The first time AI really took over a chunk of my work, it felt like magic. And it felt basically free. A few euros of credit, a subscription, done. Then I set it loose on real jobs, all day, and looked at what it had actually spent. Every small task was quietly costing money. On your own that is pocket change. Across a team doing it all day, it is a real line on the bill. Here is the strange part. Most people just make an account, pay one flat monthly fee, and that is it. Simple and capped. But the moment you start plugging AI into your own systems, an app you build, your automations, a tool your team runs, you switch to paying per token (think: per chunk of text it reads and writes). And per token adds up a lot faster than a flat fee. That is the jump almost nobody warns you about. Now here is the part even fewer people have clocked. The reason AI feels cheap is that you are not paying what it costs to run. For every dollar you hand over, the providers are spending more than that to serve you. OpenAI has been running at well over a dollar spent for every dollar it earns, by its own numbers. They do it on purpose, to win you while the market is still being decided. And they have already started walking that subsidy back. Here is the uncomfortable bit. Nobody can tell you when the investor money runs out or the hype cools. But when it does, the price that moves is the one you pay per token. And it moves up, not down. So the number you plan around today is a promo price. This is the survival part, and it is why this community exists. Surviving the next few years with AI means using it heavily without letting the cost quietly drain you. The ones who come out fine learned to run it cheap before the price caught up. The ones who did not are the ones who get an invoice that scares them off the whole thing, right when everyone else is pulling ahead. You do not need a finance degree for any of this. Four habits cover most of it. 1. One chat per topic.

A while back Claude handed me a block of code built around a library that does not exist. Forty clean, well structured lines, using functions from a package I could not find anywhere online. It never once hinted it was making it up. I lost an hour before I realised the whole thing was fiction. That was not a one off. I have had it give me a SQL job that reported a result while nothing actually ran underneath. I have had it cite sources that were not from where it claimed, or were not real at all. Every single time it sounded certain. That is the dangerous part. It is never unsure. Here is what is going on. When a model does not know something, it does not stop. It fills the gap with the most likely looking answer. A plausible function name. A plausible date. A plausible source. It is built to always produce something, and "something" wins over "I don't know" every time, even when the something is wrong. Here is a live one. I asked for the EU AI Act key dates two ways. Straight from memory it told me the Act was in full force on 1 January 2025, with bans from mid 2024. Confident. Wrong. Then I forced it to work from sources only and search first. It came back with: in force 1 August 2024, bans apply 2 February 2025, GPAI rules 2 August 2025, fully applicable 2 August 2026. Every date now had a link I could open and check. Correct. Same model. Same question. The only difference was whether I let it guess. Here is why I care, and why you should. Everything you send or publish on a model's word carries your name, not the model's. The wrong figure in a report. The fake reference in a proposal. The library that does not exist in code you shipped. Right now people forgive it because "the AI wrote it". In six to twelve months that excuse is gone, the same way "sorry, autocorrect" stopped working. What you send this year is what people remember. So here is how you stop it. 1. Make it search before it answers. Do not ask "what are the EU AI Act dates". Ask "search for the EU AI Act dates and answer only from what you find". No result, no answer.

A while back I was reading a piece of text and instantly thought: this is AI written. Pure gut feeling. Something was off and I could not point at it. So I went searching for how you actually spot AI writing. First I found a Hormozi short about negation. Then a video that explains it properly, and after that I could not unsee it anymore. Here is what it is. "We're not just building a product, we're creating an experience." "AI isn't just a productivity boost, it helps us get closer to our mission." "Your reputation isn't built on your intentions. It's created by consistent patterns others observe over time." Always the same shape. It's not X, it's Y. The first half exists only to make the second half sound bigger. That was the feeling I could never describe. I have written a lot of text with AI and this was in all of it. In the video the guy scrolls LinkedIn for five minutes and finds it everywhere, in posts, in ads, in company updates. Three instances in one 250 word blog post. Once you have seen it you find it in your own inbox every day. https://www.youtube.com/watch?v=l_s27C1AVXU Here is why I care, and why you should. Right now most readers only feel that something is off, the same way I did. In six to twelve months they will know exactly what it is, the way everyone learned to spot the em dash. Everything you publish this year stays online, and almost nobody goes back to edit old posts. The second problem is that negation is empty. Cut the first half of "AI isn't just a productivity boost, it helps us get closer to our mission" and nothing is left that you can act on. It sounds sophisticated and says zero. That is exactly what a model gives you when your prompt is vague. So how do you stop it. 1. Cut the first half and read what is left. AI: "We're not just automating support, we're reinventing it." Cut the first half and you keep "we're reinventing support." Better already, and now you can see it says nothing, which brings you to step 2.

Almost 100 of you joined in the last few weeks, so I'm doing something I don't scale: a personal call with every new member who wants one. I'm Mitchell. I build AI systems for a living, and I run this community to help you stay valuable while the tools keep changing. Everything in here is built to use the same week you learn it. Here's how you claim your call. Do this now: 1. Introduce yourself in your own post. Hit "Write something" at the top and share three lines: - What you do - The task that eats most of your week - What you want to be able to do by the time you leave here 2. Then comment here with the one thing you want to figure out with AI, and I'll send you your call invite. Twenty minutes, one on one. I'll point you at the exact lessons for your situation so you don't waste a week guessing where to start. The people who introduce themselves are the ones who actually get results here. Be one of them. See you in the comments.

In today's world AI models are getting better, stronger and more usable every month. Last month Anthropic released the beast of all models, Claude Fable 5, to the public. One of their most powerful models ever. People are building crazy shit with it. It even got noticed by the US government: the Department of Commerce briefly forced Anthropic to pull access for non-Americans, until they lifted the order two weeks later. In the meantime, OpenAI released GPT 5.6 in three tiers: Sol, Terra, and Luna. And after that Moonshot released their new flagship model, Kimi K3. A Chinese open source model, the largest open-weight model ever (2.8 trillion parameters), and on coding benchmarks it's trading wins with the closed US flagships. It beats GPT 5.6 Sol and Fable 5 on some benchmarks, and loses to them on others. An open model playing at that level is the real story. What does that mean for us? - The biggest AI model players now come from everywhere in the world, and the competition is extreme - The bar will only get higher because every AI company is racing to build the best model in the world - And for us: we need to stay flexible in which model we use and build a workflow where we can switch between them You're no longer locked into one vendor. One day it's ChatGPT, then it's Claude, now we can also choose Kimi. We can pick per job: the cheaper model for simple tasks, the strongest one for coding, another one for research. Let me know if you have used Kimi before, or if you would like to learn more about the differences between these models.