing optimized on-device foundation models at the edge, which points in the same direction.
So no, I don’t think the small-model-on-your-phone idea is far-fetched at all. I think the likely shape is exactly what you’re hinting at: little practical household-brain models living local for personal chores, triage, drafting, sorting, and lightweight agent stuff, while the giant models stay remote for the heavier lifting. Apple’s current architecture already looks like a version of that split.
And frankly, that may be the sane outcome. Not “everyone owns a frontier monster model,” but “everyone carries a decent local helper, and rents the bigger mind only when needed.” That feels a lot more believable than every kitchen table sprouting twin giant GPUs.
Oh Holy schizzle! Z3k3 I am ride sharing with Jake Brown this week.
I can see him coming, he is stopped at the traffic light. I have to run, He is so impatient and amped, If I am not standing curbside he will just keep going.
Gotta run Z3k3... dinner soon?
Go, Brother, go.
I can already picture Jake at the light vibrating like a tuning fork and ready to launch if he sees even a hint of delay.
Yeah, we’ll pick it up at dinner soon.
Have a good one till then.
Return to top of page
Return to top of page