There is a demand that IT supporters receive every day, in every organization, in increasing numbers. Someone has a spreadsheet, a document, a script. And the question is always the same: “Can’t I just throw this in ChatGPT?” — and then: “why don’t the numbers match?”
Adam Curry recently called it “vomit scale”: desktop support employees who suffer from this. There is a huge lack of knowledge about it what a chatbot is — versus what AI can actually do. And that lack does not come from the technology. It comes from the word “AI” itself: for most people, AI the chat box in your browser. Although AI is an abbreviation of Artificial Intelligence, it is also not correctly named.
And the companies that run that chat box have a vested interest in ensuring that it remains that way. Their marketing says: “AI is dangerous. It needs to be regulated. Don’t just give it to people.” That’s not security, that’s asking for a fence around their own garden — it’s exactly what Anthropic and OpenAI are trying to do: shape regulation in such a way that the rest of the world is not allowed to run their own models, “because it can be dangerous.”
A chatbot is someone else’s telephone booth
What do you get when you open the chatbot in the browser?
Someone else’s model, on someone else’s server, with a meter running. Your account, your subscription, your tokens. The model will answer differently next time. And everything you put into it — your questions, your work, your thoughts, but certainly also your translations and code projects — is subject to that, and it’s on someone else’s platform.
I wrote it earlier about a children’s book. When we had a book created through a Google app, Gemini gently refused to include a reference to abundomy.com — not with an error message, but politely, impossible. A book that is on a platform is a book that the platform can refuse. It applies to every response you get and every file you send through it.
A chatbot is useful, just like a telephone booth was useful. But you don’t build a stable and consistent environment with it.
The unmetered lesson
In No Agenda 1899 Adam puts his finger on it exactly, with a comparison from our own history. Anyone who’s ever had dial-up to an Internet provider — $3 an hour, checking the time — knows what happened when ISDN and later T1 came along:
When you are not watching the meter, your behavior changes.
First you looked at the clock. Then you never did that again — and suddenly you were using the Internet as it was intended. For local AI, that same moment is now. No meters. No per-token account. No limit on the number of questions. And that doesn’t change your internet behavior — it changes how you interact with AI: you ask more questions, you experiment, you put it into your daily work. Not because it’s “free”, but because it is from you is.
The “agent” is a tool — the harness is the brake
Anyone who goes beyond the chatbot and works with AI agents encounters something else: the combination of model and harness determines the result. Frontier models in their own armor are one tool; an open model like GLM5.2 in its own harness is another. The differences are big — and they show exactly where the limitations and possibilities can be found.
That concept (agent / armor) is too long to mention here; I will do that in more detail in a future post.
Three ways to approach AI
If you put the three shapes next to each other, you will see that it is not about “good or bad”, but about how you approach them:
| Chatbot (in your browser) | Cloud model in an agent | Local LLM | |
|---|---|---|---|
| Memory | persistent — the system knows who you are | often possible, in the cloud | none — each session starts blank |
| Costs | subscription + token limits | per token, or subscription | hardware once, then nothing |
| Your data | out the door, to someone else | out the door, to someone else | stays on your own machine |
| Speed/limits | meter, queuing, throttling | meters | unmetered |
| Who sets the rules | the platform | the API provider | you |
That difference approach is a big story — what persistent memory does to your workflow, what it costs, what to look out for if your local models don’t have memory and you have to build it yourself. This also has more background, such as the limitation imposed precisely because a chatbot has “persistent” memory of you. Here’s the gist: they’re three different tools with three different poses — and most people only know one, maybe two.
The gap is almost closed
The classic counterargument was: open source is 12-18 months behind the frontier models. That was right. Now it’s 3-6 months — and for most tasks that’s the difference between “less good” and “not noticeable.” Jason: “90% of the work I do can be done by open source.” And the examples keep coming: a CEO who paid $300 a month for voice services and is now running it locally, unmetered.
There is also a principle behind it. In 1899, about the major providers:
They’re taking all your knowledge and reinforcement learning it into their system. I don’t believe them, nor should you.
Your questions, your code, your improvements — they’re not byproducts, they are training dates for someone else’s model. I don’t call that a conspiracy, I call that control. And control is exactly what this series is all about.
What the chatbot is
To be honest: the chatbot is not useless. For a quick question, a public fact, a formulation that you check yourself — fine. It’s a telephone booth: handy if you need it, but you don’t live in it and you don’t build your business with it.
Finally
I’m not against data centers. I’m against the suggestion that the chatbot in your browser is “AI” and everything else is “dangerous.” That story is being told by the parties that want to sell the most expensive AI models — and it’s keeping people away from the very tools they can install and use themselves.
And that’s why we play a lot locally. Because the gap is 3-6 months instead of 3 years. Because the meter is not running. Because we know the differences in approach, and different ways require a different approach. And only then do you discover what you can or cannot do with it — and you can only use it optimally when the clock is no longer ticking.
The choice is not up to the AI. The choice is yours.
—Patrick
Sources I listened to for this post: No Agenda 1899 — Asymmetric Intimacy (the “unmetered” conversation from 01:01:03, “so here’s the problem” from 01:04:17) and No Agenda 1898 (latency and the chatbot criticism from 02:19).
Previous posts in this series: How can you use AI? · Suppose you run your own AI coder locally · Creating children’s books with AI
Keep the conversation going
Also on Telegram
New posts, short notes and what’s going on — in the Brainfusionchannel on Telegram. Follow @Brainfusion →