Spent all night "benchmarking" Fable - and for the first time, I think software is pretty much dead at this stage.
My usual benchmark for model capability is "building a pixel-perfect Notion clone, with full-feature parity, without using any framework or library. No Next.js. No tanstack router. No Zustand. No shadcn. No tailwind.". I spent a lot of time in plan mode, then let the model one shot it.
No model ever got close to a working product - the closest ones were Opus 4.6 (not 4.7 or 4.8) and GLM 5.1. Building a custom block editor without using tiptap or blocknote is the component all models notably struggle with.
Fable got the whole thing done in 5 hours; inference costs are through the roof, I've hit my weekly limit to generate it - but who cares, it built the whole thing, and it was perfect. It even added support for local AI models with Notion AI, built up Notion Calendar, and built a PWA app so I can use it on my phone - all without me asking. The end result is so realistic I couldn't even tell it was a clone when I first opened it.
I'm not re-subscribing to Claude Max though.
As capable as Fable is, I see it as a warning sign that SaaS will soon be dead, along with knowledge work (if it wasn't already).
I think the next step is to start working on robotics, infrastructure and hardware - as it's one of the few areas AI cannot truly help with. LLMs can help you understand the basics, but human skill remains the sole variable leading to genuine breakthroughs in these areas.
Fable is terrifyingly good at software. But as soon as I asked for help on a hardware project I'm working on, responses and intuitions were off. While it was incredibly helpful to help me design my motherboard, it couldn't truly help me with motion, power control and system performance monitoring. In fact, Gemini 3.1 and Grok 4.3 - as bad as they are for agentic coding - had slightly better intuition for these kinds of problems than Fable, even if they were also largely useless.
AI cannot plug cables, tighten a screw, source hardware components, or even less assemble them. LLMs cannot travel to Asia, meet suppliers, do quality control, request iterations with a clear rationale, cannot assess prototypes - even less accurately define prototype development budgets.
Time to build tangible things.
A week ago, I thought this day would never come.
But here we are. Lots of mixed feelings.
This may age like milk - but I don't think I'll be back.