I just had a longer session with Claude Fable 5, and I’m pretty impressed. The model simply does its job. It doesn’t chatter, it handles and tests so much in the background that you can lean on the result more and more. Never completely, of course. I still have to step in all over the place. But not because the coding agent got something wrong. Because I keep finding new complexities.
While the agent codes, I think about how the whole thing would run if not only I used it but other people too. And how it all scales. Those are questions I have to work out for myself first. Usually I head out and walk down to Queens Wharf and back — about five kilometres along Newcastle’s lovely harbour foreshore — before I can even discuss them with Claude and land on a solution.
So I was done with a coding session and wanted to wrap it up, the way I do after every iteration, with the Simplify and Code Review skills. And halfway through /simplify, Claude Code tells me Claude Fable is gone. Matter-of-factly: I could drop back to Opus 4.8, or just wait. No problem, it’s Saturday. But what if I had to have something finished tonight, and then for some reason the connection drops? How embarrassing would it be to confess to a client that the model wasn’t available and I couldn’t get it done without AI? It used to be: sorry, the email didn’t arrive, the power went out, the DSL was down. That seemed acceptable. But “AI interrupted”? Come on.
But it happens. Power users know it. We’re getting ourselves rather dependent on the providers. So what now? I can’t exactly have a limp Gemini model fix up Fable’s code! And I’ve got my own doubts about letting Opus 4.8 near it. So I’m stuck — and thinking, once again, about my favourite topic: local LLMs. And, oddly, Apple.
Maybe Apple wins in the end after all
I’ve come to believe that of all companies Apple — the absolute latecomer — ends up in front. Namely once all the intelligence no longer hangs off an API but sits hard-wired in the computer; once the local models get good enough to be fast and smart enough on a powerful machine. That’s plenty for most of the people paying Anthropic, OpenAI or Google today. Right now everyone seems to reach for the beefiest models for the simplest third-grade maths homework. Soon the listed Anthropics and OpenAIs won’t be able to afford subsidised plans anymore. And the power users will think hard about when the AI in the Mac, or some other box, is enough — and when you actually need the super-power.
So: everyone still reckons Anthropic or OpenAI will be the big winners. But maybe it’s smarter to bet on Nvidia and Apple. Or on some other hardware company that simply builds the AI into the computer and bundles a few tools to stay current — or slips new models in via an update.
And then what? The whole AI hype dissolves like a fizzy tablet in a glass of water. Because then the artificial intelligence in your computer is nothing special anymore. AI becomes like a browser. Just there. Thirty years ago a browser was still something special. Today nobody can picture a computer without one. That’s how AI will soon just sit in the machine. Like the radio in the car. Nothing special, just there. Nobody would need all the tech bros drooling next to a president, wanting to take over the world. There’s a catch, mind you: the local LLMs are currently coming, graciously, from Google, Meta and others. Should Apple head that way, I’m curious how long Google’s models stay free to download. Maybe we should quickly stash the current Gemma models and the fat Chinese ones — Qwen, Kimi and co. — while we still can.
It’ll be a while yet, though. Like right here. While I’ve been writing this, still nothing has moved. I’m still sitting in front of a blinking cursor in Claude Code, no idea how to get any further. Maybe I really should buy a MacBook Pro with 32 gigs. Or bet on the next Mac Mini and run local models on it. Then again, a blinking cursor can also slow you down — in a good way. I’m heading out to the harbour now.
This post started in German on reinergaertner.de — yes, 1997, I’ve been doing this a while. The translation was AI-assisted. Any remaining awkwardness is authentically bilingual.